Concept · compute

Compute

The raw processing work (chips, time, and energy) consumed to train or run a model.

◷ Share view, as of June 2026: confidence, rationales and priors are hidden. The questions, stances, claims, tests and evidence stay visible.

Settled

Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.

What it is

Compute is the physical input of the AI era, usually measured in floating-point operations (FLOPs). Training a frontier model concentrates enormous compute in clusters of specialized accelerators for months; serving it consumes compute again on every request.

Much of the AI-economy debate is really a debate about compute: who can buy it, what it costs, whether the build-out of chips, power, and data centers that supplies it is justified by returns, and whether controlling its flow (export controls) can steer the trajectory.

Builds toward

inference-time-computeInference-time compute (test-time compute) · Additional computation spent during inference (e.g., generating and evaluating many candidate answers) to improve output quality without retraining.pre-trainingPre-training · The compute-heavy first phase where a model learns general capabilities by predicting tokens across vast text corpora.scaling-lawScaling law · The empirical regularity that model performance improves predictably as compute, data, and parameters grow.training-efficiencyTraining efficiency (algorithmic efficiency) · The reduction in compute required to reach a given capability level, achieved through better algorithms, architectures, or data curation rather than more hardware.