Huawei's AI-chip sprint, and the moat it has not yet reached

This week Shanghai plays host to Huawei Connect, the annual showcase where the sanctioned Chinese firm parades its latest hardware before the officials and telecoms buyers who keep it alive. Any investor holding Nvidia would have felt the familiar flutter of dread in the headlines that followed: another leap for the Ascend line of AI chips, and a roadmap that keeps tightening. The numbers deserve their billing. The Ascend 950PR, launched in April, processes inference—running a model once it is trained, rather than building it—at 2.8 times the speed of Nvidia's H20, the chip the American firm was long permitted to sell to China. DeepSeek, the lab whose frugal models rattled the industry, tuned its latest V4 release to run on Huawei's gear, and ByteDance, Tencent and Alibaba have been scrambling for orders. Huawei planned to ship 750,000 of the new chips this year, with demand running ahead of supply.

It is tempting to read this as the first crack in the largest artificial-intelligence trade on earth. To do so is to confuse a political rivalry with an engineering one.

A lead that widens, not closes

Measure Huawei against the true yardstick, and the arithmetic is less dramatic. The H20 was a deliberately weakened export product; beside Nvidia's flagship, the Ascend 950 delivers about half the arithmetic performance of the H100, a chip Nvidia has shipped since 2022. One careful estimate has Huawei trailing by three to four years in per-chip capability and producing less than 4% of Nvidia's total computing power this year—a deficit projected to stretch to a seventeen-fold gap by 2027, because Nvidia keeps compounding advantages in manufacturing that Huawei cannot buy.

The reasons the gap persists are the reasons it will not soon close, and each is a constraint rather than a choice. The binding one is high-bandwidth memory, the fast memory that AI chips depend on and China cannot yet make at scale. Only about 240,000 of the 1.5m Ascend units Huawei was expected to produce in 2026 would have used domestically made HBM; the rest lean on stockpiles or slower substitutes. Then there is the forge itself: Huawei's designs rest on SMIC's 5–7nm process, while Nvidia builds on TSMC's leading nodes and enjoys a roughly twofold transistor-density edge that compounds across generations. Finally comes software. Nvidia's CUDA ecosystem is a moat in its own right, and Huawei's rival stack, CANN, is years less mature. The sharpest evidence arrived when DeepSeek tried to train its R2 model on Ascend hardware: Huawei's own engineers could not finish the run, and the lab fell back to Nvidia.

Rents, and the rival that policy built

None of this is accident. Export controls are supposed to defend the American lead, and at the frontier they plainly do. Yet they also built the rival. By starving China of Nvidia chips, Washington handed Huawei a captive and enormous home market that now finances its climb—and the policy has wobbled as officials weigh the cost of that bargain. Beijing itself banned the H20 that had once been the permitted export, and Washington approved the more powerful H200 in December only for it to sit in regulatory limbo, approved but unshipped, while the two governments haggle over terms.

Seen this way, the question for an investor is not whether Huawei is getting faster—it plainly is—but whether it is getting faster at the activity that earns Nvidia its extraordinary rents. On that, the record is reassuring, and the reassurance is structural rather than sentimental. Nvidia just reported quarterly revenue of $96bn, up more than 100% from a year earlier, at a gross margin of 75%—the signature of pricing power, not of luck. The near-unanimous Street has kept the stock a buy. That profitability rests on a moat—CUDA lock-in, unrivaled access to high-bandwidth memory, and the best foundry on the planet—defended by export controls but not dependent on them.

What the holder should actually watch

The risk implied by Huawei's sprint is therefore real and distant, and it turns on a short list of conditions rather than on any single product launch. It materialises if Chinese HBM production scales and unlocks Huawei's volume; if CANN matures enough to train frontier models rather than merely run them; or if process-node parity ever arrives. Each of those is years away, and each is a constraint that Washington principally controls. The uncomfortable irony is that the same policy which preserves Nvidia's rents is also, by nurturing a protected and well-funded rival, planting the only plausible seed of their end. Investors would do better to watch the constraint than the conference.