China Mobile has switched on a 1,024-chip Ascend 950DT supernode in Beijing with no named tenant, price or utilization. The same week, Beijing asked ByteDance and Alibaba how many Nvidia RTX Pro 5500s they want, and ByteDance is weighing about 1 million. When private buyers get a choice, they still pick Nvidia.
China Mobile has installed a 1,024-chip Huawei supernode at its Beijing computing hub, a machine its cloud unit rates at 1 exaflop of FP8 compute with 256 terabytes of unified memory. It is the largest named deployment yet of Huawei's Ascend 950DT, and it arrived weeks after Huawei said its 950 supernodes had entered volume commercial use. As Poe Zhao, a China tech analyst, pointed out on X, the announcement listed training, inference and agent development as uses and gave no price, no customer and no utilization figure.
That makes it evidence of supply. A state-owned carrier can buy a thousand Huawei chips for reasons that have little to do with anyone's cost per token. Zhao argued in July that supernode buyers are mostly state-owned enterprises, telecom operators, banks and energy and transport groups, and that compliance, data sovereignty and local deployment drive those purchases more than unit economics. China Mobile fits the profile exactly. The buyers whose choices reveal commercial demand are the private platforms that pay for compute out of advertising and subscription revenue, and this week they told us which chip they want.
China's industry ministry has asked ByteDance and Alibaba to report how many Nvidia RTX Pro 5500 chips they would buy, and for what, according to The Information, as reported by Seoul Economic Daily. ByteDance is weighing about 1 million. Nvidia plans to ship roughly 500,000 a quarter to China from late December and told customers to order by Sept. 30, today. The part is a workstation card with GDDR7 memory rather than the high-bandwidth memory of data-center accelerators, and it sells in China for ¥85,000 to ¥90,000, about the same as Huawei's Ascend 950PR. One prospective buyer plans to put eight in a server to run larger models.
Given an Nvidia graphics card and a Huawei AI chip at the same price, China's largest private AI spender wants the Nvidia card, and Beijing, which spent the past year steering companies toward domestic silicon, is surveying the demand itself.
Huawei's position looks contradictory until you separate the fab from the cloud. Rotating chairman Eric Xu said on Sept. 17 that Huawei cannot make enough AI computing equipment for China and is limiting sales abroad: "Since we don't have enough capacity to even satisfy the demand in China, we don't have a plan to expand into the international market in a fully-fledged way," he told reporters. Yet in March, Shanghai Securities News, a state-affiliated financial daily, reported that at one national computing center, domestic AI servers rented far less than Nvidia ones. Chips are scarce at the factory gate. Capacity is idle in some of the clouds that bought them. What sits between the two is software.
That's why DeepSeek's announcement on Sept. 30 matters more than China Mobile's. The model developer said it has open-sourced programming infrastructure for Ascend, including compute and communication libraries built with Huawei's support, and jointly advanced a supernode of 128 Ascend 950 chips. It is pushing TileLang, an open-source language it says offers "a simpler programming model" than Nvidia's CUDA. "To build a new generation of independent, self-controlled GPU software ecosystems, the first priority is establishing a high-level language that is universal, easy to program, and still capable of reaching the hardware's full performance potential," DeepSeek said. That is a model lab naming the bottleneck in plain terms. The most important Chinese model lab optimized for 128 chips. The carrier bought the 1,024-chip showpiece.
Huawei's defense is substantial. Xu said he believes Ascend's China market share "should be bigger than Nvidia's," and expects many Chinese model developers to train on 950DT systems next year. The company has deployed more than 1,000 systems on its older Ascend 910C, counts more than 5,200 developers working monthly on Ascend software and says more than 40 models have been trained on its platform. It is pulling the Ascend 960DT forward to the first quarter of 2027, three quarters early, and plans a 960 supernode linking up to 4,096 processors, with a new architecture, Peerium, meant eventually to run a million chips as one system and a new chip generation every year through 2029.
None of that answers the utilization question. A supply shortage at Huawei's fabs says Huawei can sell every chip it makes to someone; much of that someone is the state and its carriers. Xu gave no data behind the market-share claim. And the software gap he is racing to close is the one that decides whether installed Ascend clusters get rented. Huawei itself says communication between machines can eat more than 40% of training time in conventional systems, which is why its pitch leans so heavily on interconnect. DeepSeek's libraries attack that problem directly. A carrier announcing a bigger box does not.
There is a policy reason the Nvidia survey matters, too. China's sovereign-compute build rests on the assumption that domestic demand will fill domestic capacity. If the ministry clears a million-unit order for a foreign workstation chip, it is conceding that private inference will not wait for the software to mature. Washington's approval is still an open question: U.S. Trade Representative Jamieson Greer has said export controls weren't discussed in recent talks with China.
We expect the split to become explicit over the next six months. By the time Nvidia's first RTX Pro 5500 shipments land in late December, the useful signals will be whether MIIT approves purchases at anything like ByteDance's 1 million and whether China Mobile names a single tenant for its Beijing node. If ByteDance gets close to what it asked for, Ascend's 2027 growth will come where Xu points, in model training and state and carrier clouds, while the private inference business, the part of Chinese AI with paying customers and a meter running, stays on Nvidia hardware. Huawei will keep selling out. The harder test is who keeps its supernodes busy.
