Huawei CloudMatrix 384: Chinese AI giant with high price and high computing power

Huawei recently started delivering its new CloudMatrix 384 AI clusters to Chinese customers – and is making no secret of its goal: technological independence from Western suppliers, particularly NVIDIA. However, with a price tag of around eight million US dollars per system, Huawei’s in-house development is anything but an economy model. In return, it promises raw computing power in petaflops, manufactured exclusively with in-house hardware.

Chinese answer to NVIDIA GB200: more chips, more performance, more power consumption

At the heart of the cluster are 384 Ascend 910C chips, which are interconnected in an “all-to-all” topology. The system operates internally under the abbreviation CM384. According to Huawei, the configuration achieves up to 300 PetaFLOPS in BF16 calculations. For comparison: NVIDIA’s GB200 NVL72, currently the most powerful AI system from Santa Clara, “only” achieves around half that. However, Huawei buys this dominance through brute force. While the GB200 NVL72 relies on 72 GPU units, Huawei scales its system to five times the number of chips – which inevitably increases not only the space requirement but also the power consumption. According to industry sources, the power consumption of the CM384 is around 3.9 times that of the NVIDIA system. The efficiency in perf/watt is also clearly lower in direct comparison – a fact that Huawei is probably willing to accept.

Independence comes at a price: 8 million dollars per system

According to industry information, the price for a complete CloudMatrix 384 system is around eight million US dollars, which is around three times the price of an NVIDIA GB200 NVL72. This clearly indicates a strategic objective: Huawei does not want to offer a low-cost alternative, but rather establish an independent, export-independent high-performance platform for the Chinese market. Ten major Chinese customers are said to have already adopted the system and integrated it into existing data center infrastructures. Names are not mentioned, but according to reports, these are long-standing Huawei partners – presumably including state-funded cloud providers, telecoms groups and research institutions.

Technological assessment: efficiency problem remains

Although scaling beyond the number of chips alone allows Huawei to achieve impressive computing power, it does not solve the fundamental problem of energy efficiency. The Ascend 910C may shine with tensor calculations, but in direct comparison to NVIDIA’s Hopper or Blackwell GPUs, the chip lags noticeably behind in modern workloads – both in terms of memory coherence and software integration. Another shortcoming: While NVIDIA relies on a mature ecosystem with CUDA and a deep software stack, Huawei’s AI software remains largely proprietary and less widespread. Whether this can prevail in the long term remains to be seen. In China, however, the geopolitical factor is likely to be more decisive than open standards.

Conclusion: A politically driven system with a willingness to compromise on technology

With the CloudMatrix 384, Huawei shows that China is willing and able to develop complex AI infrastructures independently of the West. The choice was deliberately made to maximize autonomy – with all its advantages and disadvantages. Technically, the system is by no means efficient or modular in the Western sense, but its raw performance is certainly competitive – provided that power consumption and acquisition costs are of secondary importance. The goal is clear: in a market where US sanctions make access to NVIDIA GPUs difficult or impossible, Huawei is to secure the national computing base for AI applications. However, the offering is likely to be too expensive and too power-hungry for smaller companies – which is why CloudMatrix 384 will remain a prestige product for the really big players for the time being.

Source: Financial Times