What was the training compute of the largest training run completed by any developer between 2025-08-01 and 2026-08-13?

While raw historical trends suggest a mid-2026 frontier run might exceed 2e27 FLOP, public databases lack any estimates for closed models after July 2025 epoch.ai. The true scale of the largest 2025-2026 training runs splits into three distinct paths. In a steady scaling scenario (45% probability), runs land near 1e27 FLOP, a clear step up from GPT-4.5. A compute-dampening regime (35%)—where capacity routes toward post-training and inference 2 sources, driving modest downstream capability leaps —keeps the largest runs in the 4e26 to 7e26 FLOP range. Conversely, a 20% chance of recursive acceleration envisions massive, unannounced gigawatt-class runs exceeding 1.5e27 FLOP, highly correlated with surging AI R&D uplift . The mixture yields a median of 9.5e26 FLOP, constrained by energy limits but reflecting continued sub-exponential growth. Reflecting the divergence between constrained scaling and massive unannounced compute runs, the distribution captures wide extremes: the lower tail (a p10 of 3.9e+26 FLOP and p25 of 5.5e+26 FLOP) accounts for the possibility that no single run significantly outsized GPT-4.5, while the upper percentiles (a p75 of 1.5e+27 FLOP and p90 of 3e+27 FLOP) cover scenarios of vast unannounced scale-outs.