I believe anyone following the technology space will not forget January 27, 2025—the day Nvidia lost $590 billion in market capitalization, the largest single-day loss in stock market history. The catalyst was a Hangzhou startup called DeepSeek, which had released a reasoning model that matched OpenAI's best work at roughly one-twentieth the training cost. Within hours, the entire premise of the American AI boom—that whoever had the most GPUs would win—looked suddenly fragile.

The "DeepSeek Moment," as it came to be called, was immediately understood as a technical achievement. What has taken longer to understand is that it revealed something more basic and important between the lines: a divergence between two different theories of how AI value gets captured. American AI is built to win at the cash register. Chinese AI is built to win on the cost curve. These are not the same competition, and treating them as such obscures more than it illuminates.

The Constraint That Became A Weapon

The story of Chinese AI over the past 3 years is one of brilliant adaptation to constraints. When the US restricted exports of advanced Nvidia chips in October 2022, the conventional wisdom or expectation was that China's AI ambitions would stall. Instead, Chinese labs did what constrained competitors often do: they innovated around the bottleneck.

DeepSeek's breakthrough was architectural. Rather than building dense models in which every query activates every parameter—the brute-force approach that defined GPT-4—DeepSeek built sparse models using a Mixture-of-Experts architecture. You can think of it as the difference between turning on every light in a building versus installing motion sensors that illuminate only the rooms you're actually using. The result was a model with hundreds of billions of parameters that activated only around 37 billion per query, dramatically reducing computational requirements – this was unheard of until then. 

This was genuine innovation, and it worked really well. DeepSeek-R1 matched OpenAI's o1 reasoning model on key benchmarks. The training cost was reportedly $5.6 million, compared to $100 million or more for comparable American models. Chinese engineers had proven that algorithmic efficiency could substitute for raw compute, at least temporarily and partially.

But here's the thing about efficiency breakthroughs in competitive markets: they benefit everyone. DeepSeek's innovations weren't trade secrets; they were published papers and open-source code. Within months, every major AI lab—American and Chinese alike—had incorporated similar techniques. The gap that DeepSeek had briefly closed through cleverness began reopening as American labs applied the same logic to their vastly larger compute clusters.

Here’s something that needs important consideration. What didn't reopen was the open-source gap. By September 2025, Alibaba's Qwen had become the most downloaded model family on Hugging Face, dethroning Meta's Llama. Downloads of Chinese open-weight models now exceed American ones. China didn't just catch up on efficiency—it took the lead on diffusion.

More concerning for the containment strategy: in early 2026, Zhipu AI announced it had trained GLM-Image, a multimodal model, entirely on Huawei's Ascend accelerators—no American silicon involved. This is a milestone, not proof of parity. Independent assessments haven't yet confirmed whether domestically-trained models match the performance of those trained on Nvidia's latest hardware. But the strategic significance may outweigh the technical maturity: if Chinese labs can train competitive models without Western chips, the leverage provided by export controls erodes regardless of whether they've fully closed the performance gap. The algorithmic innovations that DeepSeek pioneered—Mixture-of-Experts, sparse attention—reduce the performance threshold required to make domestic training viable. Even partial success here shifts the calculus.

The real effect of the DeepSeek Moment wasn't that China caught up. It was that everyone got cheaper. And when prices fall, the advantage goes to whoever has greater demand, better distribution, and stronger pricing power. On that axis, the American ecosystem is undeniably ahead.

But ahead on what dimension? This is where the standard narrative starts to mislead.

The Business Model That Isn't Universal

To understand what's actually happening, you have to understand something elemental about enterprise software markets: they are not natural. They are constructed.

When Marc Benioff launched Salesforce in 1999, he wasn't just selling software—he was training an entire generation of CFOs and CIOs to think about technology spending in a drastically new way. The subscription model, per-seat pricing, automatic updates, and predictable monthly costs: none of this was obvious or inevitable. It took two decades of sustained effort to make American enterprises comfortable paying for software as a recurring expense rather than a one-time purchase.

This infrastructure doesn't exist in China—at least not in the same form. For every dollar Chinese companies spend on hardware, they spend about 50 cents on software—compared to over $3 in the US. Enterprise buyers demand perpetual licenses, not subscriptions. State-owned enterprises insist on on-premises deployments rather than cloud services. And while WeChat Pay and Alipay technically support automatic recurring payments, the implementation reflects a different relationship with subscriptions: users must explicitly opt into auto-deduction agreements for each service, platforms are selective about which merchants can access the feature, and the cultural norm treats recurring charges with suspicion rather than convenience. The technical capability exists; the behavioral infrastructure doesn't.

The numbers tell the story with brutal clarity. OpenAI reached $20 billion in revenue by the end of 2025, with roughly 15-20 million consumers paying $20 per month for ChatGPT Plus and another 1.5 million enterprise customers on higher-tier plans. Anthropic, buoyed by the success of its Claude Code agent, raised its 2026 revenue forecast to $18 billion and is currently raising $20 billion at a $350 billion valuation. In the twelve months to October 2025, ChatGPT generated $1.7 billion through Apple's App Store alone. The top ten Chinese AI chatbots combined? $500,000. This is a ratio of 3,400 to 1—not a gap, a chasm.

But here's what the revenue numbers obscure: OpenAI is projecting a $14 billion loss in 2026. $20 billion in revenue, $34 billion in costs. The cash register is ringing, but the money is flowing right back out to fund the next generation of models. Anthropic and Google are in similar positions—racing to spend faster than they earn in hopes of achieving escape velocity before the capital runs out. The American advantage isn't profitable unit economics; it's access to capital markets willing to fund staggering losses at unprecedented scale.

The structural disparity with China runs deeper than consumer behavior. Chinese companies spend roughly $50 billion annually on software—less than a tenth of what American companies spend. Again, this isn't a pricing problem or a marketing problem. It's a market that fundamentally doesn't value software as much as the American market does.

Now consider MiniMax, the Chinese AI company that completed a landmark IPO in Hong Kong in January 2026—shares surged 80 percent on debut, valuing the company at $11.4 billion. The company had 200 million users globally for its Talkie companion AI app. Its 2024 revenue was $30 million. Its 2024 losses were $465 million. That's a loss multiple of fifteen times revenue.

Zhipu AI, the Tsinghua-backed company positioning itself as China's OpenAI, raised $558 million in its January 2026 IPO at a $4.4 billion valuation—despite being added to the U.S. Entity List, which restricts its access to U.S. technology. Its 2024 revenue was roughly $45 million, compared with $410 million in losses. 85 percent of that revenue came from on-premises deployments for government clients—labor-intensive, low-margin project work rather than scalable subscriptions.

If you evaluate Chinese AI companies by Silicon Valley metrics, they look like failures. But here's the question that evaluation begs: what if Silicon Valley metrics are measuring the wrong thing?

The DeepSeek Distortion

Before we can answer that question, we need to understand the force that's making it impossible for Chinese AI startups to find any sustainable economics: DeepSeek doesn't need to make money.

Most AI startups are backed by venture capital, which means they eventually need to generate returns for their investors. DeepSeek is backed by High-Flyer Quantitative, one of China's most successful hedge funds with roughly $8 billion in assets under management. The AI lab is effectively a research division funded by trading profits—a strategic asset rather than a profit center.

When DeepSeek released its V2 model in May 2024, it priced API access at $0.14 per million input tokens—five to ten times below prevailing market rates. Within months, average LLM prices in China collapsed by 92 percent. Alibaba slashed Qwen pricing from $1.10 to $0.07 per million tokens. ByteDance launched Doubao at $0.04 per million tokens.

The price collapse was amplified by extraordinary market crowding. By September 2025, there were more than 500 Chinese AI models on offer—up from just 14 two years earlier. DeepSeek's efficiency breakthrough triggered what analysts called a culling of weaker players, but even after the shakeout, the market remains intensely fragmented. Too many models chasing too little willingness to pay.

At these prices, the API business isn't a business. It's closer to infrastructure provisioning or national capability building. Once you frame it that way, DeepSeek's behavior no longer looks irrational; it becomes a deliberate strategy: commoditize the model layer so that value can be captured elsewhere.

This is devastating for venture-backed startups like MiniMax or Zhipu, which need to show revenue growth to justify their valuations. But it's not necessarily bad for Chinese AI as a system. It depends entirely on where you expect value to accumulate.

The Android Precedent

This is where historical analogies become useful—not as predictions, but as stress tests for assumptions.

Consider the Android OS. Google released it as open-source software in 2008, effectively commoditizing the mobile operating system layer. Microsoft, which had spent years trying to build a profitable mobile OS business, was devastated. So were the carriers who had hoped to control the software layer. From a traditional software business perspective, Android lost—Google never charged licensing fees, never built the kind of per-device revenue stream that Microsoft had with Windows.

But Google didn't need Android to be a business. It needed Android to be a distribution channel. By ensuring that the default search engine, maps application, and app store on 70% of the world's smartphones were Google properties, the company captured value at a different layer than the OS itself. The loss at the software layer was offset by a win at the services layer.

Chinese AI may be following a similar logic. If foundation models become commoditized—and DeepSeek is actively working to ensure they do—then the value migrates elsewhere: to the applications built on top, to the hardware the models run on, to the data generated by billions of users, to the infrastructure deals that bundle AI with telecommunications and power systems.

The solar industry tells a similar story. Chinese manufacturers lost for years, operating at thin or negative margins while American and European competitors chased profitability. But by accepting those losses, they achieved scale that eventually made them the default supplier globally. Today, China controls over 80% of solar panel manufacturing. The companies that won the margin game in 2010 are mostly gone. The companies that won the volume game own the industry.

Telecom infrastructure shows the same pattern. Huawei spent decades offering equipment at prices Western competitors couldn't match, often bundled with financing that made the total cost even lower. From a pure-margin perspective, Huawei was losing. From a market-share perspective, building the installed base now makes it the dominant infrastructure provider across the Global South.

In each case, the Western frame—focused on margins, ARR, and per-unit profitability—declared victory, even as the actual market shifted beneath it.

The Global South Isn't A Consolation Prize

The standard narrative treats Chinese AI's strength in emerging markets as a fallback—the places they're forced to compete in because they can't win in America or Europe. This framing misreads the opportunity.

The Global South represents roughly two-thirds of the world's population. It's where smartphone penetration is still growing, where first-time AI adoption is happening, and where the next billion internet users will come online. Dismissing these markets as low-value mistakes, current revenue for future potential.

More importantly, it mistakes the nature of the competition. If Chinese AI becomes the default stack—embedded in Transsion phones across Africa, integrated into BYD vehicles across Southeast Asia, bundled with Huawei-built telecommunications infrastructure across Latin America—then monetization doesn't need to look like SaaS subscriptions. It can look like Android: value captured through distribution, data, and ecosystem lock-in rather than direct software revenue.

The investment follows this logic. Goldman Sachs estimates that China's major cloud providers—Alibaba, Baidu, ByteDance, and Tencent—will invest $70 billion this year in data center infrastructure, with particular focus on Asia, Latin America, and the Middle East. This isn't defensive spending; it's building the physical layer of an alternative AI stack that will be formidable. 

The numbers already hint at this. 73 percent of MiniMax's revenue comes from international markets. Nineteen of the twenty-three Chinese AI products in the global top 100 by revenue generate most of their income overseas. In November 2025, Singapore's national AI program announced that its Sea-Lion model had switched from Meta's Llama to Alibaba's Qwen, running on Alibaba's cloud infrastructure—a small but symbolic defection from the American stack.

Even in the West, Chinese models are gaining ground, though unevenly. Airbnb now uses Qwen. Startups prize Chinese models for their efficiency and zero licensing costs. But enterprise adoption remains shallow—Chinese models account for only about 10 percent of open-model usage among American enterprises, according to Menlo Ventures. Large companies cite data privacy concerns and the risk that Chinese models could be banned, which would force costly system migrations. They're willing to experiment, not to depend.

Chinese AI companies are building their businesses outside China, not because they've failed at home, but because the home market was never going to support venture-style returns regardless of their technical achievements.

The question isn't whether this constitutes a real business by Silicon Valley standards. The question is whether it constitutes a durable strategic position. Those are different questions with potentially different answers.

The Startups Are Losing; The System Is Advancing

This brings us to the reframe that the standard narrative misses: Chinese AI startups are failing, but Chinese AI as a national capability may be succeeding through entirely different vehicles.

The Six Little Tigers—DeepSeek, MiniMax, Moonshot, Zhipu, 01.AI, Baichuan—are indeed in crisis, though their responses vary. MiniMax and Zhipu rushed to Hong Kong IPOs in January 2026 as survival mechanisms. Moonshot AI, backed heavily by Alibaba and Tencent, just released Kimi K2.5, featuring Agent Swarms—still competing on the frontier. Baichuan pivoted to medical vertical AI, seeking a sustainable niche away from general LLM competition. And 01.AI, founded by Kai-Fu Lee, has reportedly stopped training new foundation models entirely, instead building applications on open-source bases like DeepSeek.

The pivot signals are unmistakable. When Kai-Fu Lee—one of the most prominent AI investors in the world—concludes that competing on foundation models is a losing proposition, that clarifies the strategic landscape. The startups that survive will be those that find defensible verticals or get absorbed by the giants.

But the platform giants—Alibaba, ByteDance, Tencent, Baidu—can sustain AI losses indefinitely. They don't need AI to be a standalone business; they need AI to enhance their existing businesses. Alibaba's Qwen models drive users to Alibaba Cloud. ByteDance's Douwithyin integrates into TikTok and its Chinese equivalent, Douyin. Tencent embeds AI across WeChat, reaching its billion-plus users. These companies don't show up in AI startup analyses, but they may be the actual vehicles through which Chinese AI achieves scale.

Similarly, hardware manufacturers are integrating AI in ways that bypass the software monetization problem entirely. BYD, Xpeng, NIO, and Huawei's Aito vehicles feature increasingly sophisticated AI systems—but the AI isn't sold separately. It's a feature that makes the car worth more. Xiaomi, Oppo, and Vivo are building AI directly into their phones. The AI business here is invisible because it's subsumed within product margins rather than reported as a line item.

This is the strategic logic that the DeepSeek Moment actually revealed. By commoditizing the model layer, China ensures that AI capabilities become cheap and ubiquitous, benefiting players who monetize through hardware, platforms, and infrastructure rather than software subscriptions. The startups die. The system advances.

The Cash Register Isn't Permanent

None of this means American AI companies are in trouble. OpenAI's $20 billion revenue is real. Anthropic's enterprise traction is real. Microsoft's ability to charge $30 per user per month for Copilot across the Fortune 500 is real. The SaaS business model works, and it works well, for now.

But for now is doing meaningful work in that sentence. The cracks are already visible.

OpenAI's share of AI traffic fell from 86 percent to 64 percent over the past year as competition intensified. In January 2026, Google launched Google AI Plus at $7.99 per month—deliberately undercutting OpenAI and Anthropic by 60 percent to capture the mass market. This is segmentation, not surrender; enterprise pricing remains intact. But it's early pressure on the assumption that Western AI can sustainably command premium consumer prices. If the American giants are willing to compete on price with each other at the consumer tier, the margin compression that devastated Chinese startups may eventually reach Western shores as well.

Several other forces could accelerate the erosion:

  • Enterprise fatigue with AI subscriptions. Companies that rushed to adopt AI tools in 2023-2024 are starting to ask hard questions about ROI. If the productivity gains don't materialize at the promised level, the willingness to pay $30/seat/month may soften.
  • Open-source erosion of premium features. Every capability that OpenAI or Anthropic ships is replicated in open-source within months. The feature gap that justifies premium pricing keeps narrowing.
  • Regulatory pressure on AI pricing. Governments are increasingly interested in ensuring AI access isn't stratified by ability to pay. Mandates for public-sector AI, educational AI, or healthcare AI could pressure commercial pricing models.
  • Sovereign AI demands. Even Western governments are growing uncomfortable with dependence on a handful of American AI providers. The push for on-premises deployments—which China already dominates—could spread.

If inference costs keep falling, value migrates upward from models to workflows, integration, and ownership. That's a shift that could weaken centralized SaaS players faster than their current growth rates suggest.

The Implications

So, what does all this mean? The answers depend on where you sit.

  1. For Chinese AI companies, the strategic imperative is to stop fighting for a business model that doesn't fit their market. The startups that will survive are those building for hardware integration, vertical specialization, and Global South distribution—not those trying to replicate OpenAI's subscription machine. The pivot patterns are already clear: vertical niches like healthcare, applications built on commoditized open-source models, and aggressive international expansion. Fighting DeepSeek on price is suicide; finding adjacent value pools is survival.
  2. For Western competitors and investors, the lesson is that technical moats are temporary but business model moats—while more durable—are not permanent. The comfortable assumption that "we own the cash register" deserves scrutiny. Meta's recent move to acquire Manus, a Chinese AI lab that had relocated to Singapore, for over $2 billion signals that even American giants recognize the value being created in the Chinese ecosystem—and are willing to pay handsomely when geopolitics permit the transaction. If Chinese AI achieves ubiquity through hardware and infrastructure, the installed base and data advantages could eventually support alternative monetization paths. The Android precedent suggests that winning the margin game and winning the market are not always the same thing.
  3. For policymakers, the DeepSeek Moment should prompt a more nuanced assessment than simply assuming that containment is working or that it has failed. The export controls have maintained a meaningful compute gap—roughly 8:1 in deployed AI accelerators. But they have not prevented China from reaching technical parity on key benchmarks, nor will they prevent Chinese AI from becoming the default infrastructure across the Global South. The question is whether that outcome is acceptable, and if not, what alternative strategies might address it.

Two Theories Of Victory

The framing that best captures this moment is not who's winning the AI race. It's that there are two different races happening simultaneously, with two different finish lines—and neither has yet produced a profitable business at scale.

This isn't necessarily a crisis. The internet wasn't profitable in 1999 either, and the companies that survived the shakeout became the most valuable in history. But it does mean that declarations of victory from either side are premature. We're watching two different bets play out, each with its own logic and risks.

The American theory of victory is value capture at scale: build the best models, charge premium prices, generate massive revenue, and use that revenue (plus investor capital) to fund the next generation of even better models. This is the OpenAI/Anthropic playbook. It's generating $20 billion in revenue—and $14 billion in losses. The bet is that scale will eventually bend the cost curve, that enterprise lock-in will enable price increases, and that the capital markets will keep funding the gap until profitability arrives.

The Chinese theory of victory is value diffusion: commoditize the model layer, achieve ubiquity through low prices and open-source distribution, and capture value through hardware, platforms, and infrastructure rather than software margins. This is the DeepSeek/Alibaba/Huawei playbook. It's generating minimal direct revenue—but it's building installed base, distribution channels, and ecosystem dependencies that may monetize differently over time.

Neither theory has been proven yet. The American approach is winning the revenue race while losing money at historic rates. The Chinese approach is winning the adoption race while also losing money at historic rates. The difference is that American companies are losing billions to fund frontier research, while Chinese startups are losing billions just trying to survive a market where the model layer has been commoditized to near-zero margins.

These theories aren't mutually exclusive in terms of outcomes. Both can succeed in their own domains. The American approach will likely dominate enterprise knowledge work, frontier research, and high-margin professional applications. The Chinese approach will likely dominate edge devices, consumer applications in price-sensitive markets, and infrastructure across the developing world.

The question is which domain matters more in the long run. The answer isn't obvious. The high-margin enterprise market generates more revenue today, but the high-volume consumer and infrastructure markets shape a larger share of the world's installed base. Revenue is one form of power. Ubiquity is another.

The DeepSeek Moment didn't resolve this question. It clarified that the question exists—that we're not watching a single race with a single winner, but a bifurcation into two different AI ecosystems with two different logics of success.

China won the cost curve. America owns the cash register—but the cash register is hemorrhaging money too. Both statements are true. The question isn't which side is winning today—that framing assumes a single race with a single finish line. The real question is which theory of victory proves correct when the capital runs out and someone, somewhere, finally has to demonstrate that artificial intelligence can be not just transformative, but sustainable.

That question remains genuinely open. And that's what makes this moment so consequential.