Clean power PPAs reprice as AI load meets subsidy rollback

Clean power purchase agreements are repricing sharply higher. The Financial Times reports that PPAs favored by Google and Meta face material increases as Biden-era subsidies roll off. The timing is poor: AI inference load is scaling at exactly the moment the cost floor rises.

Hyperscalers built their GPU expansion plans on energy assumptions that no longer hold. The subsidy structure gave them a wedge between sticker and effective cost. That wedge is closing. The impact flows directly to datacenter operating expense, which flows directly to the cost per inference token. When energy cost per kWh rises, inference cost per million tokens rises in lockstep.

The providers who locked long-term PPAs at old pricing have a margin buffer. The providers who did not are now negotiating into a higher base. This is not a temporary spike. The subsidy regime has changed. The new contracts will carry the new cost for years.

The implication for model economics: the gap between hyperscaler-served inference and third-party API providers narrows. If Google and Meta face higher energy costs, their internal serving cost rises. If they pass that through to Cloud customers, AWS and Azure face the same input cost pressure. The entire stack reprices upward unless demand destruction intervenes.

This is a structural headwind. AI capex assumed cheap clean power would scale with GPU count. It is not scaling. The power is available, but the price is not what the models assumed. Watch for hyperscaler capex guidance revisions in the next two quarters. The energy line is no longer a rounding error.