The Power Grid Reckoning

The power problem comes in three layers, each with different time horizons.

Layer 1: generation supply. Data centers globally will consume an estimated 1,000 TWh in 2026 — equal to Japan's entire national electricity consumption — per the IEA. By late 2027, US AI data centers alone will need 20–30 GW of additional power. PJM (the grid operator serving Virginia, the largest US data center market) projects a 6 GW capacity shortfall by 2027. PJM's capacity market clearing price for the 2026–2027 delivery year jumped from $28.92/MW to $329.17/MW — a tenfold increase.

Layer 2: transmission and distribution infrastructure. Power transformer lead times have stretched from 40 weeks pre-pandemic to 2.5–3 years today. Wood Mackenzie models a 30% deficit in power transformers and a 10% deficit in distribution transformers for 2025. Switchgear, busbars, and medium-voltage gear are also on multi-year lead times. In Northern Virginia, utility connection wait times exceed 3–5 years for large-scale deployments. Nearly half of planned US data center projects are delayed or cancelled specifically due to this equipment shortage.

Layer 3: the response. Hyperscalers are executing three parallel bypasses. First, on-site generation — primarily natural gas turbines — deployable in 12–18 months versus 5-year grid waits. Second, battery energy storage systems (BESS), growing at 28–38% CAGR with a 20 GW hyperscaler opportunity through 2035; BESS enables peak shaving and grid arbitrage without adding generation. Third, nuclear PPAs for the long-tail: Microsoft's $16B Three Mile Island restart (835 MW, 2028), Amazon's $20B X-energy SMR deal (5 GW by 2039), Google/Kairos Power fleet (500 MW by 2035), Meta's 6.6 GW nuclear commitment. Constellation Energy has emerged as the dominant counterparty for existing nuclear PPAs; Oklo, NuScale, and BWX Technologies lead SMR development.

Imagine you're a hyperscaler CEO. You've signed contracts to deploy a gigawatt of AI compute by 2027. You go to your local power company and ask for a grid connection. They tell you: "Sure. Six years." You ask why so long. They tell you: "Because the transformer you need used to take us 40 weeks to order. Now it takes 3 years. Same for the switchgear, the substations, the high-voltage cables. Everything on our side of the meter is backordered."

That is the power grid crisis in one sentence. The problem is not that there isn't enough electricity being generated in theory. The problem is that the physical equipment that moves electricity from power plants to racks — transformers, switchgear, circuit breakers — has lead times that stretched from months to years. The entire AI buildout is being paced by this shortage. Roughly half of all planned US data centers are being delayed or cancelled specifically because operators cannot get the electrical gear they need.

So what are the hyperscalers doing? Three things at once, on three different timelines.

Short term, over the next 12 to 24 months, they're building their own power plants on-site. Mostly natural gas turbines. Not glamorous, but you can deploy one in about a year and it bypasses the grid entirely. This is why turbine manufacturers like GE Vernova are also sold out.

Medium term, over 3 to 5 years, they're installing grid-scale batteries. These don't generate power, but they let you store cheap off-peak electricity and discharge during the day, so you can run heavy workloads on a smaller grid connection. Battery energy storage for data centers is growing at 28–38% a year — one of the fastest-growing segments in the entire infrastructure stack.

Long term, over 5 to 15 years, they're going nuclear. Every major hyperscaler signed a nuclear deal in the past 18 months. Microsoft is restarting Three Mile Island. Amazon bought into a 5 GW small modular reactor program. Google is backing Kairos Power. Meta signed a 6.6 GW commitment. These plants won't come online until 2028 at the earliest, and mostly after 2030. But they represent a permanent answer to the power problem — clean, firm, 24/7, and independent of grid queues.

The winners are spread across all three layers — equipment makers who can deliver transformers faster than competitors, battery systems integrators, turbine makers, nuclear operators, and uranium fuel suppliers. Every big AI data center announcement is secretly an announcement about who is selling them electrons.