Key points
- Huawei lays out an 11-chip AI portfolio
- Ascend roadmap pulled forward into 2027
- Huawei says demand already tops its supply
Huawei outlined an 11-chip portfolio for AI infrastructure at its Huawei Connect conference in Shanghai on Thursday. Rotating chairman David Wang said its Ascend accelerators are now officially supported as a PyTorch backend. He described Ascend as the fourth hardware platform listed alongside Nvidia, AMD and Intel, and the first Chinese platform to reach that point.
Huawei also says it cannot make enough Ascend chips to meet demand at home. Eric Xu, another rotating chairman, told reporters that Chinese demand is "significantly outstripping supply." Huawei is supplying limited quantities overseas but does not plan a major export push for now, Nikkei Asia reported on September 17.
What Huawei announced
The lineup covers more than 10 chipsets: Ascend AI accelerators, Kunpeng central processors, chips for its UnifiedBus high-speed interconnect, storage and storage-controller chips, and a network interface controller. "Among these chips, Ascend is the most critical one in the entire system," Wang said.
Huawei plans to introduce a new Ascend generation each year. The Ascend 960DT is due in the first quarter of 2027, three quarters earlier than previously planned. Huawei says its compute performance will be twice that of the preceding generation.
| Chip | Timing | Role |
|---|---|---|
| Ascend 950PR | In commercial use | Inference |
| Ascend 950DT | Q4 2026 | Training |
| Ascend 960DT | Q1 2027 | Training |
| Ascend 960PR | Q3 2027 | Inference |
| Ascend 970 | 2028 | Next generation |
| Ascend 980 | 2029 | Next generation |
Wang said more than 1,000 of Huawei's Atlas 900 A3 SuperPoDs are already deployed, and its Atlas 950 SuperPoD is in large-scale commercial use. SuperPoD is Huawei's AI cluster that links many accelerators so they can work as one computing system, the way Nvidia sells racks rather than single chips. UnifiedBus, the new interconnect, is Huawei's answer to Nvidia's NVLink, and last week the company submitted a proposed near-packaged optics specification to the Optical Internetworking Forum, an alternative to the co-packaged optics that Nvidia (NVDA) and Broadcom (AVGO) are pushing.
Why Huawei says it can't keep up with demand
Xu said Huawei is holding back on exports because the company cannot make enough chips for buyers at home. "We are doing some testing ... providing some supply to those countries, but the volume is quite limited," he said, adding that Huawei has fielded inquiries from clients in several countries "who have a very strong demand." The demand at home is real enough that Chinese AI lab DeepSeek is building around Huawei silicon; we covered DeepSeek's plan to run 160,000 Ascend chips at a single data center.
Xu tied the shortage to the whole supply chain, not just Huawei. "When we look at those suppliers, whether it is for optical modules, storage or memory products, they are selling at a very high premium price, not necessarily because they have very, very good products, but because of a very severe supply shortage," he said. That squares with the broad rise in chip prices this year. Xu said he expects global supply and demand to balance around 2029, and China around 2030.
What it means for Nvidia, Intel and AMD
Xu said that, based on Huawei's own data, its Ascend products already hold a bigger share of the China market than Nvidia. That lines up with what Nvidia's own chief executive has said. Jensen Huang has called Huawei a formidable rival and said U.S. export controls cost Nvidia market share in China, where he put the company's share at 95% before the curbs took hold and much lower after.
The PyTorch line is the sharper signal for Intel (INTC) and Advanced Micro Devices (AMD). Wang said Ascend is now an officially supported PyTorch accelerator backend, listed on the framework's site alongside the three American chipmakers. That is a real foothold in the software ecosystem Nvidia has defended hardest, though it does not by itself mean Ascend matches them on maturity or performance. PyTorch has said work toward a native, out-of-the-box Ascend experience is still continuing. Chinese labs are already testing that ground; one, Zhipu, says its domestic chips match Nvidia on cost per token. None of this is a US market event on its own. Nvidia closed Friday at $222.08, AMD at $559.80 and Intel at $108.62.
How much Huawei is spending
Xu said Huawei has invested 1.8 trillion yuan ($268.4 billion) in research and development since 2007, a company-wide total, and framed the whole effort as a national one.
"China is a country with a very strong sense of crisis. We cannot let our fate be determined by others' willingness to sell chips to China or not," Xu said. "Even though our chips may not be as advanced, even though our technology is not as developed, at the very least, you have a chip supply from China, so that you don't have to worry day in and day out."
