Picture an engineer in a Shenzhen office, somewhere past midnight, staring at a build log. The model trains. Not on an H100, not on an A800, not on any of the trimmed-down parts that used to arrive through export carve-outs. It trains on Ascend silicon, through a toolchain nobody outside China had much reason to learn two years ago. The loss curve descends like it always did. Nothing dramatic happens. That is the whole story.
Bernstein Research put numbers on that quiet moment. Nvidia held roughly 40% of China’s AI chip market in 2025, sitting nearly level with Huawei. By 2026, the firm projects Nvidia’s share drops to about 8% while Huawei’s climbs to around 50%. A forty-point swing in a single year is not a market responding to price or performance. It is a market responding to policy, and then discovering it can live with the answer.
Why the Software Layer Mattered More Than the Die
For the better part of fifteen years, the argument for Nvidia was never really the transistors. It was CUDA: the kernels, the profilers, the fused attention implementations somebody else already debugged, the Stack Overflow answer that exists for your exact cryptic error. That accumulated sediment of other people’s suffering is what made the hardware feel inevitable. Buy the GPU and you inherit a decade of solved problems.
Huawei’s CANN framework had to be built against that. And the single most instructive data point in this whole shift is not a market-share figure at all — it is DeepSeek moving off CUDA and onto CANN. DeepSeek is not a company that treats its training stack casually. Its published work has consistently lived close to the metal, squeezing throughput out of communication patterns and precision choices rather than just buying more compute. When a team like that migrates frameworks, it means the new stack cleared a bar that matters: you can do non-obvious, performance-critical work on it without the floor giving out.
That is a different claim than “the chips are fast enough.” Bernstein’s read is that Ascend parts now sit roughly comparable to Nvidia’s H200. Fine. Hardware parity has been achievable by determined national programs before. Software parity is the harder thing, because it is not an engineering target so much as an accumulation of habit. Habit transfers when enough competent people are forced to transfer it.
What This Means for Agent Architecture
For those of us building agent systems, the interesting consequence is not geopolitical. It is architectural, and it is uncomfortable.
Agent stacks have quietly grown dependent on a single hardware assumption in ways their designers rarely acknowledge. Consider what sits underneath a typical production agent:
- A serving runtime tuned for specific GPU memory hierarchies, with KV-cache paging strategies written against one vendor’s bandwidth characteristics
- Quantization kernels that exist in optimized form for one instruction set and in reference form everywhere else
- Speculative decoding and batching heuristics calibrated against a particular ratio of compute to memory throughput
- Orchestration logic that assumes a known, stable latency distribution per inference call
None of that is portable in the way a Python dependency is portable. Change the accelerator and the orchestration layer above it starts making bad decisions — scheduling too aggressively, caching the wrong things, timing out on calls that would have returned. The abstraction leaks straight up through the stack into agent behavior.
So a bifurcated hardware base means something concrete: agent frameworks are going to need a real portability story, not a nominal one. Not “we support multiple backends” in a README, but measured performance envelopes per backend, and orchestration policies that adapt to them. The teams building for both the Ascend and Nvidia worlds will be forced into that discipline first. They will end up with more adaptable systems as a side effect, which is an odd gift to receive from export controls.
The Part Worth Watching
I am skeptical of reading an 8% projection as the end of anything. Market share in a policy-shaped market tells you about procurement rules more than about engineering merit. What I take seriously is the migration itself — the fact that a frontier-capable team ported its training work and kept shipping.
Ecosystems were supposed to be the durable moat precisely because they could not be legislated into existence. Two parallel ones now exist, which means the next generation of agent infrastructure gets designed under an assumption we have not had to carry in a long time. The hardware underneath you is a variable, not a constant. Build accordingly.
🕒 Published: