DeepSeek's Liang Wenfeng, Unfiltered: Inside the Leaked Four-Hour Investor Meeting

A rare four-hour session behind closed doors

In July 2026, a roughly 34,000-word document began circulating in investment circles: a transcript of a closed-door investor meeting held by DeepSeek founder Liang Wenfeng on May 20, 2026, transcribed on July 16. The session ran nearly four hours and covered 118 questions. Major Chinese outlets — including Tencent Technology and Yicai (China Business Network) — have reported on the transcript, though DeepSeek has not officially confirmed its authenticity. Read with that caveat, one word captures the whole conversation: restraint.

A financing that broke the mold

The meeting took place around DeepSeek's first external funding round. According to reporting based on corporate filings, the round totaled about 51 billion yuan, and the post-money valuation landed in the range of roughly 325–351 billion yuan.

The round was invite-only, opened to a small number of institutions under strict terms. One investor who joined the talks told Yicai they ultimately passed, citing limited disclosure on commercialization and exit paths.

"Restraint": the core strategy

If one word summarizes Liang's stance, it is restraint. He argued the company deliberately avoided pushing consumer products, even when a Lunar New Year surge brought a flood of users. He framed near-term consumer traffic and B2B revenue as "sesame seeds" next to the "watermelon" of AGI.

On pricing, the internal rule is simple: recover hardware costs in about ten months, and that is a reasonable profit. He recalled cutting a model's price to a quarter of the original — a decision the team cheered.

The "won't-do" list

Liang was explicit that anything off the AGI mainline gets cut:

Multimodality, he said, is treated as a component rather than intelligence itself — V4 and later versions support native multimodality, but it does not define the ceiling of capability.

The AGI roadmap

Liang described AGI as a staircase, each step unlocking a core capability:

"A model shouldn't rely forever on one-off training," he said. "Like a person, it should keep learning at work — and then it can help design the next generation of AI."

Compute is the real gap

Asked about the gap with the United States, Liang was blunt: "Our only difference from the US is resources" — above all, compute. He still believes in scaling ("we haven't touched the ceiling") and sees a historic opening for domestic chips as the CUDA moat erodes. On building DeepSeek's own chips he was non-committal: he would rather buy chips at a fair price than make them, and expects the next year to prove domestic chips are viable.

Open source and revenue don't conflict

Liang insisted open source and commercialization are compatible. Even with an open model, matching DeepSeek's deployment cost is hard. And because the AI market is enormous — potentially 10% of GDP — no single company can monopolize it. "You must build a mechanism that caps the benefit you take," he said. "Only then are you more likely to actually pull it off."

Toward an IPO

Separately, multiple outlets (including Bloomberg and The Wall Street Journal) report that DeepSeek is preparing a mainland listing on Shanghai's STAR Market, aiming to file by the end of 2026 for a 2027 debut, with a rumored valuation of up to about 71 billion US dollars. In product terms, DeepSeek's V4 line is documented at up to a one-million-token context window on a trillion-parameter MoE architecture.

"A group of ordinary people"

Liang's management philosophy is unusual for a high-efficiency tech firm. He admires Jack Welch's view that a large company runs on vision, not rules or KPIs. Formal, top-down tasks should take under 50% of an employee's time, leaving the rest for free exploration. The company's one non-negotiable interest, he said, is keeping the core team stable.

"We really are just a group of ordinary people," he said. "The story I like is ordinary people doing something extraordinary — not geniuses doing it." He also predicted that base models will converge to a handful of players once the assumed high margins prove unrealistic.

The bottom line

Even in the worst case, Liang said, "we'll go all-in on selling API and doing the service well — and I think that's enough." Restraint, open source, low prices, an AGI mainline, domestic compute, continual learning — presented less as tactics than as conviction.

CATAI note: figures here come from a widely circulated but officially unconfirmed transcript and related media reporting. Treat them as reported, not settled fact.