Build AI in America: Anthropic Energy Report

Source: Anthropic Date: 2025-07-21 URL: https://www.anthropic.com/news/build-ai-in-america

Summary

Anthropic published a policy report arguing US AI leadership requires massive domestic energy infrastructure investment. Key figures: Anthropic alone needs 2GW by 2027 and 5GW by 2028 for single models; total US frontier AI training needs 20-25GW by 2028; training + inference combined minimum 50GW by 2028 (~2x New York City peak demand). China context: 400+ GW expansion vs. several dozen GW in US. Policy recommendations: open federal lands for AI infrastructure, accelerate NEPA reviews, fast-track geothermal/natural gas/nuclear permitting, establish transmission corridors.

Implications

• Capital / compute infrastructure / energy dependency made public. Publishing the 5GW single-model training figure for 2028 is an extraordinary disclosure — it makes Anthropic’s compute ambitions (and energy requirements) concrete and public. The figure also contextualizes the Amazon $100B+ compute commitment and the Google 1GW TPU partnership.
• 50GW by 2028 = national infrastructure problem. Anthropic framing the combined training+inference need as 50GW positions AI energy as a national infrastructure challenge equivalent to major cities, which requires federal land and regulatory changes — the exact asks the report makes.
• China’s 400GW as the competitive frame. The China comparison (400+ GW vs. several dozen US GW) is the geopolitical urgency argument. It frames energy policy as AI competitiveness policy, which is the message that lands with both Republican (national security) and Democratic (industrial policy) audiences.
• Open federal lands. The specific ask to bypass local zoning via federal land access is politically controversial but operationally important — data center permitting in established geographies is multi-year. Federal land access is the speed mechanism.
• Watch: how the 2GW (2027) and 5GW (2028) targets tracked against actual Anthropic deployments; whether the federal land access recommendations were adopted; how the 50GW forecast compared to actual AI energy consumption by 2028.