Graphics processing unit

Graphics processing units are, on ToTheMoon, less pieces of hardware than the yardstick of the whole AI race. The hosts reach for the term when they discuss NVIDIA's valuation, data-center construction, export controls or a country's chances, and the underlying question is nearly always the same: who owns the compute and whether it pays for itself.

In 2024 the hosts looked at GPUs mainly through economics. Episode 002 noted that GPU sales in generative AI far exceeded the revenue earned with them, and that companies competed by announcing how many cards they had bought. Episode 012 added the Sequoia arithmetic about six hundred billion dollars and the observation that every new GPU generation sharply devalues the previous one. Episode 013 explained why a GPU remains a general-purpose device, and episode 028 described the xAI cluster of a hundred thousand cards as an industrial project of power, cooling, networking and people.

By 2025 the conversation had moved to the geopolitics of compute. In episode 035 the hosts argued that the raw GPU gap had stopped being decisive because new methods were leveling it, and that nobody knew what China had. After DeepSeek, episode 043 questioned whether training on a few thousand cards could be taken at face value, concluding that a more efficient model widens demand for compute instead of killing it. Episodes 058 and 074 treated GPUs as a matter of sovereignty, with the US opening NVIDIA to the world and China shifting to Huawei accelerators, and episode extra09 named the lack of modern chips as a reason a local model cannot close the gap.