Teravolt looks to cannibalize older industries to meet AI power demand

AI datacenter power demand is projected to rise from 104 GW in 2025 to 132 GW in 2026 and 290 GW by 2030, while Morgan Stanley estimates a potential 49 GW US supply shortfall by 2028. Teravolt forecasts AI demand of 410 GW by 2036 against only 170 GW of available grid capacity, implying a 240 GW deficit and a material constraint on planned AI buildouts. The company expects developers to repurpose thermal plants and industrial sites, shift workloads to energy-abundant regions, and potentially prolong coal and gas generation as AI compute economics outbid traditional industrial power users.

Analysis

The investable scarcity is not aggregate electricity generation but deliverable, firm power with transmission rights and an executable interconnection path. That favors CEG, VST and TLN near existing load centers, while PWR, GEV and ETN monetize the multi-year grid build irrespective of which AI platform ultimately wins. The second-order risk for AI infrastructure suppliers is that GPU clusters without secured power become delayed revenue rather than canceled demand; this would shift the bottleneck from chip supply to deployment schedules over the next 1-3 quarters.

Claims that AI-compute economics justify displacing industrial load should be treated as promotional until utilization, contract tenor and customer credit quality are disclosed. Higher wholesale power prices and curtailment risk would pressure energy-intensive operators such as AA and CF, but only in merchant or weakly hedged regions; regulated-rate structures can socialize much of the cost instead. The more likely near-term outcome is a widening regional power-price and capacity-value spread, not a uniform national power shortage.

Consensus is increasingly long merchant generation on the AI narrative, creating material valuation and regulatory risk. A reversal would come from hyperscaler capex guidance cuts, lower realized data-center utilization, adverse FERC/state treatment of dedicated-load contracts, or a fall in forward capacity prices; these matter within months, whereas transmission and generation earnings accrue over 6-18 months. MS has modest upside through financing, restructuring and private-capital activity, but AI-related debt issuance is not by itself a material earnings driver absent sustained transaction volumes and credit spreads.

Key Decisions for Investors

• Prefer a 6-18 month basket long PWR, GEV and ETN over a pure merchant-power bet: these names capture grid-capex backlog with less exposure to hourly power prices. Add on post-earnings guidance confirmation; cut if backlog conversion or utility capital-spending guidance weakens materially.
• Use a relative-value trade: long CEG or VST versus short an equal-dollar broad technology proxy (XLK) for 3-6 months only where forward power contracts and capacity pricing continue to tighten. The thesis fails if data-center load commitments slip or capacity-price auctions/forward power curves decline.
• Place AA and CF on a regional power-cost watchlist rather than shorting broadly. Initiate only if disclosed power hedges roll off while local wholesale prices rise; aluminum and nitrogen prices can otherwise offset energy-margin pressure.
• Do not add directional exposure to MS solely on AI debt issuance. Reassess after quarterly investment-banking fees, leveraged-finance pipeline and private-credit marks demonstrate that financing activity is translating into durable earnings rather than balance-sheet leverage elsewhere in the ecosystem.