Deepseek's low costs may have been a trick

When Deepseek presented its language model Deepseek R1 V3, one of the details that raised eyebrows in the industry was that it had cost 6 million dollars to develop and used 2048 AI GPU chips. Now Semianalysis publishes information that this figure may not be particularly close to reality.

According to the report, the Chinese hedge fund High-Flyer was behind the initial investments in what would later become Deepseek, and they invested an estimated equivalent of 500 million dollars in 10,000 GPUs as early as 2021 to develop AI. When it was time to spin off Deepseek, these followed, and through High-Flyer, they reportedly invested a total equivalent of 1.6 billion dollars in hardware, including 50,000 NVIDIA Hopper GPUs, which gives a total number of 60,000 processors, and not 2048 as stated. They are also said to have spent up to 944 million dollars solely on operating costs.

According to Semianalysis, the alleged cost of 6 million dollars in the report refers to the GPU cost for the so-called "pre-training run," which is only part of the total cost of the model. According to the report, this can be compared to pointing to a specific part of a bill of materials for a product and claiming it is the entire cost, when in fact it is only a small part of the total cost.