Can AI Scaling Continue Through 2030?

Epoch AI asked a concrete question: can the roughly 4x-per-year growth in training compute keep going to 2030, given physical limits. They check four constraints: electric power, AI chip manufacturing, training data, and latency. Their answer is that training runs around 2x10^29 FLOP will likely be feasible by 2030, a jump over today comparable to the gap between GPT-2 and GPT-4. The supporting estimates are specific. A single data center campus at 1 to 5 GW could support 10^28 to 3x10^29 FLOP; a geographically distributed network at 2 to 45 GW pushes higher. They estimate around 100 million H100-equivalent GPUs producible for AI by 2030, with a wide range, and 400 trillion to 20 quadrillion available training tokens once multimodal data is counted. The binding constraint is power, followed by chip manufacturing, while latency is the least pressing. The framing matters: this is about what is physically possible, not what any lab will choose to spend.

Why it matters

Most later arguments about datacenters, energy, and AI capex use this as their baseline, so it is worth reading the original rather than the secondhand version. The headline for planning: power, not data or money, is the wall you hit first.