Liang Wenfeng's Arithmetic, and Why Restraint Is a Strategy

In mid-July 2026 a 42-page PDF began circulating among investors: "Liang Wenfeng Investor Meeting, Recording Transcript." Liang Wenfeng is the founder of DeepSeek, the Chinese model lab. The recording is dated May 20 and runs about 3 hours 44 minutes. It was auto-transcribed by speech recognition and tidied by AI, with speakers not distinguished, and compiled on July 16. Around July 23 it spread widely through media retellings. One detail at the end of the transcript is telling: as the meeting closed, the host reminded investors "never to leak any numbers or details, including the GPU counts, and never to screen-record and share it." The transcript itself exists because that request was broken. First, what kind of material this is. It is not an official release, and the transcript itself says "some proper nouns and numbers may be misrecognized, refer to the original audio." Everything quoted here should be read as highly credible but unconfirmed circulating material. All quotes below are translated from Chinese. One correction first. The best-known figures in secondhand retellings, a raise of more than 50 billion, a pre-money valuation of 367.5 billion, and 20 billion of Liang's own money, do not appear in the full text. The "20 billion" that does appear is this year's GPU purchasing budget ("If I can spend 20 billion this year, our procurement department will have done a superb job"), and "50 billion" is the compute assumption in "even if I spent all of it I couldn't train the largest model." Both were mistaken for fundraising figures. The transcript confirms only that the financing closed, that "the option amounts everyone received are fairly large," and that team attrition, the biggest risk, has been removed. The transcript does not state a currency for these figures. The ten points below are picked from the full text, keeping the original wording as far as possible. The one interest he won't give up: team stability He narrows his "core interest" to a single item: "Our biggest core interest is keeping the team stable, and you could even say it's the only core interest." Money isn't a problem and neither are resources. Any other setback will delay the company by half a year or a year, "but won't mean we can't do it." On that account, the first job of this round is very likely issuing options and not buying GPUs, which removes the one risk. One to two years behind, on one-twentieth of the compute The original is a string of ranges: "12 months behind the US, maybe 12 to 18 months, or 6 to 12 months, anyway in short about two years behind," but doing it "with only one-twentieth of America's compute." The next version of the story he wants to rewrite compresses the gap to 6 months, then 3. With an order-of-magnitude gap in total compute, overtaking across the board is unrealistic, but overtaking "in a few focused places where we choose to trade off" may be possible. Every gap is a compute gap