The Six Million Dollar Lie

The number that ate Nvidia's market cap — $6 million, cost to train DeepSeek R1, proof that America has been wasting money, proof that the GPU buildout is a fraud, proof of seventeen different things depending on which podcast you were listening to Monday morning — is, in a technical sense, accurate.

It's also the kind of accurate that gets people killed.

$6M is the cost of one training run. Compute time, job queued, model trained, bill settled. What it does not include is the $1.3 billion DeepSeek spent acquiring the GPUs that ran the job. The capex. The hardware. The actual physical objects that made the $6M training run possible.

This is like a restaurant claiming a meal costs three dollars because that's what they paid for the ingredients, having spent forty million on the kitchen.

More people are arriving at this conclusion, which is nice, though arriving four days after the panic doesn't really help the investors who sold Nvidia at the open on Monday. The number moved markets. The correction is a blog post.

The frustrating part isn't that DeepSeek gamed the framing — they didn't, exactly, the $6M training cost is a real number for a real thing — it's that everyone just accepted it without asking what wasn't in it. A $6M training run sounds like a company you could fund with a Kickstarter. A $1.3B GPU fleet sounds like, well, exactly what you'd expect a well-funded lab operating at the frontier to need.

The story isn't "AI got cheap." The story is "AI got cheaper per run while the capital requirements stayed enormous." Those are completely different stories. One of them was worth several hundred billion dollars in market movement. The other one is, more or less, fine.

Still sorting out exactly how misleading the full picture is. But "misleading" seems like the right word.