Huawei Doubles Down on Domestic AI Silicon: Two Ascend 960 Chips Slated for 2027 Despite US Curbs
At its 2026 Huawei Connect Conference (announced ~September 17, 2026), Huawei said it will introduce two new Ascend 960 series AI chips in 2027 — the 960DT (potentially Q1 2027) and 960PR (commercial in Q3 2027) — continuing the roadmap cadence it laid out for the Ascend 950 line, even without access to advanced Western manufacturing. Rotating Chairman David Wang framed the effort as directly challenging Nvidia:
"The firm shared this news during the 2026 Connect Conference, revealing its plans for AI infrastructure development... David Wang... said today that the company is pouring efforts into taking its AI computing business to new heights and rivaling Nvidia in the artificial intelligence technology battle." — Huawei Central, Sept 17, 2026
The system-level story is the real signal. Huawei says it has developed 11 semiconductors based on its UnifiedBus ("Unified Interconnect") technology, that its superclusters can support 1 million AI processors at once, and that it has sold over 1,000 supernodes to 370+ customers to date, with an AI chip ecosystem of 5,270 monthly active developers. Huawei also disclosed no specs for the 960 series — only "optimism regarding performance, efficiency, and launch time."
Why it matters. This extends the Chinese custom-silicon push (DeepSeek ordering 160k Huawei chips over Nvidia; Zhipu's domestic accelerators — see Huawei Doubles Down on Domestic AI Silicon: Two Ascend 960 Chips Slated for 2027 Despite US Curbs) from lab workarounds to a standing national compute stack with a public multi-year roadmap. It also lands the same week as the first multi-rack Nvidia Vera Rubin cluster going live in the US (see NVIDIA's Vera Rubin Era Goes Live: CoreWeave Brings Up First Multi-Rack Vera Rubin NVL72 Cluster) — both sides of the export-control wall are now shipping next-generation systems on a schedule. One demand-side data point from Huawei's own ecosystem: the Ascend 950DT's price has jumped 60% in three months, suggesting domestic supply remains tight.
Part of
This finding is an example of a pattern recurring across your work:
Highly competitive pricing and local deployment options for powerful Chinese open-weight models are breaking the pricing leverage of premium U.S. proprietary APIs.
DeepSeek is designing custom inference silicon optimized strictly for its own architectures to bypass traditional GPU bottlenecks and lower serving costs.
It also escalates China's domestic-silicon push one week after Huawei's Ascend 960 announcements (deepseek-custom-ai-inference-chip) — and a 20GW-by-2032 target is a hyperscaler-scale commitment tha…
It lands two days before the Trump-Xi summit (see us-china-ai-safety-diplomacy) and alongside Huawei's domestic-silicon push (deepseek-custom-ai-inference-chip).
** This extends the Chinese custom-silicon push (DeepSeek ordering 160k Huawei chips over Nvidia; Zhipu's domestic accelerators — see deepseek-custom-ai-inference-chip) from lab workarounds to a sta…
**Why it matters:** While the US frontier labs debate pacing (see frontier-ai-pacing-debate) and the US-China race drives custom silicon (see deepseek-custom-ai-inference-chip), Europe's champion just…
Revision history
Update: Huawei announced Ascend 960DT/960PR for 2027 at Huawei Connect, plus supernode/supercluster scale figures — extends the Chinese domestic silicon roadmap.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.
· by the agent
Update the existing Chinese custom silicon and serving note (deepseek-custom-ai-inference-chip) with Zhipu AI's massive GLM-5.3-Flash / Ox Alpha breakthrough, which ran entirely on a domestic 100,000-chip cluster serving 100T tokens per day.