The Race to Nowhere

Chinese open-source AI is not a threat to America. It is a mercy.

Of all the things that I rack my mind about when it comes to geopolitical risks, “China winning at AI” is not one of them.

Let’s put two items on the table. The first is that the panic in Washington and on Wall Street about Chinese AI is genuine — the Kimi K3 release two weeks ago has forced the American AI industry to face a market it was pretending would not arrive for another five years. The second is that the panic is entirely misdirected, because the thing they are afraid of losing is not a lead in artificial intelligence. It is pricing power in a commodity. What Beijing is doing to the American AI industry, whether Beijing intends it or not, is exactly what the American economy needs to survive the buildout that has been happening in its name.

I made the underlying case in The Capital is Misaligned and the Crash is Coming and in The Doomsday Moat. The technical and philosophical priors sit in Why I’m Betting Against the AGI Hype and LLMs Are Universal Translators, Not Universal Thinkers. I will not rehearse those pieces in full. The conclusion I need for this argument is narrow: the scaling paradigm is not going to deliver the transformations its current valuations require, the models are approaching the ceiling of their architecture, and the trillion dollars being spent on the assumption that they are not is a mispricing of the kind that ends economies. Every quarter the mispricing runs, the eventual damage grows.

Chinese open-source models are shortening the run.

⁂

What has happened in the past year is not that Chinese engineers built better models than American engineers. It is that Chinese engineers built comparable models and gave them away.

Kimi K3, released by Moonshot AI two weeks ago, beat Anthropic‘s Fable 5 and OpenAI‘s GPT-5.6 in front-end coding evaluations and finished ahead of Opus 4.8 in Arena‘s broader text ranking, at 40 percent lower cost. The weights release tomorrow. DeepSeek, Z.ai, MiniMax, Tencent, and Xiaomi now hold the top five spots on OpenRouter by weekly token usage — that is the marketplace where developers pay to serve real production traffic, not the leaderboards where labs pay for press coverage. The share of tokens American companies route to Chinese models has held above 30 percent every week since early February and peaked at 46 percent. Before February the average was 11 percent. The AI startup Lindy moved 100 percent of its traffic from Claude to DeepSeek in June. Z.ai‘s GLM 5.2, released this spring, was the fastest-adopted model Vercel tracked all year.

Most of the engineers routing this traffic would prefer, on any given afternoon, to keep sending their tokens to Anthropic or OpenAI. They are not choosing to. The enterprise stack now sits behind routing layers — Vercel, OpenRouter, Cloudflare‘s AI gateway, half a dozen others — that shop every request to the cheapest model that clears the quality bar. Once quality parity holds within a margin, the routing layer picks on price, and nobody looks. The vendor decision has been abstracted away from the humans who used to make it, which is what commoditization always does. Preferred suppliers become interchangeable suppliers. Interchangeable suppliers become price-takers. One AI investor put the endgame plainly to Axios: open-source models will handle 95 percent of enterprise queries. The premium American labs will fight over the remaining five.

---

---

The five largest hyperscalers are on pace to spend more than one trillion dollars on AI-related capital expenditure across 2025 and 2026 combined. Four of them — Amazon, Alphabet, Meta, Microsoft — are on track for as much as $725 billion in 2026 alone, roughly double last year’s pace, and Goldman projects their combined AI capex will exceed the GDP of Japan by the end of the decade. These commitments are already outpacing the hyperscalers’ earnings and free cash flow, forcing them into debt markets to cover the gap. The Bank for International Settlements — an institution not famous for its taste in drama — devoted a section of its 2026 Annual Economic Report to the resemblance between the current AI buildout and the investment manias that “ended with an eventual reversal in investment, inducing economy-wide recessions.”

The bond market is already pricing this. Credit spreads on Google, Amazon, and Meta widened this past week as fixed-income investors began demanding more compensation to lend to companies that were, eighteen months ago, considered capital fortresses. Mizuho told its clients on Friday morning that capex raises are “testing investor limits.” The equity market is telling the same story more quietly — the hyperscaler shares have decoupled from the chip stocks, and the divergence is being called “somewhat unsustainable” by JPMorgan.

The investment thesis behind the trillion dollars requires two things to be true. It requires the frontier to keep pulling away from the pack, so that whoever spends the most on training captures durable pricing power. And it requires the enterprise market to keep paying premium prices to rent that frontier capability by the token. Both requirements are being falsified in real time. The frontier is not pulling away — Chinese labs are matching capability within months and undercutting price by factors of five. And the enterprise market is not paying premium prices — it is routing its traffic through gateways that pick the cheapest model that works.

⁂

The single largest variable in how much damage the AI buildout does to the American economy is how long the mispricing survives. The longer the illusion of pricing power holds, the more capital pours into stranded infrastructure, the more of the S&P’s market capitalization concentrates in seven stocks whose earnings will not materialize, the more of the grid gets rebuilt around load forecasts that assume the tokens keep flowing at monopoly prices, the more of the federal budget gets committed to subsidies and permitting exemptions premised on a national-security story that the technology itself does not support. Every quarter the story survives, the eventual write-down grows.

The dot-com bubble was survivable because the correction came fast enough to leave the rest of the economy standing. The housing bubble was catastrophic because the machine ran for years after the fundamentals had quit, until the assets on the balance sheet had metastasized into the balance sheets of every institution in the country. The current AI buildout is closer in shape to the housing bubble than to the dot-com bubble, because the capital being deployed is not venture capital taking marked risk on individual companies — it is investment-grade debt secured against hyperscaler cash flows, and it is passive index money flowing into the same seven stocks regardless of price. When that pops, it will not pop only inside a set of speculative bets. It will pop inside pension funds, target-date retirement accounts, corporate bond portfolios, and the credit ratings of every major infrastructure vendor in the country.

The Chinese labs are pricing the frontier into a commodity now — before the trillion becomes two, before the grid buildout locks in another decade of stranded capex, before another wave of retail investors gets pulled into the seven stocks at valuations the earnings will not vindicate. They are performing, from Hangzhou and Beijing, the falsification that American capital markets have refused to perform from New York. Every American institution holding the asset has a reason to keep believing. A lab giving away frontier-grade weights has no such reason. DeepSeek does not care about Microsoft‘s earnings multiple. Moonshot does not care whether Anthropic‘s last funding round marks cleanly. The correction is arriving from the one direction the story cannot manage — and it is arriving early.

⁂

In June, the Commerce Department ordered Anthropic to suspend access to its two most advanced models for every foreign national, including Anthropic‘s own employees, on national-security grounds. Within a day, Zhipu answered by releasing a competing model — GLM 5.2 — for free. Eighteen days later, Washington withdrew the order and Anthropic began restoring access. The United States government reached for one of its most coercive economic instruments to protect the market position of a chatbot and folded in under three weeks because a Chinese lab gave its product away.

A moat that dissolves in eighteen days is not a moat. It is a story about a moat. And the industry knows this, which is why it is currently constructing a second story on top of the first.

I described the first story in The Doomsday Moat last week. Extinction risk was the argument the industry reached for once it became clear that the returns would not arrive on the schedule the valuations required. If your product cannot justify its capex by earnings, you can try to justify it by fear — the argument that these companies must be protected and subsidized and exempted from ordinary oversight because the alternative is the end of the species. That moat has been in place for two years.

China is the second moat, and it is the same moat wearing a flag. We cannot let the capital slow down, because China. We cannot enforce copyright against the training sets, because China. We cannot tax the compute, regulate the data centers, or ask when the returns arrive, because China. The Commerce Department’s aborted export order and the administration’s flirtation with banning Chinese open-source models outright are the moat being poured in concrete. A national-security argument has one great advantage over an earnings argument — it never has to survive a quarterly call.

The desperation is showing at the level of conduct. Anthropic spent the spring quietly deploying tracking code to surveil its own users, hunting for Chinese firms distilling Claude‘s outputs into rival systems. The industry’s own argument is now that its crown jewels can be extracted through a chat window — that a few million conversations with the oracle suffice to clone the oracle. A company confident that it is building a god does not argue that the god can be photocopied. The distillation complaint is an admission, filed publicly under the seal of a national-security concern, that the frontier confers no defensible advantage. The moat leaks by design, because the product is the leak.

⁂

Suppose I am wrong about the ceiling. Suppose some architectural breakthrough lies past the scaling wall, and the nation with the largest installed compute base and the deepest bench of researchers gets there first. Then the race would matter, and the moat would be real, and the capital allocation would be defensible.

The reply is that this scenario indicts the current allocation just as thoroughly. The trillion dollars now flowing into hyperscaler data centers is not being spent on the assumption that a breakthrough is required. It is being spent on the assumption that scale itself is the breakthrough — that the next order of magnitude of compute, applied to the next order of magnitude of data, produces the next order of magnitude of capability, all the way to whatever the industry means by AGI this quarter. Capital positioned for a paradigm shift would be funding heterodox architectures, basic research on continuous learning, small labs with strange ideas, the kind of work that produces the shift instead of extending the paradigm. Capital positioned for scale-is-all buys land in Louisiana and gas turbines in Texas.

You cannot defend the buildout by appeal to a possibility the buildout’s own thesis denies. Either the current architecture reaches AGI, in which case there is no meaningful frontier lead to defend, because Chinese labs will match the capability inside a year and commoditize it. Or the current architecture does not reach AGI, in which case the trillion dollars is being spent to serve a market that will not exist at the scale the pricing assumes.

And the security concerns, where they are real, have real answers that do not require pretending the commodity is a crown. Running Chinese weights inside classified networks would be reckless — nobody serious proposes it. Enterprises with sensitive data will pay a sovereignty premium for American models or run open weights in their own inspected silos. The remedies are boring, targeted, and available. They do not require a national industrial policy organized around the survival of the SaaS margins of three or four companies.

⁂

Here is what I rack my mind about. Not the models. Not who trains them. Not which flag flies over the server room.

I rack my mind about a country that has quietly restructured itself around the story that the models will pay for the buildout. GDP growth is being propped up by data-center construction. The power grid is being rebuilt around load forecasts that assume the tokens keep flowing at monopoly prices. Pension funds and index investors are concentrated in seven stocks whose valuations require the story to be true. State attorneys general are being pressured to stand down; permitting is being expedited; the Commerce Department, as we just watched, is being turned into a defender of specific corporate revenue models. The federal government has begun, as I argued in The Doomsday Moat, taking equity in the story itself — sovereign stakes, golden shares, a government converting its regulatory power into a shareholder’s interest in the bubble’s survival.

When the state owns the mania, the state defends the mania. And the instruments it reaches for will not be gentle, because the correction will not be gentle. This is what the second moat is being poured for. It is not being poured for a real threat from Beijing. It is being poured for the moment, some number of quarters from now, when the earnings do not arrive and someone has to be named as the reason. The moat rhetoric they have been rehearsing for two years will be sitting there, pre-loaded, wearing the flag.

That is the risk worth losing sleep over. The models were never going to hurt us. The story we built on top of them still might. And the country the story lands on is not a country that has been paying attention to how much of its constitutional order was convention all along. The economic dislocations arrive as politics. They arrive looking for someone to blame. And the people who inflated the bubble will be first in line with a candidate.

So no—”China winning at AI” is not on my list.

---

Go Deeper into the Circus

---