Huawei CloudMatrix on par with Grace Blackwell for the first time

A rare admission from NVIDIA: In an interview with Bloomberg, CEO Jensen Huang confirmed that Huawei’s “CloudMatrix” AI infrastructure can now compete with the latest Grace Blackwell systems. This indicates a paradigm shift in the global AI arms race – from a country that is denied high-end chips such as the H100 and H200 by the USA under export controls.

“They are unstoppable”

“They are moving very fast,” said Huang in the interview, which is tantamount to a diplomatic capitulation. While Huawei was seen as a lagging challenger in the past, the NVIDIA boss now speaks openly of a “comparable level” to the Californians’ current high-end products – namely H200 and Grace Blackwell. The CloudMatrix platform, which was mentioned publicly for the first time, is said to be capable of scaling even beyond the capacities of the current GB200 clusters. The admission comes as no surprise, but at the worst possible time. NVIDIA is under pressure – not only due to falling margins in the gaming segment, but above all due to regulatory uncertainty in the lucrative Chinese business. The USA prohibits the export of high-performance chips, which NVIDIA is trying to circumvent with specially restricted models such as the H20 or B40. However, Huawei has apparently created its own architecture – without Western IP, but with a great deal of in-house expertise.

Ascend 910C: The new rival to the H200

In detail, we are talking about Huawei’s latest AI accelerator, the Ascend 910C. According to Huang, its performance is roughly equivalent to the H200 – NVIDIA’s most powerful Hopper derivative. This would not only put Huawei on an equal footing, but also on a comparable product cycle. The lead that NVIDIA has built up over the years seems to have shrunk noticeably in less than 24 months. Just a year ago, the consensus among Western analysts was that China was at least a generation behind in AI accelerators. This applied to both raw performance and software integration. However, with CloudMatrix and Ascend 910C, Huawei seems to have not only made up for this lag, but also embarked on an autonomous development that is no longer dependent on US technologies. Jensen Huang himself speaks of a company “that is looking for ways to compete – and is doing so very successfully”.

Geopolitics and technology: NVIDIA on the losing end?

The technical details of the CloudMatrix clusters are only fragmentarily known so far. Internal sources and leaks speak of a close integration between Ascend hardware, MindSpore software stack and a distributed computing cluster that can be scaled horizontally – similar to NVIDIA’s Grace Blackwell superchips. The key difference: Huawei can offer these systems on the Chinese market without export restrictions, while NVIDIA is only allowed to supply slimmed-down versions – if at all. This could prove to be a strategic disadvantage for NVIDIA. Data centers in China are one of the biggest growth areas in the AI market. At the same time, NVIDIA is gradually losing market share there. The US sanctions act as a double-edged sword: while they restrict China’s access to Western cutting-edge technology, they are also forcing it to build up its own semiconductor structures – with apparently considerable success.

The new balance of power

The fact that Huang now publicly acknowledges the performance of Huawei’s systems is not only a signal to investors, but also an indication of an intensified competitive situation. Until now, NVIDIA was considered unassailable in the field of high-performance AI infrastructure. But this dominance is crumbling – first regionally, soon possibly globally. It remains to be seen how much of Huawei’s solution actually consists of in-house development, but the facts speak for themselves: the West does not supply chips, but Huawei does. And at a level that even commands respect from NVIDIA.

Source: Youtube