We’re at another peak in the AI hype cycle. If the hype of the past few years was a rising tide that lifted all (tech) boats, what’s different now is that the technology for the first time since it debuted in 2022 seems to be dividing the industry into clear winners and losers. And, as of the start of this year, software—the category as a whole—has been branded the AI economy’s first clear loser.

Over $800 billion in software stocks has been wipe from the market. Salesforce’s shareprice dropped over 30% from its peak this year. Atlassian’s price halved. Even Microsoft—with one of the most diversified portfolios and a major AI player itself—is down 15%. All this, even despite many of the hardest-hit software companies showing strong earnings throughout 2025.

This sell-off is shocking because software has been, for the past few decades, a key input of American dynamism. It has powered productivity gains across industries, from logistics to finance to retail to healthcare, creating some of the country’s most valuable firms as well as empowering some of the most powerful legacy players across the economy. Entire sectors have reorganized production around the software revolution. And it has become a central feature of white-collar work as a whole.

The problem with software, however, is that it is also the domain in which AI is most straightforwardly competent. With the release of Claude Cowork, Google’s Antigravity, and most recently OpenAI’s Codex, the power of AI generated code has been put on full display, making investors worried that these advancements would make software much cheaper to build.

Software has historically been expensive because coding skills have long been in short supply. So when software firms aggregate that labor, build a product once, and sell it many times over, they spread development costs across thousands of customers willing to pay a hefty price for the convinience of not having to build it themselves. If AI tools make it significantly cheaper to write and maintain code, it may undermine the economics of this business model.

From here, several related outcomes are plausible. First, large enterprise customers may decide it is cheaper to build “good-enough” in-house solutions rather than rely on pricey vendors for their software needs. Second, software makers facing new cost pressures, might engage in a race to the bottom, cutting costs and selling their software on the cheap. Third, a new industry of vibe-coding consultancies may emerge, specialized at building budget solutions for individual clients rather than general-purpose software. These scenarios aren’t mutually exclusive. But in each case, dedicated software makers will lose their leverage, and the ecosystem they collectively sustain will begin to thin and erode.

I won’t speculate on more on what the future will look like. But what I plan to do in this post is explain what it would mean for AI to eat our software ecosystem; and why we may actually want to prevent that from happening.

The Positive Externalities of an Ecosystem

The obvious purpose of an ecosystem of dedicated software makers is that it supplies buyers with a vast portfolio of software to use. For a single company who has a particular software need, the diversity offered by the ecosystem can seem superfluous and may well be replaced by a half-competent in-house vibe-coder. But an ecosystem has many positive externalities beyond providing clients with basic software.

One externality comes from the fact that software ecosystems comprise different products that can integrate with each other and so their combined utility is greater than the sum of its parts. These products will often have shared standards and technical conventions, allowing members of the ecosystem to specialize deeply while remaining compatible with one another. Improvements in one corner of the ecosystem can even ripple outward and enable advances elsewhere. A better database architecture can make new analytic tools possible. A new breakthrough in distributed computing may reshape how we design high-performance computing workloads.

Another externality is that the software-dedicated members of the ecosystem will also invest in shared infrastructure. Open source is the obvious example of this dynamic. Projects like PostgreSQL (database), Spark (big data processing), or Kubernetes (infrastructure management) are actively supported by software companies because, even though no single firm can fully capture the returns from maintaining them, these shared foundations expand the overall software market and provide a stable substrate on which each company can build differentiated products.

In fact, part of the reason AI is so good at coding is because it can draw from the vast stock of open source software that has accumulated within the ecosystem over decades. A vibe-coder can quickly develop a working application because their coding agents draw from existing frameworks, mature libraries, and best practices. Ask an AI to spin up a webapp and it will default to tools like React and Django—projects of sustained collaboration and collective investment. Without an ecosystem of software-dedicated firms, software progress will slow and will, past a certain inflection point, hamper AI coding progress with it.

That is not to say that all software companies are valuable contributors to the ecosystem. Many do not contribute to open source and are freeriders of these positive externalities. But the problem with AI eating software is that it will eat with it the field of software engineering.

Imagine a world where software-dedicated firms no longer exist and all software is created on-demand by in-house vibe-coders. Building software will then simply be an exercise in using the innovations of the past rather than creating them anew. And that is because the incentives for advancing the field of software will also vanish. A piece of software that works only for one firm is relatively easy to build. But building a system that can operate across industries, regulatory environments, and millions of users is not. And it is precisely the pressure to build highly scalable general-purpose software—with near perfect uptime, stringent security, and impeccable performance standards—that drives the technical advances and novel innovations in software.

Consider a small regional bank that decides to build their own software solution in-house. If the software can meet the basic needs of the small bank, the project is a success, and the incentives stop there. The bank, who is itself hardly interested in advancing the frontier of software engineering, does not need to design for 10X growth and extreme reliability. It does not need to architect for thousands of software integrations or defend against the full spectrum of security attacks. Yet without these pressures, there is also no need to produce truly exceptional software that pushes the frontier forward.

Software is a Discipline, Not a Cost-Saving

The AGI boosters would offer a counter: software innovation in the age of AI does not require an ecosystem of dedicated software makers. With the self-recursive improvements it can make on its own, code generation will be so powerful that it could conjure entire frameworks or generate novel algorithms without the prompter even needing to know about it. Perhaps so, in some more distant future. But that’s far from the capabilities we have today.

In the meantime, what is to be done? How do we make sure that AI doesn’t flood the economy with mediocre webapps at the expense of dismantling the ecosystem (and with it the discipline of software) needed for advancing software innovation?

As I’ve suggested, one of the possible futures is that enterprise customers start building in-house solutions rather than buying from software-dedicated firms. Software-dedicated firms, for their part, may start selling their products on the cheap to remain an attractive option. In this scenario, AI is primarily used as a cost-cutting function—something to replace engineers, drive down wages, and churn out code faster. The aggregate outcome is the thinning of the discipline of software engineering and ultimately a blow to software innovation in the long run.

But AI is itself not necessarily the danger. Even in keeping AI central to software’s future, another outcome is possible. What’s key is how firms choose to deploy it: whether AI is used to deepen engineering expertise or to discipline engineers and drive down costs. To walk the latter path is of course much easier and a surer gaurantee of short-term profits. That also means avoiding it will require making a deliberate choice.

For that, firms on both sides of the market have a role to play. Enterprise customers should resist the temptation to retreat into cheap, in-house builds simply because AI makes them feasible. Software companies for their part should not freak out and cut technical staff in the name of cost savings but should instead continue to invest in their engineers. And even as AI coding is used by these engineers, the goal in using it should be to deepen their expertise and tackle ever-harder software problems. AI should, in other words, raise the ceiling of engineering not simply lower the cost of doing it.

As Marc Andreessen predicted, software has indeed eaten the world. But software’s spread was never automatic. Just as Moore’s Law depended on generations of engineers and firms relentlessly pushing the limits of chip design and production, the ubiquity of software rests on the cumulative work of hundreds of thousands of engineers who developed new algorithms, best practices, and even programming languages—each contributing to software becoming more scalable, more reliable, and more useful year after year. If software’s capabilities had frozen in 2011—the year of Andreessen’s prediction—it would not have eaten much of anything.

In the end, whether AI compresses the software industry into a thinner, cheaper version of itself or pushes it into a new era of ambition is not technologically determined—it is a political choice. It will require employers to continue investing in engineering staff and investors valuing long-term capability over short-term profits. It will also require policymakers to put in the appropriate guardrails to incentivize the correct behaviors that can keep the ecosystem (and the discipline) intact. None of it is easy, but that choice is up to us.

My article with Kathleen Thelen, “Cloud Capitalism and the AI Transition,” is now out in Politics & Society! The paper actually came out in December, so this post is a bit overdue. I’ll blame some of that on a stretch of near-constant travel for the past month—Shanghai, Guiyang, Chengdu, Shenzhen, Hong Kong—and the fact that I somehow managed to get sick twice along the way.

Of course I’m biased, but I really think that readers of Value Added, especially those academically inclined, will find the article quite useful. It’s an effort to make sense of the AI transition by starting, not from the cutting-edge language models themselves, but from its underlying political economy. It looks at the industrial structure of the AI economy, the entanglement of the cloud hyperscalers in state power (what we call in the paper the techno-nationalist alliance), and new forms of competition which we argue are quite different from the preceding era of platform capitalism.

Below, I’ll sketch a few of the paper’s main themes. I’ll resist the temptation to summarize the whole thing, but sharte just enough to motivate you to click through to the article and read it in full.

After Platform Capitalism?

The very first draft of my article began as a response to Sabeel Rahman and Kathleen Thelen’s influential 2019 paper on “The Rise of the Platform Business Model and the Transformation of Twenty-First-Century Capitalism.” Their paper walks us through platform capitalism’s core features—its consumer-investor coalition, focus on horizontal dominance, and reliance on a precarious and increasingly gigified labor regime. It also explains why the institutional configuration of the United States was uniquely conducive to this model’s ascent.

When I read their piece, it immediately stood out as one of the most useful text for understanding the political economy of the platform era. It became my go-to recommendation to friends and colleagues who wanted to read something about platform capitalism (alongside, of course, Nick Srnicek’s Platform Capitalism book). However, as much as I admired it, I couldn’t help but notice that my experience having worked in Microsoft’s Cloud and AI division was completely overlooked by their paper.

From this perspective, it was obvious that the most consequential developments at the time were not happening in consumer-facing apps or gigified labor platforms. Instead, they were taking place under a different business model that was organized around expensive infrastructure, long-term planning, and often much deeper dependencies between firms, customers, and even governments. In fact, the common features of platform capitalism—footlooseness, asset-light strategies, and regulatory arbitrage—didn’t match what I was seeing at all.

For some context, I read their paper when I took a class taught by Professor Thelen. And when I argued for my term paper that her platform business model paper with Sabeel completely missed the developments that now underpin the AI boom, we decided to develop the cloud business model paper together.

Our aim was not to overturn the platform capitalism framework. But we felt that the prevailing theories used to describe the knowledge economy were not sufficient for understanding how such an asset-light business could lead up the asset-heavy underbelly atop which the ChatGPT moment relied on. To make sense of this, we needed to explain another development in the knowledge economy: the rise of the cloud business model, and its central role in the AI transition.

The Cloud Business Model

The cloud business model is immensely important not only because it is the foundation on which the development of large language models was made possible but also because it is now the primary vehicle by which AI can diffuse to other sectors of the economy. So to understand the AI transition, we need to understand how the cloud differs from the platform model that came before it.

In this table, I’ve summarized some of the notable shifts. I won’t walk through each of them. The paper already does that well, so I’ll encourage you to read it directly. But one feature is worth emphasizing: the cloud’s extreme asset intensity. Unlike the largest platform firms, which are famously asset-light (Airbnb, for instance, controls all the lodging without owning it), cloud providers are investing extraordinary sums in physical infrastructure, above all in data centers. And in doing so, they have become much more vertically integrated across the stack. This, of course, is a well-known fact by now. Everyone has seen the graphs of Amazon, Google, and Microsoft’s CAPEX spending shoot up over the years.

However, this feature of the cloud business model is particularly worth highlighting because this new asset-heavy approach in what was a traditionally asset-light sector carries important political consequences. During the platform era, firms’ asset-light structure made venue arbitrage a key strategy: operations could be shifted across jurisdictions to minimize taxes, evade regulation, or exploit legal gray zones. Cloud firms, by contrast, are physically anchored. Their dependence on massive data centers, energy infrastructure, and long-term fixed investments ties them tightly to particular places—and, by extension, to the states that govern them.

As a result, the relationship between leading tech firms and the state has begun to change. Rather than treating regulation as something to be avoided, cloud providers—and especially their leaders including Sundar Pichai, Larry Ellison, Satya Nedella, and Jeff Bezos—increasingly seek accommodation and partnership with the government. This has given rise to what we call a techno-nationalist alliance: a political alignment in which the state facilitates rapid data center build-outs at home (and in allied countries), while simultaneously deploying export controls and sanctions to hobble geopolitical rivals abroad.1

This shift produces quite a different political coalition. Platform firms are anchored in a consumer–investor alliance: investors pursuing winner-take-all returns frame platforms as protectors of consumers against “crippling” regulation, casting state capacity as an obstacle to innovation and choice. The state, in this view, is something to be evaded or politically neutralized. The cloud business model points in the opposite direction. Its asset-heavy character underpins the techno-nationalist alliance described above. In this configuration, the central coalition is between the state and investors; consumers are cut out entirely.

Concluding thoughts

I hope this brief post is enough to entice you to read the piece in full. Theres much more in it on the hyperscalers’ competitive strategies, their relationships with suppliers and customers, and even labor/distributional implications. My hope is that it also offers a useful way of thinking about the AI transition: not as a story about an individual technology or the most cutting edge labs, but as a broader shift in the underlying political economy of the digital economy.

I recently attended a workshop about gov-tech relations in the age of digital capitalism and had a co-authored paper about the cloud computing sector circulated as part of a discussion about the political economy of AI (and for the most comprehensive reading list on this topic that I’ve seen, check out Henry Farrell’s AI syllabus).

My paper, which I’ll share more about in a future post, discusses the rise of Microsoft, Amazon, and Google as the three cloud giants underpinning the AI revolution. One question that kept coming up during the workshop was why was it these three and not, for instance, Meta or Apple that had pursued building a public cloud. I didn’t have an immediate answer at the workshop, but have been thinking about this question since, and so that’s what my short post today will try to explain.

What I’ll also argue in this post is that in the AI era, tech players that do not have their own cloud platform (e.g., Meta or AI labs like OpenAI) represent a markedly different kind of tech firm than those that do. I think, in fact, they represent two starkly different visions for how AI will diffuse and who the winners and losers will be—I’ll return to this point in the second half of this post.

Why did Amazon and not Meta or Apple become cloud giants?

Today, we take for granted that a small handful of firms—Amazon, Microsoft, and Google, trailed by Oracle and IBM—dominate the US cloud market. But fifteen years ago, as the cloud business model was just beginning to mature, it was far from clear which companies would prevail. A quick look at this shortlist reveals two key prerequisites for entering the cloud business: deep pockets and established enterprise relationships. Cloud infrastructure is capital-intensive, demanding massive upfront investment in servers, networking equipment, and data centers. During the brutal price wars of the early 2010s (which I wrote about here), only firms with vast cash reserves could afford to compete; the rest were quickly driven out.

Equally important was an existing base of enterprise customers. Microsoft’s decades of experience selling IT management tools like Windows Server and productivity software like Office gave it a built-in distribution network and credibility among corporate clients. Google, though better known for consumer products, was also aggressively expanding its Workspace suite—Gmail, Docs, Spreadsheets, and Calendar—into enterprise offerings. Oracle and IBM, meanwhile, were already synonymous with enterprise software and database systems.

So even as Meta and Apple were among the most cash-rich companies of the 2010s, neither had meaningful experience serving enterprise clients. Apple never viewed other businesses as part of its customer base; its strategy was to dominate the consumer market as a tastemaker and premium hardware producer. Even as Macs gained popularity among individual users, the corporate computing world remained firmly tied to Windows. Meta, for its part, repeatedly tried to break into the enterprise market throughout the 2010s—with workplace collaboration tools, internal communication platforms, and enterprise versions of its social network—but each effort fizzled. Eventually, the company abandoned the idea altogether. This reveals an important point about the cloud computing business—that its success depends not on access to millions (even billions) of individual users but on landing the top (ideally Fortune 500) companies as customers.

If being cash-rich and deeply embedded in enterprise networks were the key ingredients for success, then why was it Amazon—an online retailer that was neither cash-rich nor enterprise-oriented at the time of AWS’s founding—that pioneered and now dominate the cloud computing industry? Part of the explanation lies in Amazon’s retail DNA. Unlike Microsoft or Google, which built their businesses on high-margin software and advertising, Amazon came from a world of razor-thin profits and relentless scale. Retail had trained the company to operate on low margins and high volume—to make money not by maximizing profit per unit, but by driving operational efficiency and selling vast quantities.

That discipline—obsessing over cost control, logistics, and customer satisfaction—translated naturally into running a lean, scalable cloud operation. Selling books might yield only a few cents of profit per transaction. For Amazon, then, selling virtual servers was considered an upgrade: a capital-intensive but still higher-margin business than selling books. This is in part why Amazon—not Google or Microsoft—was most likely to pioneer the cloud. For Google or Microsoft, the cloud business looked comparatively less attractive from the outset; why chase a lower-margin segment when your core products are already cash cows? Only after Amazon had established the market and demonstrated its potential did the cash-rich, enterprise-oriented giants begin to pile in.

It’s worth noting, as a brief aside, that China operates quite differently in this realm. Huawei and Alibaba had started their cloud business early on (just a couple of years after Amazon pioneered the whole sector). Yet new players are constantly showing up to rival them. JD.com entered in 2020. ByteDance in 2021. The reason, much like Amazon’s own trajectory, is that Chinese tech firms have a far higher tolerance for low margins. And because the competitive environment is so cutthroat, selling compute at slim or even nil profits is considered an acceptable business venture. In other words, unlike their US counterparts, which tend to discipline themselves around “core competencies” and high-margin lines of business, Chinese firms routinely expand into down or upstream, or even entirely unfamiliar sectors, as long as there is money to be made.

Divergent AI diffusion trajectories: enterprise workflows vs killer app

Meta/Apple’s decision not to enter the business was made well before the AI boom years. If they had anticipated the rise of LLMs and the compute-intensity of AI models, they may have chosen to build public cloud services in preparation for this new era. Even today, Apple continues to hold back on entering the AI-capex game, so as far as being a major player, it is effectively out of the running. But Meta, even without a cloud business, has kept up with the hyperscalers in buying NVIDIA GPUs and building powerful data centers across the country. But the window of opportunity for building a cloud platform with a robust ecosystem has closed, and this has major downstream effects on how each of them will position themselves in the AI era.

Specifically, whether cloud computing sits within a firm’s arsenal will determine how they see AI diffusion happening: whether it can happen broadly and heterogeneously or whether they must instead bet on a “killer app.” For Meta, without a cloud to monetize downstream usage, the burden for such a killer app is far greater since it has no other way to make good on its infrastructure costs. So its best bet is to use AI to further entrench its position on its existing product lines and extract even more from its already captive userbase. Newer AI labs—OpenAI, Anthropic, and xAI—face a similar pressure as they race to build flagship applications that can be widely used.

Cloud hyperscalers, by contrast, do not need a killer app at all. By providing the infrastructure—and along with it, the vast ecosystem of cloud services—through which the rest of the economy adopts AI, they can offload the burden of experimentation to their customers and partners. Microsoft, for example, does not care whether it invents the defining AI application of the decade—so long as whoever does builds it on Azure. So rather than intervening only at the application layer, Microsoft’s key value add is providing the scaffolding that would empower their vast ecosystem of SaaS companies to build AI into their own products.

Cloud-first companies might, in fact, even be hostile to the idea of a single killer app. A single dominant AI application would concentrate value at the application layer, compressing the diversity of downstream demand and reducing their leverage as cloud providers. Such a future would risk turning cloud providers into commodity infrastructure rather than a strategic platform on which innovation can happen. As such, cloud firms are much more likely to push a model of AI diffusion where thousands of enterprises integrate AI in idiosyncratic, business- or industry-specific ways—each requiring custom workloads, proprietary models, and a mix of different cloud services.

In short, the biggest divide in today’s AI landscape is between two fundamentally different diffusion strategies: betting everything on a killer app or quietly embedding AI across the long tail of enterprise workloads. Those choices will produce very different industrial structures—and very different winners.

Western coverage tends to frame China as an “AI boogeyman,” highlighting its breakthroughs (e.g., DeepSeek) and geopolitical ambitions (e.g., the “AI Plus” initiative) as a justification to plow ahead with AI advancement at all costs. But seldom is the material basis that underpins those advances closely examined.

That’s what I’ll do in this post. A closer look at the recent history of China’s data center buildout reveals a far more complex story. Behind the dazzling progress of its frontier AI models is an infrastructure story rife with inefficiency—hundreds of facilities built in a frenzy, now sitting idle or chronically underused. Much like the ghost cities that have long plagued the country’s real-estate sector, China’s zealous data center buildout have resulted in the proliferation of what I call ghost infrastructure.

Data center boom years

China’s data center boom began well before the ChatGPT moment. In 2020, Beijing announced a “new infrastructure” initiative that focused on seven key infrastructures to drive the next phase of the country’s modernization—among which were data centers. The announcement triggered a wave of investment and hundreds of firms rushed into the sector.

The ensuing build out was driven more by political enthusiasm than by market logic. Local governments competed fiercely to attract firms to build, offering subsidies, discounted land, and other kinds of financing. Data centers were erected across the country almost overnight. This enthusiasm was further intensified by another policy directive issued in early 2022 called East Data, West Computing, which encouraged the construction of more data centers in western provinces where land and energy are abundant.

The release of ChatGPT—and the realization that AI progress would hinge on ever-larger models demanding exponentially more compute—lit an even greater fire under the ongoing building spree. Dozens of startups rushed to develop their own LLMs, each needed vast computing resources to train them. Local governments and firms once again raced to expand capacity, retrofit older facilities, and position themselves as the backbone of China’s AI economy.

However, the facilities required to train today’s frontier AI models are far more complex than the commodity data centers of the past. They demand an enormous and steady power supply, advanced cooling systems, and round-the-clock operational stability—and, above all, access to vast clusters of GPUs capable of training large-scale models. In the West, such AI-oriented facilities are often called “neo-clouds”; in China, they’re known as zhi suan (智算), or “smart compute” centers. Many of the firms that had entered the market in the initial 2020-rush could make do essentially as real-estate ventures that focused on easy financing, sourced commodity equipment, and built cheaply at scale. But this new generation of AI-capable data centers demanded a level of technical sophistication that most players didn’t have.

This of course didn’t stop companies from building. In the years that followed, hundreds of new projects were announced as state-owned enterprises, publicly listed firms, and government-backed funds all sought a piece of the AI infrastructure buildout. Many started construction long before securing any reliable sources of demand. An MIT Tech Review article hilariously pointed out that “among them were companies like Lotus, an MSG manufacturer, and Jinlun Technology, a textile firm,” hardly the type of firm with relevant technical competencies.

Without market discipline, however, this reckless rush started to unravel. By 2024, hundreds of new AI-focused centers had been built, yet many fell far short of the technical and operational standards required to train modern AI models. Meanwhile, China’s AI model ecosystem began to consolidate. Dozens of smaller LLM startups folded as the market coalesced around a handful of dominant players—the major internet giants among them—leaving much of the newly built facilities without customers.

These more sophisticated users also had far higher technical standards, demanding facilities capable of 24/7 uninterrupted power and flawless operational reliability. This meant that inexperienced players—many of whom had treated data centers as real-estate construction projects rather than highly technical systems—were in no position meet them. And when DeepSeek’s R1 model was released in early 2025, showing that cutting models could be trained on a fraction of the computing capacity, demand contracted even further. Soon, it became clear that the hundreds of facilities built in the preceding years were heavily underutilized, with many operating well below capacity, some running at just 10%.

China’s industrial model

There are two underlying dynamics worth pointing out in the story above. First is the gold rush. Once Beijing designated data centers as a national priority, the policy itself became a signal of future demand—prompting both state and private actors to pile in, often well before any actual demand had materialized. This leads to the second related dynamic: the rapid influx of new players—many with little relevent competency beyond access to financing, sometimes from local governments—makes it difficult for more serious players to pursue longer-term, quality-oriented investments. These speculative entrants drive prices down, forcing capable firms to compete on cost rather than capability, and leaving them unable to reinvest in R&D, technology upgrades, or other improvements to the production stack. The result is overcapacity at the low end, underinvestment at the high end, and an industry trapped in a race to the bottom.

This of course is not unique to data centers. In fact, to make sense of China’s current data center crisis, we need to understand it within the country’s broader industrial model, which goes something like this: A new national priority is announced—or a promising technology emerges—and firms rush in to stake their claim. Local governments, eager to meet central targets and stimulate growth, each back their own “champion,” showering them with subsidies, tax breaks, and cheap financing. Production ramps up rapidly, competition intensifies, and prices spiral downward, creating a deflationary race to the bottom. Yet rather than allowing weaker firms to fail, local officials continue to prop them up, unwilling to see their homegrown projects collapse. The result is a proliferation of “zombie” firms that survive not on their commercial strength, but on policy support.

China’s electric vehicle (EV) manufacturers offers the clearest illustration of this dynamic, and is perhaps the most talked about case in recent years as they threaten incumbent automakers around the world. China now produces some of the world’s most competitively priced EVs and the top firms have started to export them across the world. Yet competition is so cutthroat that few automakers are profitable. Many persist only through a lifeline of local subsidies, as officials refuse to let their regional champions die. The industry is thus trapped in chronic overcapacity, with too many firms vying for an already saturated domestic market. (I’ve written more about this dynamic in the EV sector here.)

This pattern of hyper-competition has become so severe that it has drawn concern from the very top. Earlier this year, President Xi Jinping publicly warned automakers to exercise “self-discipline” to curb involution—a term borrowed rom Chinese internet culture that is now used to describe the self-defeating cycle of competition and overproduction. Indeed, this combination of industrial overcapacity and price wars has long plagued China’s economy, from steel and cement to solar panels and EVs.

And as the story of the data center boom shows, the same logic has also taken hold in the AI economy. The rush to build data centers has replicated the very pattern that once defined China’s manufacturing ascent: rapid expansion without market discipline, leading to impressive capacity on paper but widespread inefficiency in practice. Yet unlike manufacturing that can be easily exported, its much harder for domestic data center firms to sell excess capacity abroad.

Fighting involution

The AI data center boom hasn’t gone bust for everyone. Over the past few weeks, I’ve started to take a closer look at the data center firms that have managed to survive, and have noticed some important patterns about what distinguishes them from the many that didn’t.

First, however, it’s worth understanding briefly where AI compute demand in China has shifted to today. Model development has consolidated around a handful of major internet companies—Alibaba, Tencent, ByteDance, and Baidu—and a few well-capitalized AI startups such as DeepSeek and Moonshot. These firms, flush with capital and wary of security and reliability risks, increasingly build and operate their own data centers. As a result, there is relatively little demand for independent, third-party providers of AI-grade facilities. Most other companies that need compute lease capacity from these third-party players; these third-party players also supply overflow capacity for the top firms when they exceed their in-house resources.

In this environment, success hinges on a very different logic than the kind that characterized the earlier data center boom years. Data center firms have to focus not on quantity but quality—serving large, stable (mostly big tech) clients with real compute demand and focusing relentlessly on operational efficiency. Companies like Chindata Group’s success rests on cultivating deep relationships with major customers and maintaining an exacting operational model that guarantees uptime, stability, and energy efficiency.

This shift has implications for the industrial structure of the data center industry. Many of the failed entrants treated data centers as real-estate projects that can be cheaply replicated anywhere. The more successful firms, by contrast, know that the industry is fundamentally an engineering business—one that demands capabilities in efficient energy use, cooling, networking, and operations. To achieve this, firms like BCI Group have pursued tightly integrated supply chains, investing in in-house energy management, bare metal servers, and other IT equipment. Even the internet giants are increasingly vertically integrating down the stack and investing in their own capabilities in data center construction.

Successful firms have also benefited from long-term policy stability. The founder of BCI Group described choosing to concentrate operations in Datong, a city with abundant and diversified energy sources, a reliable regulatory environment, and proximity to Beijing. But most importantly, rather than chasing short-term subsidies, he emphasized the importance of consistent, long-horizon industrial policy:

Achieving a vertically integrated model requires certain conditions at the local level: long-term policy stability and resistance to excessive commercialization; not just access to electricity, but also a complete and sustainable set of resources—energy, surface water, land, and labor. (translated with AI)

Data centers and the compute layer on which the AI race is being waged represent just one layer of a much larger technology stack. And already, the biggest players are looking to take control over this layer to enable tighter integration across their value chain. But the rise and fall of China’s data center boom, and the few success cases illustrated above, offers a lesson that can be carried forward to other domains. It is that success doesn’t always come from speed and scale and capturing short-term profits, especially when demand is untested. Rather, it depends on a much longer-term vision where firms embed themselves in stable ecosystems, cultivate deep customer relationships with actual demand, and invest in the technical and organizational depth that can turn commodity infrastructure into value-added capabilities.

I’m excited to share a new research article that I co-authored with Natalia Luka and Emily Mazo—two close friends, collaborators, and longtime interlocutors on everything related to labor in the US tech sector. The article draws from years of conversations, a dataset of collective actions in the tech sector that we’ve collective built and maintained, and a mutual interest in understanding why tech workers came to be organizers. In this short post, I’ll highlight some of its key findings.

Argument in short

This basic premise of this article begins with a puzzling trend: since 2018, organizing efforts have grown within the tech sector, led not by the typical ranks of exploited or low-wage workers, but by highly paid, technically skilled professionals—engineers, designers, data scientists, etc.

Traditionally, labor organizing is a way for the most exploited workers in society to fight for better pay and working conditions—for example, factory workers in the auto industry have long organized around issues like low wages, dangerous working environments, and long hours, leading to landmark union contracts that reshaped industrial labor standards in the 20th century.

Tech workers sit far outside the traditional image of the exploited laborer. They earn some of the highest salaries in the country and enjoy comfortable jobs with lavish perks—free meals, game rooms, on-site massage therapists, nap pods, and even cooking classes. And unlike the typical factory worker, tech workers also have a high degree of autonomy, often setting their own schedules, working remotely, and facing minimal direct supervision.

All of this makes tech professionals seem like unlikely candidates for labor organizing. And indeed, labor relations in the US tech sector was relatively stable and pro-worker for most of the 2000s and 2010s.

Yet, over the past few years, tech workers have increasingly turned to organizing—and not just in isolated cases. In 2019, software engineers contracted through HCL Technologies at Google successfully formed what became the first recognized union of its kind in the United States. The next year, workers at Kickstarter also won their union drive. The tech division at the New York Times also voted overwhelmingly to unionize. Today, there are more than two dozen unions representing tech workers that have been officially recognized.

To understand the rise of labor activism in tech, we analyzed all publicly documented instances of collective action by US tech workers (which we’ve made public here) from 2014 to 2022. We found that tech workers engaged in both traditional labor activism and what we call social activism—actions over issues like military contracts that don’t directly affect their own pay or conditions. Our article then uses a logistic regression with company fixed effects to test how these different forms of activism relate to each other and how organizational characteristics shape their likelihood.

The central finding of our article, and what our regression analysis confirms, is that participation in social activism often served as a catalyst for more durable labor organizing within the sector. When employees faced retaliation or clashed with management, many began to recognize their shared vulnerability—not as individual professionals, but as workers. These experiences pushed them toward organizing for stronger protections and better working conditions. In other words, tech labor organizing builds on earlier waves of struggle. It’s a reminder that engaging in conflict can clarify class position and beget further conflict.

Or, as we say quite nicely in the article:

Participation in [social/political] activism generates solidarity among employee-participants. It can also create conflict with management, who see such protest strategies as tarnishing the company’s brand image. In this way, social activism exposes the previously hidden divide between workers and management, masked by the industry’s techno-utopianism.

Backlash against tech labor uprising

One of the recurring themes of this newsletter is tech labor—and, more broadly, how certain labor arrangements can shape innovation. I’ve argued, in fact, that stable, inclusive, or even empowering labor arrangements, can benefit innovation outcomes. But just as revealing are the breakdowns. The most extreme cases—the moments when labor relations collapse—offer critical insight into what happens when the foundational social contracts that underpin innovation go through an upheaval. In a sense, this article sets the stage for one of those breakdowns.

If the years between 2018 and 2022 marked a flowering of labor organizing in the tech sector, the years since have seen an unmistakable backlash. Rising inflation in 2022 meant the end of the zero-interest-rate-policy, giving tech employers a chance reset labor relations in the industry and dismantle the very conditions that gave rise to the tech labor movement. As we show in our paper, this hasn’t led to the death of labor organizing in the industry. Still, its impacts are far reaching.

Since then, mass layoffs have wiped out tens of thousands of jobs. The very departments that once embodied a more socially conscious tech culture—trust and safety, AI ethics, etc—have also been purged. Diversity and inclusion programs have been dismantled. Internal communication channels once used for open debate have been restricted or shut down altogether. The latest iteration of this hostility comes from the AI startup world, where some founders have openly taken a page from China’s infamous “996” work schedule—9 a.m. to 9 p.m., six days a week (and I’ve written here about how this kind of work model has shaped innovation in China’s internet sector). All of this has made workers more precarious and less willing to organize.

This shift could have a major impact on innovation in the tech sector. A culture of fear and overwork can yield short bursts of productivity, but it in the longer term, it corrodes the trust, creativity, and critical exchange that breakthrough innovation requires. As I’ve written about before, DeepSeek was actually among the only Chinese tech companies that didn’t enforce 996 and other draconian labor practices that were ubiquitous in China’s tech scene, contributing in part to its breakthrough success.

It’s too soon to know exactly how this new age of tech labor austerity will play out. The effects aren’t uniform either—AI researchers get far more leeway than most software engineers. Still, it’s a dynamic worth watching, especially as the US economy bets so heavily on continued breakthroughs in AI.

From Ezra Klein and Derek Thompson’s Abundance agenda to Dan Wang’s lawyerly society, we’re increasingly told that America has lost the ability to build. Restrictive zoning laws, the story goes, choke off housing supply and underpin today’s housing crisis. Environmental reviews tie up new construction (like high speed trains) for decades. And the liberal impulse to give communities a say in shaping their neighborhoods has often enabled them to halt new projects in the name of preserving local interests. The result is a maze of local interests and procedural hurdles—reinforced by an army of lawyers and bureaucratic agencies—that has crippled America’s capacity to build, and especially to build big.

But if America has lost its ability to build, how do we explain the breakneck buildout of data centers? Underlying the AI revolution is the gargantuan building spree of data centers across the country. And in large part, such a buildout is powered by the cloud hyperscalers like Amazon, Microsoft, and Google, who have each poured hundreds of billions of dollars (and billions of tonnes of concrete) into data center construction.

And in the truest sense, these data centers are about building big, with some now exceeding a million square feet—spanning the equivalent of more than 25 football fields—and commanding some of the largest construction budgets in the country. And unlike housing or high-speed rail, the United States is far ahead of China in data center construction and, with countless new projects already in the pipeline, is poised to maintain that lead well into the future.

You might argue that data centers are easier to build than new housing or transit lines because they have more flexibility in where they can be sited. There’s some truth to this, but hyperscalers aren’t building data center facilities randomly across the map either—they are converging on a limited set of geographies with the right mix of resources: abundant energy and water, cooler climates, and close to key high-speed fiber connectivity. These constraints drastically narrow the field of viable locations. And often, this means building takes place in exactly the kinds of places Abundance theorists insist are impossible—such as densely populated, liberal-leaning1 Loudoun County, Virginia (known as the “data center capital of the world,” with the highest concentration of facilities anywhere on the planet).

Herein lies the paradox. If America is unable to build—chasing software and finance instead of concrete and steel—then how has it managed to erect data centers at such speed and scale, far outpacing China, in what is now one of the largest infrastructure buildouts in the nation’s history?

Why corporate elites thrive in the American political economy

AI is now attributed to much of the buildout of data centers across the country. We should remember that data center has steadily expanded since the advent of cloud computing in the mid-2010s, long before generative AI went mainstream. But even if we accept that the AI boom has added fresh urgency, the Abundance thesis would insist that political motivation alone cannot overcome local interests and procedural gridlock—otherwise America’s housing crisis, which is nothing if not politically salient, would have been addressed long ago.

So the question then is why haven't data centers been stymied by the very forces that Klein, Thompson, and Wang insist has crippled America’s ability to build? To answer that, we need to step back and examine the country’s political system. Here, the emerging field of American Political Economy (APE)2 is especially useful for helping us understand why some actors are able to build while others are not.

APE highlights two features of the American political system that deserve close attention. First, it is exceptionally fragmented: authority is divided across multiple layers of government, and political contestation is spatially dispersed, decentralized, and played out across a wide range of venues. This creates a multitude of veto points and conflicting preferences, especially for projects or policies that span multiple jurisdictions. So even when there is broad political support for something like housing affordability, coordinating across venues can be prohibitively difficult, making ambitious national-level policy agendas hard to advance. Second, governance emphasizes local self-rule, which empowers local bodies like neighborhood associations, zoning boards, and city councils to wield outsized authority over decisions that carry national consequences. And when it comes to building new things, each of these local bodies can stall or even kill construction outright (which Abundance theorists also point to this as a root cause of the country’s building woes).

To navigate this political environment, APE identifies two decisive capacities. First, the ability to sustain long time horizons allows actors to outlast shifting coalitions, election cycles, and procedural delays that would exhaust if not outlast actors with shorter time horizons. Second, the ability to operate across multiple venues enables actors to coordinate in a federalized political system where coordination is otherwise difficult but essential for getting policy agendas passed. The capacity to operate across multiple venues also allows actors to engage in “venue arbitrage” such that if one jurisdiction resists, they can (threaten to) shift their investment to another. This creates competition among venues—states, counties, or regulators—which actors can exploit to extract better terms (e.g., bigger subsidies, looser permitting, etc).

These capacities are the hallmark of organized, well-resourced groups, and they explain why such actors thrive while disorganized, and poorly-resourced ones—renters, small developers, individual voters, and even local governments—rarely prevail. So while mass politics may have some influence, building large-scale policies requires sustained efforts by organized actors. This is why business and economic elites, above all the tech giants, have amassed such disproportionate political power in the United States.

How big tech sidesteps the red tape

Seen this way, we can start to understand why data centers (and not housing or high-speed rail) are bypassing restrictive zoning laws, circumventing red tape, and getting built. Unlike the individuals who voted for California High-Speed Rail or the diffuse (though national) constituency for more housing, tech giants are organized and well-resourced. They are armed with vast financial resources, professional lobbying operations, and long planning horizons. They also have the capacity to influence subnational governments across the country allowing them to play the multi-venue game with ease.

When it comes to building data centers, this dynamic plays out in two ways. First, tech giant hyperscalers can move between different arenas of decision-making and can thus exploit America’s fragmented political economy by engaging in venue arbitrage. If one county shows resistance, tech giants can threaten to exit and shop the project across localities for the best deal. And because of their largely oligopolistic market structure, they can operate across venues without competition. The result is a race to the bottom in incentives, as states and municipalities bid against one another with ever-larger tax abatements, cheaper land, and subsidized infrastructure hookups for the “privilege” of hosting data centers. Tech giants know this and use it to maximum effect, extracting the most favorable terms wherever they decide to build, allowing them in-effect to build at a discount.

Second, tech giant's ability to coordinate across the full spectrum of the American political system enables them to outmaneuver local opposition but shape the agenda at the federal level as well. They wield influence simultaneously at the state and county levels, as well as in Washington, where they lobby federal agencies and the White House. This ability to coordinate across venues has allowed their policy agenda to rise to the top—most clearly demonstrated by Trump’s executive order on “Accelerating Federal Permitting of Data Center Infrastructure,” which makes federal lands available, fast-tracks permitting, and encourages federal agencies to provide financial support for constructing data centers.

Combined, these capabilities have allowed the tech giants to build data centers at a staggering pace, even when such facilities are deeply unpopular with local communities. Marketed as engines of regional development, data centers in reality provide few permanent jobs, attract little ancillary high-tech industry, and consume vast amounts of water and energy—often at the expense of nearby residents. Unsurprisingly then bottom-up resistance has not been uncommon. Data Center Watch has documented over 140 activists groups across 24 states, but only a half-dozen cases of data center projects have been blocked by local resistance, suggesting how easily local resistance is outmaneuvered by the tech giants.

(This new Business Insider mini-documentary shows how homeowners in Northern Virginia have been powerless to stop the wave of new data center construction happening in their own neighborhood.)

Concluding Thoughts

The Abundance agenda frames America’s crisis as one of excess proceduralism—that endless red tape has made it impossible to build. But the data center boom is evidence of an America that can still build, and at staggering speed, when powerful, well-organized actors are behind it.

Zoning laws, permitting processes, and environmental reviews bend easily for hyperscalers like Amazon, Microsoft, and Google, even as they may block projects backed by disorganized/poorly-resourced constituencies such as renters, voters, or subnational agencies. What looks like procedural paralysis is, in fact, a system that privileges powerful actors.

The problem then isn’t that America has lost the capacity to build—it has lost the capacity to build for the public. And until that changes, the country will likely keep getting more data centers instead of affordable homes and high-speed trains.

Dan Wang’s Breakneck is all over my feeds, and rightly so—its a must-read for anyone interested in more than the usual caricatures of China, whether as an authoritarian menace or a hyper-builder state the West ought to envy. I'll spare you a recap of its key arguments since readers of this newsletter have, given its topic, likely already encountered ample coverage of this book. So let me jump right in with some of my own thoughts.

Development is the driving logic

Wang's overarching framing that the United States is a lawyerly society and China is an engineering state comes from his complaint that the labels capitalism and socialism—and most of all neoliberalism—do not adequately explain the two countries and often, in fact, leads people to mischaracterize them. A so-called socialist state might, for instance, be expected to provide a robust safety net for the poorest members of its society, yet China spends far less on social transfers than the United States, the self-styled champion of capitalism. For from a self-proclaimed leftist state, China is also extremely socially repressive—labor unions are illegal, queer culture is squashed, gender roles are hardened, and activists are disappeared—to the extent that even the most right-wing conservatives in the United States would salivate at China's policies. I emphatically share his view on the inadequacy of these labels, thus the need for new ones.

However, although socialism and capitalism may have been the dominant way to describe a nation in the 20th century (a Cold War habit perhaps), they were hardly the only ones. Coming from the development literature, another binary mode of comparison that comes to mind is the developmental versus the regulatory state, something that Chalmers Johnson first wrote about in his analysis of Asia's then-rising giant, Japan.1

As Johnson argues, the regulatory state (of which he includes the United States and other early developers) concerns itself with rules of economic competition but not with substantive matters of development—what the economy ought to look like, which industries boost, etc. By contrast, the developmental state (of which post war Japan is his main case) concerns itself with substantive matters like which industry matters for the national growth. The tools of intervention differ accordingly. The regulatory state erects strong legal institutions to protect markets (e.g., property rights). The developmental state, by contrast, empowers state agencies to pursue the agreed-upon goals—picking sectors or firms to support, mobilizing labor through targeted policies, and even setting export targets. The stuff of classic industrial policy.

When it comes to understanding the political economic logic of a nation, this distinction is more useful than the capitalist-socialist binary since it doesn't fixate on government size. Both regulatory and developmental states depend on a robust, professional bureaucracy. What matters is not whether the state intervenes, but how it intervenes.

To me, Wang's characterization of China as an engineering state is quite similar to Johnson's developmental state. In both postwar Japan and present-day China, we see a largely authoritarian state with strong views about the substance of the economy, backed by a capable and hierarchical bureaucracy able to act on those priorities. Both states also "engineered" labor suppression in order to become better building societies. And the outcomes (if we adjust for scale) are strikingly similar: extensive rail networks, robust urban infrastructure, and—most importantly—formidable manufacturing capacity. The similarities are striking, so much so that I might, and perhaps rather uncontroversially, say that China is a developmental state.

This distinction matters because whereas Wang attributes China's engineering state to the occupational choices of professional class, Johnson sees Japan's developmental state as a product of late development. On this point, I'm on Johnson's side. Japan's development is itself perhaps the best example. First of all, in the post-war years of massive build out, Japan's bureaucracy was staffed by lawyers, not engineers. And now that it has largely "caught up," it has shifted its orientation to something quite a bit more like the regulatory state: greater emphasis on strong legal institutions, less industrial policy of the classic sort, and more of a social safety net for its once heavily suppressed workers.

Things like social safety nets are less so the outcomes of the occupational leanings of the political elite; rather, they are more closely correlated to stages of development (even more so than political orientation). China is itself the best case-in-point: despite President Xi's fear that a social safety net promotes laziness (which needs to be socially engineered away), the country has nevertheless recently put in place new measures to expand its social safety net. This is quite uncharacteristic of Wang's "engineering-state" but perhaps unsurprising from a development perspective.

A key difference between post-war Japan and present-day China—and perhaps the reason such a comparison has eluded Wang's book—is in how each has presented themselves to the world during their era of breakneck building-driven growth: Japan a parliamentary liberal democracy vs China a Marxist-Leninist communist state. Both characterizations are smokescreens for what really matters: a hierarchical bureaucratic state with a clear view of what development requires—and a willingness to intervene directly in the substance of the economy to achieve it.

In the end, it is neither the Chinese government's Marxist-Leninist ideology nor its engineering-heavy occupational tendencies that has pushed it into one of the most capable building-nations there today. Lawyers or engineers, capitalist or socialist, which ever it may be, my money is that China will continue to build to drum beat of development.

Why does China have an abundance of process knowledge

This brings me to my next point about process knowledge and learning by doing. The point Wang makes with these concepts is that Chinese workers being the producers of most things in the world are more engaged in the process of technological learning than any one else; and when you have a whole city doing that, you end up with a community of engineering practice which makes the overall system much more adept at building stuff.

Part of what I love about Wang's writing is his ability to make even the most mundane ideas feel compelling. Learning by doing is an arcane academic concept that I've struggle to make sound like a profound insight when telling my friends about it. The idea is that the capacity for innovation improves through the cumulative experience of producing goods or services (that capacity is what Wang calls process knowledge).

This idea, as far as I know, was first written about by economist Kenneth Arrow in 1962, then brought into the context of development studies by way of Alice Amsden in her 1989 book about the rise of South Korea (also then a developmental state).2 What she points out is that the best way for late developers (e.g., South Korea, also Japan and Taiwan) to catch up is not by inventing new technologies, but learning from advanced countries by doing their manufacturing. Thus domestic firms must first enter labor-intensive, low-value-added industries, where they can work with more advanced foreign firms, learn from them, build process knowledge, and eventually climb the value chain.

This probably sounds a lot like its describing China as well. In fact, the first few chapters of Breakneck really reminded me of many of the insights found in Amsden's book. Amsden, in her introductory chapter, even emphasizes that engineers were key actors in the development of South Korea. But her reason is quite different. In her telling, the opposite of the engineer is not the lawyer but the scientist. Advanced countries can afford to fund the risky research of scientists that may or may not become commercially viable technologies. Late developers, by contrast, must prioritize engineers, whose task is to serve as a vessel of technology transfer, absorbing process knowledge from foreign firms and diffusing it across the domestic economy. In other words, engineers matter not because they can impose an "engineering mindset" on society, but because they hold the keys to learning by doing.

When it comes to process knowledge, what Wang quite correctly observes about China is its complete dominance in process knowledge at every layer of the value chain—a result of its willingness to keep even low-value added segments within its borders. This seems to break from most patterns of development where production tends to move wherever labor is cheapest. That was true for the United States, and was true too for Japan and South Korea which quickly offloaded lower–value-added production to less developed economies (like China in fact) in order to focus on higher-value industries. Within the East Asian countries, this pattern became known as the "Flying Geese" model, after the cascading "V" formation of the migrating birds. China, by contrast, has largely kept production—and its associated process knowledge—within its borders, rather than letting it spill out.

In Wang's telling, this is because the engineering state obviously loves to build and thus wants to keep that capacity domestic. This may well be the case. But another explanation is that the country is itself huge. With so many still living in the rural countryside, and provinces the size of entire countries, an entire flying geese formation exists inside of the country. Each time a Chinese firm climbs the value-added ladder, lower value-added segments aren't offshored abroad but absorbed by the country's own vast supply of rural labor eager for development opportunities. This has enabled China to keep all production domestic allowing it to accumulate process knowledge at every layer of the value chain. What this suggests is that China’s process knowledge abundance stems less from the sensibilities of its "engineering state" and more from the vastness of its rural population.

While I disagree about the source of China's process knowledge abundance, I agree that the concept itself has never been more important. In the fading neoliberal order and a post-pandemic push for self-sufficiency, many countries are now racing to domesticate or near-shore their supply chains. In this context, process knowledge is critical not only for late developers but also for advanced economies like the United States, and Wang is absolutely right to put this front and center of his agenda.

Concluding thoughts

I'm a PhD student in the international development group at MIT's planning school, so it would've almost been irresponsible not to have written a review of Breakneck. But with this training, what stands out to me is how many of the core lessons of late development—from Chalmers to Amsden to the developmental state—have been forgotten and pushed out of the mainstream; and that's apparent in Wang's book as well.

This is hardly Wang's fault. In the 1980s, Washington Consensus economists (wrongly) cast Japan and Korea’s success as free markets triumphs. And as both countries successfully developed and transitioned toward something resembling a liberal market economy, the very idea of the developmental state faded into obscurity. Wang's book, however, has been such a joy to read precisely because it has rediscovered and now also popularized many of these ideas in his account of China's political economy—and for that, Breakneck is worth celebrating in full.

At the same time, however, his framing of China as an "engineering state" rather than a developmental one further obscures this rich body of work on the developmental state, and in fact, risks overstating China's uniqueness in development. While his label is indeed memorable, it fails to put China’s trajectory within the longer lineage of late developers that used a similar playbook—state led industrialization, learning by doing, and building a whole lot of stuff.

This is not to say China is no different from Japan or Korea. It obviously is, and in some previous pieces (here and here) I've argued why. But to disregard how China fits into the larger story of global development would risk producing yet another caricature—a task that this book has worked so carefully to avoid.

(While working on this piece, I came across Jonathon P Sine's excellent review of Breakneck—data-filled, extremely sophisticated, and complements the broad strokes I've sketched out in my review. It’s long, but well worth read.)

Cogitations

Litigation Nation, Engineering Empire

This essay analyzes the big idea at the heart of Dan Wang’s new book Breakneck: that China is an “engineering state” facing off against America, the “lawyerly society.” Breakneck is well-informed, creative, and filled with wit. It also hinges heavily on vibes and intuition. Much of the intuition is compelling, b…

Read more

8 months ago · 166 likes · 27 comments · Jonathon P Sine

As Value Added has grown, it’s attracted a variety of audiences—policy folks, academics, and China watchers alike. While my PhD research focuses on the cloud sector and the AI transition, longtime readers will have noticed that for Value Added, I look at a much wider range of topics; from enterprise software, to the EV transition, to other developments in the knowledge economy—sometimes in a Chinese context, sometimes in the U.S., and occasionally in Europe.

This variety can of course be disorienting and feel unfocused, but I believe that putting different sectors, technologies, and countries in direct conversation opens up rich opportunities for comparison.

Comparison is itself an act of analysis. It sharpens our understanding of a political economy, an industry, or even a single firm by revealing features we might otherwise take for granted. The uniqueness of the American political economy, for instance, comes into sharper relief when set alongside other advanced economies; just as China’s distinctive approach to technology development becomes more apparent when we contrast it with the U.S.'s experience.

That’s why I write across what may seem like disparate topics and contexts: examining different sectors, technologies, and countries in parallel offers a powerful analytic lens. I believe you can’t fully understand how China or the U.S. innovates by looking at just one sector or technology in isolation. Whatever first brought you to Value Added, I hope you’ll find value in exploring these connections—and join me in using comparison as a tool for deeper understanding.

To help draw out these connections, this post highlights four key cross-cutting themes that I tend to return to across sectors, technologies, and countries—along with several pieces that best exemplify each theme.

1. State-business relations and industrial policy

Innovation is not the activity of a lone genius or even a single firm, but unfolds always within a political economy. And so, the basic premise of this newsletter is to understand how states and firms interact to generate the best or newest innovation outcomes. Indeed state institutions (education systems, labor regimes, trade associations, etc) and increasingly direct state coordination, as seen in the recent resurgence of industrial policy, are central in establishing the conditions under which innovation happens. A key theme of this newsletter is to examine how different institutions and different models of state–business coordination (whether in China’s AI ambitions or America’s manufacturing revival) structure the incentives, capacities, and trajectories of innovation.

1. The Political Economy of China's Cloud

2. Explaining overcapacity in China's EV sector

3. Industrial Policy for the Knowledge Economy: The case of Hangzhou

4. The "China as superpower" vs "China in crisis" narrative

2. Labor relations

Labor relations are an often overlooked aspect of innovation, but in an increasingly knowledge-intensive economy they play a crucial role in shaping how new technologies are developed, adopted, and deployed. From the internal incentive structures within the firm to the employment protections and labor laws of a country, the way labor is organized can determine whether innovation is extractive or generative—whether it enables creativity or simply disciplines workers to meet KPIs. In both China and the U.S., I’ve explored how different labor regimes condition not just productivity, but the very forms that innovation takes.

1. DeepSeek part 1: How new labor practices propelled an unknown AI firm to the top

2. Why layoffs are the new normal in tech

3. The labor foundations of China’s consumer-first internet

4. "Boss culture", 996, and KPI-driven management in China's tech sector

3. Firm strategies and corporate governance

If state–business relations shape the broad parameters of industrial development, firm strategies and corporate governance determine how companies navigate those parameters on the ground. Even within the same policy environment, firms can make very different choices about how to finance their activities, organize production, manage supply chains, and structure governance. And with new technologies with new business models, firm strategies are constantly forced to change. These decisions ultimately influence whether firms pursue short-term gains or long-term innovation, vertical integration or outsourcing, high-road or low-road models of growth.

1. From Fordism to the Platform Economy: The Evolution of Firm Strategies

2. The Return of Vertical Integration

3. The evolution of the cloud computing business model

4. A brief history of firm strategies in the auto sector

4. Tech transfer and diffusion

When we think about innovation, we tend to spotlight cutting-edge breakthroughs. But the real economic impact of a technology depends much more often on how it spreads. The extent in of its diffusion determines whether a new technology can transform an entire economy so that its productivity gains are felt widely or remains confined to a handful of firms (or labs) or sectors. Similarly, technology transfer—how technologies move from one country to another—is fundamentally about how capabilities move across borders or firms, shaping which countries benefits from innovation and who gets left behind.

1. How Every Superpower Stole Its Way to the Top

2. Innovation without invention

3. Rethinking AI diffusion in China

4. Learning from the learners

I’ve referenced Fordism quite a bit in previous posts, so it’s fitting this time to focus on the company that gave the term its name. The big unveiling this week was Ford's new plan for a $30,000 electric pick that would give the company a new lifeline as the industry turns to an EV future. Such an announcement is not entirely unexpected. Without a clear response to China’s fast-moving EV makers, legacy automakers like Ford they risk being left behind.

More than a price-competitive EV, what makes Ford’s announcement stand out is its new Universal EV Platform—a major reimagining of the overarching organizational innovation that made Henry Ford a household name over a century ago: the assembly line.

The new platform will use a so-called “assembly tree,” where the front, middle (housing the structural battery pack), and rear of the vehicle are built independently and in parallel, then brought together for final assembly. Ford also says the platform will cut parts count by roughly 20%, eliminate about 4,000 feet of wiring to make vehicles significantly lighter, and boost production speed by an estimated 15%. All of this is aimed at helping the company that once pioneered mass car production now build EVs at a price low enough to compete in today’s EV race.

Understanding which kinds of firms are capable of innovation and where within the market they belong has long been a focus of mine—and a recurring theme of this newsletter. We often think of breakthrough innovations as emerging from outside the circle of incumbents, driven by entrepreneurs on the periphery. This is a dynamic Joseph Schumpeter famously described as creative destruction. Ford’s latest move, however, is a striking example of an incumbent making an aggressive play to stay relevant.

Market leaders rarely have strong incentives to radically overhaul their products or processes when they already dominate—and this certainly applies to Ford, whose F-Series trucks have been among the best selling vehicles in the U.S. for decades. Newcomers, by contrast, have everything to gain by challenging the status quo, making them much more likely to adopt new and riskier technologies. Tesla’s popularization of the EV is the most recent and striking example in the sector. The company likes to characterize itself not as an carmaker, but as a Silicon Valley tech company that just happened to make cars—a clear outsider to the inner and much more established core of American car makers.

This is why Ford's newest move is worth paying attention to. Ford might never have undertaken such an aggressive overhaul of its product line and production methods without Tesla’s (and now China’s EV makers’) disruption to the market. But that doesn’t change the fact that such drastic changes within a company that continues to produce one of the nation’s best-selling vehicles is worth paying attention.

How can incumbents innovate?

New innovation can be extremely challenging for incumbents in any sector. Instigating processes that propel new innovation within a large established organization can be a bit like steering giant ship already moving at full speed. Their internal processes, hierarchies, and norms carry with them a kind of inertia, often designed—consciously or not—to reinforce the status quo.

The big Detroit automakers of the 20th century, Ford among them, excelled at mass-producing cars in an execution-focused environment built to crank out reliable products at scale. But that very strength meant also that these firms didn't encourage or reward unconventional or creative thinking—let alone the adoption of risky new technologies. And in the case of EVs, this makes sense as it can mean cannibalizing their existing and highly profitable product lineup.

How, then, can legacy automakers pivot to new technologies like EVs when so much inertia is working against them? How this happens is a key theme in my newsletter. In Ford’s case, the company took two important steps. First, it created a skunkworks project—the initiative that ultimately produced the new Universal EV Platform. The term "skunkworks" originated at Lockheed Martin during World War II, when a small, team was tasked with rapidly developing advanced aircraft free from the bureaucracy of the larger organization. The key idea is to carve out a part of the organization that can operate autonomously and take on much more experimental projects.

And that's exactly what Ford did with its new EV unit: set up an isolated, autonomous unit that operates outside the company’s usual structures, thereby giving it the freedom to bypass entrenched processes, break with conventional thinking, and take on much riskier approaches. As Ford CEO Jim Farley put it last year, the skunkworks team was "separate from the Ford mothership" and operated like "a start-up" within the broader company.

Bringing Silicon Valley and EV talent to Ford

Step two is not only to adopt the start-up ethos of Silicon Valley, but to quite literally bring in Silicon Valley veterans who've worked on EVs to lead the new effort. The first key hire by Ford was Alan Clarke, who spent 12 years at Tesla and played a key role in engineering Tesla's cars. He was then joined by Doug Field, another tech industry veteran who worked at Segway (the company's second hire), Apple (where he oversaw the best-selling MacBook Air) and Tesla (leading production of the Model 3 sedan).

Now serving as Ford’s Chief of EVs, Field has also become a talent magnet (here’s a nice profile of Field). Under his leadership, the legacy carmaker has brought in dozens of engineers from top EV players to their new EV initiative—last year alone hiring 50 from Rivian, over 20 from Tesla, about a dozen from Canoo, 10 from Lucid, and several from Apple’s now-canceled Project Titan.^[1]

These hires play a key role in transferring the tacit knowledge of EV production into Ford. They carry with them the lessons, best practices, and problem-solving instincts honed over years of designing and building EVs at scale. And now, by pairing this with the company’s manufacturing muscle, Ford is attempting to marry the innovation of the EV upstarts with the production capabilities of one of the world’s most experienced automakers.

Of course, it’s far too early to know whether this will propel Ford into the ranks of leading EV makers. The first $30,000 models aren’t slated to hit the market until 2027—a lifetime in a fast-moving industry where much can change before then. Still, the early production gains (e.g., the 15% boost in production speed) Ford reported last week are promising—and worth watching closely.

Global EV race

By now, China's BYD has prevailed as the largest and most relevant EV maker. And with it, its model has become the one to beat. But its clear from this week's announcement that Ford’s EV team isn’t trying to out-BYD BYD.

The Chinese automaker’s competitive edge comes from its massive scale (employing some 700,000 employees) and deep vertical integration—especially in battery manufacturing, where it controls everything from raw material processing to final assembly and now even recycling. That scale gives BYD an almost unassailable cost advantage. And Ford knows it can’t win at that game (especially while manufacturing in the U.S. where labor costs are so much higher).

Instead, Ford's Universal EV Platform team has focused on engineering innovation. In one example that Field described, they built a propulsion system "like Apollo 13," engineered down to the watt so that the vehicle’s battery could be dramatically smaller than BYD’s while still delivering competitive performance. If you can’t match a BYD’s cost structure on the battery, make the rest of the system so efficient that you don’t need as much battery in the first place. As Ford CEO Jim Farley said in a Wired interview, “We can’t beat them on scale. We can’t beat them on vertical integration. But we can beat them on innovation.”

This difference strikes me as one of incremental vs radical innovation. BYD’s approach is more incremental, building on its existing strengths—scale, vertical integration, and battery manufacturing prowess—making steady, cumulative improvements at every stage of their production line to lower costs and optimize performance.

Ford’s strategy, by contrast, feels much closer to radical innovation: rather than refining existing production processes and retooling their factories for the EV age, it’s rethinking core elements of car design and manufacturing from the ground up. This kind of leap, while riskier, has the potential to shift the competitive landscape in ways that incremental improvements alone rarely do.

Whether legacy carmakers can use this riskier approach to catch up in the EV race remains to be seen—and by the time Ford’s results are clear, BYD may have pulled even further ahead. Worse still, Trump’s “Big Beautiful Bill” is putting an end to EV tax credits at the end of September, making the timing of Ford’s Universal EV Platform unfortunately quite unlucky. Still, with Ford’s new EV initiative, it’s far too early to count the America's top legacy carmaker out of the game.

Last week’s earnings from Microsoft and Meta had AI boosters cheering. The former reported an 18% revenue jump, the latter a 22% surge—both far exceeding expectations. The big takeaway is that the multi‑hundred‑billion‑dollar gamble the industry is making on AI data centers is finally paying off.

Not long ago, investors worried that this massive outlay might not generate meaningful returns and result in a bubble. Last week’s earnings seemed to allay those fears. As a result, Microsoft’s share price jumped nearly 6%, enough to push it into the ultra-exclusive $4 trillion club with Nvidia. Meta's jumped over 11%. Together, the two firms helped propel the S&P 500 to all new highs. And with renewed confidence, the race to build ever-more AI data centers is back on full throttle.

The problem with this takeaway, however, is that the reported performance from Microsoft and Meta are not all that indicative of the productivity gains from generative AI. Instead, the structure of their earnings function more as an accounting smokescreen—designed to keep investor confidence high.

Let’s start with Microsoft. I used to work in the company’s cloud division, Azure. Back then, the way the company reported on its cloud earnings felt intentionally opaque. When investor attention was focused on Azure’s cloud revenue, Microsoft never reported Azure sales cleanly. Instead, it bundled them together with server products like SQL Server and other server software products under a broad umbrella called Intelligent Cloud.

Here’s how Microsoft defined "Intelligent Cloud" in its earnings reports:

Our Intelligent Cloud segment consists of our public, private, and hybrid server products and cloud services that can power modern business and developers. This segment primarily comprises:

• Server products and cloud services, including Azure and other cloud services, comprising cloud and AI consumption-based services, GitHub cloud services, Nuance Healthcare cloud services, virtual desktop offerings, and other cloud services; and Server products, comprising SQL Server, Windows Server, Visual Studio, System Center, related Client Access Licenses (“CALs”), and other on-premises offerings.

• Enterprise and partner services, including Enterprise Support Services, Industry Solutions, Nuance professional services, Microsoft Partner Network, and Learning Experience.

And it’s worth noting here that Microsoft’s server products, especially SQL Server, are absolute cash cows, making it especially tricky to really gauge how well Azure is doing. Even when revenue was broken out by product group, Microsoft still bundled "server products and cloud services" with "Microsoft 365 commercial products and cloud services."

On the one hand, this isn't completely non-sensical. Microsoft’s enterprise server and Office businesses are tightly linked to Azure. Customers already using these services are much more likely to also use Azure. On the other hand, this bundling also means that we don't ever get a clear breakdown of Azure's revenue. In fact, it was only last week at the company's most recent earnings call, more than a decade after Azure's performance was seen as central to the company's future, that Satya finally disclosed Azure revenue (surpassing $75 billion).

Microsoft is now reporting on AI revenue in the same way it did Azure. Whatever revenue Microsoft is making from its various AI services is bundled into this much larger Intelligent Cloud basket (i.e., you’ll see that “AI consumption-based services” is just one item on the list). So there's a good chance that the stellar performance from its most recent earnings call has nothing to do with the key technology driving the company’s massive splurge in data centers: generative AI .

Instead, the recent surge in the company's revenue is most likely attributable to increased usage of other cloud computing services on Azure. And given that cloud migration among the world's largest companies have been underway for years—well before the ChatGPT moment—it’s a safe bet that much of the revenue jump from last week reflects prior technological bets that Microsoft made rather than a sudden wave of generative AI adoption.

To be clear, having a strong cloud means that Microsoft is likely to be in a strong position to capture the gains from spreading AI across the economy once the technology is truly ready—but that doesn’t mean it’s happening yet. The technology driving the hundreds of billions of dollars in data centers investment is still very much in untested territory.

Meta’s case is probably even less indicative of generative AI usage. Its even strong earnings performance comes from its ad business, which does use AI—but the traditional machine‑learning variety, not the generative kind driving their massive spending on data‑center build‑out.

The diffusion potential of Meta’s AI services is also far more limited. Unlike Microsoft and other cloud providers, Meta lacks the ability to sell AI services across industries—a key channel for widespread adoption. Yet the company is building data centers with the entire economy in mind. So even if Meta fully embedded generative AI into its ads and social platforms, it’s hard to see how that would produce an ad‑revenue jump large enough to justify the tens of billions in new infrastructure. To make that math work, Meta would have to create entirely new markets. In other words, there’s still no concrete sign that all this AI investment will pay off.

Last week’s blockbuster earnings say more about the past than the future. Microsoft and Meta are reaping the rewards of much earlier bets—enterprise cloud services in Microsoft’s case, and ever‑more‑sophisticated advertising for Meta. That doesn’t mean today’s AI data‑center splurge won’t bear fruit. But whether it will is a question we’ll have to look elsewhere to for answers.

My recent fascination with understanding the auto sector comes from the observation that a country’s position in the global economy seems often to be linked to the competitiveness of its auto industry. This was certainly true when the United States first pioneered the mass production of cars and became the world's industrial powerhouse in the early-to-mid 20th century. It was true again with Germany developing its manufacturing dominance in the postwar decades, and then with Japan's rise in the 1980s as a new global leader on the market.

If dominance in automotive manufacturing is a good heuristic for a country's economic strength, then the EV revolution must be considered one of the defining contest of the 21st century. And in this new race, China has outpaced previous incumbents, position itself at the forefront of the EV revolution.

Each time a new leader emerged, they did so not only because of new technological innovations but because they pioneered new organizational strategies that allowed them to build cars more efficiently, flexibly, and competitively than their rivals. In this aspect as well, EVs are no exception. As I've written in previous posts, the top EV makers (most of all BYD) today depart from the more fragmented supply chains of the past few decades and are instead vertically integrated behemoths. But as new players enter into the market, greater variation in firm strategies are also emerging. There seems to be, at this time, no one best strategy that EV makers are conforming to.

This post is traces how firm structures in the auto sector have evolved over time—in particular how automakers puruse different approaches to labor and supply chain governance in order to stay competitive. It offers a brief (and mostly American) overview of that history, from Fordist mass production to lean manufacturing, modularization, and finally to today’s unsettled EV landscape.

Fordism and the rise of mass production

The mass production of automobiles began in the United States in the early 20th century under the Fordism mode of production. Since I've written about Fordism in a previous post, we can forego a lengthy discussion of it here.

In short, Fordism was defined by massive, vertically integrated supply chains housed within geographically anchored corporations that were both labor- and capital-intensive. These production sites carried high fixed costs and employed large numbers of workers, making them especially vulnerable to labor unrest. To mitigate this risk, Fordist firms maintained stable relationships with labor unions and shared the gains from automation and rationalization, ensuring a degree of labor peace and long-term productivity growth.

Under this arrangement, unionized workers enjoyed high wages, job stability, and generous benefits, while employers had strong incentives to invest in machinery and production techniques to increase productivity over time. Labor and capital thus held each other in check, creating a stable and relatively equitable model of industrial production. The problem with the Fordist model, however, was that production was highly standardized and relied on relatively low-skilled, assembly-line tasks—leaving little room for bottom-up innovation.

The lean manufacturing revolution

The Fordist model broke down in the post war decades, as German and Japanese automakers emerged as formidable global competitors. The rise of Japanese firms, in particular, marked a decisive shift away from rigid vertical integration toward what came to be known as lean manufacturing—most famously described in James Womack, Daniel Jones, and Daniel Roos's 1990 landmark book, The Machine That Changed the World, which remains the definitive account of Toyota’s transformative production system.1

Whereas Fordist production prioritized the mass output of standardized, homogenous products with minimal variation in design, Japanese automakers like Toyota pioneered a more flexible approach that could adapt to diverse market niches and shifting consumer preferences. Emphasis was placed on continuous, incremental process innovation at every stage of the supply chain. And workers played an active role in identifying bottlenecks, solving problems, and proposing improvements—breaking down traditional hierarchies and embedding innovation into daily operations.

Under the new lean manufacturing paradigm, supply chains became more decentralized and relational, composed of networks or "families" of specialized firms. These systems relied on long-term collaborative partnerships and extensive information sharing, particularly with competent first-tier suppliers who could independently innovate while remaining closely coordinated with lead manufacturers.

One best way?

The rise of lean manufacturing in Japan hit American automakers with billion in combined losses, and by the early 1990s, Ford, Chrysler, and GM began implementing lean practices. This new paradigm touched nearly every part of these American firm’s production strategy—from the factory floor to supplier relations, workforce organization, and engineering design.

A 1992 Wall Street Journal article documented Chrysler’s efforts to reform its internal labor structure.2 Taking cues from Honda, the company dismantled organizational silos separating engineering, design, and marketing, and introduced mechanisms for bottom-up innovation. Employees at all levels and in all departments were now expected to contribute to process improvements—a radical departure from Chrysler’s previously more top-down model.

To Womack and his co-authors, these changes were expected as they saw lean manufacturing as a universally superior system that all automakers would eventually converge on. But while these efforts produced some tangible gains, wholesale adoption of the lean paradigm proved difficult for these U.S.-based automakers.

In the United States, shareholder pressure for short-term returns (a product of the Shareholder Value Revolution, which I've written about here) undermined efforts to build durable, trust-based supplier relationships. Firms like Ford and GM reverted to outsourcing strategies built on inter-supplier competition, maintaining arm’s-length ties that prioritized cost-cutting over co-development.

Since then, new evidence challenged the convergence assumption. Auto manufacturing continued to vary significantly across countries—a divergence shaped by deeply embedded national institutions, including labor laws, vocational training systems, corporate governance structures, and long-standing social compacts between labor and capital (see One Best Way? by Freyssenet et al., 1998).3

Modularization

By the 1990s and 2000s, as the industry began to design vehicles around standardized and interchangable parts, the auto industry embraced yet another paradigm: modularization.

Under this model, cars were broken down into distinct modules—such as the chassis, powertrain, and interior systems—that could be developed and assembled semi-independently. Each module followed standardized design interfaces, enabling different suppliers to work on components in parallel while ensuring compatibility at final assembly.

Crucially, modularization also made it easier for lead firms to outsource key parts of the production process to a broad network of arm's-length suppliers without requiring the tight, long-term relationships characteristic of the Japanese supply chains. But in doing so, it also forfeited some of the dynamic learning and process innovation potential that lean manufacturing had harnessed.

The diagram below offers a useful typology to trace this historical evolution in supply chain governance.4 The Fordist model, situated on the far right, is best described as hierarchical or captive, with tightly controlled, vertically integrated supply chains concentrated within a single firm. Then, over the next century or so, production systems have moved progressively to the left. Lean manufacturing introduced more relational forms of coordination, emphasizing long-term collaboration and trust-based ties with a smaller number of key suppliers. This eventually gave way to modular supply chains, where standardized interfaces allow for loosely coupled relationships and greater supplier autonomy—prioritizing flexibility, cost efficiency, and ease of substitution over deep inter-firm integration.

The EV Transition

Since the era of modularization, frontier corporate strategies in the automotive sector had remained relatively stable. But with the rise of EVs, automakers are once again forced to rethink their production strategies.

From a supply chain governance perspective, what’s most striking about the EV transition is the revival of vertically integrated production models reminiscent of the Fordist era. Leading the way in this are BYD and Tesla—both of which control a substantial portion of their value chains, from battery cell production and vehicle assembly to proprietary software development.

Yet as the battery technologies mature and legacy automakers enter the market, alternative strategies are also gaining traction. Companies like Ford and Volkswagen have maintained modular supply chains, especially with battery manufacturers. Batteries are also now the most complex, valuable, and strategically critical component of EVs; they are also highly modular. This has tilted power away from traditional automakers and toward a new class of specialized battery markers, particularly in Asia, where firms like CATL, SK Innovation, LG dominate.

Between the vertically integrated first movers (Tesla and BYD) and the more modular approach of legacy carmakers (Ford and Volkswagen), what the best model is remains an open question. It is also one that will also be contingent on the national institutional contexts in which the lead-firm is situated in.

If a country’s position in the global economy is linked to the competitiveness of its auto industry, then understanding how firm strategies are evolving in the EV age must be considered one of the most consequential—and still unfolding—questions in industrial strategy today. And it’s an evolution I’ll continue to follow closely and write about here on Value Added as the sector matures and the contours of the next industrial era take shape.

It’s becoming increasingly clear that China is now competing head-to-head with the United States at the frontier of large language model (LLM) development. Models like DeepSeek’s R1 and V3 have demonstrated capabilities rivaling, and in some cases exceeding, those of leading U.S. models. Meanwhile, Alibaba’s Qwen3 has climbed to the top of global benchmarks in reasoning and language tasks. These achievements are all the more notable given the constraints Chinese firms face under U.S. export controls, which limit access to state-of-the-art Nvidia hardware.

By some measures, China is no longer playing catch-up in AI model development—it is already there. But reaching the technological frontier is only part of the story. As Jeff Ding has argued, what really matters for long-term economic advantage is diffusion: how widely and effectively these innovations are adopted across the broader economy.

This post weighs in on the ongoing debate around AI diffusion in China. Broadly speaking, I've encountered two contrasting views. On one side is the more common view which argues that China is well positioned to harness the benefits of AI and, as with other critical technologies, will eventually surpass the United States in both adoption and impact.

On the other side is Jeffrey Ding, who argues that China faces a diffusion deficit—that is, a gap between cutting-edge AI development and its widespread adoption across the economy. And because diffusion, at least in Ding's analysis, is key for economic prowess, China's AI diffusion deficit means that we're overestimating the economic impact that AI will have in China.

From there, I'll add my own position—that China does have a diffusion deficit but for reasons quite different from what Ding has argued.

The diffusion advantage argument

Most accounts fall into the camp of seeing China as having an edge in the diffusion of AI—and perhaps unsurprisingly. After all, the country does have a strong track record for rapidly diffusing critical technologies. Mobile payments, for example, reached mass adoption far more quickly than in the U.S.—by 2019, 35% of Chinese consumers were using mobile payments compared to just 8.8% in the U.S. High-speed rail is another example with the state pushing for its adoption and building some 30,000 miles of rail since Xi took China's top position in 2013.

The most recent and arguably most salient example is EVs with China going from a marginal player to the global leader in EV production and adoption in the span of several years. Thanks to a combination of state subsidies, local government incentives, infrastructure build-out, and industrial policy targeting both supply and demand, China accounts for two-thirds of global EV sales as of 2024. These examples suggests that China can diffuse new technologies throughout its economy at quite a remarkable speed, and the idea is that the same capacity for diffusion will be applied to AI.

As Kaiser Kuo has argued, part of the reason China has an edge in the diffusion of AI is cultural. Chinese society, he suggests, tends to have "a techno-optimistic orientation where innovation is associated with improved material well-being, state capacity and national rejuvenation." By contrast, the U.S. has what he calls a more "Black Mirror" mindset: a cultural tendency to view new technologies, especially AI, with suspicion or dystopian anxiety. As a result, U.S. domestic demand may be more restrained. Indeed, the field of AI ethics and the concerns it raises about the unrestrained pursuit of AGI is virtually nonexistent in China, meaning the country are free of the speed bumps that would normally slow the rollout of AI in the West.

Beyond cultural factors, Kuo has also pointed to the structural advantages of China’s tightly coordinated innovation ecosystem—specifically, the close alignment between public institutions, academic research, and private firms. He cites examples like Zhipu AI, which emerged directly from a university lab and was funded by state-affiliated venture capital, as evidence of this state-academia-industry nexus in action.

What he’s describing reflects a broader pattern across many of China’s high-tech sectors: first, the state provides patient capital to support long-term, high-risk projects that would struggle to survive under the short-term pressures of profit-driven private investment; then, rather than sheltering these ventures, it exposes them to intense domestic competition such that market pressures impose discipline and drive up performance. The result is a system where firms, pushed by fierce competition, bring down costs and scale rapidly—accelerating the mass diffusion of technological breakthroughs, as seen in China’s electric vehicle and solar manufacturing sectors.

The diffusion deficit argument

The "China is good at technology diffusion" is certainly convincing, but if we look carefully at the types of technologies that have successfully diffused, we can see that they largely fall into two categories—consumer-facing tech or large-scale infrastructure. Mobile payments, EVs, and AI apps like DeepSeek or Yuanbao installed on smartphones are all examples of consumer technology. These tend to diffuse quickly because the barriers to adoption are low. More importantly though, their spread doesn’t necessarily translate into substantial productivity gains across the broader economy. On the other hand, infrastructure projects like high-speed rail succeed not because of organic market uptake, but because the state can orchestrate their rollout from the top down and absorb short-term losses if these ventures aren't immediately profitable.

These two types of technologies differ quite a bit from the type of technology that characterizes AI—or what Jeffrey Ding, among others, would call a general purpose technology (GPT). Unlike infrastructure-based technologies, the diffusion of a GPT can’t be mandated from the top because adoption depends on market viability and firm-level incentives. And unlike consumer tech, its transformative potential doesn’t come from widespread consumer adoption, but from its capacity to enhance productivity across a broad range of industries. What truly matters in the diffusion of AI, then, isn’t how many people are chatting with DeepSeek, but how many firms are embedding AI into their business processes, reshaping workflows, and leveraging it to gain a competitive edge.

Indeed, AI diffusion in China may not be has robust as it first appears. This, in fact, is one of the core argument Ding makes in a recent article. While China has clearly produced a number of frontier innovations—evidenced by the growing roster of cutting-edge, homegrown LLMs—the more critical, and often overlooked, factor in translating technological breakthroughs into economic power is diffusion. And on that front, Ding argues, China is behind.

Technology diffusion is tricky to measure and any attempt to do so is requires creativity. In Ding’s case, he counted the number of universities in each country that have placed at least one research paper at one of leading AI conferences. This, he argues, is a good proxy for understanding the geographic diffusion of AI because it represents how widely AI expertise is distributed. His findings was that China had 29 such universities; the U.S. had 159. This gap suggests that beyond China’s elite universities and tier one cities, the country lacks the institutional depth needed to support widespread AI diffusion such as the kind of skilled AI engineers who can integrate and deploy this general-purpose technology across diverse sectors of the economy.

To explain China’s weak AI diffusion—and the U.S.’s relative strength—Ding zeroes in on the role of education and training institutions in expanding the talent pipeline. Skills/education are key because AI is highly technical and requires sophisticated engineering know-how in firms that are adopting AI (e.g., knowing how to monitor the system, how to update the system, how to benchmark between models, and how to integrate these models into existing workflows, etc). And so, it’s not just about producing elite researchers at top universities or firms, but about cultivating a broad base of engineers across regions who can work with, adapt, and implement AI technologies throughout the economy.

You have to walk before you can run

Ding is a political scientist, and his argument, in classic political science fashion, is essentially an institutional one: China’s AI diffusion weakness lies in its lack of a broad base of training and skills institutions beyond the handful of elite ones in top tier cities. And without this broad foundation, China’s ability to diffuse AI widely—and to translate frontier innovation into economy-wide productivity gains—is fundamentally constrained.

I’m quite sympathetic to Ding’s argument and agree that China does—and likely will continue to—lag behind the U.S. when it comes to AI diffusion. But if what we care about is whether firms across sectors and across China's geography can adopt AI for productivity-enhancing purposes, then skill institutions are really just a small part of the explanation.

As I see it, what matters more is the level of digitization across China’s economy—which, compared to the U.S., still lags significantly behind. The intuition is simple: for firms to adopt AI, they must have basic IT practices in place e.g., storing data in a structured format, ability to process large quantities of data, etc. Without these basic IT capabilities, adopting AI would be like learning to run before knowing how to walk.

[Updated July 22nd, 8pm: Several readers have mentioned that Jeffrey Ding does talk about IT spending lagging in China as a cause for slow adoption of AI in China. That’s my mistake for overlooking this point.]

To assess how far along China is in terms of digitization, we can turn to several proxy indicators—but perhaps the clearest is IT spending, both in absolute terms and as a share of GDP. As the chart below shows, China ranks lowest among this group of seven advanced economies in IT spending relative to GDP. Even in absolute terms, China spends just one-tenth of what the U.S. does, trailing not only the U.S. but also Germany, the U.K., and Japan. On top of that, Chinese firms devote less of their IT spending toward software services (as opposed to hardware/infrastructure) compared with those in other large economies. Combined, this suggests that China’s digital foundation remains comparatively underdeveloped.

Spending on IT/digitization is one issue—the weakness of China’s tech sector when it comes to delivering IT infrastructure and services is another. As I’ve argued in previous posts, China’s cloud computing sector significantly trails that of the U.S. The more robust the cloud is, the easier it is to adopt AI solutions because it lowers the barriers to entry: firms don’t need to invest heavily in their own computing infrastructure (which is both capital and human-capital intensive) to access AI tools. The gap is even more pronounced in enterprise software (or SaaS), where China lacks the kind of mature, integrated ecosystem found in the U.S. (as of 2024, the U.S. had over 9,000 SaaS companies; China only had 443).

These limitations—namely, low levels of digitization on the demand side and an underdeveloped supply of digital infrastructure (cloud computing and enterprise software)—pose serious constraints on the broader diffusion of AI. In this sense, the challenge runs even deeper than what Ding suggests. It’s not just a question of skills institutions; it’s also about the availability of patient capital, the structure of labor markets, and the incentives firms face to invest in long-term digital transformation. And there's also the simple fact that digital technology adoption has a certain path-dependency to it: already using digital services makes it easier to adopt new digital services such as AI.

Concluding thoughts

It might be a little odd to argue that China is not digitized in 2025. After all, Chinese users seems to be far more advanced than their western counterparts in e-commerce, food delivery, mobile payments, and essentially living on their phones. But that's because the adoption of digital technologies in China has largely happened in the realm of consumer products. The story is quite a bit different when we look at how firms are bringing digitization into their business processes. With the exception of the tech sector (and increasingly biotech too), China is still far behind on this dimension.

For AI, this means that in the near term, adoption will likely be driven by China’s sophisticated and fast-moving consumer-facing internet giants, which are already integrating AI into their expansive portfolios of apps and services. As a result, we can expect broad adoption of consumer AI tools. But as I’ve argued, consumer-facing adoption alone doesn’t generate the kind of productivity gains needed to significantly boost the broader economy. The longer-term question is whether China’s SMEs and other non-tech industries will have meaningful AI adoption. The fact that many of firms across the country have yet to adopt even basic digital services makes adopting AI a far greater challenge—after all, you have to walk before you run.

Given the interest in last week's post about workplace culture at Alibaba, I'm reposting this piece on the same topic that looks at more broadly at 996 culture across China's tech sector; and why internet companies have tried to curb the practice that brought them so much success. I originally wrote this piece for Dissent Magazine in 2022, but I think it’s worth revisiting as China’s internet firms pivot to AI. It sheds light on the organizational and labor hurdles these companies face—and speaks to one of the core themes of this newsletter: that labor relations are a crucial lever of innovation. The link to the original piece is here.

The term 996 has come to denote a culture of overwork in China’s tech sector. The numbers refer to a work schedule that runs from 9 a.m. to 9 p.m. six days a week, though they represent less a fixed schedule than an expectation that tech workers will consistently put in grueling hours. Even with a seventy-two-hour workweek, jobs in the country’s fast-growing internet industry are among the most sought after by young professionals. Skilled tech jobs are on average among the highest paid in the country, with university graduates at top firms able to earn salaries of up to 600,000 renminbi (roughly $70,000) out of college, and more than twice this amount with just a few years of experience. But not all tech workers are treated equally. Contract workers, non-technical workers, and workers with niche technical skills, such as front-end web development, earn only a fraction of that income but are still expected to work 996.

Although 996 has been practiced by internet companies for more than a decade, the term itself was only popularized in 2019 by a viral worker-led protest campaign known as the 996.ICU movement. (The idea is that working 996 will put workers in the ICU.) The campaign, which demanded that tech firms obey laws against excessive overtime, took shape as a project on the code-sharing platform GitHub and quickly gained traction on social media. The movement ignited a nationwide discussion about punishing work schedules, but most tech firms continued to practice 996.

In 2021, however, Chinese tech workers finally began to catch a break. Over the past year, Meituan, Tencent, Kuaishou, JD.com, ByteDance, and many other Chinese internet giants, which had alternated between five- and six-day workweeks (known as the “big-small week”), announced that they were bringing back the two-day weekend. Last year, a Tencent-owned gaming studio rolled out a policy that would require employees to clock out by 6 p.m. on Wednesdays, and by 9 p.m. other days of the week. Another Tencent division, which runs WeChat, is piloting a “1065” schedule (10 a.m. to 6 p.m., five days a week) this year. Staying at the office beyond those hours requires submitting a formal request.

Why is China’s booming internet industry putting an end to a practice that has fueled so much of its growth? One reason might be because workers have mobilized to push for humane working conditions. But it is also important to recognize how ending 996 might align with the Chinese government’s broader goals for national development—and how the new regulation might even serve the interests of the industry’s leaders.

Worker Pressure and Public Outcry

Just months before the viral 996.ICU campaign, the Chinese economy began to slow down, partly as a result of a trade war between the United States and China. Venture capital dried up in what is now remembered in China as tech’s “capital winter.” Many internet companies started to lay off employees en masse, slash bonuses, and skimp on office perks. After years of industry expansion, the myth of limitless growth was crumbling, and programmers—once regarded as a professional class—began to shed their allegiance to the entrepreneurial elite. A post on social media written by a seasoned programmer in 2020 summed up the new mood: “The real winners [of 996] are the bosses. They pay the same sum of money to buy more hours of labor. The losers, on the other hand, are workers because they have to work many more hours for the same salary.” After the viral campaign, tech workers have remained vocal about the hardships of working in the industry. Many have taken to Maimai, a LinkedIn-like service that allows employees to post anonymously, and Zhihu, a forum website similar to Quora, to vent their frustrations.

Tech workers’ grievances have led to the use of new self-deprecating terms: workers call each other manong (coding peasant) and dagongren, which loosely translates to “worker” but is typically reserved for describing rural migrants in low-paying and low-skill jobs. While working conditions for white-collar and blue-collar workers are considerably different, the use of dagongren—even if only jokingly—points to a collapse of these class identities. Office workers also often call themselves and each other shechu—a combination of the terms huishe (company) and shengchu(livestock). The term was originally used in the 1990s by disgruntled workers in Japan.

This new vocabulary refers not just to long hours but also to working conditions at offices, which tech workers now call hulianwang dachang (big internet factories). Despite their companies being housed in modern skyscrapers or at beautiful campuses, they have started to liken their workplaces to factory floors—and for good reason. In 2020, Renwu magazine released a widely circulated exposé of various tactics that internet companies use to prevent employees from spending too much time in the bathroom. At Kuaishou, a Chinese video-sharing social media platform, office bathrooms are equipped with timers that track how long employees are in the stall. According to workers, ByteDance, Kuaishou’s biggest competitor, blocked Wi-Fi and 4G connections in its bathrooms to discourage employees from dawdling on their phones.

Beyond tracking productivity, companies have also demonstrated an utter lack of concern for employee well-being. According to Renwu, there was a severe shortage of bathrooms at the offices of Pinduoduo, one of China’s fastest growing e-commerce companies. Some floors of Pinduoduo’s headquarters were crammed with more than a thousand employees yet had as few as eight bathroom stalls. Employees regularly lined up for over twenty minutes to use the restroom. If the lines were too long, many would sneak into neighboring office buildings or to public bathrooms in a nearby mall. The company responded by hanging a sign in the bathrooms: “Hope everyone can hurry up.”

The cost of overwork has also been paid in blood. At the end of 2020, a young Pinduoduo employee died after collapsing during her commute home, at 1:30 a.m. Shortly after, another employee committed suicide after returning to his hometown. These deaths—said to be related to harsh working conditions—rekindled the public backlash against the industry’s work culture. Many even compared them to the Foxconn suicides of 2010, when eighteen young workers attempted to take their own lives, leading to fourteen deaths, as a result of excessive working hours and a regimented factory culture. Adding fuel to the fire, Pinduoduo, less than a month after the deaths of its two workers, allegedly fired an employee for “defaming” the company by posting about an unconscious coworker being rushed to the emergency room, presumably owing to overwork. Afterward, the employee recorded another video to publicly share the horrible working conditions inside Pinduoduo: food served by the company was often spoiled, employees worked 300 to 380 hours each month, and days off for national holidays were illegally shortened.

Ending 996 for National Development

Amid the public outcry over 996, the Chinese internet sector became the target of an onslaught of legislation. China’s “tech crackdown,” as the period is popularly known, started in 2020, when Beijing took aim at the tech entrepreneur Jack Ma, the founder of Alibaba and Ant Group. Days before Ant Group’s planned IPO in Shanghai and Hong Kong (which would have been the biggest in world history), Ma had publicly criticized the country’s financial regulators for stifling innovation, declaring that their rules were antiquated and in need of reform. Shortly after, the Chinese government blocked Ant Group from going public. The New York Timesdescribed the pulling of the IPO as a move to let “private companies know where they stand next to the state.” Ma went missing, only to reemerge in Hong Kong several months later.

The action against Ant Group was just the tip of the iceberg of plans to rein in China’s internet giants. In the ensuing months, the government embarked on a sweeping attack on the country’s biggest and most powerful internet companies on issues such as antitrust, data security, and algorithmic control. In some regards, Chinese officials were merely catching up to European standards for internet regulations and trying to preempt the kinds of data-privacy and security scandals that have tarnished American politics. But the crackdown was nevertheless catastrophic for investors, resulting in fines of millions of dollars for companies and wiping out billions in company valuations.

The new regulations also included labor protections. Food-delivery platforms were required to guarantee drivers a minimum income and implement more lenient delivery deadlines. The e-commerce platform JD.com and the ride-hail giant DiDi set up labor unions for their staffs—a peculiar management-led arrangement meant to appease regulators who have criticized these companies for exploiting and mistreating their hourly workers.

China’s top court also delivered a blow to 996, declaring the practice illegal. Many workers saw this as an indication that the government stood on the side of the anti-996 campaign. Skeptics doubted whether condemnation from the state would result in any actual change; 996 had long been considered illegal, and no concrete policies were created to regulate the schedules of office workers.

Given the Chinese Communist Party’s history of crushing attempts at worker organization and disappearing labor activists, it’s safe to assume that the government’s new stance didn’t stem purely from a deep concern for labor rights. Instead, these new protections should be situated as part of President Xi Jinping’s new emphasis on “common prosperity,” a framework he announced in the summer of 2021 following a proclamation of the end of absolute poverty in China. With its focus on tackling the country’s gaping inequality and moving toward an economy where the middle class owns the majority of wealth, common-prosperity rhetoric has been framed benevolently around uplifting working people. The move has even won the admiration of a number of leftists in the West, who sometimes go so far as to interpret these changes as the party’s return to socialism. In an article that Xi wrote in Qiushi, the official theoretical journal of the CCP, he makes clear that common prosperity is not an initiative with the sole goal of uplifting working people. It is intended as an equalizer of opportunities to foster entrepreneurship and ultimately grow the economy. Prosperity, as Xi says, can only be achieved through “ingenuity and effort.” For Xi, growth is the primary goal, and greater equality is just a means of getting there.

Common prosperity should be considered alongside the strategy of dual circulation, which Xi announced the previous year. The idea is that Chinese goods and services should circulate both internally and externally: external circulation keeps the Chinese economy open to trade, while internal circulation, arguably more important, boosts domestic consumption and sheds reliance on export-led development. The latter would require curbing inequality and increasing the incomes of everyday people.

If the goal is to boost domestic consumption, 996 presents a problem.

Because tech workers log nearly double the hours required of a regular full-time job, the industry offers fewer jobs than other sectors but also higher pay. The result is starker inequality, which is more than just a social problem. In Trade Wars Are Class Wars, Matthew C. Klein and Michael Pettis argue that Chinese inequality, characterized by the low share of national income that goes to ordinary households, is the culprit behind the country’s low rate of consumption. An additional dollar given to a poor person will likely be spent immediately, whereas the same dollar given to a rich person is more likely to be saved. In other words, those without work have no money to spend, while those who do work end up saving their money.

Worse still, the typical 996 worker is single, lives away from home, rents a cheap room to sleep in, and spends the bulk of their waking hours at an office where every daily necessity, from food to entertainment, is provided to them. Even with salaries well above the national average, some tech workers have found that there isn’t much opportunity to spend the money they earn.

The 996 schedule also worsens the problem of China’s declining birth rate, a longtime concern of the government. At many internet companies, 996 has created an unspoken “up or out” policy. If a worker hasn’t progressed to a managerial position by the time they have children, they risk getting replaced by someone who is able to keep up with the grind. The problem has become so endemic among tech companies that the typical career path of a programmer has been turned into a joke: at twenty-five, you get hired as a software engineer; at twenty-eight, you are promoted to a senior software engineer; at thirty-two, you are promoted to architect; at thirty-five, you become a delivery driver. Men who can maintain the 996 schedule are often unable to participate in domestic labor, which typically is then left to women, in turn pushing them out of China’s already hypercompetitive job market. In response, many tech workers are opting to forgo parenthood altogether.

Given these developmental concerns, ending 996 makes a whole lot of sense for the state. Older workers won’t have to worry about keeping up with the schedule, which could increase the industry’s age ceiling and expand the job market. More workers could have children without risking their jobs, thereby alleviating the problem of the country’s aging population. And the potential to reduce inequality, however minimally, furthers the goals of common prosperity.

Winner-Take-None Competition

In studying rice production in Java, Indonesia, the anthropologist Clifford Geertz found that an increase in the number of farmers on each plot of land resulted in diminishing output for each farmer. While the total output increased, productivity per worker stagnated. To describe this phenomenon, he coined the term “agricultural involution.”

“Involution” found its way to China in 1986, by way of a Princeton-educated scholar, Huang Zongzhi, who studied the agricultural economy in Northern China. But it wasn’t until the late 2010s that the term started to circulate on the internet, alongside a series of images that depicted the toil of university students. The most famous of these showed a student studying while cycling, with his laptop delicately balanced on the handlebars.

The term’s Chinese translation, neijuan (inward roll), conjures a picture of a tightening spiral, which has resonated for young people across the country. Each year, students study strenuously for the gaokao, China’s mandatory higher-education entrance exam, and compete against millions of others for coveted spots in top universities. The harder everyone works, the more you have to study just to keep up.

By 2020, involution had also struck a chord with China’s tech workers, capturing a widespread feeling of meaningless competition. Managers pitted employees against each other for vanishingly few spots higher up the hierarchy, and promoted an ideology of meritocracy: for those who made it to the top, their success justified their hard work; those who didn’t had only themselves to blame. The only answer was to work harder.

More recently, workers have begun choosing to tangping (lie flat). Initially, the term was used as a humorous euphemism for quitting one’s job. Similar to manong and shechu, tangping embodies a defeatist attitude toward long hours and poor working conditions. But in the summer of 2021, tangping zhuyi (tangping-ism) emerged online as a concerted and coordinated call to opt out entirely. The movement was a rejection of China’s ruthless economy, from grinding at a 996 job to the social pressure to buy a house. Instead, tangping zhuyi celebrated giving up: quitting a high-salaried job to work 1065 for a modest wage; moving to a lower-tier city; reducing one’s expectations. As tangping zhuyi spread, the government censored several WeChat groups where it was discussed, out of a fear that the ideology would undermine the state-backed ethos of hard work. Elites condemned the movement as a symptom of laziness.

Although neijuan is commonly used to lament overwork at China’s hypercompetitive internet firms, it has also been used to describe the sector as a whole. For most of the past decade, Chinese internet companies have operated with an assumption that growth is virtually limitless. At first, competition wasn’t as fierce, and working overtime wasn’t as common. But in 2009, Google, Facebook, and Twitter were ousted from China, following the circulation of posts about Tibetan protests on the platforms. Chinese companies had a chance to fill the resulting vacuum and seize the country’s rapidly growing digital market. Baidu replaced Google in search. Tencent snagged Facebook’s market in messaging and social media. Weibo, which didn’t even exist at the time of Twitter’s ban, quickly became the go-to micro-blogging website. In other cases, American companies weren’t banned but simply outcompeted, including eBay and Amazon, which lost out to Alibaba’s Taobao.

The mid-2010s were years of opportunity, or fengkou (wind tunnels). The term came from Xiaomi founder Lei Jun, who in 2015 offered a famous description of Chinese tech: “even a pig could fly if it stood in the wind tunnel.” What he was pointing to was how even a fool could make it big as long as they picked the right opportunity—social gaming, AI, live-streaming, blockchain, community group buying. In the era of fengkou, Chinese start-ups pursued money wherever it flowed, moving into market segments that became crowded with dozens—if not hundreds—of copycats. Copying wasn’t stigmatized; in fact, it was a prerequisite to enter the race. When Groupon first entered the Chinese market, there were literally thousands of competing companies that delivered near-identical services. The sheer density of firms meant that the slightest inefficiency or not having the latest functionality gave competitors the edge. As a result, workers were tasked with constantly replicating, tweaking, iterating, and experimenting. The only way to come out on top was to outwork the competition.

As 996 became ubiquitous among China’s internet companies, markets grew even more saturated, and growth slowed. The problem of excessive competition was also compounded by demographic trends. Until 2012, China had double-digit year-over-year growth in the number of internet users, a huge blessing for consumer internet companies. By the mid-2010s, the annual growth of internet users slowed to 5 or 6 percent. It is expected to continue falling.

With slowing user growth and overcrowded consumer markets, internet companies shifted from value creation to extraction. After spending years competing for users, and subsequently attaining some degree of monopoly status, they shifted to taking rents from users while providing no additional value. After wiping out the competition, the Chinese food-delivery platforms Meituan and Ele.me were able to arbitrarily raise their commissions from vendors, who had no choice but to stay on the platforms due to the lack of alternatives. E-commerce sites have similarly hiked up vendor fees. In the years leading up to the internet crackdown, big-data-enabled price discrimination, which allowed internet companies to take advantage of customer loyalty, also became standard practice.

This stagnation was a precondition for the involution of tech-sector workers. As companies shifted from expansion to extraction, the demands of work have lessened. Employees are still toiling to the tune of ruthless competition, but now with even less to go around—winner take none. With little to gain, workers are finding ways to slack off while maintaining an industrious facade, a delicate balance called moyu (touching fish). The term originates from the traditional idiom hunshui moyu (fishing in muddied water), used by Sun Tzu in The Art of War to describe taking advantage of a chaotic situation. In addition to taking extended bathroom breaks, some workers stay in the office playing games on their phones late into the night, clocking out only after a certain number of hours have been logged. The idea is to “work” long hours while actually getting as little done as possible.

Moyu points to a larger problem: many workers at these internet companies are redundant, performing what David Graeber called “bullshit jobs”: “a form of paid employment that is so completely pointless, unnecessary, or pernicious that even the employee cannot justify its existence [while feeling] obliged to pretend that this is not the case.” Most 996 workers will tell you that the 996 regime has nothing to do with generating increased output or concrete results. They spend hours writing weekly reports and attending pointless meetings. The industry has become so bureaucratic and process-heavy that only a fraction of time is actually spent being productive. And most of that time is used to steal profits from competitors in a zero-sum game. (Sometimes, the company itself is redundant.) In practice, 996 became less an indicator of how much work there was to do than a way to demonstrate a willingness to work hard—even if it’s bullshit work.

996’s End Isn’t 1065’s Beginning

While the official end of 996 has been welcomed by most, workers remain skeptical that new policies will actually improve their lives. Some say that canceling 996 won’t reduce their workloads or the KPIs (key performance indicators) they have to keep up with. Instead of working out of their offices, where they can pick up free snacks or use the office gym to squeeze in a workout, many now have to take their work home, off the clock. All these factors are contingent on the whims of a worker’s direct supervisor, who may or may not respect the new policies around overtime. For most workers, not responding to their supervisor’s WeChat messages at any time, day or night, weekday or weekend, is virtually unthinkable.

As companies cancel 996, workers are also taking significant pay cuts. According to overtime-pay policies, workers receive double wages if they work weekends, and triple if they work over national holidays. When ByteDance ended the big-small week in August 2021, salaries were slashed by up to 17 percent. For workers with mortgages to pay off or schooling to put their children through, a pay cut simply isn’t worth it. Younger tech workers also complain that ending 996 impinges on their freedom to make extra money. Typically new to the cities where their jobs are located, and without a family to return to or care for, these college hires are more willing to trade their evenings and weekends for overtime pay.

The move to curb excessive overtime work looks good in theory, but many workers fear that they’ll get the worst of both worlds: a 1065 salary with a 996 workload. Some companies have already relapsed into old habits, forcing employees to work weekends. And even at places where weekends are off-limits, the twelve-hour workday has yet to be significantly reformed.

The move to end 996 belies a deeper problem for China’s internet giants: in a digital economy that is more and more reliant on extractive practices, profits are becoming increasingly detached from the industry’s relentless work schedule. Some companies have begun to catch up to these changes and are getting rid of surpluses in their workforces. Kuaishou and iQiyi (China’s answer to Netflix) both announced massive layoffs at the end of 2021, cutting up to 30 percent of their staffs. The move to cancel 996 satisfies the same aim: slash wages to cut costs.

Still, something has changed in the culture of white-collar overwork, and the end of 996 should serve as a call for celebration. Some tech workers will now have the time to catch a break, and maybe even to mobilize for a greater say in their workplaces. They are still far from the days of 1065, but it’s a starting point nonetheless.

Last month, a veteran employee at Alibaba marked his departure with a 10,000-word internal letter that delivered a pointed critique of the company’s cultural and organizational decline. The letter quickly went viral in China, resonating with tech workers well beyond Alibaba, and even drawing a public response from Alibaba founder Jack Ma, who seemed to agree with many of its observations.

The author highlighted a series of internal/organizational issues such as the erosion of collaboration in favor of cutthroat internal competition; the dominance of short-term KPI-driven metrics over longer-term innovation; the rise of “boss culture”; and the normalization of long work schedules like 996 (9am to 9pm, six days a week). The result, he argues, is mediocrity, short-term thinking, and growing dysfunction at Alibaba.

For close observers of China’s tech sector, the criticisms in the letter aren't new. Many who have closely followed labor developments in the sector (myself included) have seen these problems emerge across the industry. In 2022, I wrote about tech workers “lying flat” and how Chinese tech companies, facing growing backlash, attempted to reform their management practices in response. Some companies even pledged to end 996 in hopes of fostering greater efficiency and innovation—albeit with limited success. What is apparent, as highlighted in this letter, is that the organizational problems within China's tech giants persists to this day.

Although the letter is focused on Alibaba, the issues it raised are far from unique; they are also severe in other Chinese tech companies ranging from ByteDance to Pinduoduo to JD.com. Even non-Chinese firms aren't immune. I've written about a similar organizational decline among tech giants in the U.S. with the rise of “bullshit tech work,” which you can find here (but I should emphasized that the organizational issues among big tech firms in the U.S. are very different from those in China). The point, nevertheless, is that these issues aren't just about firm culture but are structural—rooted in the very business model of the platform firm.

This post is structured in two parts. First I’ll summarize some of the key themes from the viral resignation letter. Second, I’ll explore how these organizational problems stem from the platform business model itself, which systematically prioritizes short-term commercialization—and how the organizational logic that once fueled Alibaba’s commercial dominance may now be the very thing holding it back from more innovative domains like AI.

Key themes from the resignation letter

The resignation letter opens with a nostalgic fondness for what Alibaba used to be, with the author proudly talking about the immense impact the company use to have in the economy. But over his 15-year tenure, he witnessed a steady decline in the company's position and a shift away from the values and vision that made the company great. The bulk of the letter is a detailed account of how Alibaba's internal culture is at the root of the problem. Rather than walk through it point by point (which an AI translator could surely do better than I), I'll highlight a few key themes from the letter.

First is the company’s obsession with short-term metrics and KPIs. Because promotions are tightly bound to these metrics, employees compromise on the integrity of their work in order to hit their numbers. They resort to all kinds of “operational” hacks: running marketing promo campaigns for their product, massaging the metrics, moving the baselines, etc. They even fake the entire product life cycle—hyping it up to get resources, manipulating data to simulate success, and shifting blame once it inevitably collapses. In this environment, workers who cut corners, pursue vanity metrics, and even sabotage their colleagues are the ones that get ahead.

Second is the rise of “boss culture”—a hierarchical system in which upper management holds all the power and can determine the fate of lower ranked employees. In this system, promotions depend less on actual contributions or innovation and more on loyalty to one's manager. This makes going against the decisions of leadership unthinkable—sometimes at the cost of doing what's best for the customer or the product.

Third, and relatedly, is that the company has become captive to market trends. When a new business line seems promising, teams rush in and reorgs cascade through the company as managers scramble to stake out more territory. The letter describes a cycle of endless reshuffling—frequent leadership changes, shifting priorities, and KPIs that are constantly being updated. Strategic direction grows increasingly vague, and long-term product vision becomes untenable. With this pattern of trend-chasing, employees know that today’s projects may be scrapped tomorrow—so instead of building for durability, they build to fast-track their promotion.

Fourth is a culture dominated by internal competition rather than collaboration. Where individual KPIs are more important than shared goals, employees stop working together and start working against each other. In the letter, one passage describes how “wild dogs”—the kind of employee who plays dirty and sabotages teammates to further their own careers—will advance their careers far faster than the “gold cows,” the dependable and high-performing employees who once formed Alibaba’s backbone. As internal competition becomes the driving motivation for work, cross-team cooperation breaks down, and with it, the ability to harness bottom up innovation from the company's talent base.

Finally, the letter paints a stark picture of how forcing workers to work overtime undermines innovation. Reflecting on Alibaba’s 996 culture, the letter points out that tech work is not manual labor and longer hours do not translate to better outputs. In fact, it often breeds a culture of performative busyness, where workers compete over who can work the longest to appear the most committed. The result is a system optimized not for real product development, but for the appearance of productivity.

An organizational culture oriented towards short-term commercialization

The labor relations described in the letter—the focus on KPIs, top-down decision making and managerial control, “boss culture”, and cutthroat internal competition—aren’t accidental nor are they unique to Alibaba. They’re the result of the business model that Alibaba and other platform firms operate under.

To understand the roots of the problem, we need to look at the changing dynamics of China’s platform economy. Since the late 2010s, China’s digital economy has become hypercompetitive. In earlier years, the rapid rise in internet users was the rising tide that lifted all boats, creating enough demand for competing firms to expand together. But today, that pie has stopped growing. With user growth plateauing, China’s internet giants are locked in zero-sum competition, where growth requires the direct cannibalization of a rival's userbase.

In response to this hypercompetitive landscape, internet companies reshaped their internal cultures to stay ahead—prioritizing speed, adaptability, and execution. Strategic direction appeared to be in constant flux; not for a lack of leadership but because firms had to be able to react quickly to new market pressures or consumer trends. Especially in China's environment where copying competitors is the accepted norm, success depended less on novel ideas, and more on executing new products, scaling up quickly, and responding constantly to what competitors are doing.

In this context, the obsession with KPIs and the use of 996 made some sense as a managerial tactic. KPIs offered a way to discipline workers around commercial goals that were immediately measurable: user growth, engagement, conversion, sales. Similarly, the enforcement of 996 had for years enabled leading platforms (including Alibaba) to outcompete challengers and retain their dominance in several markets. These practices made sense in an era of hypercompetition where rapid execution, maximizing short-term gains, and maintaining market share were critical, as even a slight slip up could give competitors the momentum to activate powerful network effects.

Today, as China’s internet economy slows down, the opportunities for short-term commercialization have narrowed. Most of the easy gains have been squeezed dry. And tech firms are now pivoting towards robotics, enterprise software, biotech, and most of all, AI—all of which demand more thoughtful innovation, sustained experimentation, deep technical collaboration, and empowered teams incentivized to take risks. But as the Alibaba resignation letter makes clear, the current workplace culture isn’t built for that future.

Alternative models of labor relations are beginning to appear in China’s tech sector. DeepSeek is one example. As I’ve written in a previous post, the small AI company's workplace culture stands out for its collaborative nature, bottom-up approach to decision making, and the high degree of autonomy it grants employees. And it also isn’t weighed down by a workplace culture that was shaped during the hypercompetitive internet boom years.

As the Alibaba resignation letter writes: AI is here, embrace this era. But the question is whether China’s internet giants—most of which rose to prominence during those boom years—can change their workplace culture, or whether the inertia of the old labor model will persist.

In recent years, as Western countries have lost ground in critical manufacturing sectors, the U.S. and Europe have turned to attracting greenfield foreign direct investment (FDI) from leading Asian firms to revive industrial competitiveness.

In the U.S., the most notable example is Taiwan's TSMC, the world’s leading chipmaker, who has committed billions into bringing the most advanced chip fabs to the U.S. The initial commitment came in 2020 amid rising concerns over supply chain security. The company committed to more investment in U.S. fabs with the passage of the CHIPS and Science Act in 2022. And in 2025, they pledged an additional $100 billion in U.S. semiconductor capacity, bringing its total planned investment in Arizona to $165 billion.

A similar pattern has also emerged in Europe, particularly within the electric vehicle (EV) sector. In December 2024, Stellantis partnered with China’s CATL to announce a €4.1 billion joint venture, creating one of Europe’s largest battery plants in Spain. This followed CATL’s earlier €7.5 billion investment in a battery facility in Hungary. BYD, China's top EV maker, has also invested in a plant in Hungary.

The assumption is that such FDI projects will automatically foster technological transfer and industrial competitiveness. But this is far from guaranteed. Foreign investors enter new markets to maximize profits—not to generate development or share their technologies. In fact, lead firms in high-value-added sectors often have strong incentives to guard proprietary technologies and keep core capabilities anchored in their home countries—a pattern well documented throughout the history of industrial development. Even when firms aren’t actively blocking knowledge spillovers, FDI can still produce enclave-style growth with foreign firms remaining isolated from the host economy and offering minimal integration or long-term benefits. This raises a critical question: under what conditions does FDI actually support meaningful technology transfer and industrial upgrading?

One way to conceptualize the challenge of technology transfer is to imagine the foreign firm as the teacher and the host country as the student. In this metaphor, failure to learn can occur on either side. The teacher may be a bad one—withholding knowledge, offering only limited instruction, or lacking incentives to engage. At the same time, the student may lack certain prerequisites needed to absorb what’s being taught. Drawing on the experience of successful late industrializers including China, this post argues that for FDI to drive meaningful growth and technology transfer, the teacher must be willing to teach and the student ready to learn—and it is in large part up to the state to ensure that each plays their part.

Setting the student up for success

Let’s start with the student. In the context of FDI, a "good" student is one that has a high capacity to actively learn from its foreign teachers. Just as a teacher will not be able to teach differential equations to someone who hasn’t yet grasped the basics of calculus, learning in the context of technology transfer also requires a foundation of prior capabilities. In many cases, local firms lack the scale, skills, or capital to engage meaningfully with foreign investors or to integrate into their supply chains. This absence of pre-requisite capabilities is one of the key constraints on the host country’s ability to benefit from FDI.1 As a result, when lead firms invest in new regions, they often have to rely on imports rather than sourcing locally—foreclosing opportunities for domestic suppliers to upgrade and limiting the potential for broader industrial development.

Chen Ling’s comparative study of the divergent upgrading trajectories of Suzhou and Shenzhen offers a nice illustration of this dynamic.2 In Suzhou, local officials believed that simply attracting major multinational firms like Samsung and Nokia would catalyze automatic upgrading. But the gulf between these foreign lead firms and local suppliers was too wide. Domestic firms lacked the technical and organizational capabilities to meet the quality, scale, and reliability requirements of their foreign counterparts. And as a result, these foreign firms bypassed the local supply base entirely, relying instead on imported components. The student, in this case, lacked the foundations to understand the lessons being taught.

By contrast, Shenzhen’s local government took a more adaptive and strategic approach. Rather than courting top-tier foreign firms off the bat, the local government focused on building capabilities from the bottom up—attracting smaller, lower-tier "guerrilla" investors from Hong Kong who operated in only slightly higher stages of the value chain. These firms brought capital and technical know-how that were closer in scale and complexity to what local partners could absorb, resulting in greater knowledge transfer and the gradual accumulation of industrial capabilities. Over time, Shenzhen’s firms developed the skills and technological base to move up the value chain—at which point the state began inviting higher-end foreign firms. In this case, the student was set up to succeed: the lessons were matched to the learner’s level, and the state made sure that the learning environment was structured for long-term growth.

Forcing the teacher to teach

The problem can also come from the teacher. Firms, after all, provide FDI, not to generate growth or learning in the host countries, but to maximize their own profits. Foreign firms may also want to keep key technologies from spreading abroad, so they may limit foreign activities to low-value segments like basic assembly, while retaining higher-value functions—such as R&D, design, and strategic decision-making—in their home countries. They may also inhibit local development by using patents, proprietary standards, and intellectual property rights to raise entry barriers and block local firms from climbing the value chain. In more extreme scenarios, they may even pressure local suppliers into costly investments, only to squeeze them later through aggressive price negotiations, creating what Williamson (1983) described as a "hostage situation."3 This unwillingness to teach curtails the possibility of meaningful technology transfer or skill upgrading.

In such scenarios, only the host state can force the teacher to teach. In his 1979 study of the country’s industrialization, Peter Evans coined the term "dependent development" to describe a Brazil's development model in which FDI played a central role—but under close state guidance. Rather than allowing multinationals free rein, the state forced foreign investment to align with domestic developmental goals. One of the key mechanisms for this was to require foreign firms to operate in joint ventures with state-owned enterprises. This arrangement encouraged foreign firms to localize their operations, facilitated technology transfer, and ensured that domestic firms become embedded in the high-value segments of production, thus maximizing domestic firm's chances to learn. This corporate governance arrangement resulted in a more symbiotic relationship between foreign capital and national development objectives.

Even though Brazilian development was stymied by the debt crises of the early 1980s, this practice played out in full among the East Asian developmental states. Indeed, one key aspect that set the East Asian development states apart from the rest of the developing world during the 1970s and 1980s was their high state capacity, which allowed them to create controlled environments where foreign firms was were forced to participate in technology transfer.4 Part of this involved setting up local content requirements, which mandated that foreign businesses use domestically produced goods, services, or labor in their operations. These states also forced foreign firms to source from domestic producers. Taiwan's government even pressured foreign firms to focus their sales on exports to prevent the exploitation of domestic markets.5 Without these measures in place, foreign teachers may end only exploiting their host countries without doing any teaching.

Concluding thoughts

The student–teacher metaphor was made explicit in China’s electronics industry, when Huawei founder Ren Zhengfei explicitly referred to Apple as his company’s teacher. As Patrick McGee writes in his new book Apple in China, Apple served as the "great teacher" not just to Huawei, but to China’s electronics sector more broadly.6 The book argues that Apple invested deeply in China’s manufacturing base over the past two decades—building factories, installing advanced equipment, and deploying teams of engineers to work with and even train local suppliers to meet its exacting technical standards. This ultimately led to the massive transfer of valuable manufacturing know-how, as local suppliers learned to produce cutting-edge products at high volumes and low costs.

There's no doubt that Apple extracted immense value from its suppliers—driving prices down and pushing many to operate on razor-thin margins, sometimes even at a loss. Yet for Chinese suppliers, the long-term payoff was worth the thin margins. The technical knowledge and credibility gained from working with Apple became a stepping stone to securing more lucrative contracts with other electronics or phone makers, many domestic. Beijing even actively encouraged these relationships, recognizing that Apple was driving a massive transfer of industrial capability.

Today, the roles have reversed. In key manufacturing sectors, advanced Western economies are now seeking to learn from leading Asian firms. This review shows that FDI can be a powerful engine for technological upgrading and regional development—but only when the teacher is willing to teach and the student is prepared to learn. While the challenges facing the West are different (higher labor costs and in some cases already on the path of deindustrialization), many of the lessons from past late industrializers still hold valuable insights for how to make foreign investment work.

In Europe’s battery sector, for example, underdeveloped local supplier networks and a shortage of skilled labor have left foreign firms like CATL reliant on imports in the early stages of production.7 In this case, the student needs to be made a better learner (e.g., sustained state investment in skill training and upstream suppliers). The experience of successful late industrializers shows that FDI only delivers real gains when guided by a high capacity state that not only attracts investment, but also ensures that foreign firms are embedded into local ecosystems, and that domestic actors are equipped to learn.

Over the past few years, more and more people have begun to write about the resurgence of vertical integration as viable corporate strategy, pointing to cases like BYD, Tesla, even Ford. I too, just a few weeks back, wrote about this very subject. What’s often missing from these discussions is a closer look at why firms across different contexts of development pursue vertical integration in the first place. BYD and Tesla, for instance, may both be vertically integrated firms, but coming from different developmental contexts, the impetus behind their strategies is entirely different.

In advanced countries, vertical integration is something that firms are returning to. In the west, it is considered novel because it signals a retreat from globalization and the pattern of offshoring that has dominated corporate strategies since the 1980s. Today, there are many reasons for its return: concerns over the fragility of global supply networks (especially since the pandemic), geopolitical risk, and—perhaps most importantly—the need to restore domestic production capacity to revitalize broad-based growth. For firms in advanced economies, vertical integration usually means reabsorbing lower-value-added activities such as assembly and component manufacturing that were offshored to the developing world. (Or, as is the case of Tesla, it constructed an entirely new supply chain out of necessity and retained hierarchical control over each link.)

By contrast, in developing countries—specifically export-led countries that have exhibited high rates of development since the 1980s—vertical integration is not a retreat from globalization but one of its logical outcomes. To appreciate why and when vertical integration occurs in developing countries, we first need to understand the evolution of global value chains (GVCs) since the onset of globalization.

In the early years of globalization, the dominant strategy for lead firms in the west was dis-integration: break up the production process, offshore labor-intensive stages, and focus on high-value activities like product design, branding, and finance. This allowed firms to reduce costs, increase flexibility, and offload risk, while still capturing the lion’s share of value through control over intellectual property, branding, and other consumer-facing functions. Importantly, dis-integration was a cost saving strategy for such firms who would otherwise have to rely on a more costly (and often unionized) domestic labor force.

To economists, this fragmentation of production has long been attributed to the theory of comparative advantage. The argument is that, in a world of free trade and falling coordination costs (thanks to technologies like the internet and containerized shipping), countries with cheap labor should focus on labor-intensive production. And firms in advanced countries (where labor is comparatively more expensive) should offshore labor intensive activities in order to maximize the overall efficiency of the system. This theory implies a stable division of labor. Developing countries are expected to remain in low-value-added segments of GVCs—assembly, basic processing, component manufacturing—while firms in advanced economies would retain control over high-value functions like design, branding, and R&D.

In most cases, this is exactly what happened. Many countries that have entered GVCs at the bottom remain stuck in a low-level equilibrium trap. But some have steadily moved up the value chain, capturing ever more sophisticated and profitable segments of production.

To understand how this happened, we must go beyond traditional economistic explanations such as comparative advantage. As Ha-Joon Chang has argued, successful development often hinges on comparative advantage defiance.1 So rather than entering a GVC at the bottom and staying there, the key task for states of developing countries is to turn GVCs into stepping stones for industrial upgrading and economic development. This is the essence of “upgrading”: the process by which firms move from basic assembly into more complex, higher-value-added segments of the value chain.

There are many mechanisms that states can use for precisely this task. Joint ventures, forced technology transfers, investment in R&D to increase absorptive capacity, among many other mechanism have been used to help suppliers in developing countries upgrade. This is, in fact, the subject of a rich body of literature at the intersection of development studies and GVCs.

In successful cases—such as China, South Korea, and Japan—domestic suppliers have steadily acquired new capabilities, enabling them to capture larger portions of the value chain and negotiate more favorable terms with lead firms. Over time, more stages of production have become concentrated within national borders. This has led to a form of regional vertical integration, where entire supply chains are now localized into dense, interlinked ecosystems of suppliers. Today, many so-called “global” value chains are no longer truly global, but operate at a regional or even local scale, with multiple stages of production housed within a single region, city, or industrial zone.

In certain sectors, this has gone even further with vertical integration taking place at the level of the firm. This is tends to happen in sectors where knowledge is tacit or where coordination is complex. In such contexts, firms vertically integrate to reduce coordination costs and eliminate markups from intermediaries. The electric vehicle (EV) industry is a case in point. China's top EV maker, BYD, began as a battery maker (manufacturing for cellphone companies like Nokia and Motorola) and incrementally over decades captured more of the EV value chain—electric powertrain, semiconductors, transmission, cockpits, breaks, etc. And by keeping all stages of production under its control, the company has been able to keep costs down making it perhaps the most price competitive EV maker today.

A key difference in China’s development path is that—unlike earlier fast-developing economies such as Taiwan, South Korea, and Japan, which followed the lead of advanced countries by offshoring labor-intensive segments of the value chain to less developed nations—China has largely retained the entire value chain within its borders. In some cases, it has even consolidated multiple stages of production within a single firm. This has remained true even as China has successfully upgraded into higher-value-added activities. One reason it has been able to do so is its vast supply of cheap labor, which reduced the pressure to outsource lower-value-added production abroad.

What this story of vertical integration in developing contexts reveal is that vertical integration can come about for different reasons. In advanced economies, vertical integration is often about reclaiming what was lost to globalization: reabsorbing outsourced activities or building new supply chains from scratch to regain manufacturing control. In certain developing countries, by contrast, it’s about climbing upward—gaining capabilities, capturing more value, and reducing dependence on external partners.

The stylized diagram below captures this asymmetry: firms in advanced vs developing contexts must move in opposite directions to achieve vertical integration.

In this way, offshoring and vertical integration aren’t opposing strategies but mirror images of each other—each aimed at cost control, but from opposite ends of the value chain. For lead firms in the West, cost savings came from offshoring and dis-integration. For firms in developing countries like BYD, cost savings and competitiveness come from putting the chain back together again.

Just last month, Microsoft announced over 6,000 job cuts—its second largest layoff ever. And they’re not alone. Since the start of the year, Amazon, Alphabet, Meta, Salesforce, and many other household tech names have also slashed headcount. What makes these layoffs especially striking is that they’re happening not during a downturn, but amid a surge in tech stock valuations. Microsoft’s layoffs, for instance, came in the same month that its stock price approached all-time highs.

Since 2022, layoffs have become a new normal for the tech industry with 2023 being worst layoff year in the modern tech industry. Then, layoffs were seen as a sensible correction to overzealous hiring in the preceding two years, where tech companies went on a hiring spree as they realized that the pandemic would accelerate the pace of digitization.

What’s striking now is that, two years later, the layoffs haven’t stopped, even as the tech sector surges ahead. If anything, they’ve become routine; embedded in quarterly planning cycles, not just as emergency responses. Layoffs, initially framed as a necessary correction the pandemic hiring spree, seems to have far outlived the conditions that justified it.

Today, executives are pointing at a new explanation: AI automation. They are claiming that AI is not just as a productivity booster, but a new reason to cut jobs. But we should be skeptical of this narrative. The scale and timing of recent layoffs suggest that AI may be more pretext than cause—an excuse to further cut costs without alarming investors.

This post argues that what’s unfolding is not just a reaction to macroeconomic shifts or technological change, but a deeper reconfiguration of labor relations in tech. And as longtime readers of my newsletter will know, labor relations has everything to do with innovation. But to understand how we got here, we need to go back in time to the industry’s high-flying ZIRP years to understand labor relations in this industry.

Tech's high-flying ZIRP years

For many years, tech workers had tremendous labor market power. For most of the 2010's, computer science graduates were easily making six figures right out of college. Big tech firms courted engineers, designers, and program managers with hefty compensation packages, beautiful offices, and endless perks. These companies turned their workplaces into some of the best places to work, setting a new standard for what good work looked like around the country.

All this was underwritten by an extraordinary financial backdrop shaped by over a decade of loose monetary policy. In the wake of the 2008 financial crisis, the Federal Reserve slashed interest rates to near zero to stimulate the economy. And for most of the 2010s, rates remained near zero, creating what economists call an era of Zero Interest Rate Policy, or ZIRP.

During these years, when capital flowed freely into Silicon Valley, firms rapidly expanded into new digital segments, and demand for tech talent far outpaced supply. This gave workers significant leverage. Job seekers often collected multiple offers and used them to negotiate higher compensation packages. Inside companies, talent scarcity gave individual workers substantial autonomy.

All this was kicked into high gear when the pandemic supercharged the digital economy. Kitchen tables became office desks, and the digital tools that once complemented in-person work became the lifeline for businesses around the country. Already one of the most powerful forces in the U.S. economy, the tech sector suddenly became the infrastructure underpinning every aspect of life. Demand for technical skills exploded, and tech companies scrambled to snatch up engineers, designers, and data scientists as fast as they could, often without knowing what to do with them. Between 2019 and 2022, Amazon and Meta more than doubled their headcount, growing by 106 and 103 percent, respectively. Major firms like Microsoft and Alphabet each hired tens of thousands of new workers between those years; 80,000 and 60,000 respectively.

Then inflation hit. Strict pandemic protocols and intermittent shutdowns slowed production in factories around the world. The result was a series of major bottlenecks to the global supply chains. Adding fuel to fire, the Russian invasion of Ukraine sent energy prices through the roof. By 2022, inflation peaked at 9.1 percent, making it the largest 12-month increase since the early 1980s. The Federal Reserve responded to rising inflation with one of the fastest tightening campaigns in decades, raising rates from near zero to roughly 5.5 percent after ten consecutive hikes in just a 15 month timespan.

Why layoffs are the perfect cost-cutting tool

When interest rates are low, capital is cheap. Investors, gaining minimal returns from idle cash, become eager to deploy money into new ventures. And while low interest rates encourage economic expansion across all sectors, they are especially favorable to the logic underpinning the tech industry’s business model. Tech firms thrive on rapidly acquiring users and scaling their services. Their success hinges on achieving strong network effects such that the value of their platform increases as more users join it. This dynamic creates incentive for firms to operate at a loss, burning through cash to acquire users, scale their platform, and entrench their position—all in hopes of eventually dominating winner-takes-all markets.

But if one winner takes all, the rest are destined to fail. In other words, tech firms are extremely risky investments. This is why a zero-interest-rate environment is critical for tech’s expansion. For investors, it makes tech’s high failure rate acceptable. After all, when idle cash earns virtually nothing, even long-shot bets on future dominance can seem worthwhile. And so, venture capitalists are willing to underwrite a tech firm’s scaling costs, not for guaranteed returns, but for the possibility of spectacular ones.

This was the financial environment that supercharged tech’s expansion since the 2008 financial crash. But in 2022, with the onset of inflation, the industry built on cheap capital ground to a halt. In a high-interest rate environment, a high-risk, high-reward logic investment model is far less tenable. The opportunity cost of risky investments increases dramatically because cash no longer sits idle but accrues real value in safer assets. As a result, investors become more cautious, demanding safer returns and more predictable outcomes. For tech firms, this means that executives can no longer justify burning cash in the name of growth. Lavish spending on moonshot projects, generous perks, and fast growing headcount starts to look less like visionary leadership and more like fiscal irresponsibility.

For asset-light tech firms, labor is often the largest cost and the easiest to cut. During the ZIRP years, keeping an extra engineer on payroll was a small price to pay compared to the potential upside of a breakthrough they might help deliver. But in an environment of rising interest rates, every employee represented a financial trade-off. Firms had to weigh the cost of each new hire against the guaranteed return of simply holding that cash in the bank. In this context, tech workers were no longer treated as future value-creators by default but had to prove their value by generating short-term returns.

Yet unlike traditional workers, whose labor is tightly coupled with the daily revenue a firm generates, most tech employees do not generate recurring or short-term revenue but are instead tasked with building products aimed at future returns. Because of this, laying off a tech worker wouldn’t result in revenue loss since the value of their past labor has already been congealed into codebases on which digital products are built. And as a result, the absence of their labor may not be felt until much later, if at all.

This was made clear when Elon Musk, after acquiring Twitter, slashed 80 percent of the company's headcount without fundamentally undermining the stability or usability of the social media platform. So, whereas a traditional firm would see a direct reduction in revenue when cutting their workforce, layoffs in tech are an immediate cost saving (and thus a boost to profitability) without any short-term penalty.

This is precisely the mechanism that tech firms took advantage of. At the end of 2022, Meta—one of the most aggressive hirers during the ZIRP years—cut 11,000 jobs that November, amounting to over 10 percent of its workforce. Amazon followed with axing 10,000 corporate employees. By year’s end, nearly 93,000 U.S.-based tech workers had been let go, marking a historic high for the industry. Each layoff announcement triggered a bump in stock price—Meta’s first major round sent shares up 13 percent overnight. In a world designed to maximize shareholder profits, layoffs were handsomely rewarded.

Tech's addiction to layoffs

What began as a response to a deteriorating macroeconomic outlook and overzealous hiring during the Covid boom has since solidified into a lasting pattern. Trueup.io projects that nearly 200,000 tech employees will face layoffs in 2025—a figure that approaches 2022 levels. While it falls short of the staggering 400,000 layoffs recorded in 2023, it remains dramatically higher than anything seen during the ZIRP years, when tech layoffs were rare and job security was largely taken for granted.

What has become clear is that tech layoffs, initially framed as a necessary correction, has far outlived the conditions that justified it. Throughout 2023, the U.S. economy repeatedly defied expectations. Consumer spending remained strong, job growth beat forecasts month after month, and inflation gradually slowed.

By mid 2023, as inflation dropped to three percent (from nine percent in the year prior), the Federal Reserve managed a so-called "soft landing," cooling the economy without tipping it into recession. Tech giants like Meta, Alphabet, Microsoft, and Amazon all reported stronger-than-expected earnings in the first quarter of 2023, signaling a return to profitability. Meanwhile the release of OpenAI’s ChatGPT injected a new jolt of boosterism—and with it, a massive injection of new capital—into the sector. And yet, even as recession fears have subsided and earnings rebounded, tech layoffs seem now to be a permanent fixture of the sector. What gives?

Although inflation has quelled, we're still far from zero-interest rate territory (4.5 percent at the time of writing, down about one percent from one year ago), meaning that we can't expect the tech labor market to bounce back to ZIRP era highs. But two others reasons for continued tech layoffs have emerged. The first is the political backlash against DEI initiatives, led in large part by Trump's re-election. In response, nearly every major tech company has scaled back or eliminated DEI programs, slashed related funding, and deprioritized inclusion efforts. Once framed as core to company culture and innovation, DEI have been recast as luxuries in the low-interest rate environment and something that they can no longer afford. (DEI cuts, however, can only account for a relatively minor portion of layoffs.)

The second is AI. As tech executives tout the potential of AI to boost productivity and streamline operations, they are supposedly "dog fooding" these technologies (testing it first within their own firms) and framing it as a rationale for workforce reduction. AI is indeed appears to be good at coding tasks. Anthropic CEO Dario Amodei has even said that AI has the potential to write 90 percent of the code for big tech firms within the next year. Mark Zuckerberg also stated that AI could potentially replace entry positions up to mid-level software engineers as early at this year. Despite this hype, many engineers still say that the current capabilities of AI are nothing more than glorified search engines for them (and as a former engineer, I would largely agree). In other words, AI seems more like an excuse to continue cost-cutting rather than the driving force.

Resetting labor relations

There is another reason, one that I think is central to understanding why tech layoffs are continuing till today: the reassertion of managerial control. Over the past decade, tech workers had a unique degree of labor market power that—when combined with their techno-utopian ideology—underpinned an unprecedented wave of workplace activism. From the 2018 Google Walkout to protests against military contracts, surveillance technologies, Big Oil, and providing technology to authoritarian governments, tech workers increasingly organized not just for wages or perks, but for a say over how the technology they built was being used. Collective Action in Tech, a non-profit that documents employee activism in tech and also a group that I'm a part of, shows that over 50 instances of tech worker activism took place in 2019 alone.

To tech executives, this was a dangerous trend that needed to be squashed. As the tech industry expanded into new lines of business and grew into the monopolies that it initially sought to undermine, the old mission-driven rhetoric central to Silicon Valley’s ideology was no longer enough to keep workers aligned with management. And as tech workers took to major media outlets to attack their company's then-still pristine brands, employers began to closely monitor employee dissent, limiting internal communications to work-related discussions, and even hiring union busting consultants—in effect doing whatever they could to bring the balance of power back to their side.

Quickly, a new consensus among tech's billionaire class was being formed. Marc Andreessen, a prominent venture capitalist, railed against the previous era of employee activism, saying that "the [tech] employee base is going feral. There were cases in the [first] Trump era where multiple companies I know felt like they were hours away from full-blown violent riots on their own campuses by their own employees."

With new urgency to claw back power from an increasingly mobilized workforce, the tech slump in 2022 and the subsequent raising of interest rates came as a blessing to tech's billionaire class. These macroeconomic conditions gave tech employers an easy justification for sweeping layoffs and thus an opportunity to discipline vocal employees and regain managerial control in a sector that has long given workers their way. Return-to-office mandates and performance improvement plans (popularly known as PIPs) have also been increasingly used to reinforce a climate of compliance.

Tech's new labor relations regimes carries consequences not just for workers, but for innovation itself. The tech industry has long relied on the creativity and self-initiative of its workforce—qualities that thrive in environments of trust and autonomy. As firms double down on these austerity measures, they risk undermining the very conditions that once enabled workers to again and again generate breakthrough ideas. When workers are subject to narrowly defined KPIs, daily scrutiny, and fear of losing their job, the industry doesn’t just become less humane; it becomes less innovative. With layoffs the new normal, tech companies may ultimately be stifling the ingenuity that made them powerful to begin with.

An earlier version of this essay was originally posted to AI Supremacy on May 8, 2025. Thank you Michael Spencer for inviting me to contribute.

Cloud computing is often likened to a basic, fungible utility such as electricity or water. Back in 2014, Matt Wood, former chief data scientist of Amazon’s cloud business, even explicitly said that it was their goal to deliver “computing power as if it was a utility.” And like a utility, the cloud quietly powers much of our digital lives—largely invisible yet completely essential. Just as a power outage can paralyze a city, a cloud outage can bring the digital world to a standstill.

But comparing the cloud to a utility doesn’t feel quite right. The firms that dominate it are nothing like water, gas, or electricity providers. (How many utility companies can you name off the top of your head?) Even telecom giants like AT&T or Verizon, which offer a more sophisticated kind of utility, feel different. Rather than acting like sleepy utility companies, quietly collecting revenue from every customer that touches their cloud, cloud providers are tech companies in the most classic sense—defined not by stability but by relentless disruption.

To be clear, these firms do also offer the most basic, utility-like servers and likely always will. But where they diverge from traditional utility providers is in their relentless push to create new value beyond these foundational services, from AI-powered computing clusters and custom-designed silicon chips to sophisticated software platforms.

In this essay, I trace the evolution of the cloud computing business model—from its early days as a utility-style service focused on providing low-cost, commoditized storage and compute power, to its transformation into a high-margin, value-added platform. These differentiated services, ranging from AI tools to data management platforms, define their competitive advantage and justify their premium pricing.

The cloud-as-utility business model

Before the cloud, running a business in the digital age meant becoming a part-time infrastructure company. Organizations had to build and maintain their own IT backbone—buying physical servers, networking equipment, and storage drives, then installing them in on-site server rooms. Even basic functions like email, file sharing, or internal software systems required significant investment in hardware and a dedicated IT team to keep everything up and running.

For large enterprises, the burden scaled dramatically. Maintaining infrastructure often resembled running a miniature data center, complete with redundant power supplies, cooling systems, 24/7 security, and specialized staff for maintenance and support. And the costs didn’t stop there. To remain competitive, firms had to keep pace with the rapid hardware refresh cycles driven by Moore’s Law, replacing and upgrading systems every few years. This meant physically swapping out servers, migrating data with minimal downtime, and managing increasingly complex software environments—all while trying to avoid disruptions to the business’s day-to-day operations.

Cloud computing entered the corporate mainstream with the promise of alleviating these costs. Rather than owning and maintaining their own hardware, businesses could now rent their IT infrastructure. No more server rooms, no more costly upgrade cycles. Cloud providers handled the physical hardware, maintenance, and updates behind the scenes, allowing companies to stay current with the latest technologies without the usual costs and complications.

Where businesses once needed to make large investments in servers and infrastructure, even buying more capacity than needed to account for future growth, the cloud allowed businesses to rent exactly the computing power they needed, when they needed it. This pay-as-you-go model turned IT infrastructure from a fixed capital expense into a flexible operating cost, dramatically lowering the barrier to entry. For many businesses, this created new kind of agility—scaling resources up or down with a few clicks instead of months of planning and procurement.

What makes cloud computing more cost-effective is its ability to operate at scale. Its no surprise that cloud providers benefit from basic economies of scale. The large fixed costs of running a data center—such as land, cooling systems, security personnel—can be spread across more infrastructure, which lowers the per unit cost of computing. But scale also unlocks another, more powerful advantage: higher utilization rates. In the traditional model, companies had to build out enough infrastructure to handle peak demand, even if that demand only occurred a few times a year. And as a result, servers sat idle most of the time. By pooling demand across many customers with different usage patterns, cloud providers can smooth out spikes and troughs, keeping the shared infrastructure running closer to full capacity at all times. The result is higher efficiency (and lower costs) that get passed on to customers.

Cost savings was the cloud’s primary selling point. Unsurprisingly, then, providers competed aggressively on price, focusing on offering the lowest rates for the most basic, commoditized server resources. The cheaper the compute, the more competitive it became in the marketplace. Eventually, this competition led to a price war in the early years of cloud computing, where any price cut announced by one vendor was quickly matched, if not undercut, by another within days. According to Data Center Knowledge, Amazon, Microsoft, and Google made 25 price cuts on basic cloud resources like compute and storage between early 2012 and March 2013. This trend accelerated in 2014, with price reductions becoming both larger and more frequent. CRN, an IT trade publication, reports:

Google cut pricing for its Compute Engine Infrastructure-as-a-Service by 32 percent and storage by 68 percent for most users. A day later, AWS (Amazon’s cloud) slashed pricing for its cloud servers by up to 40 percent and cloud storage by up to 65 percent. A few days after that, Microsoft said it's introducing a new entry-level cloud server offering that will be up to 27 percent cheaper than its current lowest service tier.

In a paper published by the National Bureau of Economic Research, Byrne et al. (2017) conducted a systematic analysis of compute and storage prices among the top cloud providers, finding that between 2009 and the end of 2016, cloud processing costs dropped by approximately 50%, while storage prices declined by 70–80% over the same period.

In many ways, this race to the bottom mirrored the logic of the Silicon Valley playbook: aggressively subsidize costs to capture market share, then rely on scale and network effects to create market dominance. The strategy sacrificed short-term profits to achieve long-term control. And by driving prices down, cloud providers could expand their market share, knowing that once customers came on board, they were likely to stay due to high switching costs. After a business migrates to a specific cloud platform, moving to another provider is a costly and complex process.

During this time, cloud providers focused on filling data centers with off-the-shelf, commodity hardware and prioritizing universal compatibility such that any software could run on any hardware. The focus on was creating interfaces between each layer of the stack such that suppliers could plug-and-play at any level of the stack. In this way, cloud firms also resembled the classic asset-light Silicon Valley firm—leasing or contracting out whatever they could lower in the supply chain.

This approach is not unfamiliar to the cloud giants. All three had previously found tremendous success using this very strategy—monopolizing markets by temporarily slashing costs or even offering services for free—in their previous businesses. Google did it with search, Amazon with online retail, and Microsoft with its productivity software and operating system. And so, armed with billions in cash and a playbook that had worked before, the cloud giants dove headfirst into a tit-for-tat price war, aggressively cutting costs throughout the early to mid-2010s.

In some ways, a price war is a natural part of the industry’s maturation process. For the cloud sector, it lowered the barrier to entry, encouraging businesses that might have been hesitant to shift their IT infrastructure to the cloud. For providers, however, it was an expensive gamble—an aggressive attempt to gain market share.

Admittedly, evaluating the efficacy of price cuts on increasing the market share of any individual vendor is much more complicated than can be assessed here. However, from looking at overall shifts in marketshare in the 2010s, it would appear that this strategy didn’t play out quite as planned. Amazon, which was the most aggressive in its price cuts, held its market share steady at just over 30 percent for most of the decade. Microsoft, on the other hand, saw a consistent rise during the same period. Meanwhile, smaller cloud vendors such as Dimension Data, Joyent, and GoGrid were priced out of the industry—so in that sense, the strategy had some success. But when it came to dramatically reshaping the competitive landscape among the top players, the results were far less impactful.

One reason for this is that, by the mid-2010s, customers started to take a multi-cloud approach rather than relying solely on a single vendor for all their IT needs. This approach helped companies avoid vendor lock-in, giving them greater flexibility and bargaining power. Another reason was that some companies had built-in path-dependent advantages. For instance, Microsoft was able to leverage its long-standing enterprise connections from its Windows Server business, making it a natural choice for many corporations already embedded in its ecosystem.

But the core issue was that, unlike classic digital platforms, cloud computing doesn’t benefit from strong network effects. Platforms like Amazon.com or Google Search become more valuable as more people rely on it, creating a self-reinforcing loop that drives growth. As a result, Amazon and Google has monopoly shares in their respective e-commerce and search markets. In contrast, the inherent value of a cloud provider isn’t directly tied to how many customers it has, besides the fact that scale translates to cheaper costs.

By the late 2010s, the cloud providers—at least those still in the game—seemed to recognize that price competition alone wasn’t sustainable. Even though the top players were flush with cash, none of them were truly profitable under the existing model. AWS, the market leader, didn’t turn a profit until 2015, nine years after its launch. The cloud needed a new model, and as we shall turn to next, that’s exactly what happened.

The cloud-as-innovation-platform business model

If the cloud, like electricity, is a fungible utility, then the prerogative of the cloud industry would be as straightforward as moving as much of the world's existing IT infrastructure into the cloud. In a way, this isn't a bad business to be in. According to Gartner, global IT spending in 2010 was already at $3.4 trillion—by all means, a very attractive market to disrupt. Here, the cloud’s total addressable market should in theory be equal to global IT spending, or less—since the cloud, with its higher utilization rate, would reduce IT costs thereby stunting global spending. And since the cloud business model is all about taking advantage of scale to increase IT infrastructure efficiency, the growth of the cloud sector should also be pegged to the growth of global IT spending.

This, however, hasn’t been the case. Instead, the global cloud computing market has far outpaced global IT spending growth, surging from about $24.63 billion to $156.4 billion between 2010 and 2020—a 635 percent total increase or an annualized increase of roughly 20.3 percent. By contrast, global IT spending only saw a 2.14 percent annualized increase (to approximately $4.2 trillion in 2020) over the same time period.

This tenfold difference is obviously due, in part, to the rate at which traditional businesses are converting their IT infrastructure into cloud spending. But the other part of that number is that businesses are moving to the cloud, not to save on IT costs but to enable new revenue-generating activities. In other words, they’re not just looking for cheaper infrastructure—they’re willing to pay a premium for cloud services that create new value.

In a 2016 interview, Microsoft’s top cloud executive Scott Guthrie said that Azure (Microsoft's cloud business) was no longer competing with Amazon on price but was instead “competing more in value.” And by “value,” Guthrie was referring to “the higher-level services, the features, the performance, and the ability to differentiate or deliver true innovations.” He noted that this marked a shift from “two or three years ago, where I think it was more about cost per VM (virtual machine) or cost per storage.” Indeed, around the mid-2010s, the cloud price wars largely subsided.

Whereas the cloud-as-utility business model emphasized interoperability between software and hardware, this flexibility-first paradigm started to hit its limits as the industry started to move into higher value added workloads. Performance bottlenecks, latency issues, or sheer limitations in hardware performance pushed cloud firms to pay closer attention to the underlying hardware. And in response, cloud providers are vertically re-integrating, moving away from hardware-agnosticism and back to embracing full-stack control. Today, a key aspect of each cloud’s competitive advantage lies how its has recoupled software and hardware—developing specialized infrastructure tailored for specific workloads.

Concretely, what do these differentiated, higher-level services look like? Let's walk through a few examples.

Platform-as-a-Service (PaaS) + Software-as-a-Service (SaaS)

Traditionally, when there’s talk about providing higher-value services in the cloud industry, people are usually referring to platform-as-a-service (PaaS) or software-as-a-service (SaaS)—and that’s in contrast to infrastructure-as-a-service (IaaS), or what I’ve been calling the cloud-as-utility business model. Simply put, PaaS and SaaS are built on top of the three basic components of IT infrastructure (compute, storage, and networking). And because each PaaS and SaaS provide additional value on top of the bare infrastructure, they are priced at a higher rate than the underlying compute, storage, and networking they consume.

The easiest way to understand the differences between IaaS, PaaS, and SaaS is to think of running a restaurant. IaaS is like leasing a restaurant space and equipping it with everything you need to start cooking—ovens, stoves, refrigerators, tables, and chairs. You’re still responsible for hiring staff, designing the menu, and managing the kitchen. It gives you the raw infrastructure, but you handle most of the operations. PaaS is like walking into a fully equipped and staffed kitchen where everything is set up for cooking. You don’t need to worry about maintaining the equipment or hiring dishwashers—your job is just to focus on creating new dishes and running the menu. SaaS is like going to a restaurant, reading the menu, and ordering your meal. Everything is ready-made for you. You don’t worry about the kitchen, the staff, or the ingredients—you just enjoy the finished product.

Each of these services also caters to a different audience. Whereas IaaS is primarily designed IT operators, PaaS is built for software developers. It handles the behind-the-scenes infrastructure—like provisioning servers, connecting them to storage, and implementing security protocols—so developers can focus on writing code and deploying applications without getting bogged down by system administration. Meanwhile, SaaS is aimed at end-users, who interact directly with the finished software product through a web or desktop application. Whether it’s sending an email, managing sales leads, or editing a document, SaaS users engage with the software itself, not the underlying code or infrastructure.

Microsoft, Google, and Amazon operate in all three layers at varying capacity. All have robust IaaS and PaaS offerings. And among the three, Microsoft also has notable SaaS services, such as its Office 365 suite. But given the modularity of the cloud stack, other firms can nevertheless plug in at the PaaS or SaaS layer and build on top of the infrastructure that is provided by these three firms. For instance, Salesforce, a large SaaS provider, runs on Amazon’s cloud platform.

Advanced hardware

If PaaS+SaaS is about adding value at the top of the cloud stack, another way cloud providers offer higher-value, differentiated services is by enhancing the bottom of the stack. One way to do this is by investing in specialized hardware—servers that are difficult for traditional IT departments to purchase, deploy, or manage on their own due to the complexity, cost, and expertise required.

One example of this is the specialized hardware required for AI clusters. Building an AI cluster isn’t as simple as stacking a few high-powered servers together; it involves assembling a highly complex, technically sophisticated system designed to handle massive parallel processing workloads. These clusters rely on specialized components like high-end GPUs (Graphics Processing Units) and ultra-fast networking technologies (e.g., InfiniBand technology) to ensure seamless data transfer between processors.

Configuring advanced hardware can require deep expertise in data centers, which cloud providers have invested heavily in over the years. They have highly trained staff and R&D teams dedicated to improving data center uptime, energy efficiency, and cooling technologies. For most businesses, doing all this in-house would be cost-prohibitive and operationally overwhelming.

Infinite capacity

The cloud also enables businesses to operate as if computational limits don’t exist. Since cloud resources can be scaled up or down on demand, companies can offload their computationally expensive workloads into the cloud whenever their on-site infrastructure falls short.

Imagine a team of researchers at a biotech firm needs to run a protein folding simulation—a task that’s notoriously computationally intensive. (You could easily swap this out for a team of AI researchers training a new AI model.) Their on-site servers—represented by the dotted line in diagram—are sufficient for handling day-to-day operations, but running these protein folding simulations on that infrastructure would take an impractically long time (as represented by the blue line).

Rather than investing in additional hardware just to speed up one-off simulation projects, the biotech firm can offload the extra workload to the cloud, renting as many resources as they need to get the simulation done in a timely manner (represented by the red line). This not only accelerates research timelines but also empowers businesses to take on more ambitious, resource-intensive projects without being constrained by their existing infrastructure.

Concluding thoughts

This isn't an exhaustive list of the ways in which cloud providers are transitioning to higher-value, differentiated services. But these examples should begin to give us a sense of why cloud spending is far outpacing IT spending.

In the cloud-as-utility model, the cloud’s growth was a function of the growth of traditional IT. When businesses expand their IT needs—whether that's increasing the amount of data storage or the number of employees who need access to basic enterprise software—the demand for cloud infrastructure grows in parallel. Under this model, the cloud didn’t enable fundamentally different types of business activities, but was simply a utility that offered a more efficient way to power existing operations.

By contrast, the cloud-as-innovation platform model flips this on its head. Instead of merely being a byproduct of traditional IT expansion, the cloud itself becomes the driver of new business activities and technological breakthroughs. In this model, the cloud isn’t just a more efficient way to run existing software—it’s the foundation that enables entirely new capabilities that wouldn’t be possible in the traditional infrastructure model. In other words, the cloud has transformed from a traditional cost center to a powerful source of competitive advantage and a force multiplier for innovation.

As should be clear by now, AI is one of the biggest examples of the kind of value-added service that the cloud-as-innovation platform business model has enabled. Unlike traditional IT workloads, AI requires vast computational resources, specialized hardware, and sophisticated software frameworks—resources that most businesses simply cannot build or maintain on their own.

Had the cloud remained a pure utility—focused only on providing basic, fungible infrastructure—we might not have the AI capabilities we see today. Training today’s generative AI models requires thousands of high-performance GPUs running in parallel, supported by complex data pipelines and finely tuned algorithms. These capabilities didn’t emerge overnight. They were gradually developed and made accessible, in large part, through the evolution of the cloud business model.

Yet since the release of ChatGPT, the cloud has often been portrayed as nothing more than an appendage to the generative AI boom. Discussions about its role have largely centered on AI training, inference workloads, and the computing power required to sustain large-scale models. What’s often overlooked, however, is that many of these trends began long before generative AI took center stage.

This brief history of the cloud’s evolution should remind us that (1) the cloud existed long before the AI boom, and (2) it serves as a platform for a vast range of innovations beyond AI. From scientists using it for protein folding research to advance drug discovery, to quants running complex financial simulations, to climate scientists modeling weather patterns, the cloud-as-innovation platform model has empowered industries across a wide range of fields with faster experimentation, deeper analysis, and more sophisticated simulations.

The AI boom has pigeonholed the cloud and its vast infrastructure as nothing more than the platform on which AI is trained and deployed, creating a kind of myopia over the breadth of innovation that is possible with today’s technologies. If we continue to see the cloud solely through the lens of AI, we risk overlooking its broader potential—and in doing so, limiting our collective imagination for what the future holds.

China’s digital economy is among the most advanced globally, second only to the U.S. and even surpassing it on some metrics. Yet missing from China's digital economy are homegrown counterparts to firms like SAP, Oracle, Salesforce, ServiceNow, Atlassian, and Databricks. Instead of building enterprise software, the country's leading tech firms are overwhelmingly consumer-facing, specializing in businesses like e-commerce, digital entertainment, ride-hailing, and food delivery.

To be sure, there are Chinese tech firms that have ventured into enterprise services, especially in recent years. Kingsoft produces office software. And even Alibaba and ByteDance have launched B2B SaaS products and expanded into cloud computing. Still, if we look at SaaS-based revenue or the number of SaaS firms (see graph below), its clear that China lags significantly in this segment.

Why has China’s digital economy, built on the same foundational technologies as the U.S.'s, developed along such a distinctly consumer-facing path?

One reason is protectionism. Since the late 2000s, China’s Great Firewall has blocked many foreign consumer platforms such as Google, Facebook, and Twitter, from China's market, giving domestic firms a clear runway to grow and dominate in China. (This undoubtedly played a role, but it alone cannot explain the pattern. After all, some foreign consumer platforms like Uber and Amazon were permitted to operate in China, but failed to gain a foothold against domestic competitors like Didi and Alibaba.)

Another reason is that Chinese enterprises tend to spend less on IT due to its industrial composition. As Grace Shao and I wrote in a previous post, enterprise SaaS benefits white-collar productivity making it most valuable in knowledge-intensive sectors. The U.S. has a mature knowledge economy with high IT-spending industries like finance, insurance, and professional services making up a large share of its GDP. By contrast, the vast majority of Chinese GDP comes from labor-intensive sectors like manufacturing, construction, agriculture—sectors that are less likely to spend on IT.

In this post, we'll turn to a less frequently discussed factor: China's labor regime.

To see the connection, we first need to understand what makes consumer-facing platforms tick. Consumer platforms thrive on rapidly acquiring users and scaling their services. Their success hinges on achieving strong network effects such that the value of the platform increases as more users join it. This dynamic creates incentive for firms to constantly pivot in new markets in search for exponential growth.

Given this model, tech firms must be able to swiftly reallocate its resources and personnel. High employee turnover and employment volatility are thus not merely incidental but central to the consumer-facing business model. These firm benefit significantly from employment flexibility, enabling them to hire quickly when expanding and shed staff when dropping unprofitable initiatives.

Complementarily, consumer platforms also benefit from general rather than firm- or domain specific skills. Laszlo Bock, Google’s former senior vice president of people operations, said that when it comes to hiring, "general cognitive ability" was the top sought-after attribute, followed by leadership skills, humility, and a sense of ownership and responsibility. The least important attribute, according to Bock, was expertise. This preference for general skills is reflected in the way most consumer platforms interviews candidates, focusing not on specific technical questions but on applicants’ ability to solve math or logic puzzles as a way to demonstrate this "general cognitive ability" (e.g., Design an algorithm to sort a list in ascending order). The result is a workforce built for speed and flexibility—precisely the kind of labor engine that consumer platforms need to move fast and break things.

Fast hiring and firing

China’s labor regime is uniquely suited to the needs of consumer platforms. Over the past few decades, China’s employment relations have shifted dramatically away from stable jobs with extensive benefits that the country was once known for (the so-called “iron rice bowl” model) towards highly precarious employment structures that grant firms extraordinary employment flexibility.

This flexibility allows tech companies to rapidly scale up their workforce in boom periods and scale down just as quickly during downturns. For example, Alibaba’s workforce grew from around 20,000 in 2014 to over 250,000 by 2021, while ByteDance exploded from just a few thousand to over 100,000 employees in less than a decade. At the same time, Chinese tech companies frequently conduct mass layoffs, quickly shedding thousands of employees at the first sign of instability. This happened, for instance, during the 2018 U.S.-China trade war or after the 2021 government regulatory crackdowns. As a result, employee turnover is exceptionally high. Between July 2021 and March 2022 alone, major Chinese internet companies shed over 216,800 jobs but also added nearly 300,000 new roles.1 (As a point of comparison, large-scale hiring and firing is far more difficult under the EU's stricter labor regime, making it more challenging for consumer platforms to thrive there.)

The ability to rapidly hire and scale opens up business strategies that are often unavailable to internet firms in other countries. One example is ByteDance’s launch of Douyin, the Chinese predecessor to TikTok. Douyin started as a latecomer to the short-form video format, which acccording to conventional wisdom about platform firms, should've precluded it from success since it did not have a first- or early-mover advantage.2 Being a late-comer, however, did not stop Douyin's user base from exploding.

In the mid-2010s, the main constraint in the short-form video market was not user demand but a shortage of content creators. To overcome this, Douyin hired nearly 20,000 people in 2016 alone to form an ad-hoc "operations" team that would be tasked with cultivating the supply side of the platform. These workers engaged in daily outreach to creators, maintained ongoing relationships with users and promotional partners, and actively scouted new talent across the country. A key discovery was that art school graduates were a rich source of potential influencers.3 The operations team invited them to join the platform and supported their growth—sometimes acting as informal PR agents to help them reached their audience where the platform’s algorithm couldn’t. This ability to creatively make use of China's flexible labor regime gave the platform a competitive advantage, eventually propelling the Douyin to the top of the short-video sector. Douyin's case shows that market deficiencies (the lack of content creators in this case) was remedied by being able to rapidly hire a team of digital foot soldiers that could, through labor-intensive brute-force methods, generate the supply side of their intended market. Today, these "operations" teams are common across consumer platform firms in China.

Preference for general skills

The lack of employment protections in China also encourages young, educated workers to develop broad, transferable skills instead of deep, firm- or domain-specific expertise. Chinese tech companies frequently operate under an unspoken "up-or-out" policy, where employees who fail to attain a particular rank by their mid-thirties (usually 35) are quietly forced out. This practice is intertwined with China’s notoriously grueling “996” work schedule (9am to 9pm, six days a week). Younger workers, often unmarried and without family obligations, are far more willing—and physically able—to tolerate long work hours. Consequently, job security diminishes with age, and companies systematically prefer younger, more expendable workers.

On the flip side, younger workers also tend to leverage their youth to job hop, negotiating for larger compensations with each move. Some also burn out and leave the industry well before hitting their mid-thirties. As a result, employee turnover is high and leaving Chinese tech workers with some of the shortest average employment tenures—2.6 years, compared to 3.65 years in Silicon Valley.4

This churn means that there is little payoff for workers to invest in firm- or domain-specific skills, since most won't be able to reap its benefits in the long haul. Instead, workers hedge against obsolescence by cultivating transferrable, general skills to job hop and remain competitive. For consumer platforms, who are constantly pivoting, testing new markets, and pushing employees onto new projects, this emphasis on general skills is ideal.

Labor cost

While consumer platforms greatly benefit from flexibility and general skills among its white collar workers, their greatest advantage comes from the highly permissive labor regime that governs low-wage gig work such as food delivery, logistics, and ride-hailing. These sectors have ballooned precisely because tech companies can tap into China’s enormous migrant worker population, estimated at over 220 million.

The Hukou system further amplifies worker precarity. Migrants in urban centers, without local Hukou status, lack full access to education, healthcare, and other social benefits, forcing them to accept precarious and low-paid gig jobs provided by digital platforms. This source of cheap and precarious labor is indispensable for China's consumer platforms. Unsurprisingly, many of China's largest tech companies leverage cheap labor in some aspect of their business model.

Low labor costs likewise suppress demand for SaaS and other enterprise software in China. When office wages are so low, firms have little incentive to invest in tools that automate tasks or digitize office workflows. Instead, Chinese enterprises may often find it more cost effective to substitute human labor for software, scaling headcount rather than buying or subscribing to costly enterprise software. So as long as labor remains abundant and inexpensive, Chinese enterprises face little pressure to adopt SaaS solutions, preferring instead the flexibility and low capital outlay of cheap office workers.

Concluding thoughts

In 2021, Beijing cracked down on its homegrown consumer tech champions. Alibaba was slapped with a record RMB 18.2 billion antitrust fine, Didi was forced to delist from the New York Stock Exchange and submit to exhaustive data-security audits, and limits on video games sent Tencent into its slowest revenue growth in nearly two decades. These measures did not only rein in the tech giant's monopoly power, but also signaled a strategic move to steer capital, talent, and R&D away from consumer platforms.

Yet nearly four years after the crackdown, China’s digital ecosystem remains overwhelmingly skewed toward consumer platforms, and that's largely because the underlying labor regime has remained unchanged.

As AI becomes mainstream, China’s distinctly consumer-oriented digital economy will have path dependent effects on the diffusion of this new technology. Rather than being deployed to boost productivity and support enterprise upgrading, it may well end up reinforcing the existing logics of China’s consumer giants—maximizing user retention, monetizing attention, and exploiting the country’s low-cost and flexible labor regime.

The AI news cycle about China can be overwhelming and confusing. The narrative can swing wildly with each model breakthrough. One day, the U.S. is portrayed as maintaining an unshakable lead. The next, China is hailed as the world’s new AI superpower. These seemingly contradictory claims, however, are often both true. But that truth hinges on what we mean when we say "ahead" and how we choose to measure it.

Indeed, evaluating how advanced a country is in a broad technical domain like AI is not trivial. For instance, whether a country has the capabilities to produce new AI model paradigms requires vastly different capabilities from making incremental improvements on existing models. And both differ again from a country’s ability to diffuse AI across an economy and drive productivity gains. The point is that the "AI race" isn't being run on a single lane.

In this post, I use Epoch AI's dataset1 of notable models to take a systematic look at the past decade of AI innovation and to see where China and the United States stand and what each of their strengths are. And as we shall see, where the two countries stand in AI innovation is much more nuanced than what most media outlets tend to portray.

Before diving into our comparison, let me first highlight that the following analysis (unless otherwise stated) is limited to AI models that are published by industry players. The reason I do not focus on academic institutions is because industry is overwhelmingly the primary producers of notable AI models according to the database2, so if the following analysis were to have any predictive value, then including academic contributions may end up only creating noise.

The U.S. is the clear leader but China is rapidly catching up

For several years, scholars and journalists have pointed to China’s rapid catch-up with the U.S. in the sheer volume of STEM research output. However, they’ve also noted that the quality of this research (often measured by acceptance into top conferences or citation counts) has continued to lag behind.3

The pattern holds true in AI as well. While the volume of AI research coming out of China has surged—surpassing the U.S. by some measures—the overall quality still lags behind. When we rely on Epoch AI’s classification of notability (which I’ll explain shortly), the U.S. remains the clear leader. Over the past decade, the development of high-impact AI models has been overwhelmingly concentrated in a small number of U.S.-based tech firms.

But a decade is a long time. When we look at the year-by-year trend, what we see is that since around 2019, Chinese firms have steadily increased their share of notable AI models, gradually chipping away at U.S. dominance in the field. By 2024, Chinese firms represent nearly 30 percent of notable models, up from just 6 precent in 2019. If the trend continues, we should expect to see China playing a far more dominant role in producing notable AI models in the coming years.

This is a bit of a side point, but I think its also worth emphasizing that innovation doesn’t have to be a zero-sum race. Recent research shows that cross-country collaboration can in fact yield the best research outcomes in AI. Using data since 2000, a recent Nature report shows that American and Chinese researchers (including academic researchers) produce more impactful AI research on average when they collaborate with each other.4

Varieties of AI innovation

So far, the data shows that China is emerging as an increasingly important player, not only in the overall volume of AI models it produces, but also in its growing share of notable models. However, even within what are considered “high impact” or “notable” AI models, there is important variation that we should pay attention to. Epoch AI uses four different criteria to satisfy their standard for notability. The first is whether it has received more than 1,000 citations. The second is whether it holds indisputable historical significance (e.g., AlexNet, GANs). The third is whether it has more than one million active users. And the fourth is whether it represents a clear improvement on the state of the art (SOTA).

We can use the first two indicators (high citations and historical significance) as proxies for radical innovation, since these typically reflect paradigmatic breakthroughs that shift some aspect of the trajectory of the field or opens up entirely new research directions. The latter two (high usage and improvements on the SOTA) can serve as proxies for incremental innovation since they are about optimizing and improving upon existing models and building widely adopted applications on top of that.

The dataset also tracks another category of notable models: models with training costs exceeding $1 million. This metric does not strongly indicate whether a model represents radical or incremental innovation. Instead, it reflects scale and whether an organization has the financial and technical resources to support large AI training efforts.

When we split up the number of notable AI models into these categories, we see that the vast majority of China’s growth in AI innovation comes from incremental innovation and scale. The number of AI models in the domain of radical innovation remains low with no sign of growth.

One important limitation of this dataset is that using citation counts introduces a bias against more recent models. Whereas an improvement on the SOTA is normally immediately apparent (and is usually articulated in the model’s corresponding research paper), reaching 1,000 citations takes time, so recent breakthroughs no matter how notable may be excluded simply because they haven’t had the time to accumulate citations. Moreover, we may also expect a shift from radical to incremental innovation in AI since there may simply be less room for continued radical innovation as the technology matures.

That said, when we examine the breakdown for U.S.-based firms, we see that while the number of models representing incremental innovation has increased over time, there remains a relatively stable foundation of radical innovation, with around 20 highly cited or historically significant models produced each year. This suggests that the U.S. maintains a unique capacity for radical innovation in AI that China lacks.

China's focus in diffusion-oriented innovation

The relative lack of radical innovation in China shouldn’t be mistaken for a lack of creativity or ambition. Rather, China’s capabilities in incremental and process-oriented innovation reflects a broader strategic focus on the diffusion (rather than the invention) of new technologies. And as a result, this approach enables more actors across the economy to share in the productivity gains that these technologies make possible. In other words, this mode of innovation delivers real dividends in sustainable economic growth, industrial upgrading, and technological sovereignty.

By contrast, the U.S. innovation system has long prioritized and rewarded radical, zero-to-one innovation. Strong intellectual property protections incentivize breakthrough inventions by granting exclusive rights to early movers. Likewise, the country’s approach to industrial policy has focused on pushing the technological frontier. Agencies like DARPA, for instance, are explicitly tasked with funding projects that are 10 to 15 years ahead of mainstream commercial adoption.

Some scholars have argued that Chinese companies excel in what’s often called “second-generation” or “applied” innovation which focuses on incremental improvements to products, processes, or manufacturing techniques. Dan Breznitz and Michael Murphree, writing in 20115, observed that Chinese firms often operate at the edge of the technological frontier without necessarily pushing it forward. They find that firms don’t need to be radical innovators to succeed. Rather, it’s often enough to make products cheaper, faster, more precisely, and at higher quality.

This theory has largely played out across several strategic sectors in China. In batteries, solar photovoltaics, electric vehicles, and semiconductors, Chinese firms have become global leaders not by inventing entirely new technologies, but by mastering and refining existing ones at scale. And as the data above shows, this also appears also to be the trend in AI with Chinese firms excelling not in entirely new frontier models, but in incremental improvements on SOTA models.

DeepSeek is an example of this—not pushing our imaginary of what LLM-based AI models can do, but making it significantly cheaper and more efficient to train. And as in the case of other industrial sectors, this mode of innovation lends itself better to the diffusion of technology. Since the release of DeepSeek, we've seen uptake of their more inexpensive model across China”s digital ecosystem.

That said, China faces different challenges when it comes to AI diffusion. The main issue is that internet companies like Alibaba and Tencent are the primary engines of AI diffusion in China. As Grace Shao and I have explained, this means adoption is likely to remain concentrated in consumer-facing applications as these internet firms have built their core strengths around platform services for individual users, not enterprise clients. In contrast, the U.S. has a long history of building enterprise software, along with higher overall business investment in IT and digital tools- giving U.S.-based AI firms a clearer path to driving adoption across a wider range of industries.

There’s a bit of irony in this. Since 2021, the Chinese state has tried to pivot away from “soft tech” sectors—consumer-facing platforms like e-commerce, food delivery, and social media—and instead prioritize “hard tech” domains such as advanced manufacturing, renewables, and other drivers of industrial productivity. In this plan, advanced technologies like AI are meant to accelerate this transition from soft-to-hard sectors. So the fact that it’s only the internet giants that have the real technological capacity to innovate on and diffuse frontier AI models probably isn’t exactly what Beijing had in mind.