Huawei Pushes Ascend 960 AI chip Launch to Q1 2027 Amid Sanctions

Huawei’s Ascend 960 AI chip is now scheduled for release in the first quarter of 2027 — nine months earlier than originally planned for Q4 2027. A second variant, the Ascend 960PR, will follow in Q3 2027. The company confirmed the revised timeline despite ongoing US export restrictions limiting access to advanced semiconductor manufacturing tools.

Performance specifications and data center targeting

The Ascend 960 delivers double the compute performance of its predecessor, according to Huawei Central. Its technical profile includes 2 PFLOPS at FP8 precision, 4 PFLOPS at FP4, 288GB of HBM memory, and 9.6TB/s of memory bandwidth. These metrics are explicitly oriented toward data center workloads, particularly large language model training and inference — domains where Nvidia currently holds dominant market position.

Tau Scaling Law: Architecture over lithography

Huawei introduced the Tau Scaling Law in May 2026 as a strategic response to diminishing returns from transistor miniaturization. Rather than pursuing further node shrinks, the framework emphasizes concurrent optimization across chip architecture, advanced packaging, and software layers. Huawei asserts that this methodology can achieve transistor density equivalent to a 1.4nm process node by 2031 — without using any EUV lithography equipment manufactured by ASML.

Roadmap and infrastructure rollout

At the Connect 2026 conference, Huawei disclosed an annual chip release cadence: Ascend 970 in 2028, followed by Ascend 980 in 2029. The company also unveiled a SuperPoD cluster integrating 15,488 Ascend chips, alongside a 4,096-card super node capable of delivering 8 EFLOPS, 1PB of HBM memory, and reportedly consuming 550kW less power than functionally comparable configurations.

Market reception and production constraints

According to Analytics Insight, Nvidia’s H100 and H200 chips may still exceed the Ascend 960’s capabilities by approximately 17 times during the second half of 2027, as estimated by the Council on Foreign Relations. Production yields at SMIC — China’s leading foundry, restricted to deep ultraviolet (DUV) lithography — remain unconfirmed. Huawei’s primary customers for the Ascend series are domestic AI cloud operators, including Baidu, Alibaba, and DeepSeek, rather than Western enterprises.

Engineering scale under sanctions

Huawei has reportedly designed and tested 381 chips over six years while operating under US sanctions. This accumulated engineering effort forms the foundation for the Tau Scaling Law. The nine-month acceleration of the Ascend 960’s launch timeline indicates continued progress along this self-reliant development path — though whether it can close a 17x performance gap across three annual releases remains an open question.