Huawei Technologies is to begin mass production of its latest Ascend 910C AI acceleration chip for domestic Chinese customers as early as May, Reuters reported, as the company taps into demand released by newly announced US restrictions on chips from Nvidia and AMD.

The report said some shipments to customers have already been made.

The news comes after Shenzhen-based Huawei earlier this month also announced the next-generation Ascend 920, which it said should begin mass production in late 2025, in an announcement timed to coincide with publicity around the latest sanctions.

Nvidia competitor

The 910C represents a technological evolution, rather than a breakthrough, according to two people cited by Reuters and a third person familiar with its design.

The people said the chip combines two of Huawei’s 910B processors into a single package through advanced integration techniques to achieve performance comparable to Nvidia’s H100.

The combination means the 910C has double the computing power and memory capacity of the 910B as well as incremental improvements including improved support for diverse AI workloads, the report said.

Huawei declined to comment on what it called speculation.

Nvidia’s H100, a flagship chip announced in 2022 that preceded the Blackwell-based B200, Nvidia’s current flagship, was banned for sale to China before launch.

Nvidia developed the H20 to comply with then-current US sanctions, but this month the H20 was added to US sanctions, along with AI chips from AMD, Intel and others.

Nvidia said it expected to take a $5.5 billion (£4.1bn) write-off on lost sales of the chip, with reports indicating it had been caught off-guard by the move.

Next-generation chip

Previous reports have estimated the 910C can achieve about 60 percent of the H100’s performance on inference, or delivering AI workloads, in contrast to the more processor-intensive training procedures.

The upcoming Ascend 920 is to use a 6 nanometre manufacturing process and is expected to exceed the 910C’s performance with more than 900 teraflops per card, DigiTimes Asia reported this month.

The unit is expected to deliver 4 terabytes per second memory bandwidth using HBM3 memory modules, with a 920C variant, built for Transformer and Mixture of Experts models, improving efficiency by 30 to 40 percent.

Last October China Telecom, a major Chinese state-owned telecommunications firm, said it had trained two large language models (LLMs) entirely using processors from Huawei, an indication of the country’s progress in developing AI chips.

Last September Huawei began offering samples of the Ascend 910C to large Chinese server companies for hardware testing and configuration, domestic reports said at the time.

The chip was being offered offered to large internet firms that were also major Nvidia customers, reports said.