China tightens AI oversight as measurement becomes part of deployment

September 16, 2026

China’s Cyberspace Administration issued version 3.0 of its national AI Safety Governance Framework, and the country’s State Security minister called for special AI laws and regulations. This moves AI providers and users another step into an operating environment where access, deployment and research are shaped by state controls rather than by company policy alone. For Microsoft, that is likely to matter less as a question of model quality than of whether its cloud, copilots and partner models can be offered under rules that can be audited and adapted market by market.

Epoch AI updated its dataset on frontier model companies with downloadable tables covering revenue, funding, staff, compute spending and usage, and it also refreshed its benchmark hub to track 390 models across 80 benchmarks. That suggests comparison infrastructure is becoming a working layer of the market, not just commentary around it: buyers, regulators and rivals get more shared material for judging providers that now differ by cost, governance and deployment fit as much as by raw capability. Microsoft is exposed to that shift from both sides, because a platform that sells many models benefits when comparisons are standardised, but it also loses some room to let branding stand in for evidence.

Google DeepMind published a model card for Gemini 3.8 Audio, covering live dialogue models that take in speech, images, video and text and answer in speech or text. FHNW found Swiss online retailers use AI mainly in partially automated, human-in-the-loop processes, but only about 2% report a fully implemented AI or agentic commerce strategy. Meta began rolling out Meta One bundles that combine app subscriptions with added Meta AI usage credits and other perks.