Did DeepSeek really cost 6 million to train?
Sort of. DeepSeek's paper reported about $5.6 million for the final training run of its V3 model — 2.788 million GPU-hours at an assumed $2/hour. But that figure covers only the last compute run. It excludes earlier research, failed experiments, data, salaries, and the hundreds of millions spent on GPUs, so DeepSeek's total cost was far higher.
Why — the first-principles explanation
The '$6 million' number is real but narrow. DeepSeek published that its V3 model's final training run used about 2.788 million hours on H800 GPUs, and at an assumed rental price of $2 per GPU-hour, that multiplies out to roughly $5.6 million. That's an honest figure for exactly what it measures: the electricity-and-compute cost of the one successful run that produced the model.
The problem is what it leaves out. Building a frontier model is like the last take of a movie — the final run is cheap compared to everything before it. DeepSeek's number excludes the months of research, the dozens of failed and ablation experiments, the cost of collecting and cleaning data, engineer salaries, and — most importantly — the hardware itself. DeepSeek's parent hedge fund had already bought GPU clusters worth hundreds of millions of dollars; the $6M figure treats those as free by only counting rental-equivalent hours for the final run.
So the accurate statement is: the final training compute cost about $6 million, which is genuinely and impressively low. But the total cost of creating DeepSeek's models — infrastructure, research, staff, and prior experiments included — runs into the hundreds of millions. Both things are true; headlines just collapsed them into one misleading number.
An example that makes it click
Imagine a chef says her prize-winning cake 'cost $20 to make.' She's counting the flour, eggs, and sugar in the final cake — and that's true. But she isn't counting the $200,000 kitchen she already owned, the fifty practice cakes she threw away, or her years of culinary training.
The $20 is real, and it's genuinely impressive that the final cake was so cheap. But if you tried to open a bakery believing cakes 'cost $20,' you'd go broke. DeepSeek's $6 million is that $20: the ingredients for the final bake, not the whole kitchen and all the practice.
Key facts
- DeepSeek reported ~$5.576 million for the final training run of DeepSeek-V3 (about $6 million).
- This equals roughly 2.788 million H800 GPU-hours at an assumed $2 per GPU-hour.
- The figure covers only the final run — it excludes prior research, failed experiments, data, and salaries.
- It also excludes the cost of the GPUs themselves; DeepSeek's parent had already spent hundreds of millions on hardware.
- The low final-run cost was driven by efficiency techniques like Mixture-of-Experts and FP8 training.
An open-weight Chinese model family that matched frontier quality at low cost.
Affiliate link — we may earn a commission at no cost to you.▶ The 60-second explainer (script)
Did DeepSeek really cost just six million dollars to train? Kind of — but the headline is misleading. DeepSeek's own paper said the final training run of its V3 model used about 2.8 million GPU-hours, which at two dollars an hour comes to roughly five point six million dollars. That number is real, and it's genuinely impressive. But it only counts the one successful final run. It leaves out all the earlier research, the failed experiments, the data collection, the engineers' salaries, and — the biggest one — the GPUs themselves, which cost hundreds of millions and were already owned by DeepSeek's parent hedge fund. So the final bake cost six million; the whole kitchen cost far more. Both are true — just don't confuse the ingredients with the total.
What authoritative sources say
People also ask
So is the $6 million figure a lie?
No, it's accurate for the final training run's compute; it just doesn't represent the total cost of building the model.
What's the real total cost?
Including hardware, research, failed runs, and staff, estimates run into the hundreds of millions, though DeepSeek hasn't published a full figure.
Why did DeepSeek report such a low number?
To highlight its training efficiency; the $2-per-GPU-hour final-run cost genuinely showcases how little compute it needed.
How was the final run so cheap?
Efficiency techniques like Mixture-of-Experts (activating few parameters per token) and FP8 precision training cut the compute needed.