Inference Beyond Nvidia

The clearest evidence that US export controls backfired is a balance sheet. Cambricon, the Chinese AI-chip designer once nicknamed “China’s little Nvidia”, swung to profit in 2025, with revenue up more than 450% and its first annual profit since listing, then followed with a first quarter of roughly $423 million, up about 160% year on year.^[1] The demand it is soaking up is demand Nvidia is no longer allowed to serve.

Huawei is the bigger threat. Its Ascend 910C reached mass shipment to Chinese customers in 2025; DeepSeek’s engineers estimated the part delivers up to roughly 60% of an Nvidia H100’s inference performance, a credible substitute for serving models, if not yet for training them.^[2] (Treat the 60% as a reported estimate, not an independent benchmark.)

The real moat is software

Nvidia’s defensibility was never only silicon; it was CUDA, the software layer developers have built on for fifteen years. Huawei is attacking it directly: at its September 2025 Connect conference it pledged to open-source CANN, its CUDA alternative, with core Ascend components opened by year-end.^[3] The gap is still wide, engineers describe CANN as difficult and unstable, so Huawei has embedded its own staff inside Baidu, iFlytek and Tencent to help port code off CUDA.^[2]

The contest was never just about the chip. It is about the fifteen years of software written for the other one.

The market is pricing the shift. Biren, a GPU challenger on the US Entity List since 2023, raised about $717 million in a January 2026 Hong Kong IPO and jumped 76% on debut — a valuation the export ban helped create by handing it a captive domestic market.^[4] And software is now being co-designed for Chinese silicon: DeepSeek said its V3.1 used a numeric format built for China’s “next-generation” domestic chips.^[5]

The training-versus-inference split is the key to sizing this. Training a frontier model is a one-time, latency-tolerant, accuracy-critical job where Nvidia’s ecosystem still dominates. Inference, actually running the model for users, is a recurring, cost-sensitive, high-volume workload, and it is by far the larger share of compute spend over a model’s life. That is the beachhead a “good enough” domestic accelerator needs, and why Huawei’s embedding of engineers inside Baidu and Tencent is strategically shrewd: porting the inference workload off CUDA, one customer at a time, is how the software moat is drained.

What we watch

The training frontier still runs on Nvidia. But inference, the larger, recurring workload, is where domestic parts are becoming good enough, and where the economics of “good enough and available” beat “best and banned.” We track adoption inside the big Chinese clouds, the maturity of CANN and its tooling, and the model-makers optimising for homegrown hardware. That is where the substitution becomes real.