AIANALYSISPODCAST
OpenRouter, the routing layer most AI apps use to pick a model, now sends roughly half its traffic to Chinese open weight releases — models whose trained parameters are published so anyone can download and run them.
About half the traffic on OpenRouter, the routing layer most AI applications use to pick which model answers a query, now goes to Chinese open-weight models: the kind whose trained parameters get published so anyone can download and run them. The share crossed that line this month after Kimi K2 from Moonshot AI and Qwen from Alibaba shipped within a week of each other, according to the 20VC roundtable covering the week's news. On that panel, investor Rory O'Driscoll pegged the Chinese labs at "six to nine months behind the frontier, not further," which is not catching up on raw capability but already ahead on the distribution lane below it.
That gap is the story, and it is not a capability story. Kimi K2 is a roughly 2.8 trillion parameter model that needs large GPU clusters to run, not a laptop-friendly DeepSeek-style footprint. The labs shipping these weights have the compute, the researchers, and the capital. The US has all three too. The question is why every US lab with the resources to ship an open-weight competitor has decided not to.
The answer is a business-model choice, not a technical one. The 20VC panel of Harry Stebbings, Rory O'Driscoll, and Jason Lemkin walked through the labs that could plausibly ship: Meta, Google, xAI, Thinking Machines, Reflection. All five have frontier-class research teams. None is shipping a product into the open-weight lane. O'Driscoll named the underlying constraint directly: cap tables at the US frontier labs are priced for the closed-API business, where Microsoft, Google, and Amazon buy model access in bulk and pay per token. An open-weight release would compete with their own paying customers, compress the multiple their investors wrote checks against, and produce no obvious revenue line in return. Chinese labs do not carry that constraint. Their cap tables are priced for state-aligned strategic value, distribution scale, and downstream compute sales, not for a per-token API multiple that an open release would cannibalize.
The open-weight lane is a business the US has not chosen to enter, not one it has tried and lost. Chinese open-weight labs already fetch $50 billion to $70 billion private valuations on that logic; O'Driscoll said on the show that he would not turn down a $50 billion outcome from a convincing US entrant. The money is on the table. The cap tables are not structured to pick it up. The closest the US has come is building the infrastructure around the open-weight lane rather than the weights themselves: Fireworks AI closed $1.5 billion at a $17.5 billion valuation running roughly 40 trillion tokens a day, more than double its earlier volume. Capital is flowing to the routing and serving layer above the models, not to the model releases themselves.
The fight over whether to ban the lane landed inside OpenAI last week, and the optics exposed the underlying incentive. Dean Ball, an OpenAI policy staffer roughly two weeks into his role, tweeted that restricting the distribution of Chinese open-weight models was a matter of "AI communism", shorthand for treating frontier AI as a public good that should not be handed to a geopolitical rival. The 20VC panel reported that the tweet drew in-house pushback from David Sacks and Emil Michael, not support, because the ask read as a competitor's complaint dressed as a policy position. Business Insider's coverage of the same week describes the open-weight China shift as a structural challenge to the US industry's distribution assumptions, the same point made by a different route. OpenAI's commercial incentive is to slow the open-weight lane; that incentive happens to align with a real national-security argument, which is why the question is hard and not because the answer is obvious.
The variables that would have to flip for a US open-weight entrant are concrete. A frontier lab would need either a cap table that can absorb a revenue line below the API (advertising, device, services, or sovereign contracts) or a spinoff structure that prices the open-weight business on its own multiple. A second path is a hyperscaler with a non-API revenue stream, like search, ads, or cloud, using open weights to defend a different business. Neither path is impossible. Neither is being walked.
The watch item for the next quarter is whether the open-weight share on OpenRouter holds above half as Kimi K2 and Qwen absorb production traffic, and whether any US frontier lab announces a spinoff or open release before that share becomes structural. If neither happens, the US will have made the choice the cap tables imply: let China run the distribution layer, and compete for the closed-API dollars above it.