Huawei has said the new chips will strengthen the foundation of AI computing power, both in China and around the world
In sum – what to know:
Huawei lays out three-year chip roadmap – Ascend 950 (2026), Ascend 960 (2027), and Ascend 970 (2028), with each release doubling compute capacity.
950 series delivers major upgrades – New low-precision formats, stronger vector processing, 2 TB/s interconnect, and specialized HBMs for prefill, recommendation, training, and decode workloads.
Scaling through proprietary innovation – Huawei adds custom HiF data formats and HBMs, aiming to sustain performance despite sanctions and intensifying global AI demand.
SHANGHAI — Chinese vendor Huawei unveiled its roadmap for the next generation of Ascend AI chips during the recent Huawei Connect event in Shanghai, China, outlining plans for three new chip series over the next three years.
“As we look ahead, some of you might be wondering about Huawei’s chip roadmap,” said Eric Xu, Huawei’s rotating chairman, during his keynote speech. “We’ll keep evolving Ascend chips to strengthen the foundation of AI computing power, both in China and around the world.”
The roadmap begins with the Ascend 950 family, which includes two chips: the Ascend 950PR, optimized for prefill inference and recommendation, and the Ascend 950DT, designed for decoding and training. Both chips share the Ascend 950 Die and deliver improvements in multiple dimensions, the executive said.
The 950 series supports new low-precision formats such as FP8, MXFP8, MXFP4, and Huawei’s proprietary HiF8. Xu noted that these enhancements will provide “significantly higher training efficiency and inference throughput,” with the chips capable of 1 PFLOPS in FP8, MXFP8, and HiF8, and 2 PFLOPS in MXFP4.
Other improvements include stronger vector processing, refined memory access granularity, shrinking from 512 bytes to 128 bytes, and a 2 TB/s interconnect bandwidth—2.5 times that of the Ascend 910C.
The Ascend 950PR will hit the market in the first quarter of 2026, available in card and SuperPoD server formats. Xu explained its design responds to the growing compute needs of agent-based applications and recommendation systems: “Our Ascend 950PR chip is designed to support these two scenarios with HiBL 1.0, our proprietary, low-cost HBM. HiBL 1.0 is more cost-effective than HBM3E and HBM4E.”
The Ascend 950DT, set to be launched in the last quarter of 2026, targets training and decoding workloads with higher demands for memory and bandwidth. It integrates HiZQ 2.0 HBM, offering 144 GB memory, 4 TB/s memory access bandwidth, and a 2 TB/s interconnect.
Xu highlighted that Huawei will also introduce the Ascend 960 chip in the last quarter of 2027. Compared with the Ascend 950 chip, it will double compute power, memory, and interconnect capacity. The chip also introduces HiF4, a proprietary 4-bit precision format designed to deliver greater accuracy than standard FP4 solutions, promising a major boost to inference throughput.
The roadmap culminates in the Ascend 970, scheduled for the fourth 2028. While still under development, Xu said the target is to double FP4 and FP8 compute power compared with the 960, double interconnect bandwidth, and increase memory bandwidth by at least 1.5 times. “I’m sure its performance will be worth the wait,” he added.
“Generally, we will follow a 1-year release cycle and double compute with each release. Throughout this process, we will keep evolving our Ascend chips, making them easier to use, supporting more data formats, and increasing their bandwidth. The goal is to stay on top of ever-growing demand for AI compute,” Xu added.
“Gaps still exist between Huawei and Nvidia in terms of individual chips and ecosystems, and we are working to close these gaps. But we have the capability to build SuperPoDs and SuperClusters – which are the source of our confidence,” Xu told Chinese media.
“Huawei is different from companies that work on AI models and applications: We are a provider of ICT infrastructure and smart devices. We are committed to making the most of our strengths to provide solid infrastructure. We seek to monetize infrastructure. Our core focus is SuperPoDs and SuperClusters,” he added.
Xu also unveiled the company’s roadmap for AI computing platforms, with plans to release what it claims to be the world’s most powerful single system in the last quarter of 2026 and double the performance with a new launch a year later.
During Huawei Connect event, Xu also unveiled the company’s newest SuperPoD products: the Atlas 950 SuperPoD — with 8,192 Ascend chips, and the Atlas 960 SuperPoD with 15,488 Ascend chips. He explained that SuperPoDs are designed to operate as single logical machines, combining multiple physical units for large-scale AI workloads.
The executive said that the Atlas 950 SuperPoD will be available in the last quarter of 2026, while the Atlas 960 SuperPoD will hit the market in the last quarter of 2027.
