Where did OpenAI actually spend its 2024 compute budget?

Epoch AI's Josh You estimates that OpenAI's 2024 cloud compute spending split into roughly $5 billion in "R&D" (research and development) compute — covering "all training and research compute" — and around $2 billion in inference compute, and that "the majority of OpenAI's R&D compute in 2024 was likely allocated to research, experimental training runs, or training runs for unreleased models, rather than the final, primary training runs of released models like GPT-4.5, GPT-4o, and o3." The $5 billion figure is Epoch AI's own grouping of reporting that OpenAI spent "$3 billion on training compute, $1.8 billion on inference compute, and $1 billion on research compute amortized over 'multiple years'"; assuming a two-year amortization schedule, Epoch AI books about $2 billion of research compute expense to 2024, which added to the $3 billion training figure produces the $5 billion R&D total.

The piece Epoch AI can isolate — the cost of final training runs — comes from models OpenAI released or announced from the end of Q1 2024 through the end of Q1 2025: GPT-4.5, GPT-4o, the December preview of o3, and Sora Turbo. For GPT-4.5, Epoch AI estimates cloud spending directly from "cluster size, duration, and price per GPU-hour," reasoning there is more public evidence about the training cluster than about compute measured in floating-point operations (FLOP). For the other models, it estimates training compute in FLOP and converts to cost using "standardized assumptions about GPU utilization and the cost of GPU-hours," reporting confidence intervals on FLOP rather than point estimates. These final-training-run estimates explicitly exclude "experimental runs (often called 'derisking runs')," so whatever R&D spend is left over after subtracting them — the majority, in Epoch AI's reconstruction — falls to research, derisking, and training runs for models that were never released.

This cuts the same $5 billion differently than How much do AI staff costs drive frontier model training?, which splits a given training run's cost by hardware, energy, and staff rather than by purpose. It also complicates Did GPT-5 earn back its research and development costs?, which treats OpenAI's R&D spend as a lump sum GPT-5's gross profit needed to recoup: if most of that lump sum is experimentation rather than the run that produced the shipped model, the R&D figure being measured against revenue is not cleanly the cost of building GPT-5. Epoch AI notes OpenAI "made around $3.7 billion in revenue in 2024, and is widely reported to be substantially unprofitable, meaning it spent much more than $4 billion in 2024," a scale of unprofitability consistent with what Epoch AI later finds across other labs in Do AI labs spend far more than they earn?.

Epoch AI is explicit that "how exactly OpenAI distinguishes training and research compute is not known," so the $3 billion/$2 billion split inside the $5 billion is Epoch AI's grouping, not OpenAI's own accounting. The underlying spending figures trace to "documents shared to investors in September 2024 and shown to journalists," relayed through reports in The Information and The New York Times, and Epoch AI cautions they "should not be considered completely reliable" since they rely on projections for the last quarter of 2024 — even as it finds other evidence suggesting the totals are "approximately correct." The excerpt does not say what fraction of the majority is research versus derisking versus training for unreleased models, only that final training runs of released models are a minority slice of the $5 billion — which implies that a frontier lab's headline R&D spend is a weaker proxy for the cost of the models people can actually use than it might appear.