Compute Forecast (AI 2027) (ai-2027.com)

Romeo Dean’s April 2025 Compute Forecast projects a tenfold increase in global AI‑relevant compute to ~100 million H100-equivalents (H100e) by December 2027, driven by combined improvements in chip efficiency (~1.35x/year) and chip production (~1.65x/year) for an overall 2.25x/year growth. The report defines AI‑relevant compute by Total Processing Performance (TPP) relative to the NVIDIA H100 (~15,800 TPP) and models supply-side bottlenecks (advanced packaging and HBM) while concluding wafer capacity is unlikely to constrain growth through 2027. Compute will concentrate: the top AI firms (OpenAI, Anthropic, xAI and AGI groups in Google/Meta) are forecast to expand their share from ~5–10% today to ~15–20% of the global stock (~15–20M H100e), giving the leading company ~40x more compute by 2027 (a ~3.4x/year compound increase). Technically, the forecast anticipates frontier GPUs (Rubin/R200) reaching ~6e15 FP16 FLOP raw performance but only ~2.5x net efficiency after die‑size adjustments, continued 40% model FLOP utilization assumptions, and a shift in compute use away from external pretraining toward research automation, synthetic data generation, and internal experiments. Inference assumptions imply that, with algorithmic efficiency gains, a leading firm could deploy ~1M copies of high‑speed research AIs (50× human speed) using ~6% of its compute on specialized inference chips. Projected impacts include steep rises in datacenter spending (to ~$1T by 2027) and power demand (leading firm ~10 GW; AI total ~60 GW globally), though later sections are explicitly conditional on rapid capability progress and carry greater uncertainty.