Huawei pulls its next AI chip forward and bets on scale to close the gap with Nvidia
Huawei now expects its Ascend 960DT in early 2027, three quarters ahead of plan. The real strategy is to wire thousands of weaker chips into one machine, and that changes how China's AI market is built.
Huawei used its annual Connect conference in Shanghai on Thursday, September 17, to bring forward the launch of its next artificial intelligence training processor. David Wang, the company's deputy chairman and rotating chairman, said the Ascend 960DT is now expected in the first quarter of 2027, three quarters earlier than originally planned. A companion chip for inference, the Ascend 960PR, moves up by one quarter to the third quarter of 2027. Huawei also confirmed an annual cadence beyond that, with an Ascend 970 in 2028 and an Ascend 980 in 2029.
The chip dates made the headlines, but they were not the most important part of the announcement. Huawei's real answer to Nvidia, which cannot sell its most advanced processors in China, is not a single faster chip. It is a system architecture that links thousands of processors, and eventually far more, so that they behave like one computer. That strategy shapes the market for every Chinese company that trains or runs large AI models, and it determines how much of China's AI spending can stay at home.
A machine built from many chips
The centrepiece was the Atlas 960E SuperPoD, which Huawei describes as the industry's first to use near-packaged optics. A single pod holds 4,096 processors, delivers 8 exaflops of compute at FP8 precision and carries up to a petabyte of high-bandwidth memory. Huawei designed its own optical engine, called Hi-ONE, which moves up to 7.2 terabits per second. About 5,500 of these units replace roughly 48,000 conventional 800G optical modules. Huawei says the change cuts power draw by more than 550 kilowatts and raises system availability to 99.8%. It has submitted an implementation agreement on the technology to the Optical Internetworking Forum, a standards body.
Holding the system together is UnifiedBus, Huawei's interconnect, which links processors, memory, storage and networking hardware through a common protocol, and a new Peerium Computing Architecture built on top of it. Huawei says its SuperClusters can link up to 512,000 processors through a two-tier network, and a different topology could take that to a million. The Register, as reported by TNW, described the million-processor configuration as theoretical rather than proven.
Deployment figures were more concrete. Huawei said more than 1,000 of its Atlas 900 A3 SuperPoDs have been deployed, that the Atlas 950 is in commercial use, and that an Atlas 950 SuperCluster with 256,000 accelerator cards is being installed. The Atlas 960 system is still in testing.
Why interconnect matters more than the chip
Huawei's argument is that in very large clusters the chip is rarely the bottleneck. Yang Chaobin, who runs its ICT business group, said that in a conventional cluster of 100,000 processors as little as 20% of the compute goes to the model, with the rest waiting while data moves. A simulation by Huawei's Markov Lab found that a 100,000-processor cluster built from pods of 4,000 delivered 2.75 times the model FLOPs utilisation of one built from eight-processor servers.
The argument is self-serving, but it has a basis. Nvidia's own flagship products are increasingly sold as rack-scale systems rather than individual graphics cards. The difference is that Huawei must compensate for weaker silicon. TNW, citing The Register, put the 960DT at up to 288GB of memory and four petaflops of FP4 compute, roughly half the FP8 and a third of the FP4 performance of Nvidia's B300. Huawei's response is to use more chips, more optics and more power to reach a comparable result.
The caveats
Not every number moved in Huawei's favour. Rui Ma, an analyst who follows Chinese technology, noted that Huawei had previously said its Atlas 960 SuperPoD would scale to 15,488 Ascend 960 chips, while the pod announced this week holds 4,096. The chip is coming much earlier, she observed, but the system is considerably smaller than first described. Earlier delivery of the processor may have come at the expense of the scale of the largest configuration.
Manufacturing is the other constraint. Huawei does not have access to the most advanced chipmaking tools because of American export controls, and the clustering strategy is partly a way to work around that limitation. Announced schedules are only as good as the supply of wafers and high-bandwidth memory behind them, and Huawei has not disclosed production volumes.
Software and politics
Huawei spent as much time on developers as on hardware. CANN, its software layer for Ascend processors, which plays a role similar to Nvidia's CUDA, has moved to community-driven open source, and external developers now make up 61% of its community. Ascend supports more than 90 third-party open source projects and is now an official PyTorch accelerator backend. The Kunpeng processor ecosystem claims 4.16m developers and 7,200 partners. Software lock-in has been Nvidia's most durable advantage, and these figures are Huawei's attempt to show that it is eroding in China.
The timing was also political. The announcement came a week before a planned meeting between President Donald Trump and President Xi Jinping in Washington on September 24. Eric Xu, another of Huawei's rotating chairmen, argued that China needs to accelerate AI development to catch up with the United States, according to the Financial Times. Mr Wang put the scale of demand in stark terms, saying daily AI token consumption in China has reached about 500 trillion.
Our view
Huawei's announcement should be read as a statement about market structure rather than a technical milestone. China's AI industry is being built on a different template from America's: more chips, more interconnect, more power and a domestic software stack, all supplied by one national champion. That is less efficient per unit of compute, but it is good enough to train and serve competitive models, and it is improving every year.
For investors, the consequence is that Nvidia's China revenue should be treated as structurally impaired rather than temporarily blocked, whatever comes out of the Washington meeting. For Chinese AI start-ups, the trade-off is dependence on a single supplier in exchange for assured access. For founders elsewhere, Huawei's emphasis on optics and interconnect is a useful signal: the value in AI infrastructure is migrating from the processor to the system around it. Watch whether the 960DT ships on time and in volume in the first quarter of 2027. Huawei's record on the largest system configurations suggests some scepticism is warranted.
Sources
- TechCrunch: Huawei plans Q1 2027 launch of new AI chip as it takes on Nvidia
- South China Morning Post: Huawei unveils latest tech to boost AI power in push to break China's Nvidia reliance
- TNW: Huawei's answer to Nvidia isn't a faster chip, it's a much bigger machine
- TrendForce: Huawei speeds up AI chip roadmap, reportedly pulls Ascend 960DT forward three quarters to 1Q27
- Tech Wire Asia: Huawei unveils Ascend 960 roadmap and 11-chip AI infrastructure portfolio
- CRN Asia: Huawei outlines Ascend 960 roadmap and million-processor AI systems
Filed under Big Tech. Written by Invest Times Editorial under our editorial standards. Commentary and analysis for information only, not investment advice.
