Illustration: AI-generated · Inside China AI
AI Without Nvidia: China’s Silicon Declaration of Independence
When DeepSeek launched its V4-Pro model this week, the most important detail was not the benchmark scores. It was the silicon underneath: DeepSeek has been adapting its frontier models to run on Huawei’s chips — serious AI performance without a single US GPU. For three years, the assumption behind US export controls was that cutting off Nvidia would cap China’s AI ambitions. The evidence now points the other way: the controls are accelerating exactly the decoupling they were meant to prevent.
The numbers behind the shift
The scale of China’s silicon pivot is becoming measurable. Industry reports now project that Huawei and other domestic firms will supply around 90 per cent of China’s AI chip market in 2026 — with Huawei on course to displace Nvidia as the country’s leading AI chip vendor. SMIC, China’s largest foundry, saw its profit more than triple on AI-driven demand, per Reuters. And the demand side is real, not aspirational: Zhipu AI reports its API now runs on more than 50,000 domestically produced accelerator chips, serving nearly seven million users.
From workaround to strategy
What began as a forced workaround has hardened into industrial strategy. DeepSeek’s Huawei adaptation is the most visible example, but the pattern runs across the ecosystem: Leju’s Kuavo humanoids run a fully domestic Huawei stack — HarmonyOS with Pangu models — and Beijing is channelling capital-market funding into the chip race through IPOs and state-linked funds. Meanwhile, the pressure from Washington is set to increase rather than ease: Anthropic CEO Dario Amodei publicly urged the US in July to tighten chip export bans further, calling them “the most efficient and direct way” to slow Beijing’s frontier AI.
The uncomfortable question for Nvidia’s world
None of this means Chinese chips match Nvidia’s best hardware today — by most accounts, they do not, particularly for training the largest models. But the strategic logic has inverted: every tightening of export controls raises the incentive for Chinese labs to optimise for domestic silicon, and every optimisation success — like DeepSeek on Huawei chips — shrinks the moat that justified a $500 billion Wall Street ecosystem built on GPU collateral. The question is no longer whether China can build AI without Nvidia. It is how good “good enough without Nvidia” becomes, and how fast. Follow-up: our analysis of China’s AI pricing power — what running on domestic silicon does to the cost curve.
Why it matters for Europe
Europe is watching a masterclass in technological sovereignty under pressure — while remaining almost entirely dependent on US silicon itself. If China proves that a state-backed, vertically integrated chip-to-model stack can sustain frontier AI, Europe’s own conversations about sovereign AI infrastructure will look underpowered by comparison. The practical takeaway for European decision-makers: diversification of AI compute is no longer a theoretical exercise, and the assumption that Nvidia access equals AI capability is being tested in real time — by the world’s second-largest AI ecosystem.
No responses yet