Huawei has pulled forward the launch of its next-generation Ascend 960 AI chip to the first quarter of 2027, three quarters ahead of its original schedule, with the company claiming the processor delivers twice the performance of the current 950 series. The announcement came at the Huawei Connect conference in Shanghai and forms part of an explicit bid to displace Nvidia inside China’s AI buildout.
Wang Tao, Huawei’s rotating chairman, told the conference the 960 “has been readied ahead of schedule and delivers twice the performance” of existing products, according to Chinese outlets including Kechuangban Daily. The 960DT variant for AI model training is due in the first quarter of 2027, with the 960PR inference chip following in the third quarter. The Ascend 970 is scheduled for 2028 and the Ascend 980 for 2029, giving Huawei a published three-year roadmap in a market where it has historically announced hardware close to availability.
Chips alone are not the plan
The more revealing part of the announcement sits outside the processor itself. Huawei said it has developed 11 types of semiconductors based on its UnifiedBus interconnect architecture, covering computing, networking, storage and management. The company is betting that system-level design can compensate for the performance gap that US export controls impose on any single Chinese chip.
That bet has a name attached to it. Huawei has argued that Moore’s Law, which drove decades of performance gains by shrinking transistors, has hit physical limits, and it promotes what it calls Tau’s Law instead, focused on cutting the time signals take to travel inside a chip. Reuters noted the obvious dependency: for the approach to work, the chips must be able to exchange information at sufficiently high speeds, which is exactly what UnifiedBus is meant to provide.
The clustering strategy is already visible in deployment numbers. Wang said more than 1,000 sets of the Atlas SuperPoD have been deployed with about 370 customers, and that the Atlas 950 SuperPoD has begun large-scale commercial deployment. The newer Atlas 960 SuperPoD, unveiled at the same event, chains up to 4,096 optically interconnected neural processing units into SuperClusters that can reach 512,000 NPUs. Huawei also introduced its Peerium Computing Architecture, the foundation layer for those systems, and launched the OceanStor M900 storage cluster aimed at AI inference workloads.
The Nvidia comparison
Guo Ping, chairman of Huawei’s supervisory board, told new employees earlier that “Huawei’s goal is to become Nvidia.” The concrete meaning is software: Huawei wants the range of large language models in China to run on its platform the way they run on Nvidia’s CUDA ecosystem today. Winning that position requires convincing Chinese AI developers that the toolchain is mature enough to bet on, which is a slower process than shipping silicon.
The stakes are visible in the broader chip market. Omdia reported this week that global semiconductor revenue jumped 31.4% sequentially to a record $425 billion in the second quarter of 2026, driven by AI demand and rising memory prices, and expects the third quarter to top $500 billion. China’s share of that buildout is the market Huawei is competing to own, and the company’s decision to publish a three-year roadmap reads as an attempt to give domestic buyers confidence to plan around its hardware.
| Chip | Scheduled | Role |
|---|---|---|
| Ascend 960DT | Q1 2027 | AI model training |
| Ascend 960PR | Q3 2027 | Inference |
| Ascend 970 | 2028 | Next generation |
| Ascend 980 | 2029 | Next generation |
Constraints remain
US sanctions still cap what Huawei can achieve per chip. Without access to leading-edge fabrication, the company cannot match Nvidia’s single-chip performance, which is precisely why the roadmap leans on clustering and interconnect rather than raw silicon. Analysts have repeatedly noted that Huawei’s multi-chip systems draw more power and cost more to run than equivalent Nvidia installations, a disadvantage that matters less if Chinese customers have no alternative.
Power consumption is not a trivial concern at the scale Huawei is proposing. A SuperCluster of 512,000 NPUs competes with the largest systems in the world, and keeping that many chips synchronized depends on the interconnect performing as advertised. The 10-day weather forecast system China recently built on homegrown AI chips, reported by Seoul Economic Daily this week, shows domestic demand exists for such systems beyond benchmarks and demonstrations.
The timing of the announcement carries political weight. It came days before an expected meeting between President Trump and Chinese leader Xi Jinping, and as Washington debates how far to tighten or loosen chip export rules. Huawei moving its schedule forward signals confidence that domestic demand, driven by Chinese AI firms unable to buy Nvidia’s best hardware, will absorb whatever it can produce.
The company also used the event to position itself against Nvidia on networking, where interconnect bandwidth determines how efficiently thousands of chips train a single model. Nvidia’s own response has been to bundle networking ever tighter with its accelerators through NVLink and its systems products, so the two companies are converging on the same system-level thesis from different starting points. The difference is that Nvidia sells globally while Huawei’s addressable market is effectively China plus countries willing to accept sanctioned-adjacent hardware.
For Nvidia, the immediate commercial loss is already locked in, since export controls bar its top chips from China. The longer-term risk Huawei’s announcement highlights is different: if Chinese AI infrastructure standardizes on Ascend and UnifiedBus, Nvidia’s eventual return to the market, under whatever trade terms emerge, would face an entrenched domestic competitor with its own software gravity rather than an empty field. That is the scenario Guo Ping’s remark to new employees was really describing.