Chinese firms are actively choosing Huawei’s Ascend over Nvidia — and when buyers switch, the decoupling stops being a policy story and becomes a structural one.
What Happened
A new Bloomberg survey published July 7, 2026 shows Chinese companies are actively abandoning Nvidia’s advanced accelerators in favor of domestic alternatives — led by Huawei’s Ascend 910B and 950. This is not a government directive, not a lab experiment, and not a supply-chain workaround. These are enterprise buyers making deliberate procurement decisions to exit the Nvidia ecosystem.
The numbers corroborate what’s been building for two years. Bernstein estimated Nvidia held roughly 40% of China’s AI-chip market in 2025, with Huawei already matching that share. By end-2026, Huawei’s share is projected to reach approximately 60%. Nvidia CEO Jensen Huang acknowledged the trajectory directly, saying the company had “largely conceded” China’s advanced AI-chip market to Huawei. That is a remarkable admission from the dominant global GPU vendor.
The addressable market at stake is not small. China’s AI-chip market is projected to approach $67 billion by 2030. Every percentage point of share that Huawei gains — and Nvidia loses — is a compounding revenue hole that no other geography fully replaces. The supply story is known. The demand story, now confirmed, is what changes the calculus permanently.
The key insight: Export controls were designed to constrain Chinese AI. Instead, they forced the creation of a viable domestic alternative — and now that alternative has earned genuine buyer preference. The policy achieved the opposite of its intent at the hardware layer.
The Structural Read
The critical distinction in the Bloomberg data is that this is demand-side switching, not supply-side coercion. The supply-side story — export controls, architecture divergence, labs experimenting with novel silicon like Peking University’s neuromorphic work — was always the easier narrative to dismiss. Governments restrict. Labs experiment. Companies adapt. None of that necessarily means the incumbent loses the customer.
But when enterprise buyers choose to switch, the dynamic changes. Switching costs in AI infrastructure are enormous: tooling, software stacks, developer workflows, deployment pipelines. A buyer who rebuilds those workflows on Huawei’s CANN framework is not coming back to CUDA next quarter. The decision is effectively irreversible on a 3–5 year horizon. That is what makes the Bloomberg survey a structural signal, not a cyclical one.
The sanctions-backfire pattern here is a textbook import-substitution dynamic, accelerated by the scale of China’s AI investment. Washington applied pressure at the supply layer expecting a bottleneck. Beijing responded by treating that bottleneck as a strategic mandate. The result is a domestic champion — Huawei — that now has a captive, growing market, hard-won engineering credibility, and an increasingly coherent software ecosystem. The same playbook that built China’s domestic EV industry is now running in AI silicon.
Jensen Huang / Nvidia
“We largely conceded China’s advanced AI-chip market to Huawei.”
The deeper structural consequence is a bifurcation at the silicon layer — the most fundamental layer of the entire AI stack. The world is not splitting at the model layer or the application layer, where interoperability is still theoretically possible. It is splitting at CUDA versus CANN, at the instruction set and memory architecture level, where the cost of bridging is prohibitive. We have covered the geopolitical framing of this in The Geopolitical Fencing of Frontier AI and the downstream software implications in Open Source and the Bifurcated AI Market. What the Bloomberg survey adds is the confirmation that the bifurcation is now self-sustaining — it no longer requires policy to maintain it.
Where This Sits in the AI Stack
Hardware / Accelerator Layer
SPLITTINGNvidia (CUDA) vs. Huawei Ascend (CANN). Buyer switching is making this permanent. Every other stack layer above it inherits the fracture.
Software / Framework Layer
DIVERGINGPyTorch/CUDA toolchain vs. Huawei’s MindSpore/CANN ecosystem. As adoption scales, the toolchain gap widens and re-bridging costs rise.
Model / Application Layer
STILL FLUIDChinese frontier models (DeepSeek, Baidu, etc.) are now being trained and deployed on domestic silicon. The model layer is catching up to the hardware bifurcation.
Three Implications
FOR NVIDIA: A Revenue Hole That Doesn’t Close
China was Nvidia’s second-largest market before controls. With enterprise buyers now actively migrating to Ascend and the projected $67B China AI-chip market foreclosed, Nvidia’s TAM argument for the next decade has a structural gap in it. No amount of H200 demand from US hyperscalers fully replaces a market of this size and growth rate. Nvidia’s moat remains intact everywhere export controls don’t apply — but that is a meaningfully smaller world than it was in 2022.
FOR HUAWEI: A Captive Market Becomes a Platform
Huawei is now executing the classic platform playbook inside a protected geography. A captive enterprise customer base funds R&D, generates software ecosystem gravity, and produces the feedback loops that accelerate chip improvement. By the time Ascend is mature enough to compete internationally — if export rules ever relax — it will have years of production-scale optimization behind it. The sanctions didn’t just create a competitor; they funded its maturation.
FOR THE AI INDUSTRY: Two Stacks, Two Trajectories
The world’s largest AI market and the rest of the world are now training, fine-tuning, and deploying on fundamentally incompatible hardware. Model weights may still be shared across borders — for now — but the compute substrate they run on is diverging at an accelerating rate. Any company building AI infrastructure, tooling, or applications that requires optionality across both geographies now faces a real architectural decision it did not have to make three years ago.
