Huawei moved its next AI chip up by nine months. It still can't build enough of the current one.

Huawei used its annual Connect conference in Shanghai on September 17 to pull its next AI training chip forward by as much as three quarters. Rotating chairman Eric Xu told the audience the Ascend 960DT, previously slated for the fourth quarter of 2027, now targets Q1 2027 -- with an inference-focused sibling, the Ascend 960PR, following in Q3 2027. Two further chips, the Ascend 970 and Ascend 980, remain on the roadmap for 2028 and 2029 respectively. The timing wasn't incidental: the announcement landed one week before a planned meeting between Presidents Trump and Xi Jinping in Washington on September 24, and at least one China-watching analyst read it as a deliberate show of strength heading into those talks.

The headline engineering claim sits at the cluster level, not the single chip. The SuperPoD -- Huawei's term for a rack-scale system that wires thousands of Ascend chips into one machine with shared memory -- gets a new entry called the Atlas 960E, packing 4,096 NPUs into a unified pool of roughly 1 petabyte of memory, with a peak of 8 exaflops at FP8 precision. Huawei says a new optical interconnect, 5,500 Hi-ONE modules running at 7.2 terabits per second each, replaces roughly 48,000 conventional pluggable modules in the same system and cuts power draw by more than 550 kilowatts. Two independently reported accounts of the same keynote -- from TheNextWeb and XenoSpectrum -- arrived at matching figures for all of this, which is the closest thing to confirmation available without a Huawei-published spec sheet.

The number Huawei didn't put on a slide is market share. Xu told reporters Ascend processors have "overtaken Nvidia in domestic market share," per the South China Morning Post -- a claim he did not quantify, attach a timeframe to, or otherwise support with data. It arrives the same month Nvidia resumed shipping its China-compliant H200 chip under a new U.S. licensing arrangement, sales Nvidia itself has said account for less than 1% of its quarterly data-center revenue. Both things can be true at once: Huawei genuinely gaining share in a market Nvidia is only partly allowed back into, and "overtaken" remaining a number nobody outside Huawei can check.

Huawei's own comparison points are consistent in direction and inconsistent in specifics. The company has previously claimed the Atlas 950-generation SuperPoD delivers 6.7 times the computing power and 15 times the memory capacity of "Nvidia's corresponding generation" -- a comparison Huawei made itself, against a baseline it also chose. The Register, cited independently by both TheNextWeb and XenoSpectrum, estimates the Ascend 960DT lands at roughly half of Nvidia's B300 in FP8 compute and about a third in FP4 -- notable mainly because Nvidia cannot legally sell the B300 in China at all, which makes the comparison more a statement of ambition than a choice available to any actual buyer.

Xu was unusually candid about the gap between ambition and capacity in the same keynote, telling reporters Huawei is prioritizing Chinese customers with no broad global expansion planned (Huawei rotates its chairmanship among a small group of senior executives on a fixed schedule -- Eric Xu is this period's rotating chairman, not a permanent CEO or founder.) -- and putting the moment domestic chip supply actually catches up with domestic demand at 2030. DeepSeek alone plans to deploy Ascend 950DT chips at a scale that reportedly exceeds what Huawei can currently produce for a single customer.

160,000 (Ascend 950DT chips DeepSeek alone reportedly plans to deploy -- more than Huawei says it can currently supply)

"Now we can make our own chips, and even if they are inferior, at least we are going down the path of solving those challenges and not having to worry every single day."

That capacity math is the real ceiling on this announcement, whatever the SuperPoD's peak numbers claim.

None of that makes the roadmap fake, and it doesn't make the acceleration meaningless either. Moving a chip up by three quarters is a real engineering and manufacturing commitment, not just a slide. What it isn't, yet, is proof that Huawei has closed the gap with Nvidia in any sense an outside buyer could act on -- only proof that Huawei wants the gap to look closed a week before its president sits across from America's, at a summit where chip policy is explicitly on the agenda.