There are two very different jobs inside every AI system. Training is the long, punishing part, tens of thousands of processors linked in near-perfect step, digesting an ocean of text over months. Inference is the everyday part that comes after: answering a question, drafting an email, running the finished model. Keep that distinction in mind, because it is the hinge on which China's chip story now turns. According to reporting cited this week, DeepSeek does not plan to use its 160,000 new Huawei Ascend chips to train a single model. They are for inference only.
That is a quiet but telling admission. Huawei executives had promised its latest Ascend silicon could handle training and inference alike. But DeepSeek tried, under pressure from Beijing to show off domestic hardware, spending months last year attempting to train its R2 model on older Ascend parts, with Huawei engineers on site to help. The Financial Times reported the effort produced no successful training run at all, and the company fell back on Nvidia for the heavy lifting. A hardware generation later, the division of labour is the same: Huawei for the easy job, foreign silicon for the crown jewel.
The bottleneck is not raw compute but memory. High-bandwidth memory determines how fast an accelerator can actually feed itself, and when Washington cut China off from South Korea's SK Hynix and Samsung and America's Micron in late 2024, Huawei had to build its own from scratch. This is the first Ascend generation to lean on homegrown memory, and it is throttling output to the low hundreds of thousands of units a year. DeepSeek has reportedly had to petition Beijing directly just for a larger allocation. When American restrictions eased slightly, Alibaba and Tencent rushed to buy Nvidia H200 chips, only for Beijing to approve a fraction of what Washington was willing to license. In other words, some of the pressure pushing Chinese labs onto inferior chips comes not from the United States but from their own government.
The workaround, if there is one, may be to stop moving the hardware and start renting it. A Taiwanese analyst, speaking to The Epoch Times, argues that as the United States tightens the screws on physical shipments through third countries, Chinese firms will increasingly reach for computing power abroad, renting capacity in overseas data centres rather than smuggling servers home. "The cross-border rental of computing power could become the next battlefield," he said, describing servers routed through Singapore or Malaysia and equipment dismantled into components to slip past controls. Regulators, he suggests, should watch for buyers whose orders dwarf any plausible domestic need.
None of this has dented China's export machine. Customs data show shipments jumped 25 percent in August, with semiconductor exports up an eye-watering 130 percent, and analysts describe a country that has "moved aggressively up the value chain" in AI infrastructure and industrial automation. That is the paradox worth sitting with. China can flood the world with chips and machinery, run finished models at scale, and still not reliably train a frontier system on its own soil. The gap between selling the tools and building the very best of them is, for now, the real map of the AI race.