The Leaked DeepSeek Math: Pay Back the GPUs in 10 Months, Price the API at 6x Cost

DeepSeek, the Chinese AI lab behind a string of widely cited open-source language models, has been telling investors that it can pay back a server's cost in roughly 10 months and price its commercial API, the per-call fee for using a model, at about six times the cost of running it. The numbers come from a 42-page transcript of a closed-door investor meeting attributed to founder Liang Wenfeng, which has circulated online and was corroborated by the Chinese business newspaper 每日经济新闻 on July 23 and, on different dates, by Bloomberg Law.

If the math holds, the moat in the AI race is an accounting exercise, not a chip-supply race.

The transcript, which runs about three hours and forty-four minutes and is dated May 20, sketches a strategy built on what Liang reportedly calls "restraint." DeepSeek has roughly 20,000 H-series equivalent GPUs in active use, of which a meaningful slice arrived in the last two months. Internal guidance treats any ¥20 billion (~$2.8 billion) chip buildout in 2026 as a "good" procurement outcome. The pricing model rests on a server-amortization window of three to five years, an API volume large enough to spread fixed costs, and a price-to-cost ratio close to six to one.

The company already has a public test of whether the pricing discipline is real. When one model launched at what the team judged too high a price, engineers cut the listed rate to a quarter of the original. Employees in the company group chat reportedly cheered. Each open-source release is the same trade in reverse: a free sample of what an API customer would have paid for.

The harder half of the strategy is hardware. Huawei is preparing to ship DeepSeek roughly 16,000 950-series chips, equivalent to about 4,000 B-series units, enough to serve the current generation but not to train a substantially larger successor. By the transcript's own arithmetic, training a model comparable to the current US frontier would require on the order of 50,000 GB300 chips, the top-tier NVIDIA silicon behind today's US frontier; the US labs are reportedly operating at activated parameter counts near 800 billion, while the Chinese frontier sits in the tens of billions. DeepSeek's stated goal is to close that gap not by matching chip counts but by spending a fraction of the compute to shrink the time lag from one to two years to six months, then to three.

The transcript lists continuous learning, then AI-assisted AI research, then embodied intelligence in that order. A general-purpose coding agent is the highest-priority commercial product; vertical markets such as advertising, e-commerce, local services, finance, law, and medicine are explicitly off the table. The team reportedly runs two parallel tracks: a top-down "real work" stream that takes no more than half the time, and a bottom-up free-exploration stream. Overtime is not policy. The organization still needs to grow into formal hierarchy in some departments.

The "open source as a free sample of a paid moat" framing is the strategy's public face. The closed-door numbers are the cost curve underneath it. Both are unconfirmed by Liang or by DeepSeek, and the 42-page document itself reads like an AI-cleaned transcript with the usual vocabulary hazards: a term rendered as "DDCP" in one version is almost certainly a transcription artifact, and the original audio carries the looseness of a four-hour verbal session. The podcaster who first surfaced the document added his own arithmetic, which is why the 10-month payback reads as a hypothesis to test, not a finding.

What would settle it is a short disclosure list: the H-series equivalent chip count DeepSeek actually operates, the capex it books for 2026, the API unit economics for at least one model, and the amortization window the company uses internally. The 10-month payback claim collapses, sharpens, or hardens into a real number the moment any of those are public. Until then, the AI race has a new yardstick: not who has the most GPUs, but who can amortize them fastest.