How Can We Meet AI's Insatiable Demand for Compute Power? (www.bain.com)

A new technology report highlights a looming supply-side crisis: AI’s compute demand has grown at more than twice the historical rate of Moore’s law, and Bain’s modeling suggests global AI compute needs could reach ~200 gigawatts by 2030 (roughly 100 GW in the U.S.). Meeting that demand would require roughly $500 billion of annual data‑center capex—which Bain argues implies about $2 trillion in annual revenue to sustain—and would stress grids that have seen flat load growth for decades. Executives face a binary risk: overbuild and carry stranded capacity, or underinvest and miss the next wave of AI adoption and market share. For the AI/ML community the implications are practical and strategic. Short-term relief can come from algorithmic efficiency (Transformers, mixed‑precision math, distillation, chain‑of‑thought prompting, and newer designs like DeepSeek) and specialized silicon (training/inference ASICs vs. general‑purpose GPUs). Long-term shifts could come from disruptive hardware (quantum is likely 10–15+ years away for large‑scale workloads) or faster power/ construction buildout—both constrained by supply chains for GPUs, switchgear, cooling and multi‑year timelines to add generation/transmission. The takeaway: technical work on model and system efficiency, co‑design of HW+algorithms, and policy/market coordination on power and capex will determine whether AI growth is sustainable or becomes concentrated among players with privileged access to compute and capital.