.png){kind=logo}
Huawei has laid out plans for two new AI chips, the Ascend 960DT and Ascend 960PR, as part of a three-year roadmap meant to close the gap with Nvidia. The Shenzhen-based company shared these plans through its rotating chairman, Eric Xu, building on the roadmap it first revealed at the Huawei Connect event.
The 960 series is set to arrive in the last quarter of 2027, following the earlier 950 series launching through 2026.
Huawei wants these chips to power AI training and inference workloads both within China and for customers abroad, at a time when US sanctions have limited its access to foreign chip technology.
The Ascend 960 chips will double the compute power, memory and interconnect capacity of the 950 series, according to Huawei. The company is also introducing HiF4, a proprietary 4-bit precision format designed to push inference throughput higher than standard FP4 solutions.
Xu had earlier described the improvements across the roadmap as delivering "significantly higher training efficiency and inference throughput." The 950 series that precedes the 960 chips already includes a 2TB/s interconnect bandwidth, which is 2.5 times faster than the Ascend 910C.
Huawei's own high-bandwidth memory, branded HiBL and HiZQ, allows the company to sidestep its past dependence on South Korean and American suppliers for this critical component.
Also read: A20 Pro vs Previous Apple Chips: 7 Major Improvements
A report from the Council on Foreign Relations offers a more cautious view of Huawei's progress. It estimates that Nvidia's best AI chips could still be about 17 times more powerful than Huawei's by the second half of 2027, even though Huawei has closed some ground despite sanctions.
The report suggests the Ascend 960 could eventually match Nvidia's H200 chip in raw capability, though production limits may cap Huawei's output at around 4 million AI chips by 2027. Huawei's Atlas supernodes, built to link thousands of Ascend chips together, have already outperformed Nvidia's GB200 NVL72 system on select benchmarks, according to research firm SemiAnalysis.
