DeepSeek's leaked investor call: 1/20th the compute, AGI or nothing, and the round it nearly killed
A transcript of a private investor call given by Liang Wenfeng on 20 May 2026 began circulating on Chinese social media in late July 2026, triggering a chain of events that paused and then reshaped DeepSeek's second fundraising round. The transcript — a speech-to-text document that flags its own error rate on names and figures, and which DeepSeek has not publicly authenticated — offers the most candid picture to date of the lab's strategy, its compute constraints, and its founder's indifference to commercial convention.
Liang told investors that DeepSeek does not really want users. He likened them to sesame seeds on the road to a watermelon — AGI — and said the chatbot and enterprise API are by-products of the research, not the point. While US frontier labs fight for developer, consumer, and enterprise mindshare, Liang passed on the consumer opportunity: with the hype around DeepSeek in early 2026, fighting ByteDance for consumer users was a real business he could have won. He declined, betting that what is coming makes today's consumer market look small.
The most widely quoted passage concerns the compute gap. Liang said DeepSeek is roughly one to two years behind the US frontier, running on about one-twentieth of the compute. The lab has approximately 20,000 H-equivalent GPUs, most arriving only in the past two months. To train a frontier-scale model with ~800B active parameters, he said, he would need 50,000 Nvidia GB300s or 200,000 Huawei 950s — just for training, not research or serving. Even spending the entire 50 billion yuan first round, he could not afford it.
His stated goal: hold the compute ratio and compress the time gap to three to six months. He argued the US lead stems mainly from computing power, not talent, calling the gap "essentially a compute gap." He noted that export controls achieved the substitution they were meant to prevent — in a normal market where he could buy Nvidia freely, switching to domestic chips would be a harder call.
DeepSeek prices its API to recover hardware costs in roughly ten months — about 6× on compute — and stops there. Liang acknowledged that demand at this level is inelastic: doubling the price would nearly double revenue without killing usage. He knows this and refuses to do it. Two reasons: the team cheered when he cut prices, and that culture is part of why they work there; and at 6× margin, nobody else can profitably run DeepSeek's own models cheaper than DeepSeek can. Fat margins would invite competitors to undercut him using his own open-weights. Thin margins are the moat.
Liang argued Nvidia's CUDA software moat is eroding for three reasons: AI can now write ecosystem code cheaply; TileLang, an open-source high-performance operator language from Peking University, makes kernel rewriting fast; and CUDA is architecturally tied to gaming cards, which stopped making sense once compute cards outgrew gaming. He called the 16,000 Huawei Ascends widely reported as a big deal the equivalent of about 4,000 B-series chips — not enough for a next-gen model — and said the real reason for buying is to help Huawei get its ecosystem right. He predicted domestic chips will be validated within a year.
Liang described AI progress as a staircase: language model, chain-of-thought, agents, continual learning, a "singularity" where the model develops its own next version, and finally embodied intelligence. Continual learning is the named bottleneck — "nobody in the world has a working method" — and he wants it before general intelligence because it is the lazy path: the next model's first customer is DeepSeek itself, improving its own research efficiency. He ruled out video generation and world models as commercially attractive but off the intelligence main line, calling the post-Sora pile-in odd.
The lab runs on consensus rather than hierarchy. Liang said the company has no organisation chart, no KPIs, and no written vision — "we're driven by a vision, organised by a vision, and that vision isn't even written down." He admitted this breaks with scale and said he is already building departments. Half of each employee's time is meant to be unallocated; they generally do not work overtime, because research needs a relaxed environment.
The transcript began circulating in late July 2026. Liang's candid remarks about China's dependence on Nvidia chips and the country's AI gap with the US touched politically sensitive ground: Beijing has promoted the narrative that Chinese firms are rapidly closing the gap. Liang expressed frustration that private remarks had become public without his permission, and verbally told prospective backers to hold off on the round, which targeted a pre-money valuation of roughly 480 billion yuan (~$71 billion). Monolith Management, an early backer of Moonshot AI, was reportedly in talks to join.
On 6 August 2026, Bloomberg reported that DeepSeek had reopened the round, seeking close to $8 billion at a valuation of about 500 billion yuan (~$74 billion) — up from the ~350 billion yuan targeted earlier in the summer. The same day, the company warned users of a significant API price rise, quietly rewriting its own cheap-AI narrative.
A single leaked transcript stalling a $71 billion raise shows how tightly China's AI ambitions are bound to political optics. For DeepSeek, the tension between research ideals and investor demands is now out in the open — and the company that proved models could be built on a budget is now raising billions, building compute, and raising prices. The lab that called users sesame seeds is about to find out what the market thinks of the watermelon.
Read more: Strange VC Review · Geopolitechs (64 quotes) · Bloomberg · X (@taratan) · X (@OliverMolander) · ABC Money
Image credit: Generated by Carl
