In January 2025, one number went around the world: DeepSeek trained a frontier-class AI model for less than $6 million. Compared with the hundreds of millions that GPT-4-class models were rumored to cost, it sounded like a company in China had just rewritten the economics of artificial intelligence overnight — and markets reacted as if it had.
The number is real. It comes straight out of DeepSeek’s own technical report. But it answers a much narrower question than the headlines implied, and DeepSeek said so in the very next sentence — a sentence almost nobody quoted. So did DeepSeek really cost $6 million? The honest answer is: that figure is real, it’s genuinely impressive, and it is not what most people think it is. Let’s do the math.
Where the $6 Million Number Comes From
DeepSeek-V3 is a large open-weights language model — 671 billion parameters, of which only about 37 billion activate per token thanks to its Mixture-of-Experts design. Its makers published a detailed technical report (arXiv 2412.19437), and buried in it is a training-cost table. Here’s what it actually says:
| Training stage | H800 GPU-hours |
|---|---|
| Pre-training | 2,664,000 |
| Context extension | 119,000 |
| Post-training (fine-tuning + RL) | 5,000 |
| Total | 2,788,000 |
The model trained on a cluster of 2,048 Nvidia H800 GPUs. DeepSeek then multiplied those ~2.79 million GPU-hours by an assumed rental price of $2 per GPU-hour and arrived at roughly $5.576 million. Round it up in a headline and you get “under $6 million.”
So the figure is not made up. But look closely at what it measures: it’s an estimate of the rental-market value of the compute burned by one successful training run. That’s it. It’s not an invoice, not an accounting statement, and not a claim about what DeepSeek itself actually spent.
The One Sentence Everyone Skipped
Immediately after that cost table, DeepSeek’s own report adds a caveat. Paraphrased only lightly, it says: these costs include only the official training of DeepSeek-V3, and exclude the costs of prior research and ablation experiments on architectures, algorithms, and data.
Read that again, because it’s the whole ballgame. DeepSeek is telling you, in its own paper, that the $5.6 million covers the final run — not the road that led there. Here’s what that final-run number leaves out:
- Every failed and exploratory run. Frontier models are the survivor of hundreds of experiments. You don’t get one clean 2.79-million-GPU-hour run without burning many times that on things that didn’t work.
- The research itself — the architecture ablations, the hyperparameter searches, the data experiments.
- Data curation — collecting, cleaning, filtering, and deduplicating trillions of tokens is enormously expensive and labor-intensive.
- Salaries for a team of very well-paid researchers and engineers.
- Electricity, cooling, and datacenter overhead beyond the rental proxy.
- The GPU cluster itself. This is the big one. $2/GPU-hour is a rental comparison price. DeepSeek didn’t rent thousands of GPUs by the hour — its parent company owns them. Owning a fleet of Hopper-class accelerators is a capital expense in the hundreds of millions of dollars, and that cost is completely invisible in a “GPU-hours × rental rate” calculation.
The “$5.6 million” is best read the way you’d read “the gas for the final lap” — accurate, but not the cost of building the car, hiring the driver, or running the race team.
The Real Bill: What Analysts Estimate
The most-cited counter-analysis came from SemiAnalysis (“DeepSeek Debates,” January 31, 2025). Their reporting says plainly that the $6 million figure “is attributed to just the GPU cost of the pre-training run, which is only a portion of the total cost,” and excludes R&D and the total cost of ownership of the hardware.
Their estimates for the fuller picture:
- DeepSeek’s parent, the quant hedge fund High-Flyer, has accumulated access to roughly 50,000 Hopper-class GPUs (a mix of H800s, H100s, and Nvidia’s further-cut-down H20).
- Total server capital expenditure: ~$1.6 billion.
- Annual operating costs: ~$944 million.
In other words, the infrastructure behind the “$6 million model” is plausibly a billion-dollar-plus operation. (You may also see a viral “$1.3 billion / 216× higher” figure floating around — treat that one with caution; it’s a secondary aggregation, not SemiAnalysis’s own stated number. The defensible figures to cite are the ~$1.6B capex and ~$944M opex.)
None of this means DeepSeek lied. It reported exactly what it said it reported: the compute cost of the final run. The distortion happened in the retelling, when that narrow number got stripped of its own caveat and recast as “DeepSeek beat OpenAI for the price of a house.”
So What Is the Impressive Part?
Here’s the twist: once you strip away the misleading headline, DeepSeek’s actual achievement is still remarkable — just in a way that’s harder to fit on a chyron.
First, that 2.79-million-GPU-hour figure is genuinely low for a model of this quality. DeepSeek squeezed it down with real engineering: the sparse Mixture-of-Experts architecture (activate a fraction of the model per token), a memory-saving attention trick called Multi-Head Latent Attention, and — most striking — training the bulk of the model in FP8, an 8-bit number format that roughly halves memory traffic and doubles throughput versus the usual 16-bit. Efficient training was the whole design goal, and they hit it.
Second — and this is the part the cost debate often buries — DeepSeek-V3 was trained on H800s, Nvidia’s deliberately hobbled, export-restricted chip built to comply with US rules on selling to China. The H800 has cut-down chip-to-chip bandwidth versus the H100. That a frontier-competitive model came out of nerfed hardware is arguably a bigger story than the dollar figure. The real headline isn’t “$6 million” — it’s “world-class results under a hardware handicap.”
How to Actually Measure What a Model Costs
This is a Measure question at heart, and DeepSeek is a perfect teaching case for reading AI cost claims critically. Whenever you see a training-cost number, ask three things:
- Marginal run or total program? Is this the compute for one final run, or the full cost of R&D + data + failed runs + people + hardware? These differ by one to two orders of magnitude. DeepSeek’s own numbers show a ~$5.6M run sitting on top of a ~$1.6B infrastructure base.
- Rented or owned compute? “GPU-hours × hourly rate” is a rental proxy. It’s a fair way to compare runs, but it hides the capital cost of owning the cluster — which is where most of the real money lives.
- What’s the comparison actually comparing? “DeepSeek cost $6M, GPT-4 cost $100M+” pits a final-run number against total-program rumors. That’s apples to orchards.
Get those three straight and the DeepSeek story clicks into place: a genuinely efficient training run, reported honestly with its own caveat, that got flattened into a myth on the way to the front page.
(One note: this is a compute-economics explainer, not investment or trading advice. The market moves that followed DeepSeek’s release — including the record one-day drop in Nvidia’s value — are their own story.)
❓ Frequently Asked Questions
Did DeepSeek really train a model for $6 million? Sort of. The ~$5.576 million figure is real and comes from DeepSeek’s own report, but it’s the cost of the final training run only — about 2.79 million H800 GPU-hours at an assumed $2/hour. DeepSeek itself states the number excludes prior research, failed experiments, and data work, and it doesn’t include the cost of the GPU cluster, salaries, or datacenter overhead at all.
What’s the real total cost then? Nobody has a precise public figure, but the analyst firm SemiAnalysis estimated DeepSeek/High-Flyer’s server capital expenditure at roughly $1.6 billion, with annual operating costs near $944 million, across a fleet of about 50,000 Hopper-class GPUs. So the infrastructure behind the model is plausibly a billion-dollar-plus operation.
Was the $6 million figure a lie or marketing? No. DeepSeek reported exactly what it said — the compute cost of the final run — and included the caveat in its own paper. The misleading part happened in the retelling, when the narrow number was quoted without its context and compared against total-program cost rumors for other labs.
If it wasn’t really $6 million, is DeepSeek still impressive? Yes. The training run genuinely used far fewer GPU-hours than comparable models, thanks to Mixture-of-Experts, latent attention, and FP8 training — and it was done on export-restricted H800 chips. The efficiency story is real; only the “cheaper than a house” framing was overblown.
How can I sanity-check other AI cost claims? Ask whether the number is a single final-run cost or the full program cost, whether the compute was rented or owned, and whether it’s being compared against a like-for-like number. Most viral AI cost claims fail at least one of those tests.
📚 Keep Learning
- Learn AI and ML Concepts Through the Lens of DeepSeek — the full tour of what makes DeepSeek efficient.
- What Is Mixture of Experts? and What Is FP8 Training? — the two efficiency tricks behind the low GPU-hour count.
- The Day DeepSeek Broke Nvidia — how this cost story triggered a record market shock.
📚 Recommended Resources
Disclosure: some links below are Amazon affiliate links. If you buy through them, I earn a small commission — at no extra cost to you — which helps keep this site running.
Artificial Intelligence: A Guide for Thinking Humans
Melanie Mitchell · Book · ~$28 on Amazon
The perfect companion to a myth-busting piece like this one. Mitchell teaches you how to read AI claims critically — separating what a system actually does from what the headlines say it does. No math required, and it will make every future “AI breakthrough” story easier to weigh.
Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow
Aurélien Géron · Book · ~$70 on Amazon
Once you understand why training is expensive, this is the book that shows you what actually goes into it. It builds from the fundamentals to modern deep learning with runnable code — the single best structured reference for seeing where the compute (and the cost) really goes.
Deep Learning with Python
François Chollet · Book · ~$50 on Amazon
The clearest gentle introduction to building neural networks in code, from the creator of Keras. If this article made you curious about the machinery behind a training run — parameters, precision, and all — this is the friendliest place to start getting hands-on.