The U.S.-China AI race is being won on the invoice

The U.S.-China AI race is no longer being run on the leaderboard. It is being run on the invoice.

When switching costs are near zero and the model is a free download, the capability gap becomes academic. Open-source distribution on Hugging Face and GitHub has turned the deployment race into a per-token auction, with the next deployment going to the cheapest stack. That is the deeper read behind a 25-person San Francisco startup that just cut its single largest line item by moving 100% of its traffic to DeepSeek-V4, a Chinese open-weight model.

Flo Crivello's Lindy.ai — a small AI-agent startup — did the arithmetic last month. Anthropic, the U.S. AI lab, was bigger than payroll, bigger than rent. DeepSeek-V4 was roughly 10x cheaper. "It was a very, very simple business decision," he said.

The pattern is already wider than Lindy. Uber's Dara Khosrowshahi said on the Invest Like the Best podcast that Uber "blew through our AI budget in a quarter." Bloomberg reported Airbnb's Brian Chesky calling Alibaba's Qwen — a Chinese open-source model — "good" and "fast and cheap." So have Perplexity and Nvidia. Many will not say so on the record; the politics is hotter than the pricing.

Chinese models still trail U.S. leaders by six to 12 months on the hardest reasoning. That gap is real — and increasingly irrelevant for the bulk of the workload, where the trade is a token price, not a benchmark. Enterprise AI will be settled on the cost layer, not the frontier.