Cloud Without Nvidia

US export controls were meant to deny China the most powerful AI chips. Huawei’s response was not to match Nvidia chip for chip, it cannot, but to overwhelm the problem with scale. Its CloudMatrix 384 networks 384 of Huawei’s own Ascend 910C accelerators into a single rack-scale system, and on aggregate performance it beats Nvidia’s flagship.^[1][2]

The numbers are a study in tradeoffs. CloudMatrix 384 delivers roughly 300 petaflops of dense compute, about 1.7 times Nvidia’s 72-chip GB200 system, with 3.6 times the high-bandwidth memory.^[1] But it wins by numbers, not efficiency: each Ascend chip is only about a third as powerful as a single Nvidia Blackwell, so Huawei uses more than five times as many, and the whole system draws roughly four times the power: about 559 kilowatts against Nvidia’s 140.^[1][4]

“China has no power constraints, just silicon constraints.” The whole design follows from that one sentence.

Trading watts for autonomy

That inefficiency is the point, not a bug. As the analysts at SemiAnalysis put it, China has no power constraints, just silicon constraints — abundant domestic electricity makes it rational to burn extra watts to work around chips it cannot buy.^[1] Huawei delivers the system not as hardware to buy but as capacity on Huawei Cloud, where its cloud chief has said it is already operational.^[3] It is a sovereign AI-compute stack, rented by the hour.

Two honest caveats. The performance figures are analyst estimates and Huawei claims, not independently benchmarked on production workloads, and “beats Nvidia” means on brute aggregate throughput, not efficiency. And the real bottleneck is upstream: how many Ascend 910C chips China can actually manufacture, which these figures do not address.

How this reaches a portfolio

CloudMatrix 384 is the clearest proof that export controls have changed the shape of Chinese AI infrastructure rather than stopped it. The lesson is that China can reach the frontier on system design and sheer scale even when denied the best silicon, at a cost in power and efficiency it is willing to pay. For investors that reframes the export-control thesis: the binding constraint is no longer whether China can build competitive AI clusters, but how many domestic accelerators it can fabricate to fill them. We watch Ascend production capacity, not benchmark slides, as the number that actually gates China’s sovereign compute, and we treat Huawei Cloud as the delivery vehicle that turns scarce chips into rentable national capability.