Liang Wenfeng's 3.5-Hour Investor Talk: Restraint, Open Source and the AGI Roadmap | LaoJin's Notes

On May 20, DeepSeek founder Liang Wenfeng held a 3-hour-44-minute closed-door investor meeting. Two months later, the transcript leaked online — no PR packaging, essentially his own voice: vision, open source, pricing, the AGI roadmap and the China-US compute gap, all laid bare.

DeepSeek is the model provider behind our content engine, and the most unusual specimen of China's AI wave. I've compressed 42 pages of transcript into nine themes, each with a proofread quote, plus my own take at the end. (Note: the transcript was machine-transcribed and AI-edited; treat specific numbers as approximate to the original recording.)

1. Vision as the org: how a KPI-free company holds together

Liang says DeepSeek has "no organization": no written vision statement, no KPI reviews — just "immense goodwill toward the world" holding a group of ordinary people together. Twenty years ago he idolized Jack Welch; today he keeps only one line — a company's most important asset is its vision, and vision is not a slogan on the wall: it's what you do, not what you say.

"Ordinary people achieving extraordinary things — not geniuses achieving extraordinary things."

2. Open source is conviction, not tactic

He names the contrast: Zhipu's open source feels forced; DeepSeek's is intentional. The logic is hard: AI may eventually take 10% of human GDP — something that big cannot be monopolized, and whoever tries will be discarded by history. You need a mechanism ensuring your own share stays limited, or you can't win at all.

"Those who take more will be beaten by those who take less."

3. Restraint as strategy: the 10-month-payback pricing philosophy

API pricing targets recovering hardware cost in ten months — roughly 6x profit. Demand is inelastic in this price band: doubling the price would barely change token consumption and would nearly double revenue, yet they don't raise it. When the model's price was cut, the company chat erupted in cheers. After last Spring Festival's explosion of users, they neither monetized nor fought for them — they just tried to serve them well.

"There's a watermelon ahead; everything before it is just sesame seeds."

4. The AGI roadmap: language models → CoT → agents → continual learning → self-iteration → embodied AI

Step by step: last year was CoT, this year is agents; the next bottleneck is "continual learning" — letting a model learn over long horizons the way a new hire spends two months learning the company; beyond that lies the self-iterating "singularity" (he stresses it's gradual, not a leap); embodied intelligence comes last.

"With this roadmap, we don't have to work overtime."

5. The single core interest: team stability

Money isn't the problem, resources aren't the problem — the one thing that cannot be conceded is team stability: if nobody leaves, setbacks are survivable; everything else only costs you six months or a year. He admits it's also the biggest risk, largely defused by the latest funding round granting core staff meaningful options.

"As long as I can keep the team stable, we will definitely reach AGI."

6. The China-US gap is now just compute: 20,000 H-equivalents

The gap, he says, is resources alone: DeepSeek runs about 20,000 H-equivalent GPUs, producing frontier-class models with one-twentieth the compute and a one-to-two-year lag; the talent gap is nearly zero — "it's the same pool of people," some stayed in China, some went abroad. The largest models activate ~800B parameters versus tens of B in China — an order of magnitude apart, so they master their own scale first. Scaling is far from topping out; what stops them is compute, not laws.

"When Silicon Valley says scaling is over, that's over for Silicon Valley. For us in China, we're nowhere near that wall."

7. CUDA's moat is crumbling: the historic window for domestic compute

Three reasons: AI can write code, making ecosystems far cheaper to build; their self-built high-level language TileLang can rewrite the entire CUDA operator stack — V3 was already trained on NVIDIA cards without NVIDIA's ecosystem; and compute cards now outsell gaming cards, so dedicated chips naturally decouple from CUDA. Huawei's 950 supernode: four cards match one GB300, two years behind in time, but interchangeable in tasks and near-parity in price. His call: within a year, the ecosystem question for domestic chips will be settled by facts — what remains is capacity alone.

"On this front, NVIDIA is digging its own grave."

8. The endgame gap is only three things

When model capabilities converge, the final gap comes down to cost, time, and user experience. Stickiness is real but not fundamental — what matters is the time lead of shipping first, and the cost moat of equal quality.

"Cost is the number-one differentiator."

9. Half researchers, half data labelers

The most surprising passage: labeling high-end data in China carries no cost advantage over the US — it's so expensive they walk on two legs, labeling the cheap data first. And right now, "half the company is labeling data" — including half of the core researchers, the most important people.

"At this stage, what solves AI is data labeling."

Laojin's take

Honestly, after finishing this transcript, I sat at my desk for a long time.

Among China's new generation of AI founders, very few hold this kind of conviction and discipline — few even among AI founders worldwide. Liang Wenfeng represents the idealism of China's new entrepreneurial generation, and the true spirit of open source. I dare say this because I'm a beneficiary.

The air in China's AI startup scene these years is restless: chasing C-end users, chasing ARR, chasing narratives — a pitch deck without an ARR curve barely gets a meeting. Liang went the other way: no monetization at peak virality, company-wide cheers for price cuts, "ten-month payback is enough." This counter-commercial restraint became the deepest moat — you can't copy it, because it's not a tactic; it's a belief. Tactics can be cloned. Beliefs can't.

We ourselves benefit from that open-source spirit. This site's content engine runs on DeepSeek every day: two bilingual articles daily, the frontier daily and weekly — all on its API. Ten-month-payback pricing means small teams like ours can afford a frontier model. That's what "trading margin for ecosystem" looks like when it lands on us: he gives up profit; we build businesses on top. We sell AI adoption — and part of its foundation is the margin he chose to leave behind.

So stop treating idealism as a tax on sentiment. In a market as big as AI, idealism is the most rational long-term strategy: you can't take it all anyway — restraint isn't taking less, it's making sure you never leave the table.

Three lessons for ordinary founders

First, restraint is scarcer than ambition. Those who want everything lose to those who want one thing. Last year everyone fought bloody for the C-end — and the prize was taken by the one who didn't fight. That's the most vivid line in his talk.

Second, vision is the cheapest organizational power. A small team without KPIs or an option pool can still hold together — if you truly believe, and truly act on it. Vision isn't a slogan on the wall; it's what you do, not what you say.

Third, doing lower-tier work from a higher technical ground is dimensionality reduction. Ask first: "Is this on the main line of intelligence?" On the main line, your byproducts may beat others' main business. Off it, the hottest business is just a business — not a moat.

---

_Source: a transcript of Liang Wenfeng's investor talk circulating online (recorded 2026-05-20, ~3h44m, surfaced July 2026); transcription errors and specific figures defer to the original recording. Further reading: Reusable skills for your AI agents · LaoJin's Picks: other people's verified good work · Frontier: weekly AI digest_