Huawei's Nvidia-Rival Chip Arrives Nine Months Early

Huawei Pulls Its Next AI Chip Three Quarters Forward

Huawei's rotating chairman David Wang used the opening keynote of Huawei Connect 2026 in Shanghai on September 17 to move the launch of the company's next AI training chip forward by three quarters. The Ascend 960DT, an AI training chip Huawei had targeted for the final quarter of 2027, is now scheduled for the first quarter of 2027, and Wang told the keynote audience that demand for Huawei's chips is outstripping its own manufacturing capacity.

A second chip, the Ascend 960PR, built for inference and recommendation workloads, moves up by one quarter to the third quarter of 2027. Huawei says both chips will be followed by an Ascend 970 in 2028 and an Ascend 980 in 2029, on what it now calls an annual upgrade cycle.

A Cluster That Shrank Even As the Chip Moved Up

Huawei's own numbers for the AI cluster built around this chip do not match what the company said a year earlier. The AI cluster Huawei demonstrated on September 17, a 4,096-card SuperPoD built on the new optical interconnect, delivers 8 exaflops of FP8 compute and 1 petabyte of HBM memory, saving more than 550 kilowatts of power per pod compared with a conventional design.

In September 2025, Huawei's own newsroom described an Atlas 960 SuperPoD scaling to 15,488 Ascend NPUs, a gap independent Huawei-watcher Rui Ma flagged this week, calling the new 4,096-card system "much smaller than what they originally laid out." No independent benchmarking organization has tested any Ascend 960 chip, so both the 2025 claim and the 2026 demonstration remain Huawei's own numbers until someone outside the company measures the real thing.

The Optical Standard Behind the Chip Is Not Just Huawei's

The interconnect technology carrying data between those 4,096 cards, called Near-Package Optics, is moving through a standards process that reaches well beyond Huawei. NPO places the optical components within five centimeters of the accelerator chip itself, moving data at 12.8 terabits per second while removing the retimer chips that add power and latency to a conventional pluggable-module design.

The Optical Internetworking Forum, a 170-member international standards consortium, approved Huawei's NPO project at its Malta plenary in May 2026. More than 40 companies are now contributing to the specification, including the American chipmaker Broadcom alongside Alibaba Cloud, Tencent, Baidu and China's state-run CAICT institute, and a September 2026 NPO industry summit brought together participants from the United States, Canada, China and Japan. NPO competes with Nvidia's proprietary NVLink 4 fabric, which links up to 576 GPUs; Huawei separately claims its own Lingqu UnifiedBus fabric can link up to one million processors, a figure that, like the SuperPoD numbers, has not been independently verified.

What Changes for a Buyer Who Cannot Order This Chip

A European or British infrastructure buyer still cannot simply order an Ascend 960DT. Huawei has sat on the US Entity List since 2019, and even where EU law does not name a specific chip, most European buyers have avoided Huawei silicon for the same trust and secondary-sanctions concerns that led national governments to exclude Huawei equipment from their 5G networks.

What the OIF process changes is different from the chip itself. Near-Package Optics is a published specification, not a Huawei product, so any systems vendor, a European chipmaker included, can build hardware to the same optical interconnect standard without touching Huawei's silicon or its export-control exposure. For an infrastructure architect who has treated Nvidia's proprietary NVLink as the only realistic path to hyperscale GPU clustering, an OIF-ratified alternative with Broadcom already at the table is a genuine second option to track, even though nobody outside Huawei has yet measured whether the chip riding on it performs as claimed.

Servola Journal

We do this for everyone trying to keep up with what technology is doing to our lives. The people who build it, and the people it happens to. The Servola Journal exists so that what we learn belongs to all of them.

Nobody pays us for this. No ads, no paywall, free to everyone. We just believe that understanding what's happening to all of us shouldn't depend on who can afford to pay for it.

If it gave you something today, tell us to keep going. Follow us, leave a like, or write a positive comment. We read every one, and they are what keeps us going.

Read next: China Puts a Number on Chip Independence  |  The Merger Filing Nvidia Never Made