Intelai
DeepSeek's Real Number Was Never $6 Million
The training run was cheap and the weights are open. What changed is the price of a credible second option, not the price of intelligence.
Nvidia lost $589 billion in market value (opens in a new tab) in a single day, a record one-day loss. The trigger was a Chinese lab releasing a reasoning model that appeared to match the best American systems at a fraction of the price.
The number that did the work was smaller and more specific than the market reaction. DeepSeek's technical report (opens in a new tab) for its V3 base model put the pre-training run at $5.576 million.
That figure is real, and it is also the most misread number in the industry.
What the $5.6 million covers
The number is the cost of one training run, priced at market GPU rates, for a 671 billion parameter model with 37 billion parameters active per token. It is a remarkable engineering result.
It is not the cost of building DeepSeek. The report does not include the salaries of the research team, the data acquisition and cleaning, the failed experiments, the smaller models used to develop the approach, or the cluster itself. A serious lab spends years and many multiples of that figure to get to the point where one run can be that cheap.
The efficiency came from architecture and training choices, not from a coupon. Mixture-of-experts reduces the compute per token by activating a fraction of the parameters. Multi-head latent attention reduces the memory cost of the key-value cache. Those are real innovations, they are documented, and they are the part American labs can learn from.
The part that made the market react was the comparison. A frontier-class model trained for single-digit millions, from a lab outside the American capital cycle, broke the assumption that capability required a billion-dollar budget every time.
Open weights changed the deployment question
R1 shipped with open weights under a permissive license. Anyone can download it, run it, and inspect it.
For an enterprise, that changes the build-versus-buy calculation. For a defense or intelligence program, it changes something narrower and more interesting: the model can run on infrastructure you control, with no API call leaving the network.
That is a genuine shift. It does not make the model eligible for controlled unclassified information, a classified enclave, or any environment where provenance matters. Open weights are not provenance. You can inspect the weights and still not know what data shaped them, who touched them before publication, or what behavior a fine-tune might have introduced.
The supply chain question does not disappear when the download is free. It moves to the people who downloaded it.
Cost per token is the wrong metric
The pricing collapse is real. Inference from an open-weights model on your own hardware costs less per million tokens than a frontier API, and the gap is large.
Cost per token is still a bad way to choose a model. It ignores three things that dominate real deployments.
Attempts. A reasoning model that is wrong 40 percent of the time and cheap is more expensive than a model that is right 90 percent of the time and costly, once you count the retries and the human review.
Review. Every output that a person has to check has a labor cost. A cheaper model that produces more questionable output can cost more per accepted answer than an expensive one.
Integration. A model you host yourself is a model you operate. GPUs, capacity planning, upgrades, evaluation, and on-call are real costs that an API price hides.
The number to manage is cost per accepted answer: price per attempt, times attempts, plus review time, times the value of the reviewer's hour. That number is different for every workflow, and it is the only one that maps to a budget.
What a federal program should take from it
Three things, and none of them are "use DeepSeek."
The efficiency techniques are public. Mixture-of-experts, latent attention, and the training recipes are documented. A program buying American models should ask its vendor what efficiency work has been done, because the same compute now buys more capability than it did two years ago.
The model is not the deployment. Open weights remove a cloud dependency. They do not remove the need for evaluation, provenance, an update path, or an authorization boundary. A model you can download is a model you have to govern.
The strategic assumption changed. The frontier is not exclusively American, and the cost curve is not controlled by one capital market. That is a planning fact, not a policy preference. Programs that assume a single supplier with a durable price advantage should revisit the assumption.
The second option is the real story
The market reacted to a price. The durable change is that a credible second option exists, with published methods, downloadable weights, and a cost structure that a competitor can study.
That is good for buyers and uncomfortable for incumbents. It also raises the bar for anyone deploying AI where the answer has to survive an audit. When models are cheap, the model is no longer the scarce input. The evaluation, the record, and the integration are.
The $6 million headline will age badly. The methods behind it will not.