You open Claude or ChatGPT, type a question, get an answer in two seconds. You do this fifty times a day and assume it'll keep getting faster and cheaper — the way software always has. Reasonable expectation. One problem.

Every single query runs on physical hardware that consumes electricity the way a small city does. We've covered the individual pieces over the past two weeks — Amazon's $200 billion capex bet and the cash flow squeeze it created, the NAACP suing xAI over 27 unpermitted gas turbines in Mississippi, TSMC's record Q1 confirming that 61% of its revenue now comes from chips that all need to be plugged in somewhere. Each story made sense on its own. Together, they tell a different story — one about which companies will actually serve you in 2028 and which ones won't.

The thesis: power is the new moat

The conventional framing of the AI race is about intelligence — who has the best model, the highest benchmark score, the most capable agent. But model quality is converging. Access to electricity is not.

According to a Tom's Hardware analysis published on April 3, 2026, utilities and grid operators delayed or cancelled roughly half of the ~140 large-scale US data center projects planned for this year. Not because of money or demand — because of the electrical grid. Transformer lead times have stretched to five years. New nuclear plants take a decade. Microsoft CEO Satya Nadella admitted in January 2026 that the company has GPUs sitting in inventory because there's nowhere to plug them in.

You can ship a better model in six months. You cannot ship a power plant.

A vendor power scorecard

This is where it gets practical. If you're picking AI tools for your team — a coding assistant, an API provider, a platform for agents — here's what the power landscape looks like as of April 2026, organized by who actually secured generation capacity versus who's still scrambling for it.

Tier 1 — Power-secured (owns or leases generation capacity under long-term contract)

  • Microsoft signed a 20-year deal in September 2024 to restart Three Mile Island's Unit 1 reactor (835 MW). The company also holds long-term renewable PPAs across multiple regions.
  • Amazon locked in 1.9 GW of nuclear power through 2042 from the Susquehanna plant in a deal announced March 2024. Amazon also owns solar and wind farms directly.
  • Google signed contracts with Kairos Power in October 2024 for 500 MW of small modular reactors, with Kairos targeting delivery in the late 2020s.

These companies spent years and billions securing electrons before the current crunch. That head start is now a structural advantage no model improvement can cross.

Tier 2 — Power-dependent (buying from the grid, competing for permits)

  • Meta builds data centers aggressively but relies primarily on grid power and renewable energy credits. No announced generation ownership at the scale of Tier 1.
  • Oracle expands rapidly without disclosed long-term generation agreements comparable to the nuclear deals above.

Both can build. Neither controls the fuel.

Tier 3 — Power-desperate (cutting corners)

  • xAI installed 27 gas turbines without air permits at its Colossus 2 site in Southaven, Mississippi. The NAACP filed suit under the Clean Air Act on April 14, citing an estimated 1,700+ tons of nitrogen oxides per year pumped into a region that already fails federal smog standards. This is what power desperation looks like when it meets a community that can't afford more pollution.

A Futurum Group analysis published on April 11, 2026 found that the top five hyperscalers now consume nearly 100% of their operating cash flows on capex — up from a 10-year average of 40%. That concentration accelerates the gap between tiers. Companies with secured generation can keep building. Companies without it face a ceiling that no amount of capital alone can lift.

What this means for your stack

When you evaluate AI vendors for 2027–2028, add these questions to your checklist:

⚙️ Does the provider own or lease generation capacity? Not "renewable energy credits" — actual megawatts under long-term contract. PPAs of 10+ years matter. RECs alone do not guarantee the lights stay on during a grid crunch.

⚙️ Where are their data centers? Regions with grid headroom (parts of the US Southeast, Nordics, parts of Canada) will serve requests more reliably than power-constrained corridors (Northern Virginia, parts of Texas, the Netherlands). Ask your vendor which region your traffic routes through.

⚙️ What's their capex-to-cash-flow ratio? A company spending 95% of operating cash flow on infrastructure is one bad quarter from cutting capacity expansion. Check the last two earnings calls before you sign a multi-year contract.

⚙️ Do they have a track record of regulatory compliance on energy? Unpermitted turbines today signal operational risk across the board tomorrow. If a vendor cuts corners on air permits, ask what else they're skipping.

The smartest model in the world is useless if the company running it can't keep the lights on. The simplest model backed by a nuclear plant will still answer your queries in 2030.

The AI race was supposed to be about intelligence. The bottleneck is the electric bill. 🫶