Hugging Face's summer 2026 open-model report delivers a quiet but consequential verdict: Chinese labs have held the lead in frontier open-model parameter scale since 2026, while America's competitive center has drifted away from model labs like Meta and Google toward chipmakers Nvidia and AMD. The headline is not who trains the biggest model, but who now carries the flag.
The strategic logic is clean once you follow the incentives. For a hyperscaler lab, open weights are a cost center that erodes a proprietary moat. For a chipmaker, they are demand generation. Every widely adopted open model that runs efficiently on your silicon locks in the next hardware cycle. Nvidia and AMD backing open models is less an act of research idealism than vertical defense of their accelerator franchises. This reframes the open ecosystem as a hardware go-to-market channel, and it explains why the US baton passed to the companies that sell the compute rather than the companies that once evangelized openness.
The global risk is bifurcation. If China owns the top of the open-parameter chart and US chipmakers own the distribution rails, enterprises worldwide face a governance fork: adopt the highest-performing open weights regardless of origin, or stay inside a Western, hardware-aligned stack for provenance and compliance reasons. Neither choice is free. Performance leadership and supply-chain trust are pulling in opposite directions.
For Japan, this is a sovereignty and procurement problem before it is a technical one. Japanese enterprises and public-sector buyers have leaned toward open models for on-premise and data-residency reasons, precisely the segment where Chinese labs are strongest. SIers now have to build model-portability into architecture as a first principle: abstraction layers that let a client swap the underlying weights without rewriting the application, plus explicit review gates for the origin and licensing of any open model that touches regulated data. The firm that treats model choice as a fixed dependency will be exposed when export rules or client policy shifts.
RPA and internal dev teams inherit the same tension in miniature. As agentic features migrate onto open weights running on Nvidia and AMD hardware, the durable advantage is not picking today's best model but owning the evaluation harness, the switching cost math, and a hardware roadmap that is not hostage to a single vendor. Expect Japanese integrators to package this neutrality as a paid service, and expect procurement teams to start asking not just what a model can do, but whose compute and whose weights sit underneath it.
