The Gap That $285 Billion Couldn’t Buy
Here’s a number that should make every OpenAI and Anthropic investor lose sleep: the United States spent 23 times more than China on private AI investment in 2025. $285.9 billion versus $12.4 billion. And the performance gap between the best American and Chinese AI models? According to Stanford HAI’s 2026 AI Index, it has shrunk to 2.7 percentage points.
In May 2023, that gap was 17.5 to 31.6 points. Three years and hundreds of billions of dollars later, it’s a rounding error.
DeepSeek V4 Pro scores 87 on BenchLM. GPT-5.4 scores 88. DeepSeek openly acknowledges trailing the absolute frontier by “approximately 3 to 6 months.” But V4 is open-weight, has 1.6 trillion parameters, and costs four times less to use than its American competitors. The question isn’t whether the gap will close. It’s what happens to the entire American AI economy when it does.
“They Just Copy American Models” and Other Fairy Tales
Let’s get something out of the way. The narrative that Chinese AI models are strong because they distill American ones is not just wrong. It is mathematically absurd, and anyone repeating it in 2026 should be disqualified from serious AI discussion.
Can you do supervised fine-tuning from another model’s outputs? Sure. You can coax some latent capabilities to the surface, transfer certain reasoning patterns, calibrate response quality. But thinking you can reconstruct a frontier model from API traces is like seeing a handful of points on a billion-dimensional surface and believing you can interpolate the whole thing. You can’t. The information simply isn’t there.
What is there, in plain sight, is DeepSeek’s published research. Detailed papers on MoE architecture, reinforcement learning pipelines, the most advanced publicly available attention mechanism. These aren’t people copying homework. They’re publishing the textbook.
And here’s the part that makes the distillation accusation genuinely embarrassing: LLMs themselves are a distillation of public human knowledge. Every American frontier model is trained on the same public internet, the same books, the same papers. Accusing China of distillation while doing the exact same thing with a bigger GPU bill is, to put it charitably, a position that’s very hard to defend.
The Real Reason China Can Build Frontier Models (It’s Not Complicated)
The answer is painfully simple. They have the talent, the recipe, and the motivation.
Talent: China has a massive population that has invested heavily in technical education. The pipeline of ML researchers is enormous. While Western researchers get absorbed into corporate labs where their work becomes proprietary, Chinese researchers publish. China published more AI research papers than the US in 2025.
The recipe: The Transformer architecture is public. The training techniques are well-documented. If you look at DeepSeek V4 and look at GPT-2, you don’t see two unrelated technologies. You see a refinement of the same fundamental approach. There’s no secret sauce. There’s execution.
GPU constraints as a feature: US export controls on Nvidia H100 and A100 chips forced Chinese labs to innovate on software efficiency. That constraint produced breakthroughs in algorithmic optimization that now benefit the entire open-source ecosystem. Sometimes the best way to breed innovation is to take away the brute-force option.
Open-Weight Isn’t Charity. It’s Strategy.
Why does China release models for free? To understand this, imagine they didn’t. Imagine China had no competitive LLMs, like Europe.
The consequences would be catastrophic. Total dependency on American infrastructure for military, government, and corporate AI. Every sensitive query flowing through US servers. A massive hemorrhage of capital, hundreds of billions flowing from the Chinese economy to Silicon Valley just to keep up. And a productivity gap that would compound every year: if your competitor’s programmers generate 10,000 lines of code per hour with AI assistance and yours generate 200, you’re not competing. You’re dying.
Not having domestic LLMs would have been a death sentence. China understood this immediately.
But they also understood they couldn’t sell APIs. The Americans had more GPUs, a first-mover advantage, and the trust of the global enterprise market. So China did the only rational thing: give the models away, distribute the knowledge, and capture the ecosystem.
The result? Alibaba’s Qwen family now has 942 million total downloads, more than double the next eight competitors combined. That’s not charity. That’s the most successful platform play since Android.
The 1,000x Cost Collapse Nobody Priced In
The financial numbers around LLM inference costs are almost comical in how fast they’ve moved.
In late 2022, running a GPT-4-class model cost about $20 per million tokens. In early 2026, equivalent performance costs $0.40 per million tokens. That’s a 1,000x reduction in three years, one of the fastest cost declines in computing history.
DeepSeek V3.2 charges $0.28 per million input tokens. GPT-5.2 charges roughly $10. That’s a 35x price difference for near-parity quality on most tasks. For coding, math, and general workloads, Chinese models often offer quality parity at 1/15th the cost.
Gartner predicts another 90%+ decline by 2030. An academic paper published earlier this year identified a structural break in May 2024: the shift from technology-driven price decline (hardware, architecture) to competition-driven price decline (providers undercutting each other for market share). Technology-driven decline has physical limits. Competition-driven decline can push prices below cost. Which is exactly what’s happening.
OpenAI Is Burning. Anthropic Is Sprinting. Neither Might Win.
The financial picture for the two leading US AI labs is a study in contrasts, but both face the same existential question.
OpenAI: 900 million users, only 5.5% paying. Burning $17 billion per year in cash. HSBC projects $44 billion in cumulative losses through 2028. The company projects spending $121 billion on compute in 2028 alone, and doesn’t expect to break even until after 2030. That’s not a business model. That’s a prayer.
Anthropic: Now ahead of OpenAI in revenue at $30 billion ARR versus OpenAI’s $24 billion. Grew 30x in 15 months. Spends 4x less on training. Generates 70% more revenue per unit of compute. Projects positive cash flow by 2027.
The divergence is structural. OpenAI is a consumer company building enterprise products. Anthropic is an enterprise company that happens to have a consumer product. OpenAI gives away 94.5% of its product for free and hopes the remaining 5.5% covers the bill. Anthropic charges premium prices to professionals who can’t afford bad outputs.
But here’s the thing neither company can escape: what happens when near-equivalent capability costs 1/15th the price? You can be the most efficient American AI lab in history and still lose to “free.”
Both companies are pivoting hard into enterprise services to escape pure model commoditization. Anthropic partnered with Blackstone, Goldman Sachs, and Hellman & Friedman to form an AI services company. OpenAI raised $4 billion to create “The Deployment Company.” They’re both running from the same monster: the day when selling API access to an LLM is like selling tap water.
The Stock Market Is Standing on a Trap Door
Nvidia crossed $5 trillion in market cap, higher than the GDP of every country except the US and China. Six AI-adjacent companies account for roughly 30% of the S&P 500. The tech sector issued $428 billion in bonds in 2025, mostly for AI capex.
And the whole thing is built on circular financing. Oracle builds data centers for OpenAI. Nvidia supplies Oracle. Nvidia invests in OpenAI. OpenAI has a deal with CoreWeave, which has a deal with Oracle. The capital is eating its own tail, and every player reports the circular spending as revenue growth.
Michael Burry, the investor who called the 2008 housing collapse, is loading puts against Nvidia. Capital Economics’ John Higgins identified something even more alarming: not just a stock price bubble, but potentially an earnings bubble. “There may be one [bubble] actually in the fundamental side of things, which is quite rare.” If the actual earnings are inflated by circular AI ecosystem spending, a correction wouldn’t just reprice stocks. It would crater the fundamentals.
OpenAI missing key growth targets in early 2026 is exactly the kind of crack that could propagate through this structure. The Bank of England has warned of growing risks of a global market correction due to overvaluation of AI tech firms. When central banks start worrying, you should too.
This Is the Nuclear Arms Race of Our Generation
Both sides had to do this. The US couldn’t not build an AI industry, because falling behind China would mean strategic subordination in military capability, economic productivity, and scientific research. China couldn’t not build LLMs, because dependency on American AI infrastructure would have been a geopolitical death trap.
The parallel to the nuclear arms race sounds forced until you think about it for five minutes. The RAND Corporation and the Atlantic Council both frame it in these terms. Both nations are compelled to invest regardless of ROI, because the alternative, ceding AI dominance, is unacceptable.
But here’s a crucial difference: the Americans interpreted this compulsion through turbocapitalism. Build unicorns, attract venture capital, target $1 trillion+ valuations. The Chinese interpreted it through state-directed ecosystem building. Distribute the technology, democratize access, build soft power.
The American approach assumes the winner captures monopoly rents. The Chinese approach assumes AI becomes a commodity and plays accordingly. If the commodity thesis is right, and all evidence points that way, one of these strategies was a $285 billion mistake.
So Who Actually Wins?
Not OpenAI, unless they can pivot from “the best model” to “the best enterprise platform” fast enough. Their current burn rate is a ticking clock.
Not Nvidia, at current valuations. If inference becomes a commodity utility, the insatiable demand for cutting-edge training hardware levels off. The $5 trillion market cap assumes permanent scarcity. Commoditization is the opposite of scarcity.
Probably Anthropic, if anyone. They’ve built the highest-revenue-per-user AI business, have the strongest enterprise positioning, spend far less on training, and are closest to profitability. But “probably wins among the Americans” isn’t the same as “wins.”
The real winners are developers and businesses everywhere. Competition plus commoditization equals cheaper, better tools for everyone. The question “who has access to AI?” disappears. The question “who builds the best products with AI?” takes over.
And Europe? Europe contributed Hinton, Bengio, half the foundational research behind the Transformer, and then watched the entire industry get built by others. No frontier model, no major AI company, no strategic position. Just a very large customer base. We didn’t deserve this outcome, but we got it anyway. (I wrote more about the open-weight revolution in my analysis of GLM 5.2.)
The Clock Is Ticking
The playing field is shifting. When the boundaries of the game change, those who seem marginal today can find themselves central, and today’s giants can find themselves in trouble. The US is ahead today. The momentum belongs to China. And the gap between those two statements is closing faster than the stock market has priced in.
The next 18 months will tell us whether the American AI economy was a generational investment or a generational miscalculation. Either way, the models are becoming a commodity. The only question left is who adapts first.
Sources: Stanford HAI AI Index 2026, Council on Foreign Relations, Gartner, RAND Corporation, Atlantic Council, Brookings Institution. This analysis was also informed by a video essay by Salvatore Sanfilippo (antirez), creator of Redis, on the strategic logic of China’s open-weight AI approach.