- Biren Technology’s revenue hit $183.9 million in the first half of 2026, a nearly 2,000% year-over-year increase, as Chinese tech giants redirected procurement to domestic AI accelerators following tightened export controls.
- Huawei’s Ascend line is projected to hold roughly 60% of China’s AI chip market by end-2026, but its Ascend 910C delivers an estimated 60-70% of Nvidia H100 inference performance on LLM tasks, and only roughly one-seventh of the FP4 throughput of Nvidia’s Blackwell-generation B300.
- SMIC’s 7nm yield rates for the Ascend 910C sit at an estimated 20-40%, against above 90% at leading global foundries, and a Goldman Sachs forecast puts SMIC at 50% yield by 2030 and 75% by 2035, a gap that constrains volume output and keeps unit costs high for the foreseeable future.
Biren Technology recorded a nearly 2,000% year-over-year revenue jump in the first half of 2026, reaching $183.9 million, as Chinese cloud giants shifted procurement to domestic AI accelerators. The number is striking not just as a business result but as a signal of how decisively Chinese buyers have moved away from Nvidia hardware since export controls tightened. The performance gap is real and measurable, but it has stopped being the deciding factor.
Huawei Takes the Lead
Huawei’s Ascend line sits at the centre of this shift. Huawei‘s Ascend chips are projected to capture roughly 60% of China’s domestic AI chip market by end-2026, with Alibaba, Tencent and Baidu placing substantial orders. The Ascend 910C is estimated to reach 60-70% of Nvidia’s H100 inference performance on specific LLM tasks, according to reports from April 2025. That’s a workable number for inference-heavy deployments, good enough that domestic buyers can absorb the gap.
Against Nvidia’s current frontier, though, the picture is less competitive. Huawei’s Ascend 950PR delivers roughly one-seventh of the FP4 throughput of Nvidia’s Blackwell-generation B300. For training workloads at scale, that gap matters. Huawei’s domestic market position has been built on inference, the workload where the performance shortfall hurts least.
The Foundry Bottleneck
Manufacturing is where the constraint is sharpest. SMIC, China’s primary advanced foundry, is currently running its 7nm process (N+2); it moved its 5nm process (N+3) into volume production by December 2025, though TechInsights has reported significant yield challenges on the node. TSMC, by comparison, was already in volume production on 2nm as of late 2025, a gap of two to three process generations.
SMIC reaches these nodes using older Deep Ultraviolet lithography, relying on multiple-patterning techniques to compensate for the lack of EUV equipment. The cost of that workaround shows up in yield rates. For the Ascend 910C on SMIC’s 7nm process, yields are reportedly in the 20-40% range, against above 90% at leading global foundries for mature nodes. A Goldman Sachs forecast from August 2026 puts SMIC at 50% yield by 2030 and 75% by 2035. Until yields improve substantially, unit costs stay high and high-volume output remains constrained.
The software side compounds the hardware problem. Nvidia’s CUDA platform, built up over more than a decade, sets the default for AI development globally, and switching costs are real. Huawei has its own Da Vinci architecture, and Moore Threads is developing its MUSA stack, but both ecosystems are still years behind CUDA in developer tooling and breadth of framework support. Chinese developers running demanding training workloads tend to prefer Nvidia hardware when they can get it, precisely because the software support is mature.
Self-Sufficiency vs. the Frontier
China’s response to export controls is to build a parallel stack: domestic chips, domestic memory, domestic frameworks. On market-share terms, that strategy is working. China’s share of global computing capacity reached 21% in September 2026 and is projected to reach 30% by 2030, according to Wu Hequan, an academician of the Chinese Academy of Engineering.
Catching up at the frontier is a different question. Bernstein Research placed Huawei roughly three years behind Nvidia’s technology as of April 2026. Domestic firms are making genuine progress on inference performance and are beginning to use AI tools to accelerate chip design itself, but the manufacturing lag and the software maturity gap together limit how far that progress translates at the global frontier. What China’s AI hardware industry has built is a viable supply chain for domestic demand. Global technological leadership across all segments is a separate target, and the current evidence suggests it remains some distance away.
