How fast is AI scaling?
The single biggest driver of AI progress has been raw scale — the amount of computation used to train each new model. Plotted on a log scale, that growth is staggering and relentless. Explore the models that mark the curve, and see just how quickly the frontier is moving.
Compute isn't the whole story — algorithms and data matter too — but it has been the most reliable predictor of new capabilities. When the input to a system grows by a factor of a hundred million in a little over a decade, and no one fully understands what the next hundred-fold will unlock, “we'll figure out safety as we go” becomes a very large bet.
The trend also reframes the race: efficient models like China's DeepSeek-V3 reached frontier-adjacent performance with a fraction of the reported compute, a reminder that capability is getting cheaper, not just bigger.
Training-compute figures are estimates, primarily from Epoch AI's database of notable models, expressed in floating-point operations (FLOP). Labs rarely disclose exact numbers, so recent frontier values are approximate and shown for illustration; the robust finding is the trend, not any single point.
The doubling time shown is computed from the plotted models, so it reflects this curated sample rather than the entire field.
Capability is outrunning our ability to control it.
Every doubling of compute buys capabilities we didn't design and can't fully predict — while our tools for aligning and governing these systems improve far more slowly. Closing that gap is the whole project. Get our briefings on the risk and what can be done.