"Can AI Scaling Continue Through 2030?" is a 2025 analysis by Epoch AI, a companion to its piece on data-center power demands. It asks whether continued training-compute growth of roughly 4× per year is physically possible through 2030, and which of four inputs — power, chips, data, and latency — would bind first.
Summary of argument
Epoch frames the question around four physical inputs to frontier training and assesses how close each comes to a hard limit by 2030 if compute keeps growing at about 4× per year. Its conclusion is that scaling to 2030 at historical rates is physically possible but not guaranteed: it would require simultaneous stretch on every input, with no major geopolitical disruption of Taiwan or ASML.
Key claims
Power is identified as the single tightest constraint. Under continued scaling, frontier runs reach 1–5 GW by 2028 and 4–16 GW by 2030, and US data-center power rises from about 30 GW in 2024 to about 90 GW by 2030. Meeting that demand would require a 3× acceleration of historical US grid capacity additions.
Chips are treated as achievable but contingent on sustained capex growth of 30–40% per year from NVIDIA, TSMC, and memory vendors. Epoch identifies CoWoS-L advanced packaging and HBM3e/HBM4 memory as the tightest links in the supply chain.
Data is described as the scientific wildcard. Epoch projects that high-quality web text is exhausted around 2026–2028, leaving synthetic data and multimodal and reinforcement-learning rollouts as the three main substitutes.
Latency — the sequential-dependency limits on training wall-clock time — is assessed as not binding before 2030 under current architectures.
The analysis builds on the scaling-law framework (Scaling Laws) and shifts the continuity question from architecture toward physical inputs such as grid interconnect, packaging yield, and data exhaustion, locating the proximate chokepoints upstream of chip export controls. It quantifies the energy-bottleneck question raised in Compute Governance and identifies power and HBM/packaging as the proximate policy targets.
Caveats
Epoch notes that the 4×-per-year growth assumption is itself debatable: some models are already shifting toward test-time compute, where the scaling curve differs. The single-site power assumptions may also relax if geographically distributed training works.
Provenance
Epoch AI. "Can AI Scaling Continue Through 2030?" epoch.ai/blog/can-ai-scaling-continue-through-2030, 2025.
Relationships
- supports: Epoch AI — How Much Power Will Frontier AI Training Demand in 2030? — same numerical backbone on power.
- supports: RAND — AI's Power Requirements Under Exponential Growth (2025) — agrees on order of magnitude and direction of travel.
- depends-on: Scaling Laws — builds on the scaling-law framework.
- supports: Compute Governance — identifies power and HBM/packaging as the proximate policy targets.
- contradicts: (partial) bullish takes on unconstrained scaling in AI 2027 and Compressed 21st Century — not on the model-capability trajectory, but on the ease of the physical buildout required.