The $5.5 Million Lie is the Best Part

The number everyone is passing around — $5.5 million to train R1, a fraction of what American labs spend, proof that the moat is gone, sell your Nvidia — is almost certainly not the real number.

The speculation, and it is speculation, goes like this: Microsoft is investigating whether a group linked to DeepSeek improperly obtained OpenAI data. The specific fear is that they didn't just train on o1 outputs — distillation, which everyone does — but that they got access to the chain-of-thought reasoning itself, the scratchpad, the thing OpenAI specifically doesn't publish. Meanwhile, the Scale AI CEO, and then Elon, are pointing at a different problem: DeepSeek may have had 50,000 H100s that they couldn't disclose, because China officially isn't supposed to have any H100s. Do the math on that at cloud rates and you're not at $5.5 million anymore. You're at something closer to $1.5 billion.

None of this is confirmed. All of it is plausible.

The thing is, the $5.5 million figure isn't even the interesting story if it's true. The interesting story is what it means if it's false — specifically, if someone seeded that number knowing exactly what the financial press would do with it. Nvidia lost something like $600 billion in market cap in a single day. If you knew the narrative was coming and you shorted the right things before it landed, the model almost pays for itself.

That might be giving someone too much credit. Or exactly the right amount.

The reproducibility question is the real tell. DeepSeek published a paper. Papers are supposed to be reproducible. Hugging Face is already trying with open-r1. There are plenty of companies with $5 million sitting around who are also quietly running this experiment right now, guaranteed, because if the paper holds up, that changes everything, and if it doesn't, that also changes everything. The answer is probably six weeks away.

I believe the model is excellent — use it, it's excellent. I believe a lot of the architectural work is real. What I don't believe is the shoestring. The shoestring is a story. Stories spread faster than corrections, and someone may have known that.

The most charitable read is that $5.5 million was the marginal training run cost, not the total compute bill, and the press ran with it because it was a better headline than "Chinese lab spent a lot of money but spent it smarter." That's probably true. The least charitable read involves a short position and a very good PR team.

Both can be true. That's the thing about narratives — they don't require coordination to become weapons. You just have to point them in the right direction and let the clicks do the work.