DeepSeek Meeting Posts Removed After Huawei-Nvidia Chip Claims

A transcript said to record a closed-door meeting between DeepSeek founder Liang Wenfeng and investors began circulating in Chinese technology and investment circles late on July 22. It contained blunt comparisons of Huawei and Nvidia AI systems. By the following day, several WeChat posts carrying the document were no longer available.

China's National Business Daily reported on July 23 that the meeting took place in May and lasted about three hours and 44 minutes. An institution that has invested in DeepSeek told the newspaper that the meeting occurred and that the circulating account was broadly credible. 21st Century Business Herald also published an edited version. DeepSeek has neither authenticated the transcript nor said whether it is complete.

In a timestamped copy retained online, Liang said Huawei's Ascend 950 SuperPoD could replace an Nvidia GB200 or GB300 system for the same class of work. His estimate was that four Huawei accelerators would be needed for every Nvidia chip, with the Chinese technology about two years behind. He also said customers would still buy the domestic system at twice the price when high-end Nvidia processors were unavailable.

Huawei publicly displayed an Atlas 950 SuperPoD on July 17. The company says the system contains 1,024 accelerators, delivers one exaflop of FP8 computing power, provides 256 terabytes of unified addressable memory and has a three-microsecond round-trip latency. Nvidia's GB200 NVL72 combines 72 Blackwell GPUs with 36 Grace CPUs. Nvidia lists 360 petaflops of dense FP8 or FP6 performance and 720 petaflops when sparsity is used.

Those specifications do not test the two systems on the same model, dataset, power budget or software stack. Their scale, numerical formats and network designs also differ. No public benchmark verifies the transcript's "four to one" ratio or its claim of equal latency.

Neither company has published a complete and directly comparable procurement price, and the circulating record includes no contract. Data-centre power, networking and maintenance were not part of the comparison. Liang's comment that a system costing twice as much could still serve as a substitute described a supply-constrained purchase, not a cost advantage for Huawei.

The transcript says Huawei had allocated about 16,000 accelerators to DeepSeek and that the company was working on software compatibility. Liang estimated that training a system comparable with the largest US models might require about 50,000 GB300 chips or 200,000 Ascend 950 chips. On his own figures, the reported Huawei allocation would not be enough for that training run.

Liang gave several estimates for the gap between Chinese and American AI, including 12 to 18 months, six to 12 months and two years. He also said China was using about one twentieth of the computing resources deployed in the United States. The answers were oral and not entirely consistent. He identified access to chips, rather than research talent, as DeepSeek's immediate constraint.

Later in the meeting, Liang discussed TileLang, a higher-level programming method designed to speed the development and porting of computing kernels across GPUs and other AI accelerators. He said some workloads lost only 1% to 2% of their efficiency and argued that AI-assisted programming was weakening CUDA's software advantage. CUDA still encompasses compilers, libraries, debugging tools and distributed communications, and public evidence does not show DeepSeek moving its large-scale training to Ascend hardware.

Liang also said DeepSeek remained focused on artificial general intelligence, was in no rush to build a dominant consumer app and intended to keep releasing its strongest models as open source. DeepSeek has not publicly confirmed those remarks.

WeChat articles carrying the transcript began disappearing on July 23. Screenshots show a notice saying that a complaint had been received and that the material breached China's rules for public social-media accounts. Some users said the posts had been removed for disclosing secrets, but WeChat, regulators, DeepSeek and Huawei have not confirmed that account.

The material was never issued publicly by the meeting's organiser. It contains assessments of suppliers, hardware allocations, purchasing plans and the possible computing scale of future models. Those details could prompt a complaint from a participant, a company named in the discussion or another rights holder. China's rules require platforms to process complaints, but WeChat's notice does not identify the complainant or the passage that triggered removal.

Ruibao found the references to 16,000 chips, the four-to-one comparison, the two-year lag and TileLang in the surviving timestamped copy. It did not find the widely repeated claim that half of DeepSeek's core researchers were labelling data. It is not possible to establish whether other circulating versions were edited differently.

As of July 25, DeepSeek, Huawei, WeChat and Chinese regulators had not issued a public explanation of either the transcript or its removal. The notices show that posts were taken down after complaints. They do not establish that the document contained state secrets.

Sources: National Business Daily on the meeting and the origin of the transcript; an edited version published by 21st Century Business Herald; a timestamped copy of the transcript; a GitHub archive; official specifications for Huawei's Atlas 950 SuperPoD and Nvidia's GB200 NVL72; and China's rules for public internet accounts. Vendor performance figures are reported as company claims. DeepSeek has not authenticated the meeting record.