1. A rally in Chinese semiconductor shares should not be mistaken for a breakthrough in computing power, because stocks can be repriced overnight while engineers need far longer to build a reliable alternative to Nvidia [para. 1]. China’s domestic AI chip industry is growing, and restrictions on US technology create strong incentives to buy local, but the investment case often confuses a captive market with a competitive product; the relevant test is whether domestic chips can train and run leading AI models reliably, economically, and at scale [para. 2].
2. Industry estimates put combined 2025 accelerator-chip revenue of seven leading publicly traded Chinese suppliers at about 17.7 billion yuan ($2.7 billion), excluding Hygon’s conventional server-processor business; including smaller independent suppliers brings the total to roughly $3 billion [para. 3]. Huawei is the main complication: it does not separately disclose Ascend chip revenue, so comparisons rely on supply-chain estimates, which put 2025 accelerator chip and board sales at about $10–11 billion; including Huawei, the domestic market reaches roughly $13–14 billion [para. 4]. The estimates are uncertain and show heavy concentration: China’s computing push still depends greatly on Huawei, while independent GPU makers have yet to reach scale [para. 5]. For 2026, projections are about $5–5.5 billion for the listed group (excluding Hygon’s conventional processors) and $15.5–18 billion for Huawei Ascend, which would put combined annual sales above $20 billion; that would be substantial progress, but not proof that Nvidia’s competitive advantage has disappeared [para. 6].
3. Chinese AI developers still seek Nvidia hardware because availability on a product sheet is not availability in a data center [para. 7]. The first obstacle is supply: advanced fabrication capacity and high-bandwidth memory constrain how many competitive accelerators domestic vendors can deliver, and delivery delays can erase a design’s advantage [para. 8]. The second is software: Nvidia’s CUDA platform embodies years of developer investment, optimized libraries, and debugging tools; switching suppliers means adapting software, testing numerical accuracy, locating bottlenecks, and verifying distributed workloads, so a cheaper chip can become expensive if migration slows development or consumes scarce engineering talent [para. 9][para. 10]. The third is the cluster: frontier AI relies on thousands of accelerators communicating efficiently, and memory bandwidth, interconnects, networking, and stability determine how much advertised compute becomes productive work; a respectable chip can still disappoint inside an unreliable cluster [para. 11]. Therefore counting processors or announcing large installations is misleading; the meaningful measures are sustained utilization, training time, failure rates, and the cost of a usable result [para. 12].
4. China’s strongest response is to compete at the system level: if manufacturing limits a single chip’s performance, better interconnection and software coordination may allow many domestic accelerators to deliver competitive results, which is the logic behind Huawei’s supernode strategy [para. 13]. Export controls may accelerate domestic adoption by making foreign hardware harder to obtain, but they do not automatically make substitutes better; buyers should not build long-term plans around uncertain overseas access or opaque intermediaries, because relying on regulatory loopholes is a vulnerability, not an industrial policy [para. 14]. The practical approach is to direct more purchasing and engineering toward domestic systems while demanding measurable results, supporting competitive suppliers, investing in software migration, and insisting on transparent benchmarks; vendors need repeat deployments that work, not demonstration projects for headlines [para. 15].
5. Compatibility with widely used frameworks such as PyTorch and vLLM is especially important, because customers should not rebuild their software environment for every new accelerator; better compatibility lowers switching costs and gives smaller suppliers a chance to compete on performance rather than political preference [para. 16]. This commentary reflects the author’s own views, not necessarily Caixin’s [para. 17][para. 18][para. 19].
AI generated, for reference only