$AEP

The 2030 Compute Landscape: GPUs, Power and the Architecture of Intelligence

The constraint on AI is no longer ideas or data but compute and the power to run it. Frontier training compute has grown about five times a year since 2020, and the leading AI supercomputer of June 2030 is on course for about 2 million chips and 9 gigawatts. Shayne Heffernan sets out the 2030 arithmetic, the power problem, who has already contracted the gigawatts, and how GPUs, custom ASICs, wafer-scale engines and quantum machines divide the work.