DeepSeek raised $7.4 billion on the cheapest models in AI. Read the footnotes on both numbers.

DeepSeek closed its first-ever outside funding round this week: roughly 50 billion yuan, about $7.4 billion, at a post-money valuation reported north of $50 billion. The number got repeated everywhere within hours, usually next to the observation that DeepSeek gives its model weights away for free. A lab that open-sources its product just raised the kind of money you raise when investors think the product is worth keeping closed. That is interesting. It is also exactly the kind of round number I'm paid to distrust, so let's read the footnotes — on the valuation, on the benchmarks underneath it, and on one clause almost nobody quoted.

Compared to what? The valuation in context

$50 billion is a large number until you put it next to the companies DeepSeek is supposedly competing with. Anthropic was last valued around $965 billion. OpenAI is reported around $852 billion. DeepSeek, the lab that allegedly humbled both, is being valued at roughly a twentieth of either. So before we decide whether $50 billion is too high, notice that the market is also telling you DeepSeek is not, in fact, in the same weight class — it is priced as a serious contender, not as a peer.

Why the gap, if the models are competitive? Because a valuation is not a score on a benchmark. It is a bet on future cash flows, and DeepSeek's whole strategy — open weights, near-zero pricing — is designed to suppress exactly the cash flows a $900 billion valuation assumes. You cannot give the model away and also book OpenAI's revenue. The $50 billion is the market pricing that contradiction: real technology, deliberately uncommercial business model, large strategic value to its backers for reasons that may have nothing to do with a profit-and-loss statement. Hold that last clause. We'll come back to it.

The benchmark claim, and the seam in it

The case for DeepSeek as a frontier lab rests on its V4 model family, released earlier this spring under an MIT licence — genuinely open, weights downloadable, which I will credit loudly because it is rare and it means you can actually check the homework. The reported numbers are striking. The flagship is said to score in the low 80s on SWE-bench Verified, which would make it the strongest open-weights entry and put it level with a top proprietary model; to top a competitive-coding leaderboard; and to clear a Codeforces rating above a leading closed model's. If those hold, they are a real achievement.

"If those hold" is doing the work. Every one of those figures is vendor-reported, and the first rule of reading a vendor's benchmark is to ask three questions: which test, scored how, and was the test in the training data. On a model whose weights are public you can in principle answer all three — but the launch-day numbers came from the company, not from an independent re-run, and the variant that posts the best score is not always the variant you download. When a result is reported as coming from a 'Pro-Max' or a 'high-effort' configuration, that is a footnote, not a headline: the number describes a setting, not the model most people will actually run.

On a model whose weights are public you can check the homework. The launch-day number is the company's. Those are not the same thing.

And contamination is not an accusation; it is gravity. Coding and reasoning benchmarks leak into training corpora constantly, because the corpus is the internet and the internet discusses the benchmarks. A score in the 90s on a public coding test tells you the model is very good at the kind of problem that test contains. Whether that generalises to the kind of problem you have is a separate question, and it is the question the leaderboard is structurally bad at answering. Significant is not the same as impressive, and impressive is not the same as useful to you.

"A fraction of the price" — a fraction of which price?

The other half of the DeepSeek legend is cost. The list prices are genuinely low — fractions of a dollar per million tokens, against figures that have been quoted as tens of times cheaper than a flagship Western model on input and far more than that on output. The eye-watering multiples ('86 times cheaper') are the ones that travel. So, compared to what?

Compared to a specific competitor's list price, on a specific date, on the metered API. That is a real comparison and it is also a narrow one. It does not include the cost of running the open weights yourself, which is the option DeepSeek's open-source story is built on — and self-hosting a 1.6-trillion-parameter mixture-of-experts model is not free, it is a cluster of high-end accelerators and the power to feed them, which is precisely the bill the rest of this industry is straining under. 'The weights are free' and 'serving the weights is cheap' are different sentences. The first is true. The second depends entirely on your hardware, your utilisation, and whether you have the engineers to keep a model that size fed.

There is also a quieter caveat in the pricing history itself: the lowest figures became the standard list price after a steep discount was made permanent. A price set by strategy rather than by cost is informative — it tells you the seller wants market share more than margin — but it is not a measurement of how cheap the model is to run. It is a measurement of how badly someone wants you using it. Which, again, is a question about motive, not capability.

The footnote that should have been the headline

Here is the clause almost nobody quoted, and it is the one I'd put in the headline. According to the reporting on the round's structure, the outside commercial investors — the large strategic names attached to the deal — received no voting rights and accepted a five-year lock-up, with their capital routed through a limited partnership controlled by the founder. The single investor reported to have received direct equity with voting rights, and no lock-up, was a Chinese state-backed industry fund.

Read that the way you'd read a methodology section. $7.4 billion came in. The commercial money bought economic exposure and explicitly did not buy control. The governance — the actual say over what this lab does — was concentrated, and the one party handed a vote was the state. If that reporting is accurate, then the most important number in the deal is not the $7.4 billion or the $50 billion. It is the distribution of voting rights, because that is the line that tells you what the money is for. A lab that gives its weights away and routes control to a state fund is not optimising for the income statement. It is optimising for something else, and the cap table is where it admits so.

This reframes the valuation question entirely. I said earlier that $50 billion looks low for a frontier lab and high for one that refuses to monetise. Both are true under a commercial lens. Under the lens the cap table actually describes — strategic capability, distributed free to maximise reach and adoption, with control held close — the valuation isn't really pricing a business at all. It's pricing a position. The open weights are not a failure to monetise; they are the distribution strategy. Free is the point.

What the numbers actually support

So here is the verdict, with the doubt quantified rather than waved away. DeepSeek's models are, on the public evidence, genuinely strong and genuinely open, and that combination is rare enough to take seriously — the open weights are a real gift to anyone who wants to check the field's work instead of trusting a slide. The benchmark supremacy is plausible but vendor-reported, configuration-dependent, and exactly the kind of claim that shrinks under an independent, contamination-controlled re-run; treat the rankings as a hypothesis, not a result. The 'fraction of the price' is real on the metered API and much softer once you price the hardware to self-host. And the $50 billion valuation is best understood not as a verdict on the technology but as a price on a strategic position whose control, per the reporting, now includes a state vote.

None of that makes DeepSeek a paper tiger. It makes it a story where the impressive numbers are real and the meaningful number is somewhere else — in the footnote about who gets to vote. The chart that went up this week was the valuation. The line I'd look harder at is the one on the term sheet.