Rhodium Group's finding is blunt: seven leading Chinese AI developers together booked roughly $10.7 billion in annual recurring revenue between March and August, versus more than $100 billion combined for OpenAI and Anthropic. The capability gap has narrowed sharply; the commercial gap has not.
The reasons are structural, not accidental. China's model market has been shaped by a brutal price war and an open-weight strategy—DeepSeek, Alibaba's Qwen, and others distribute powerful models cheaply or freely to win developer mindshare. That maximizes adoption while gutting per-token economics. Export controls compound the problem by capping access to frontier compute, and the addressable market skews domestic, where willingness to pay premium API rates is thin. US labs, by contrast, monetize a global base of enterprises conditioned to pay for closed, frontier-grade inference. The lesson for the wider industry is that model quality and revenue quality are decoupling: benchmark parity does not translate into pricing power. Valuations riding on capability alone look increasingly exposed if margins never materialize.
For Japan, this reframes a procurement decision rather than a headline. Chinese open-weight models are becoming the cheapest route to competent LLM performance, and that is genuinely attractive for cost-sensitive Japanese enterprises and SIers building sovereign, on-premise, or air-gapped deployments where data cannot leave the building. NEC, Fujitsu, and NTT Data-style integrators can wrap these weights into vertical solutions at a fraction of API costs, and RPA vendors can embed local inference without recurring per-call fees.
But cheap capability carries governance freight. Reliance on models whose commercial backers are unprofitable raises continuity risk, and Chinese-origin weights invite scrutiny under data-security and supply-chain review, particularly for regulated finance, healthcare, and public-sector work. The pragmatic Japanese play is a two-track posture: use open Chinese or global open models for internal, non-sensitive workloads to compress costs, while reserving vetted US or domestic providers for customer-facing and regulated systems. The strategic signal is that the price of raw AI capability is collapsing—so Japanese firms should stop over-paying for commodity inference and concentrate budget on the integration, data, and governance layers where durable differentiation actually lives.
