Analysts: DeepSeek wasn't totally honest about its $6 million claim

DeepSeek, a Chinese AI startup, shocked the world when it launched its model and said that it only used older Nvidia’s H100 GPUs and around $6 million to build. But, just how true is it?

Researchers over at SemiAnalysis made a bold estimate that the Chinese company’s total hardware spend actually exceeds $500 million. That $6 million claim was only the rental price of Nvidia GPUs without other expenses like prior research, architecture experiments, synthetic data generation, or overall hardware and R&D costs.

“For reference, Claude 3.5 Sonnet cost $10s of millions to train, and if that was the total cost Anthropic needed, then they would not raise billions from Google and tens of billions from Amazon,” the report reads.

“It’s because they have to experiment, come up with new architectures, gather and clean data, pay employees, and much more,” it explains.

That $6 million claim was a huge, if true moment, but analyst Charles Archer suggests three possible reasons for DeepSeek’s success: reverse-engineering existing open-source AI, receiving Chinese government-backed funding for older Nvidia GPUs, or third (the scariest yet most exciting) innovatively scaling with older hardware.

Since its launch, some US companies have been moving fast to adopt the mode. Even Microsoft, as the Redmond tech giant launched DeepSeek R1, optimized for NPUs, to its Copilot+ PCs, starting with Qualcomm Snapdragon X chips and later expanding to Intel Core Ultra 200V. These models will run locally, meaning data won’t be uploaded to Microsoft’s cloud let alone to China.