All training FLOP estimates are taken from Epoch's AI models database. γ is the single training run's compute expressed as a share of what the firm's whole fleet could produce in a year: γ = training FLOP ÷ (fleet in H100e × 9.89 × 10¹⁴ FLOP/s × MFU × seconds per year). MFU figures when unknown are anchored to 20% as the most recent high-quality disclosed figure from Microsoft's MAI paper.