Huawei is looking to take on Nvidia with a domestic alternative to the chipmaker’s flagship rack-scale AI servers.

The Chinese tech giant showcased its CloudMatrix 384 offering at the World Artificial Intelligence Conference (WAIC) in Shanghai, where its servers were lined up as a rival to Nvidia’s GB200 NVL72, with the American chipmaker unable to ship its powerful servers due to U.S. sanctions on exports to China.

The CloudMatrix servers are built using 384 of Huawei’s flagship Ascend 910C hardware, which features across 16 racks – 12 for compute, and four housing networking components. It’s not clear who is manufacturing the hardware, with U.S. officials investigating TSMC last November over an alleged order of Huawei’s prior gen 910Bs.

The system features five times more Ascend chips than there are Blackwells in the NVL72s, though Huawei’s chips are considerably underpowered by comparison.

Research from DeepSeek, the starlet startup that upended the AI and technology markets at the turn of the year, revealed that the Ascend 910Cs offer around 60% of the performance of Nvidia’s prior-gen H100s.

Despite featuring underpowered hardware, Huawei claims its CloudMatrix system is capable of delivering around 300 petaflops of compute performance – that’s more than the GB200 NVL72’s 180 petaflops.

Huawei's server alternative appears to surpass Nvidia's offering in memory and bandwidth, according to SemiAnalysis, which indicates it provides over 3.6 times more aggregate memory capacity and 2.1 times more memory bandwidth.

The CloudMatrix system does, however, consume significantly more energy, approximately 559 kilowatts per hour, around four times greater than Nvidia's servers.

In addition to the 910C-backed CloudMatrix system, Huawei is working on its next-gen AI chip, the 6nm-based Ascend 920. The hardware is expected to reach mass production in the second half of 2025 and will reportedly offer more than 900 teraflops (BF16 performance) per card, while boasting 4Tbps of memory bandwidth.

CloudMatrix comes as Nvidia looks to retake China

The Shanghai showcase of the CloudMatrix 384 comes as Nvidia is gearing up to re-enter the Chinese market.

After U.S. export restrictions barred shipments of high-powered AI chips to China, Nvidia was left scrambling to come up with alternative, low-powered alternatives. However, those too would also be subject to export restrictions, effectively shutting Nvidia out of the Chinese AI market.

The chipmaking giant looks to have gotten around the rules, confirming earlier this month it plans to restart shipments of its China-specific H20 chips, having reached an agreement with the U.S. government.

Nvidia’s absence was meant to be Huawei’s gain, with the firm jumping to take advantage with a local alternative.

However, its prior generation 910B was criticized by local users who took issue with subpar inter-chip connectivity, while some of Huawei’s own staff were reported to have bemoaned its alternative to Nvidia’s CUDA, dubbed CANN, with the FT reporting some staff called it “difficult and unstable to use.”

A potential stumbling block for Huawei’s CloudMatrix 384’s success is threats from the Trump administration to take action against firms using its hardware.

In addition to easing restrictions on exports of AI chips to intermediary nations, the Commerce Department said it plans to make it a violation of U.S. export controls to use Huawei’s Ascend hardware anywhere in the world.

The agency claims Huawei’s hardware was designed using software, technologies, and semiconductor manufacturing equipment that originated from the U.S. and that firms found using Ascend chips, including the 910C, could face “substantial criminal and administrative penalties” according to guidance published in May.