AI Needs Power. Here's Where the $600 Billion Is Going.

AI + ENERGY · Intelligence Brief · Rookonomy

The Short Version

Everyone's talking about AI models. Nobody's talking about the thing that actually makes them run: electricity. Lots of it. The Big Five tech companies are spending over $600 billion on infrastructure in 2026 alone and a chunk of that is going into nuclear plants, gas turbines, and transmission lines. This is the defining capital reallocation of the decade, and most people haven't clocked it yet.

The Numbers

The Thing Nobody Told You About AI

Here's what the AI hype cycle skipped over.

ChatGPT, Gemini, Claude, all of it, where none of it runs on vibes. It runs on electricity. Massive, constant, uninterruptible electricity. A single frontier model training run can take weeks and cost tens of millions of dollars. You cannot pause it. You cannot have a blackout. The power has to be on, 24/7, no exceptions.

Now multiply that by every hyperscaler building out AI infrastructure simultaneously. Amazon, Microsoft, Google, Meta, Oracle all racing, all building, all consuming power at a scale that the global grid genuinely was not designed for.

Global data centres used 415 terawatt-hours of electricity in 2024. That's 1.5% of all electricity consumed on earth. By 2030, that number nearly doubles to 945 TWh. And the part that's growing fastest? AI-specific workloads, up 30% annually, compared to 9% for regular cloud computing. AI servers need more power, more cooling, and more reliability than anything the tech industry has ever built before.

$600 Billion. Where Is It Actually Going?

The Big Five are projected to spend over $600 billion on infrastructure in 2026. That's a 36% jump from 2025, which was itself up 73% from 2024. The numbers are accelerating, not plateauing.

Here's the breakdown of where the $450 billion AI-specific slice goes:

Each of the four biggest players now individually exceeds $100 billion in annual infrastructure spend. Amazon led 2025 at $125B. Microsoft at $95B. Google at $91B. Meta at $66–72B.

And here's the plot twist: their capex now exceeds their internal cash generation. The Big Five raised $108 billion in new debt in 2025 alone, 3.4 times their historical average. Projections suggest the tech sector may need to issue $1.5 trillion in new debt to finance all of this. This isn't just a tech story. It's a capital markets story.

The Nuclear Comeback Nobody Saw Coming

This is the wildest part of the whole thing.

Nuclear power was essentially written off in the West, too expensive, too slow, too politically toxic after Chernobyl and Fukushima. Plants were being decommissioned. The economics didn't work.

Then AI showed up and changed the math entirely.

Microsoft × Constellation Energy: In 2024, Microsoft signed a 20-year power purchase agreement to restart Three Mile Island Unit 1 in Pennsylvania, a plant that had been shut down in 2019 because it wasn't economical. It's being renamed the Crane Clean Energy Center. 835 MW of carbon-free power, online by 2028, projected to add $16 billion to Pennsylvania's GDP.

Amazon × Talen Energy: A PPA for 1,920 MW of nuclear power from the Susquehanna plant, running through 2042. Amazon is also investing $20 billion in Pennsylvania, the largest private-sector investment in the state's history and exploring new Small Modular Reactors (SMRs) with Talen.

Google × Kairos Power: The world's first corporate deal to purchase power from multiple SMRs. Up to 500 MW of carbon-free power, with the first SMR targeting 2030.

Meta × Constellation Energy: A 20-year deal for 1,121 MW from the Clinton Clean Energy Center in Illinois. Meta also launched an RFP seeking 1 to 4 GW of new nuclear capacity and received over 50 qualified submissions across 20+ U.S. states.

Amazon × X-energy: $500 million invested into an SMR developer, targeting 5 GW of new nuclear by 2039.

The pattern is clear. Hyperscalers need firm, 24/7 power. Wind and solar are intermittent. Nuclear isn't. So nuclear, the technology that was being retired, is now being rescued by the biggest companies in the world, at premium prices, on multi-decade contracts.

But It's Not Just Nuclear

Nuclear gets the headlines. The reality is messier.

The data-centre buildout needs roughly 15–20 GW of new generation capacity. That's equivalent to building around 15 nuclear power plants, plus utility-scale solar and wind on top. The supply chain is buckling under the demand.

Transformer lead times: Up 80% year-over-year. You cannot connect a data centre to the grid without transformers. They're backordered.

Liquid cooling orders: Up 200%. As AI servers push rack power densities from 15–20 kW to 100+ kW, air cooling doesn't cut it anymore. Every new hyperscale facility needs liquid cooling. There aren't enough systems.

Grid connection queues: PJM Interconnection, which covers the massive data-centre corridor in northern Virginia has queues so backed up that some developers are bypassing the grid entirely and building on-site gas generation just to move faster. Yes, tech companies with net-zero pledges are building gas plants. The operational urgency of AI is winning over the sustainability commitments, at least in the short term.

Transmission: The U.S. needs hundreds of billions in new transmission investment to connect renewable-rich regions to data-centre hubs. Permitting alone takes over a decade. So instead, developers are paying premium prices for brownfield sites, old industrial land or retired power plants that already have grid interconnection. Those assets are getting acquired fast.

Meta went a different direction too: geothermal, partnering with Sage Geosystems for 150 MW starting 2027.

Follow the Money

The Rook's Take

Here's the tension I can't stop thinking about.

Every major tech company has a net-zero pledge. Microsoft says 100% clean energy. Google says carbon-free by 2030. Amazon has the Climate Pledge. And yet, right now, in 2026, developers are building on-site gas generators to power data centres faster. The urgency of AI is directly conflicting with the timelines of the clean energy transition.

This isn't necessarily a villain story, it's a timing problem. You can build a data centre in 2–3 years. Building the generation and transmission infrastructure to power it cleanly takes a decade or more. Something has to fill the gap.

The finance angle is actually more interesting than the climate angle here. This is a massive capital reallocation. Money that might have gone into utility-scale solar or offshore wind is being redirected to nuclear restarts, gas peakers, grid hardening, and transmission upgrades because that's where the contracted cash flows are. Infrastructure funds, PE firms, and investment banks are all repositioning around this.

The $1.5 trillion in projected new tech-sector debt is not a warning sign, it's an opportunity signal. Someone has to structure those deals, finance those projects, and build those assets. That's where the career opportunities and investment theses are forming right now.

What Happens Next

Near-term: FERC (the U.S. energy regulator) is already intervening in co-located data-centre power arrangements. Expect more regulatory scrutiny on whether data-centre costs get passed onto ordinary electricity ratepayers, that's becoming a political issue.

2027–2030: The first wave of SMRs comes online, starting with Kairos Power. If they deliver on time and on budget, it validates the entire nuclear comeback thesis and unlocks a second wave of deals.

The bear case: The IEA's "Headwinds Case" projects demand plateauing at 700 TWh if AI adoption slows, local bottlenecks persist, and supply chains tighten. If AI model efficiency improves faster than expected or if inference shifts to edge devices, the demand profile underpinning these 20-year PPAs could weaken. Investors pricing in multi-decade cash flows should stress-test that assumption hard.

The bull case: The "Lift-Off Case" has demand exceeding 1,700 TWh by 2035, reaching 4.4% of global electricity. In that world, every nuclear plant, every gas peaker, and every grid upgrade looks like a smart bet made early.

Go Deeper

• IEA Energy and AI Report (2025) — iea.org

• Lawrence Berkeley National Laboratory Data Centre Report (January 2025) — lbl.gov

• Introl Hyperscaler Capex Analysis (2026)

• Constellation Energy / Three Mile Island announcement — constellationenergy.com

• Talen Energy / Amazon Susquehanna PPA announcement — talenergy.com

• Google / Kairos Power SMR announcement — blog.google

Rookonomy publishes financial intelligence for people learning to see the move. Not financial advice.