Liang Wenfeng's Leaked DeepSeek Briefing: AGI Plan 2026

DeepSeek founder Liang Wenfeng held a 3-hour-44-minute closed-door briefing for his company's first outside investors on May 20, 2026, walking them through an AGI-first strategy, a 20,000-GPU fleet, and a secret chip development program. When a recording and transcript of the meeting leaked across Chinese social media in late July, the disclosure went viral, and Wenfeng told prospective backers to pause a second fundraising round that had been targeting a pre-money valuation of approximately 480 billion yuan, equivalent to about $71 billion, Fortune and The Japan Times reported.

A First External Round and a Rare Public Accounting

The May 20 meeting was DeepSeek's first formal debrief for outside investors. The company had never taken external capital before closing roughly 50 billion yuan, or about $7.4 billion, in its maiden financing round at a post-money valuation of approximately $52 billion, with Tencent, CATL, JD.com, and NetEase among the backers, Hello China Tech reported. Wenfeng personally contributed around 20 billion yuan, or about $3 billion, of that total, funding a significant share of the round himself.

The structure of the raise reflects his stated philosophy: disciplined capital deployment, no rush to outside money, and full founder control retained as long as possible. From its founding in 2023 through early 2026, DeepSeek operated entirely on High-Flyer capital while releasing models that set international benchmarks.

StartupHub.ai data shows DeepSeek AI operating with 162 employees at the time of the funding close, implying roughly $46 million raised per employee. Among the Chinese AI companies we track, that ratio is exceptional, pointing to an operational intensity that matches the lean-team approach Wenfeng described to investors.

AGI Over Revenue: What Wenfeng Told Shareholders

Wenfeng was direct about DeepSeek's hierarchy of goals: reaching artificial general intelligence, not building the next super app or maximizing near-term revenue. He told investors the company would not pursue 3D generation, video generation, or world modeling, framing these as distractions from the core AGI path. On the question of open-source releases, which have eroded potential licensing revenue, he offered a rationale that went beyond altruism. Per a transcript summary published by Recode China AI, he told investors: "Restraint is a strategy used to increase the probability of achieving AGI." By publishing model weights, DeepSeek creates competitive pressure on its own models and forces faster internal iteration.

The daily brief

Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.

Join 12,007 subscribers.

On pricing, Wenfeng set out a concrete rule: DeepSeek targets a ten-month payback on compute hardware, implying roughly a sixfold markup on raw GPU cost. That discipline generates enough margin to fund research without the scale of a commercial sales organisation. The South China Morning Post noted that a margin analysis consistent with this rule was published by DeepSeek itself in early 2025, which showed a 545% theoretical inference margin at prevailing prices.

The session also included a briefing on DeepSeek's compute stack: a 20,000-GPU fleet, and, more notably, a confirmation that the company had been quietly developing its own inference chips for a year. The disclosure put DeepSeek alongside xAI and Google as labs building custom silicon rather than relying entirely on commercially available Nvidia hardware. Wenfeng framed the chip effort as an insurance policy against a tightening export-control environment rather than a wholesale replacement of Nvidia GPUs, according to Hello China Tech's summary of the briefing.

The Leak and a Frozen Second Round

The meeting was intended to remain confidential. When the transcript spread virally in late July, Wenfeng's response was swift. According to Fortune, he verbally informed prospective backers to hold off on the second financing round. The Seoul Economic Daily reported that DeepSeek abruptly halted the process on July 26. The paused round had been targeting a pre-money valuation of approximately 480 billion yuan, or about $71 billion, a significant step up from the $52 billion post-money figure for the first round.

The portion of the leak that appeared most sensitive, based on The Next Web's reporting, was Wenfeng's candid acknowledgment of Nvidia dependence. Confirming how exposed DeepSeek remains to a chipmaker whose hardware is subject to US export controls was a strategically awkward admission for a company seeking a higher valuation from new investors. The secret chip program addresses that exposure but is still in early development.

DeepSeek has not confirmed the authenticity of the leaked transcript. The meeting summary circulating online covers 118 numbered items across culture, open-source strategy, compute, pricing, and AGI roadmap, according to the Momentum Works transcript archive.

What It Means

Wenfeng's playbook is internally consistent: keep the team small, price for payback rather than dominance, publish model weights to accelerate the internal research cycle, and treat the current product suite as a byproduct of the AGI path rather than the destination. What the leak has done is force that philosophy into a public accounting before DeepSeek was ready to offer one. A valuation negotiation with new investors is harder once a founder has acknowledged that current products are incidental and that chip supply depends on a US-controlled vendor. The pause in fundraising is, at minimum, a reset. Whether it becomes a longer halt depends on how Wenfeng addresses the compute exposure question he inadvertently made public.

Sources

Editorial standards: every claim is sourced. Tips: [email protected]

DeepSeek's Leaked Meeting: Liang Wenfeng's AGI Strategy, Secret Chips