Huawei Connect 2025, held in Shanghai in September 2025, is best remembered for one thing: Huawei published a multiyear roadmap for its Ascend AI chips for the first time. Alongside the chips came giant SuperPoD computing systems, new Kunpeng server processors and a homegrown interconnect called UnifiedBus. This explainer covers what Huawei announced, why it matters under US export controls, and how the plan has played out through October 2026.
Huawei’s Ascend AI chip roadmap
Eric Xu presented the roadmap in his keynote on September 18, 2025. Huawei also said it had developed its own high bandwidth memory (HBM), a component that has been hard for Chinese firms to source.
| Chip | Planned timing (2025 roadmap) | Focus and key claims |
|---|---|---|
| Ascend 950PR | Q1 2026 | Prefill and recommendation; 1 PFLOPS FP8; lower cost HiBL 1.0 memory |
| Ascend 950DT | Q4 2026 | Decode and training; 144GB HiZQ 2.0 memory at 4 TB/s |
| Ascend 960 | Q4 2027 | Roughly double the 950 in compute, memory and interconnect |
| Ascend 970 | Q4 2028 | Targets double the 960’s FP4 and FP8 compute |
Splitting the 950 into PR and DT versions reflects how AI inference actually works. The prefill stage, which reads the prompt, needs raw compute, while the decode stage, which generates the answer token by token, leans heavily on memory bandwidth. Pairing each version with a different memory type lets Huawei manage cost and supply.
SuperPoDs, SuperClusters and UnifiedBus
The systems were the headline numbers. Huawei described the Atlas 950 SuperPoD as 8,192 Ascend 950DT chips across 160 cabinets, rated at 8 EFLOPS of FP8 compute, targeted for Q4 2026. The Atlas 960 SuperPoD would scale to 15,488 Ascend 960 chips in Q4 2027. Huawei also sketched SuperClusters that combine dozens of SuperPoDs into hundreds of thousands of chips.
What ties them together is UnifiedBus, Huawei’s own interconnect protocol, which it formally released at the event along with UnifiedBus 2.0 specifications. Huawei said the earlier Atlas 900 A3 SuperPoD, built on UnifiedBus 1.0, had been shipping since March 2025, with more than 300 units deployed.
Server CPU for Q1 2026 in 96 core and 192 core versions.
A general purpose computing pod built on Kunpeng 950, also aimed at Q1 2026.
A successor planned for Q1 2028, including a high density model with at least 256 cores.
Why it matters for AI chips and US export controls
US export rules block Huawei from buying cutting edge chips and from using leading foreign foundries, and they restrict China’s access to advanced chipmaking tools. That caps how powerful each Ascend chip can be. Huawei’s answer is scale: if a single chip trails Nvidia’s best, link far more of them with a fast interconnect so the whole system competes.
Washington has pushed back on Ascend’s spread abroad. In May 2025, the US Commerce Department warned that using Huawei Ascend chips anywhere in the world could violate US export controls. On the other side of the ledger, the Trump administration moved in December 2025 and January 2026 to allow Nvidia H200 sales to China in exchange for a 25 percent cut to the US government, a policy that complicates Huawei’s position in its home market.
What has happened since, through October 2026
- Ascend 950PR: Reports in spring 2026 said mass production began in April 2026, with full scale shipments planned for the second half of the year. Demand from major Chinese tech companies reportedly outpaced supply.
- DeepSeek V4: Released in April 2026, the open model was optimized for Huawei hardware, and Huawei said its Ascend 950 based systems fully support it.
- Huawei Connect 2026: At the September 2026 event in Shanghai, rotating chairman David Wang said the Ascend 960DT would arrive in the first quarter of 2027, earlier than first planned. Huawei executives also said the company cannot yet make enough AI chips to meet demand and is prioritizing customers in China.
- Still ahead: The Ascend 950DT and the Atlas 950 SuperPoD were both scheduled for the fourth quarter of 2026, so their delivery is the next milestone to watch.
FAQ
When was Huawei Connect 2025?
It took place in Shanghai in September 2025, with Eric Xu’s main keynote on September 18.
What is the Ascend 950?
It is Huawei’s 2026 generation of AI accelerators, split into the 950PR for inference prefill and recommendation and the 950DT for decode and training.
Can US companies use Huawei Ascend chips?
The US Commerce Department has warned that using Ascend chips anywhere may breach US export controls, so US firms and many global companies avoid them.
Is Huawei’s chip as fast as Nvidia’s?
Not chip for chip. Huawei’s own comparisons and outside estimates place its accelerators behind Nvidia’s newest parts, which is why it leans on very large clustered systems.
The verdict
Huawei Connect 2025 turned Ascend from a secretive program into a public, yearly roadmap built around massive clustering and homegrown memory and interconnects. A year later, the first chip is in production and the next generation has moved up, but supply limits remain the biggest constraint on how fast Huawei can challenge Nvidia, even within China.
Updated October 2026: rewritten as an explainer on the Huawei Connect 2025 AI chip announcements, US export controls and progress through Huawei Connect 2026.
Featured photo: Laura Ockel on Unsplash.
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.

