I saw TK’s Weibo post today, which mentioned that even with a fully maxed-out business environment, payment habits, talent, and capital, the U.S. can support only two and a half pure-play model companies.
My first reaction to this was: if we follow this logic and look at the domestic market, the handful of foundation-model unicorns in China will also be unable to sustain their massive R&D spending for long (barring special paths).
The core deadlock is the mismatch between extremely high and continuous R&D investment and severely shrinking monetization paths.
Foundation models are not fixed assets that pay off once and for all. For each enormously expensive model generation, the technology lead window is often only a few months. To stay in the game, a company must keep investing in the next generation of larger-scale compute clusters and top-talent compensation.
Google, Meta, or China’s big tech companies can treat large models as infrastructure, subsidize R&D with cash flow from mature core businesses, and then recoup the money through search, advertising, or cloud services.
But pure-play model companies have no other business to feed them; every dollar must be earned back by the models themselves.
More fatally, the explosion of high-quality open-source models in China has directly shut off pure-play model companies’ room for pricing premiums.
This is not to say open source is bad; rather, the entities behind open source have a completely different commercial loop.
Alibaba open-sources Qwen to earn from Alibaba Cloud’s compute rental and storage fees: as long as developers deploy and run it in the cloud, Alibaba gets its money back.
DeepSeek relies on the early accumulation of its quant parent company and extreme architectural optimization to push inference costs to the limit, turning API gross margins into thin-margin, high-volume utilities.
When users find that they can run open-source Qwen or DeepSeek locally at lower cost with an inference engine like Strata, or achieve production-grade results by directly calling extremely low-priced APIs, pure-play model companies that have neither cloud infrastructure nor a generational technology advantage have no way to recover their high R&D costs from customers.
On the cost side, compute and R&D expenses cannot be brought down; on the revenue side, under the triple squeeze of free consumer offerings, enterprise private deployment, and low-cost open source, revenue cannot grow large. The room for pure large-model companies to survive has been compressed to the limit.
Going forward, they can either expand overseas and compete with Anthropic and OpenAI for customers, completely transform into low-margin custom service providers (which basically means saying goodbye to model iteration), or be absorbed by the giants after burning through their capital.