What will AI look like in 2030?- study commissioned by DeepMind (epoch.ai)

DeepMind‑commissioned research forecasts that if current scaling trends continue to 2030, frontier AI will require unprecedented compute, capital and power—but remain technically and economically viable. The report projects training clusters costing on the order of $100B+, single training runs around 10^29 FLOP (thousands of times more compute than GPT‑4), and gigawatts of electricity. The authors examine common slowdown arguments—data exhaustion, power limits, cost, algorithmic progress, and shifts to inference—and conclude none are decisive: public and synthetic data, distributed datacenters, and rapid power build‑out (solar/batteries/off‑grid) make continued scaling plausible if revenues and returns justify the investments. Technically, extrapolated benchmark trends suggest transformative capabilities for scientific R&D by 2030: translating natural language into complex scientific software, helping mathematicians formalize proofs, answering open‑ended biology protocol questions, and improving weather forecasts. These tools could yield ~10–20% productivity gains in many tasks (with large uncertainty and variable domain timelines), though deployment and real‑world impact—especially in regulated fields like drug approval—may lag capability. The report underscores both the upside for science and the urgency for policymakers and practitioners to prepare for large‑scale compute, energy, and governance challenges as AI becomes central to R&D and the broader economy.