Unconfirmed: talk is circulating that a leaked China Mobile test report shows benchmark and economics data for Huawei's 64-card Atlas 950 PoD, but Automatica has not confirmed the document or the results.

On 18 November, @bebuu_ai on X1 wrote:

"A leaked China Mobile test report reveals some interesting real-world numbers for Huawei's 64-card Atlas 950 PoD."

In the same post, the account wrote that the "950DT delivers 1,951 tokens/s per card on GLM-5.2 and 1,451 on DeepSeek-V4-Pro" and added:

"That's 4x faster than 910C and around 86% of Nvidia H200's throughput on GLM-5.2, while using just 0.28 kWh per million tokens."

The post also made a business-viability claim, saying:

"The interesting part is the business side: GLM-5.2 generates an 18.6% profit, but running DeepSeek-V4-Pro at its current API prices results in a 34% loss."

What the post does not show is just as important as the headline numbers. The available evidence contains only that single X post, with no linked report, no published test methodology, no model settings, no batch size, no latency target, no precision format, and no explanation of the comparison basis for Ascend 910C or Nvidia H200. Without those details, the figures cannot be reproduced or fairly compared. They should be treated as an attributed claim from a social post, not as an established performance result.

The post clears the minimum account thresholds in Automatica's rumor policy, with 2,569 followers and 22 likes recorded in the available evidence for eligibility, but that is a publication standard, not evidence that the claim is true @bebuu_ai on X1. There is no independent corroboration in the material provided.

What would confirm the rumor is straightforward: publication of the alleged China Mobile test report, or on-record confirmation from China Mobile, Huawei, or both, with enough configuration detail to evaluate the workload and baseline choices. What would weaken or kill it would be a denial from either company, or a released document whose setup does not match the quoted figures. In the available evidence provided here, neither China Mobile nor Huawei has said anything.

There is also no earlier reporting in the available evidence on this specific Atlas 950 benchmark claim. The two Automatica Press links supplied are unrelated prior coverage: one on an alleged Lenovo handheld leak and one on Nvidia's Hopper shipments to China Automatica Press2, Automatica Press3. They do not independently support the Atlas 950 rumor.

As it stands, the only verified fact is that a named X account publicly posted these numbers and comparisons on 18 November @bebuu_ai on X1. The performance claims, energy figure, and profit-and-loss assertions remain unconfirmed.

Cited posts