AI and Data Center Power Demand: The New Driver of Energy Investment

Introduction

The explosive growth of AI data center power demand is reshaping energy markets more rapidly than any development since the US shale revolution. After two decades of essentially flat US electricity demand, data center construction for AI training and inference is producing the first sustained, structural increase in power consumption since the early 2000s. In the EIA's October 2026 Short-Term Energy Outlook, US consumption rises from a record 4,195 billion kWh in 2025 to new records of about 4,288 billion kWh in 2026 and 4,356 billion kWh in 2027. Data centers used about 4.4% of US electricity in 2023, and the Lawrence Berkeley National Laboratory study for the Department of Energy projects 6.7% to 12% by 2028. The power section explains why firm generation is being repriced; this article tracks the market evidence: the capex, the demand data, PJM's capacity auctions and the resulting deals.

The capital behind the demand keeps being revised upward. After second-quarter 2026 earnings, the four largest US hyperscalers guide to more than $700 billion of combined 2026 capital expenditure, roughly $720 billion to $745 billion when the company ranges are summed. Amazon raised its 2026 estimate to $220 billion in July after Alphabet lifted its range, and Meta narrowed its outlook to $130-145 billion, so the hyperscaler capex guidance keeps rising (Microsoft's lower figure reflects accounting, not less building).

These are company-wide figures for data centers, servers, chips and networking (the funding side is covered in how the AI boom is financed), so only part of the money reaches energy assets. The energy flow-through is still large: every campus needs grid connection and firm supply, and Goldman Sachs Research estimates about $720 billion of grid spending may be needed through 2030.

The demand now shows up in federal data, not just forecasts. In the October outlook, commercial-sector electricity sales (the category that includes data centers) reach about 1,549 billion kWh in 2026, edging past residential sales of about 1,541 billion kWh. When the EIA first projected the crossover in May it called it the first time on record and expected it in 2027; the October numbers put it in 2026, with commercial demand growing again in 2027 while residential stays flat. Prices are following: the EIA expects wholesale power prices in PJM to rise 41% in 2026, the largest increase of any region it tracks.

PJM Interconnection, the grid operator for 13 states and Washington, DC, including Northern Virginia's data center cluster, is where data center load meets a market price. Its December 2025 auction for the 2027/2028 delivery year cleared at the FERC-approved cap of $333.44/MW-day, came in 6,623 MW short of the reliability requirement (the first time the whole RTO fell short) and produced a record $16.4 billion of cleared value. PJM's forecast peak load had risen about 5,250 MW from the prior auction, and nearly 5,100 MW of that increase was attributable to data center demand.

The July 2026 auction for 2028/2029 repeated the pattern: price at the cap of $325/MW-day, about $16.4 billion of cleared value again, a 6,831 MW shortfall, and only 525 MW of new generation and uprates cleared. PJM now plans a backstop procurement and "connect and manage" rules that let large new loads such as data centers connect but operate flexibly when the system is tight.

Base Residual Auction (BRA)

PJM's main capacity auction, normally held about three years ahead of each delivery year, in which generators and demand response commit to be available at times of system stress in exchange for a capacity payment per MW-day. The capacity procured is measured against a reliability requirement set to limit expected outages to one event in 10 years; clearing below it is a shortfall. Since the 2026/2027 auction, a price cap and floor agreed with the PJM states has limited how far prices can move.

For bankers, the results matter twice. They are the clearest public proof of a supply shortfall: the collar caps what generators earn, but the missing megawatts show that dispatchable capacity in PJM is scarce. They also set revenue assumptions in every merchant power and capacity market valuation in the region, which is why the 2026 gas plant deals cluster in PJM.

Hyperscaler procurement has moved from buying renewable energy certificates to long-term contracts that underwrite real assets and set offtake benchmarks for whole asset classes. The headline nuclear deals: Microsoft's 20-year PPA underpins Constellation's roughly $1.6 billion restart of the 835 MW Crane Clean Energy Center (formerly Three Mile Island Unit 1), backed by a $1 billion DOE loan that closed in November 2025 and targeting 2027. Meta added agreements in January 2026 with Vistra, TerraPower and Oklo for up to 6.6 GW of existing and new nuclear capacity, on top of its 20-year Clinton deal with Constellation. Google signed a 500 MW small modular reactor agreement with Kairos Power and backs NextEra's restart of the Duane Arnold plant in Iowa.

Firm Power vs. Intermittent Power

Firm power (dispatchable or baseload) is generation that can run 24 hours a day regardless of weather, chiefly gas combined-cycle plants and nuclear reactors; intermittent power (solar, wind) produces only when the sun shines or the wind blows. Because data centers run continuously, existing firm plants are being repriced as scarce infrastructure in the power sector M&A supercycle.

Natural gas is doing the heavy lifting in the near term because existing plants can serve load now, while turbine backlogs have stretched new-build timelines and costs. ExxonMobil's proposed 1.5 GW plant for data center supply and behind-the-meter gas projects such as the one serving Meta's Prometheus campus in Ohio show the direct route, but the bigger market response has been M&A for operating fleets.

Renewables still make up much of the contracted volume, from Google's $4 billion Arkansas data center paired with new Entergy solar to Microsoft's fusion offtake with Helion. They diversify the mix but cannot replace firm generation for 24/7 load.

The clearest market response to the PJM shortfall has been M&A for operating gas plants and the platforms that own them. Three 2026 transactions set the reference points (the deal-by-deal detail is in Landmark Energy Deals of 2026):

The repricing of dispatchable generation has driven the largest power M&A wave since deregulation in the late 1990s. Constellation/Calpine and NRG/LS Power both closed in January 2026, NextEra agreed in May to combine with Dominion Energy, whose Virginia territory hosts the world's largest data center cluster, and Constellation's antitrust divestitures to LS Power have created a second round of transactions. Mid-market gas portfolios and solar-plus-storage deals add steady flow for banks with strong power practices (Goldman Sachs, Morgan Stanley, Lazard, Guggenheim).

Valuation has moved with it. Gas and nuclear assets once valued at 5-6x EBITDA on commodity-exposed cash flows now trade at 8-12x on scarcity premiums and long-term contracts, although private fleet sales still clear lower (Talen's 6.6x), leaving multiple arbitrage for listed acquirers. Beyond power, the same thesis generates mandates in three adjacent areas: