Chinese chipmakers are moving quickly to replace US technology as Washington clamps down on exports and Beijing doubles down on self-reliance.

The immediate goal is to manufacture homegrown alternatives to Nvidia’s H20 chips.

The longer-term goal is a broader push to localise the full stack of AI computing.

Alibaba, previously one of Nvidia’s best customers, has already stepped up.

Its cloud arm is rolling out a new, domestically made inference chip designed to handle a wide range of AI workloads while remaining compatible with Nvidia’s software.

The company says it will invest around $78 billion in “AI + cloud” infrastructure over the next three years to keep customers supplied with local compute.

Plenty of other contenders have also entered the competitive fray.

At July’s World Artificial Intelligence Conference in Shanghai, promising Chinese start-ups, such as MetaX, showcased substitutes for Nvidia’s chips.

These substitutes leverage larger memory and multi-GPU clustering to deliver ‘good-enough’ inference.

Cambricon, which spun off from the Chinese Academy of Sciences in 2016, is also emerging as a significant beneficiary of increased domestic demand.

The Beijing-based AI chip designer posted strong first-half results, with surging demand from Chinese firms eager to reduce reliance on imported silicon.

Nvidia’s loss appears to be Huawei’s gain

As is usually the case in China, government policy is supporting ambitious start-ups.

In January, Beijing launched a 60-billion-yuan (A$12.6 billion) national fund to invest in AI chips, software and data, part of a larger state push to close the gap with foreign rivals.

Authorities have also discouraged purchases of Nvidia’s H20, nudging buyers toward domestic suppliers.

Huawei remains the flag-bearer with its Ascend line, although its software ecosystem still lags behind Nvidia’s. The emerging compromise is to lean on domestic chips for large-scale inference while making do with multi-die systems or mixed fleets.

Rather than slowing China down, US restrictions appear to be speeding up the shift to a local AI hardware ecosystem.

Expect to see more Chinese-made, cheaper but power-hungry AI products in servers and cloud offerings.

That means buyers should closely scrutinise software compatibility claims, especially around Nvidia’s CUDA tools, before making investments in non-Nvidia silicon.

If the Chinese are concerned about catching up to the industry-leading chipmakers, they aren’t showing it.

“There’s no need to worry about the chip problem,” Huawei founder Ren Zhengfei told a Chinese newspaper a couple of months ago.