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

Training compute is growing five times a year and the power behind it is doubling. What the arithmetic says about 2030, who has already contracted the gigawatts, and which silicon will carry the load. Artificial intelligence has entered a phase in which the primary constraint on capability is no longer ideas or data, but raw computational power and the energy required to deliver it. Graphics processing units (GPUs), once tools for rendering, have become the dominant substrate for training and serving the models that now shape research, industry and public life. Their performance has scaled faster than classical Moore's Law, yet the trajectory is bending under physical and economic pressure. By 2030 the leading systems are projected to require multi-gigawatt power loads, while specialised alternatives (custom ASICs, wafer-scale engines and early quantum processors) carve out distinct roles. Compute, in the sense used here, is counted in floating-point operations, or FLOP: the total a training run performs, and the rate per second a chip or a cluster sustains. Power is counted in gigawatts, the load a site draws from the grid or from its own generation. The 2030 compute landscape is where those two numbers meet. This article examines the measured…