2012 → 2026 · training compute, log scale
AlexNet was trained on two consumer graphics cards in a bedroom-scale setup. A frontier model today occupies tens of thousands of accelerators for months. That gap is roughly 100,000,000× more arithmetic — but raw arithmetic was never the whole story. Each jump below removed a specific obstacle that was, at the time, the reason more compute wouldn't have helped.
Compute figures are order-of-magnitude estimates; for closed models they are outside reconstructions, not disclosures. Hover or tap a point to name it, click to jump to the entry.