The DeepSeek Compute-Gap Memo: What the Leaked Transcript Reveals About China's AI Ambitions

In May, DeepSeek founder Liang Wenfeng sat down with the investors backing the company's first-ever outside funding round and spent nearly four hours answering questions. A translated transcript of that closed-door meeting surfaced online in late July, spread across Weibo and Chinese tech outlets, and within days DeepSeek told prospective backers it would pause signing any new fundraising agreements.

Multiple outlets frame the pause as Liang's personal anger that confidential remarks got out, since the moderator had explicitly told attendees not to circulate figures or recordings. DeepSeek has not confirmed the transcript's authenticity, but investing institutions reportedly told a Chinese state outlet the meeting was real and the content accurate, and the granular Huawei and CUDA detail struck outside analysts as hard to fake.

The valuation and the drama are the headline. The part worth reading is what Liang told the room about compute: specific chip counts, specific gaps, and a blunt admission of how far DeepSeek actually is from the frontier.

By Liang's own account, DeepSeek runs on roughly 20,000 Nvidia H-series-equivalent processors, much of it only recently installed. His plain conclusion: even pouring ¥50 billion (~$7bn) into chips would not let them train a model at true US frontier scale.

He pegged Chinese developers at 1 to 2 years behind their US counterparts while working with about 1/20th the compute. On that hardware, he argued, China can win narrow tasks through engineering efficiency but cannot reach broad superiority. His stated purchasing wish was simple: B200s worth buying in any quantity at a fair price, and convert whatever capital is available straight into GPUs.

The most sensitive line in the whole transcript is one throwaway remark: "For us, we can buy some noncompliant cards." No chip model, supplier, or route is named, and nothing in the sourced material ties it to any specific export-control finding, so treat it as exactly what it is: unconfirmed and unelaborated.

The domestic fallback does not close the gap. Huawei allocated DeepSeek about 16,000 Ascend 950 processors, against the hundreds of thousands going to China's big internet firms. By Liang's own math that allocation is worth only around 4,000 Nvidia B-series chips, because roughly 4 Huawei processors equal 1 Nvidia in usable compute.

Liang said the Huawei chips cover DeepSeek's current model generation but not the next one, and that he buys them partly to help Huawei's hardware-plus-software stack mature. The longer-term play is a software layer that moves models across Nvidia, Huawei, and other silicon to cut CUDA lock-in. He reportedly claimed V3 was trained "barely relying on CUDA," with GPU operators rewritten in DeepSeek's in-house TileLang, and that AI-assisted coding is steadily eroding Nvidia's software moat.

This was the meeting for DeepSeek's first external raise ever. Until now the company had run on Liang's own hedge-fund capital, having held a public line of no fundraising, no IPO, no commercialization. The round pulled in roughly ¥50 billion (~$7.4bn) at a valuation near $50-54bn.

A second round was reportedly in early talks at a pre-money valuation near ¥480bn (~$71bn) when the transcript leaked. DeepSeek then verbally told backers no deals would be signed in the near term. Notably, Liang himself frames restraint as strategy rather than weakness, so the pause reads less as a crisis than as the company doing what it says it does.

Compute is not the only bottleneck Liang named. He said roughly half of DeepSeek's core researchers work on data annotation, reviewing and correcting post-training data by hand. His claim: OpenAI and Anthropic have built a lead in high-quality post-training data that money cannot quickly buy, and DeepSeek's capital structure cannot support the kind of expensive labeling US labs run.

On the roadmap, Liang put agents as this year's step, with coding agents as the top priority for China, and deprioritized 3D, video, and world models as off the main path to AGI. He also reaffirmed that DeepSeek will keep open-sourcing, saying even its most advanced models will eventually be released.