The South China Morning Post reported on 6 November that Menlo, an investor in Anthropic and other AI firms, estimates Chinese models take 89 per cent of open-source model spending. Closed models still take the great bulk of model budgets. Open-weight models can be downloaded and changed by developers. Closed ones are reached only through the owner's cloud service. Cursor, a US coding tool, said in March that its Composer 2 model is built on Moonshot's Kimi K2.5. It trained Kimi on code-heavy data, then applied reinforcement learning, a method that rewards better results.
Menlo scaled up its global figures from a September survey of nearly 400 US enterprise decision-makers, who manage over US$4 billion in AI budgets. It also used public filings and third-party reports. So 89 per cent is a spending share, not a count of models. Separately, on OpenRouter, a model platform, Chinese models used more tokens than US ones for a 27th straight week, to 1 November. That measures usage, not spending. No dollar total for open-source spending is given.
A tool sold by a US company can still sit on a Chinese base model, as Cursor shows. Buyers can ask vendors which base model sits under each product, who made it, and what licence applies. The numbers come from a US survey, so treat them as an estimate.