Huawei Is Shipping 700,000 AI Chips a Year. The Gap It Still Can't Close

In May 2019, the U.S. Commerce Department added Huawei's HiSilicon chip unit to an export blacklist, cutting it off from Taiwan Semiconductor Manufacturing Co. and Samsung's foundries, ASML's extreme-ultraviolet lithography machines, and the global electronic-design-automation toolchain that modern chip design assumes. Most observers expected attrition. Liao Heng, a Huawei Fellow and the company's chief semiconductor scientist, says most of the team did the opposite. They stayed and started pulling the entire stack, from chip design to fabrication coordination to the software around it, inside the company.

Five years later, Huawei reportedly shipped 600,000 to 700,000 Ascend AI accelerators in 2025, with plans to roughly double that in 2026. Ascend is Huawei's in-house family of AI training and inference chips, the hardware counterpart to the GPUs that power U.S. frontier AI labs. A superpod is a tightly integrated cluster of those accelerators built to train or run a single AI workload as one system. The figure comes from a five-hour interview Liao gave to podcaster Zhang Xiaojun, translated by the Recode China AI newsletter. It is single-source and hedged by the source itself. The next test is on the calendar: the Ascend 950 superpod in 2026, the Ascend 960 in 2027, with 970 and 980 already in the pipeline.

The open-weight large language model boom gave Huawei's chip team something it did not have before sanctions: a fast, public design-feedback loop. Open-weight models are publicly downloadable AI systems, the kind DeepSeek and other Chinese labs ship every few months. Each new release lets Huawei's chip team profile a real model on Ascend hardware, see which layers stress memory bandwidth, see which operators choke on interconnect latency, and fold the result into the next revision. The loop is months, not years. DeepSeek in particular has become a useful collaborator because its models are openly auditable, not just run as a black box.

Sanctions forced the second half of the arrangement. With TSMC, ASML, and the global EDA stack off-limits, Huawei had to integrate vertically. The company now designs the chip, the compiler, the interconnects, and the systems around them together, because the global supply chain will not supply the missing pieces on demand. The 950 and 960 are the first superpod generations built under that combined cadence: vertical integration on one side, open-weight workload feedback on the other.

The honest counterweight, which Liao supplies himself, is the compute gap. In 2025, by his estimate, China newly deployed a bit over 1 gigawatt of data-center capacity. The United States, on the same rough footing, added 7 to 8 gigawatts. A gigawatt is roughly the steady electrical draw of a large U.S. city, or the load of about 700,000 homes. The 1-to-7-or-8 ratio is Liao's own; outside reporting has not yet produced a tighter, independently verified number.

That ratio is what the next two years will measure. The Ascend 950 superpod is the first externally visible test of the vertical-integration-plus-open-weight-feedback playbook under production load. If it lands in 2026 with the throughput Liao describes, the loop worked. If 960 slips or 970 is shelved, the loop did not work. The ratio is the falsifier, supplied by the same executive who is selling the climb.

On Liao's account, the sanctions era did not save Huawei's chip business. It forced a structural shift. The company moved from a chip team that bought best-in-class pieces from the global supply chain to a chip team that builds, integrates, and iterates the whole stack, and that now has an open-weight model ecosystem treating it as a real customer. The climb is documented. The 1-to-7-or-8 ratio is the honest counterweight. The 950 will show which one matters more.