Did it really cost $6m to ‘make’ DeepSeek?published at 17:51 Greenwich Mean Time 28 January
Tom Gerken
Technology reporter
That widely-reported figure is at the heart of what has caused such consternation in the stock market – the idea that this impressive AI toolkit was developed for little compared to its big budget US-based rivals.
The true figure is much harder to say for sure.
Firstly, it all comes from the first table in DeepSeek’s publicly available report , externalon how the tool works. It estimates the total cost of training DeepSeek V3 at just under $5.6m (£4.5m).
Some people have challenged this figure, but if we take it at face value, that would still only cover the cost of training this version of its AI – DeepSeek-V3 came as a result of various previous iterations.
Its researchers state in the report that the $5.6m figure excludes the costs of its previous research in the space (which has been ongoing since 2023), as well as the other experiments they tried before landing on this methodology.
But more broadly, the number also ignores all the other day to day costs that add up, from renting or buying an office space to paying for staff – Chinese media say DeepSeek’s team comprises of just under 140 people, who all presumably earn a salary.
And that doesn’t consider the cost of actually deploying DeepSeek into the sleek, easy-to-use app that has exploded in popularity on app stores in the UK and US.
So while researchers say it cost less than $6m to train this particular version of DeepSeek, the true cost will have been much higher.
