Reported 17 Sept 2026 Chips

Huawei moved its next AI chip forward by three quarters and put its interconnect design in front of an international standards body.

At its Connect conference in Shanghai on September 17 Huawei said the Ascend 960DT will be ready in the first quarter of 2027 rather than the third, a three quarter acceleration, and set out a chip a year to 2029 with the Ascend 970 and 980 behind it. The larger claim was about the machine rather than the chip: an Atlas 960E system scaling to 4,096 accelerators and 8 exaflops, clusters reaching 512,000 accelerators, and an optical interconnect Huawei says replaces about 48,000 conventional 800 gigabit modules with about 5,500 of its own units.

The caveat is that everything here is a date and a specification, not a shipment: the parts do not exist to be measured, the system announced is smaller than the 15,488-chip configuration Huawei described a year ago, and Huawei has published none of it on its own site.

Until now

Huawei's answer to export controls was read as a die problem: how close a domestically fabricated chip could get to a Nvidia part, on a process two nodes behind.

Now

Read Huawei's answer as a network problem it is trying to standardise, with the chip cadence pulled forward to match and the interconnect design filed with an international body rather than kept in-house.

Huawei spent the last year being measured by its die. How dense is the transistor layout, which domestic process made it, how far behind is it. On September 17 in Shanghai the company moved the argument. It pulled the Ascend 960DT forward from the third quarter of 2027 to the first, three quarters early, and committed to a generation a year through 2029. Then it spent the rest of the keynote on everything around the chip: a system scaling to 4,096 accelerators, clusters described up to 512,000, and an optical engine it says lets one machine replace roughly 48,000 conventional optical modules with roughly 5,500 of its own. The claim is that if you cannot buy the best chip, you win by connecting worse ones better.

Two details make this more than a keynote. The first is that Huawei filed its near-package optics implementation agreement with the Optical Internetworking Forum, which is not a Chinese body and not a captive one. A company that wanted a closed stack would keep the interconnect proprietary; a company that wants its architecture to become the default files it where Broadcom and Nvidia's suppliers also sit. The second is the direction of the numbers. The SuperPoD announced is 4,096 chips, well under the 15,488 Huawei described a year ago, while the chip dates moved earlier. Smaller systems sooner is what a company says when manufacturing capacity, not design, is the binding constraint, which is also what demand outstripping capacity means when a vendor volunteers it. None of it ships for at least four months, and Huawei has published none of it on its own site. Read Huawei's answer as a network problem it is trying to standardise, with the chip cadence pulled forward to match and the interconnect design filed with an international body rather than kept in-house.

How it got here

Sep 2025
Huawei sets out an Ascend roadmap at the previous Connect, including a SuperPoD configuration of 15,488 chips.
Sep 04
Huawei publishes test results claiming a chip 55 percent denser and cooler on a domestic process. Carded in Edition 014, still Reported.
Sep 07
That chip appears in a phone going on sale the same week. Carded in Edition 015, still Reported.
Sep 17
At Connect in Shanghai, Huawei moves the Ascend 960DT to the first quarter of 2027 and files its near-package optics agreement with the Optical Internetworking Forum.
Who has to react
Nvidia, whose China position rests on the gap staying wide and whose NVLink is the thing an open interconnect standard would erode; Broadcom and the optical module makers, whose 800 gigabit business is what 5,500 units replacing 48,000 describes; SMIC and Huawei's domestic fabs, for whom a generation a year is a capacity commitment they have to meet; the Chinese labs choosing between Ascend and smuggled parts for their next training run; the United States Commerce Department, one week before the September 24 talks; and the Optical Internetworking Forum's members, who now have a Chinese submission on the table in a standards process.

The evidence

Huawei has published nothing on its own newsroom carrying the roadmap or the standards filing. The verification event is Huawei posting the Ascend 960 roadmap or the Atlas 960 specifications to huawei.com, or the Optical Internetworking Forum listing the near-package optics implementation agreement, by December 31, 2026. Until one of those lands this stays Reported and we say so.

ObservedThursday September 17, 2026, opening day of HUAWEI CONNECT 2026 in Shanghai, which runs September 17 to 19 across the Shanghai World Expo Exhibition and Convention Center and the Shanghai Expo Center on Huawei's own event page. Rotating chairman David Wang set the roadmap: Ascend 960DT in the first quarter of 2027, moved forward from the third quarter, an acceleration of three quarters; Ascend 960PR in the third quarter of 2027; Ascend 970 in 2028; Ascend 980 in 2029.

The whole observation, as recorded

Wang: we are evolving our Ascend chip series on a one-generation-a-year cycle. Systems: the Atlas 960E scales to 4,096 accelerators, 8 exaflops at FP8 precision and up to 1 petabyte of high-bandwidth memory; the SuperCluster architecture is described as connecting up to 512,000 accelerators over a two-tier four-plane Clos network, with a multi-rail topology said to extend to one million. Interconnect: the Hi-ONE optical engine at 7.2 terabits per second, with an Atlas 960E said to use about 5,500 Hi-ONE units in place of about 48,000 conventional 800 gigabit optical modules, a claimed saving of more than 550 kilowatts and 99.8 percent system availability. Huawei submitted its near-package optics implementation agreement to the Optical Internetworking Forum and said it would continue the work with industry partners. Software: more than 40 models pre-trained on Ascend and CANN, 5,200 monthly active CANN developers of whom 61 percent are external, and support for more than 90 open-source projects including PyTorch, Triton and vLLM. Context on the dates: the announced SuperPoD configuration of 4,096 chips is smaller than the 15,488-chip system Huawei described in September 2025. The United States and China are scheduled to talk on September 24. Huawei's prior actions in this brief: test results claiming 55 percent greater density on a domestic process on September 4, carded in Edition 014, and that chip shipping in a phone on September 7, carded in Edition 015. Both remain Reported, with a December 31 verification event.

Established

At HUAWEI CONNECT 2026 in Shanghai on September 17, Huawei stated the Ascend 960DT would be ready in the first quarter of 2027, three quarters earlier than previously scheduled, with the 960PR in the third quarter of 2027, the 970 in 2028 and the 980 in 2029.
Huawei described an Atlas 960E scaling to 4,096 accelerators, 8 exaflops at FP8 and up to 1 petabyte of high-bandwidth memory, and a cluster architecture reaching 512,000 accelerators.
Huawei submitted a near-package optics implementation agreement to the Optical Internetworking Forum.
The announced SuperPoD configuration of 4,096 chips is smaller than the 15,488-chip configuration Huawei described at the previous year's Connect.

Not established

That any of these parts exist or perform as stated. No independent measurement of an Ascend 960 has been published, and the earliest date is four months away.
That Huawei can manufacture at the cadence announced. A generation a year requires domestic capacity that is itself the constraint export controls are aimed at.
That the Optical Internetworking Forum will adopt the submission. Filing an implementation agreement is a proposal, and membership includes the companies with the most to lose from it.
That the gap to Nvidia has closed. FP8 exaflops and card counts are not comparable to a competitor's figures without the same benchmark, and none was published.

Why it weighs this much

Signal strength

CostModerate

CommitmentModerate

ReversibilityEasy

PersistenceYears

Cost is moderate. A keynote roadmap costs reputation if missed and nothing else up front, but the standards filing is real work with a paper trail at an outside body, and a public one-generation-a-year cadence is a commitment a fab either meets or visibly does not. Commitment is moderate for the same reason: dates announced at a conference have been moved before, including by Huawei, and this announcement itself is a date being moved. Reversibility is easy, because a slipped quarter costs an apology. Persistence is years because interconnect standards, once adopted, outlive the chips that ride on them, and because a cadence published to 2029 is the frame every Chinese buyer now plans against. We are scoring the bars low on commitment deliberately. The action here is a schedule, and schedules are cheap.

What to watch

strengthens The Optical Internetworking Forum records or progresses Huawei's near-package optics implementation agreement, or a non-Chinese vendor announces support for Huawei's interconnect, by December 31, 2027.
weakens The Ascend 960DT is not available in the first quarter of 2027, or no Atlas 960 system at 4,096 accelerators is independently measured, by June 30, 2027.

The words

SuperPoD and SuperCluster: Huawei's names for a single tightly connected group of accelerators and for many such groups wired together. The point of both is to make a large number of ordinary chips behave like fewer, faster ones.
Near-package optics: Putting the optical components that move data as light right next to the chip instead of at the edge of the board. It cuts power and latency, and it is where the interconnect fight is currently being fought.
Optical Internetworking Forum: An industry body where equipment and component makers agree how optical parts talk to each other. Filing a design there is a bid to make your way the common way.
Exaflop: A billion billion calculations a second. At FP8, a low-precision format used for AI work, the number is much larger than the same machine's figure for scientific precision, so the two are not comparable.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.