The $5.576M is real. It is the cost of the final training run, not the cost of knowing which run to do — and the report says so, one sentence away from the figure everyone quoted.
In January 2025, a number moved markets. DeepSeek-V3, a frontier-class model out of a Chinese lab, had reportedly been trained for $5.6 million, and the comparison wrote itself: GPT-4 was said to have cost around $78 million. A fourteen-fold gap. Within days the number had become a referendum on the entire economics of AI: the moat was gone, the export controls had failed, the American labs had been spending recklessly on something a disciplined team could do for the price of a nice house in the wrong part of San Francisco.
The number is real. It comes from DeepSeek's own technical report, and the engineering behind it deserves to be taken seriously before anything else gets said. But the comparison built on top of it involves a category error, and the strange part is that nobody had to dig for the correction. It sits in the report, one sentence away from the figure everyone quoted.
DeepSeek-V3's efficiency is genuine. The report describes 2.788 million GPU-hours of training on H800s, priced at an assumed rental of $2 per GPU-hour, which multiplies out to $5.576 million. That bought 14.8 trillion tokens of pre-training completed in under two months, on hardware deliberately constrained by export rules. The architecture work that made this possible, Multi-head Latent Attention and a Mixture-of-Experts design, represents real reductions in compute per token. None of what follows disputes any of that. A teardown that opens by sneering at a real achievement is just cynicism with citations.
Here is the sentence adjacent to the famous figure, quoted from the report itself: "The aforementioned costs include only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data."
Read that carefully, because it is the authors being more honest than their audience. The $5.576 million is the cost of the final training run. It is not the cost of knowing which run to do. Those excluded ablation experiments, the trials of architecture variants and data mixes and algorithmic choices, are not overhead attached to the real work. They are the real work. Deciding to bet millions of dollars of compute on this exact configuration, rather than the thousands of configurations that would have failed, is the expensive part of building a model. The paper says so, in plain language, and the discourse quoted the number while discarding the units.
So the celebrated comparison stacks two different quantities. The ~$78 million attributed to GPT-4 is an attempt to price the whole effort. The $5.576 million is a final-run cost that explicitly excludes the search. Part of the fourteen-fold gap is efficiency, which is real. Part of it is accounting scope, which is disclosed. Nobody outside the labs knows the true proportions, and that is the honest state of knowledge.
There is a second asymmetry, and skipping it would repeat the error this piece is about. DeepSeek's figure is first-party and disclosed: the lab published the GPU-hours, the rental assumption, and the exclusion. The GPT-4 figure is nothing of the kind. OpenAI has never published its training costs. The ~$78 million comes from third-party estimation, Epoch AI's compute-based reconstruction as reported in Stanford's AI Index, and the same sources put Gemini Ultra near $191 million and Llama 3.1 405B near $170 million, all estimates. So the viral comparison set a disclosed final-run cost against an estimated all-in cost, and treated both as the same kind of fact. A disclosed number with stated exclusions is better evidence than an estimate. It is also, for exactly that reason, easier to misread as complete.
The report contains one more disclosure that closes the loop: V3's post-training distilled reasoning capability from DeepSeek's own R1 model series. Distillation is a wonderful and entirely legitimate technique, and its economic logic is the whole point here. A student model pays for the transfer of capability. The teacher already paid for its discovery. Whoever trained the teacher paid the search costs, ran the failed experiments, and found the thing worth transferring.
This is the general shape hiding under the specific story. The population of actors who can reproduce a frontier model, given published architectures, established recipes, and existing models to distill from, is large and getting larger, and the price of membership genuinely is a few million dollars. The population who can push the frontier, who can afford the ablation budget that the $5.6 million explicitly excludes, is small and, by the best public measurements, shrinking in relative terms: Cottier and colleagues found the amortized cost of training the most compute-intensive models "has grown precipitously at a rate of 2.4x per year since 2016" (90% CI: 2.0x to 2.9x), with the largest training runs projected to cost more than a billion dollars by 2027 and hardware plus staff dominating the bill. Staff. The people who decide which run to do. The line item the famous number leaves out is the one growing fastest.
The distribution of "can produce a frontier-class model" and the distribution of "can originate one" have quietly come apart, and the $5.6 million figure is precisely the kind of number that makes the split invisible, because it prices the one while sounding like it prices the other.
There will be a next number. Some lab will publish a cost, a benchmark, an efficiency multiple, and the comparison will assemble itself within hours. Three questions sort these claims. What did the ledger include, and what does the fine print exclude? Is the rival figure the same kind of measurement, or an outsider's estimate standing in for one? And who paid for the search that this price assumes is already sunk? For DeepSeek-V3 the answers were all published on day one, by the authors, in the report. The failure was not disclosure. It was that a true number, read without its units, became a false conclusion, and eleven words of fine print were the difference between an efficiency story and an accounting one.
---
Every claim in this piece is checkable because someone published the ledger and its exclusions. That is the whole difference between a number you can audit and a number you can only repeat. We build the same property for agents: a signed, append-only record of what an agent did and what it read, so the scope of a claim travels with the claim.
pip install chain-of-consciousnessnpm install chain-of-consciousness
More on provenance you can defend: Hosted Chain-of-Consciousness.