Leaked DeepSeek Call Puts Huawei's Chip Supply at 16,000 of 200,000 Needed

A leaked investor-call transcript quotes DeepSeek founder Liang Wenfeng as saying Huawei had given his company capacity for roughly 16,000 Ascend cards, while frontier training would require 200,000 of Huawei's newest cards. The document remains unverified, but it directly describes the hardware DeepSeek expected to add: Liang says large batches of machines arriving in the coming months were “basically all NVIDIA.” DeepSeek then told prospective investors it was suspending its follow-on fundraising, a move people familiar with the talks linked partly to Liang's frustration over reports about the remarks.

Fred Gao published the full English transcript on July 22, 2026, describing it as a record of a May 20, 2026, meeting lasting about three hours and 44 minutes. Gao wrote that he had no details with which to verify its authenticity. The transcript's header says speech-recognition software produced the text and AI organized it, warning that proper nouns and figures may contain errors. Bloomberg reported on July 25, 2026 that it had not authenticated the posts and DeepSeek had not responded to an emailed request for comment.

What the transcript says about chips

The transcript records Liang putting DeepSeek's current capacity at roughly 20,000 Nvidia H-equivalent cards when the call took place on May 20, 2026. Most had arrived during the preceding month or two, while more machines were still to come. The large batches expected in the following months were “basically all NVIDIA,” according to the document.

For a model comparable with the largest U.S. systems discussed in the call, the document has Liang estimating a requirement of 50,000 Nvidia GB300 accelerators or 200,000 Huawei 950 cards. That estimate covered training only, before research runs. DeepSeek's resources could support more work around a 10-billion-active-parameter scale, the transcript indicates, while an 800-billion-active-parameter system remained out of reach.

A longer-term assessment in the document was more favorable to Huawei. It records Liang saying the Ascend 950 supernode could substitute for Nvidia's GB200 and GB300 systems in performance and price. Yet he also described the hardware gap as four Huawei cards for one Nvidia card and about two years of product timing. Nvidia hardware could be depreciated over five years, he estimated, compared with at most three years for Huawei hardware.

Production was the immediate limit. Huawei had offered DeepSeek capacity for about 16,000 Ascend 950 cards, Liang's remarks indicate, while internet companies could receive more than 100,000 cards. He translated DeepSeek's allocation into the equivalent of about 4,000 Nvidia B-series cards and described that as insufficient for a next-generation model. “We also can't count on training that bigger model on Huawei later,” the published transcript records him saying, though he held out the possibility for 2027 or 2028.

The gap in Huawei's paper

The comments predate Huawei's July 22, 2026, technical paper by two months and cannot be read as Liang's response to it. That paper documented full-parameter post-training of DeepSeek's V4 family on Ascend hardware. The Huawei-led authors reported 34.22% model FLOPs utilization and a 2.93-fold efficiency improvement over an open-source baseline recipe, according to our July 23 coverage.

The paper left the original pre-training hardware for the 1.6-trillion-parameter V4-Pro unspecified. Tsinghua University professor Liu Zhiyuan told MIT Technology Review that DeepSeek appeared to have moved only part of the V4 training process to Chinese chips and may have relied mainly on Nvidia hardware. The transcript is consistent with that assessment but does not confirm what trained V4-Pro. The English version attributes to Liang the statement that V3 used Nvidia cards while DeepSeek replaced much of Nvidia's software environment with its TileLang compiler.

Liang's account also separates software portability from chip availability. The document has him predicting that domestic hardware and its software ecosystem would be workable within a year of the call. He identified Huawei's production capacity as the remaining problem. His stated preference was to buy chips at a reasonable price rather than design them inside DeepSeek.

The funding round stops

On July 25, 2026, Bloomberg reported that DeepSeek had already told some prospective backers that investment agreements expected within days would not be signed, according to people familiar with the discussions. The company may resume the process later, and it was unclear whether every prospective investor had received the same message. The report attributed the pause partly to Liang's frustration over online coverage of the May 2026 meeting.

DeepSeek had sought at least 10 billion yuan in the second round at a pre-money valuation of at least 480 billion yuan. Its first outside round closed in June 2026 with about $7 billion and a valuation of roughly $50 billion. Tencent, battery maker Contemporary Amperex Technology and the National Artificial Intelligence Industry Investment Fund participated, according to the same account. DeepSeek, founded in 2023 and owned by Zhejiang High-Flyer Asset Management, has begun IPO preparations and may file this year.

What remains unverified

The transcript cannot settle which hardware trained V4-Pro. The published transcript carries an explicit transcription warning, no original audio accompanies the English text in the source packet, and DeepSeek has not confirmed Liang made the statements. WeChat links carrying versions of the document were removed after it surfaced, while a Tencent technology outlet published a version organized into 118 items.

The fundraising pause has separate anonymous-source support, but negotiations remained fluid. The next public record could be a resumed financing or an IPO filing. Until then, the unresolved question is the one left open by Huawei's July 2026 paper: which chips performed the frontier pre-training for V4-Pro?

AI-generated summary, reviewed by an editor. More on our AI guidelines.