Grok 4's training run is estimated at about $490 million and 310 million kWh. These figures show the cost and energy scale of frontier training that shapes future model availability and pricing.
Key takeaways · AI-distilled
Epoch estimates Grok 4 used 246 million H100-hours, largely from vague public xAI statements and assumptions about scaling from Grok 3. It flags significant uncertainty in this estimate, which underpins its resource calculations.
Epoch estimates 310 GWh of training electricity using H100 power draw, additional non-GPU hardware demand, cooling overhead and average draw below peak. This excludes compute for experiments and serving users.
Epoch's rental-rate and hardware-depreciation methods each estimate a median training cost of $490 million. The estimates include hardware and power, but exclude human labor, experiments and serving the model.
Epoch's water analysis estimates 754 million liters including natural-gas power generation and data-center use; its table and CSV round this to 750 million. The introduction instead claims over 1.2 billion liters for cooling, an unresolved inconsistency.
