U.S. Open-Weight AI Startups Race to Counter China — With Little VC Help

A small cohort of American startups is scrambling to build open-weight AI models capable of competing with a wave of Chinese rivals — including DeepSeek, Qwen, and Kimi — that are rapidly closing the capability gap with top U.S. systems. The effort is driven by concern in both Silicon Valley and Washington that China's freely downloadable models could undercut the profitability of American AI companies for years, while also raising security and censorship worries for enterprise users.

The most striking example is Arcee AI, a 30-person San Francisco startup that trained its flagship model, Trinity Large, in just 33 days on roughly $20 million, using 2,048 of Nvidia's Blackwell B300 chips. Founded in 2023 by former Hugging Face employee Mark McQuade alongside Jacob Solawetz and Brian Benedict, Arcee had previously raised $50 million at a $240 million valuation. Trinity Large still trails Anthropic and OpenAI on leading benchmarks, but the company is now closing a new funding round to build larger models — and in July signed a partnership with the U.S. Energy Department to develop a science-focused AI.

The funding environment has been brutal. McQuade says virtually every tier-one VC passed, with investors openly admitting they didn't want to back a model that might erode the value of their existing stakes in OpenAI or Anthropic. The dynamic reflects a stark concentration of capital: in Q1 alone, AI startups raised $255.5 billion globally, with nearly two-thirds flowing to just three companies — OpenAI, Anthropic, and xAI — according to PitchBook. Open-weight models, which are free to download and customize, have struggled to convince skeptical investors they can generate real revenue.

Nvidia has emerged as the most aggressive institutional backer of the open-weight movement. The chipmaker invested in Reflection AI — which has raised over $2 billion and plans to release its first open model later this year — as well as Poolside and Thinking Machines Lab, whose first open model, Inkling, launched in July. Nvidia has also built its own open model family called Nemotron, co-signed an open letter urging policymakers to avoid restricting open models, and formed a coalition with other AI labs to contribute to the open ecosystem. Poolside, which released its Laguna S 2.1 model family in July, argues there is massive latent demand for a capable American open-source alternative.

The debate over open models isn't purely commercial. Anthropic CEO Dario Amodei has flagged the risk of open models being misused for cyber or biological attacks, while Microsoft's venture arm M12 sees open-weight AI as a potential default for enterprise users within a few years. Whether the scrappy U.S. open-weight cohort can scale fast enough — and raise enough capital — to matter before Chinese models fully close the gap remains the central unresolved question.