Huawei's Ascend 960 AI chip arrives nine months early — and skips EUV entirely
US export restrictions were supposed to slow Huawei down. Instead, the company just moved its next-generation AI chip launch forward by nine months. The Ascend 960 is now set to debut in Q1 2027 — originally planned for Q4 2027 — with a second variant, the 960PR, following in Q3 2027. The acceleration signals that Huawei has found a way to keep moving without access to the advanced manufacturing equipment its rivals depend on.
The chip
The Ascend 960 doubles the compute performance of its predecessor, per Huawei Central. The headline specs are substantial: 2 PFLOPS at FP8 precision, 4 PFLOPS at FP4, 288GB of HBM memory, and 9.6TB/s of memory bandwidth. Those numbers target the data center workloads — large language model training and inference — where Nvidia currently dominates.
The underlying approach is called the Tau Scaling Law, a framework Huawei announced in May 2026 as its answer to slowing transistor miniaturization. The core idea: instead of shrinking transistors further, optimize across chip architecture, packaging, and software simultaneously. Huawei claims this path can deliver transistor density equivalent to a 1.4nm process node by 2031 — without ever touching an EUV lithography machine from ASML.
The competitive reality
Huawei rotating chairman Guo Ping has said the company is closing the technology gap through its own engineering rather than waiting for manufacturing access it can't get. The ambition is visible in the roadmap: Ascend 960 in 2027, Ascend 970 in 2028, Ascend 980 in 2029 — an annual release cadence designed to chip away at Nvidia's lead year by year.
At the Connect 2026 conference, Huawei also unveiled a SuperPoD cluster that bundles 15,488 Ascend chips together, alongside a 4,096-card super node delivering 8 EFLOPS, 1PB of HBM, and reportedly 550kW lower power draw than comparable configurations.
The skepticism is real, though. The Council on Foreign Relations estimates Nvidia's H100 and H200 chips could still be roughly 17 times more capable than the Ascend 960 by the second half of 2027, according to Analytics Insight. Production yields at SMIC — China's leading chipmaker, limited to older DUV lithography — remain an open question. Huawei's primary customers are domestic AI cloud operators like Baidu, Alibaba, and DeepSeek, not Western enterprises.
The bigger picture
The Ascend 960 story is less about a single chip and more about whether architectural ingenuity can substitute for lithographic access at scale. Huawei has reportedly designed and tested 381 chips over six years under sanctions — quietly building the knowledge base the Tau Scaling Law now rests on. Whether the approach closes a 17x gap in three annual releases is a very different question. But the fact that the timeline moved nine months sooner suggests the answer isn't "no" just yet.
