On Monday, Ben Thompson published a Stratechery piece titled “Who’s Afraid of Chinese Models?” The answer, judging by the 759 upvotes and 570 comments it racked up on Hacker News within hours, is: roughly everyone with a GPU allocation and a LinkedIn profile.
Thompson’s argument is characteristically crisp. The US frontier labs — OpenAI, Anthropic, Google DeepMind — will be fine, he writes. Their models are still ahead, their distribution is locked in, and enterprise customers aren’t about to rip out their Azure OpenAI deployments because DeepSeek released something clever. The real problem, Thompson argues, is that the US lacks open alternatives to compete with the wave of open-weight models pouring out of Chinese labs. Enable those, and the panic subsides.
He’s not wrong. But he’s answering a question nobody is actually asking.
The panic over Chinese models isn’t really about national security, or technological supremacy, or even the philosophical merits of open versus closed AI development. It’s about money. Specifically, it’s about the roughly $200 billion that the four largest US cloud providers are on track to spend on capital expenditures this year, much of it funneled into AI infrastructure on the assumption that cutting-edge models require cutting-edge budgets. If a lab in Hangzhou can match GPT-5 on a fraction of the compute, the entire investment thesis starts to look like a mispriced option.
The $200 Billion Assumption
The bull case for US AI dominance rests on a simple premise: frontier models are expensive, and only a handful of companies can afford to build them. That premise justifies the eye-watering capex numbers. Microsoft alone is expected to spend over $60 billion this fiscal year. Google’s parent Alphabet isn’t far behind. The hyperscalers are building data centers as if the future of computing is a land grab, and they intend to own the land.
Chinese models threaten that premise not because they’re better — they aren’t, at least not yet — but because they’re cheaper. DeepSeek’s latest model, released earlier this year, reportedly matched GPT-4-level performance at a training cost that was an order of magnitude lower than what OpenAI spent. Qwen, Alibaba’s open-weight series, has been climbing the benchmarks with a price tag that makes a single H100 cluster look like a vanity purchase.
If the cost curve bends down faster than the revenue curve bends up, the hyperscalers aren’t building moats. They’re building white elephants.
Why Open Weights Scare Wall Street
Thompson wants more open US alternatives. It’s a reasonable prescription. But it sidesteps the reason those alternatives don’t exist: the economics of American venture capital are fundamentally hostile to open-weight releases.
A US startup that releases its best model for free is a startup that just torched its own pricing power. Investors don’t fund that. Chinese labs operate under a different set of incentives — state backing, strategic mandates, a domestic market where monetization paths look nothing like Silicon Valley’s. They can afford to give away the weights because the weights were never the product.
The result is an asymmetry that no amount of “enabling” can fix. You can’t regulate your way to an open-source ecosystem that the market refuses to fund. And you can’t shame venture capitalists into backing companies that give away their only asset.
One hedge fund analyst who covers the hyperscalers put it bluntly in a note to clients Monday morning: “Every time a Chinese lab drops an open-weight model that’s 90% as good as the frontier, the NPV of every data center under construction drops by a few basis points. The market hasn’t priced this in yet because the market doesn’t know how to price it.”
The Panic Is the Point
Here’s the thing Thompson gets right: the frontier labs probably will be fine. They have distribution, brand, and enterprise relationships that won’t evaporate overnight. ChatGPT isn’t going to zero because Qwen 3.0 scores well on MMLU.
But “fine” is a low bar. The question isn’t whether OpenAI survives. It’s whether the entire capital structure behind US AI — the $200 billion in annual capex, the trillion-dollar valuations, the assumption that scale is an unbreachable moat — survives contact with a world where frontier capability is increasingly a commodity.
The panic over Chinese models is the market slowly realizing that the answer might be no. Every new release from a Chinese lab is a data point suggesting that the cost of intelligence is falling faster than anyone’s business model can absorb. That’s not a national security crisis. It’s a pricing crisis. And pricing crises have a way of making very large capital investments look very foolish, very quickly.
Thompson is right that we need open US alternatives. But the reason we need them isn’t to beat China at its own game. It’s to give the US AI industry a path to relevance in a world where the thing it sells is no longer scarce. The panic isn’t a bug. It’s the feature that tells you the old model is breaking.