The ‘death zone’: how free Chinese models are hollowing out US AI

China’s open-source AI push has created what Bloomberg calls a ‘death zone’ for rival American model makers.

The phrase captures a squeeze on the middle of the market, where US companies are too costly to match free Chinese models yet not good enough to command a frontier premium.

The pressure comes from giving models away. Chinese labs have released a wave of capable open-weight models that anyone can download and run, and that generosity is eating the market for paid American alternatives.

Open-weight means a model’s parameters are published, so anyone can run, tweak, or fine-tune it for nothing. That turns a cutting-edge model into a commodity almost the moment it ships.

The quality gap has narrowed to almost nothing. Stanford’s AI Index put China within 2.7% of the US on model performance, a gap it closed while spending a fraction of what American labs pour in.

The adoption numbers are just as striking. Roughly 80% of US AI startups now use Chinese open-source models, and DeepSeek’s R1 briefly overtook ChatGPT as the most-downloaded app in the US, a symbolic moment that unsettled the industry.

Alibaba’s Qwen family has passed Meta’s Llama in cumulative downloads, making a Chinese model the default open option for many developers. That is a role Silicon Valley assumed it would keep.

That is what makes the middle so dangerous. A US company selling a good-but-not-best model competes against a free Chinese one that is nearly as capable, and against frontier labs whose brand still commands a premium.

The frontier is feeling it too. A cheap Chinese model has been closing on Anthropic and OpenAI, pressing even the leaders on price at the top of the market.

American officials have noticed. A US congressional commission warned that China’s open ecosystem ‘creates alternative pathways to AI leadership’ and lets its labs ‘innovate close to the frontier despite significant compute constraints’.

Enterprises are voting with their budgets. Siemens’ chief executive said he saw ‘no disadvantages’ to using Chinese models, citing cost and flexibility, the kind of endorsement that turns a security debate into a procurement decision.

The economics are unforgiving for anyone in between. The same thrift-maxxing that pressures OpenAI and Anthropic’s valuations is fatal to smaller model makers who cannot subsidise their way to relevance.

China’s openness is strategic, not charitable. Giving models away wins global mindshare, sets standards, and builds dependence on Chinese tooling, all while US labs keep their best work closed.

And the releases keep coming. Firms such as MiniMax are building ever-larger models and open-sourcing them, ensuring the free tier keeps improving faster than the paid middle can differentiate.

The adoption is not frictionless. American officials warn of security and censorship risks baked into Chinese models, but cost has repeatedly won the argument inside companies weighing which to deploy.

There is an echo of earlier platform wars in all this. Free and good-enough has beaten expensive and best before, in operating systems and browsers, and the same logic is now loose in AI.

There is a counter-case, and US labs make it. The most valuable work, they argue, is moving from raw models toward agents, tools, and deployment, where being open is less of an advantage and trust still carries a price.

For the squeezed middle, the escape routes are narrow. A model maker can specialise in a defensible niche, climb into frontier research, or build a business around open models rather than trying to sell the models themselves.

The death zone, then, is a place on the price-performance curve, not a country. For US model makers who are neither the cheapest nor the best, the Chinese blitz has made the middle of the market a very hard place to breathe.