Missed The AI Boom? This Is 10 Times Bigger.

The doom-brain experiment

The video opens in a Melbourne laboratory, where roughly 200,000 human neurons grown on a silicon chip taught themselves to play Doom — in a week. Felix reads this not as a science story but as a money story: institutions are funding neuron-chip research for one reason, electricity. A human brain runs on about 20 watts. AI devours power at a scale that's becoming the industry's binding constraint. Growing brains in a dish, he says, is what desperation looks like — and desperation, properly read, is information.

Quantifying the power gap

The numbers are the argument. A single AI server rack draws roughly 100,000 watts; the largest data centers consume power at nation-scale levels. The industry uses around 485 terawatt-hours and is doubling in a matter of years. Gas-turbine waitlists stretch for years, new nuclear takes a decade or more, and a recent capacity auction repriced power 11x in two years. Gartner projects power will restrict 40% of AI data centers as early as next year. Every part of the AI buildout assumes electricity that doesn't exist yet.

Big Tech buys power directly

Big Tech's response reveals the desperation in three rungs. First, hardware efficiency — though the Jevons paradox applies: cheaper compute summons ten times the demand. Second, long-dated power contracts. Third, buying generation outright: Microsoft is restarting Three Mile Island at roughly twice the market rate for power; Amazon and Meta have cut their own nuclear deals, with Meta locking up 6.6 GW through 2035. When the world's richest companies pay double for electricity instead of waiting, the shortage is real.

The legal monopoly

The sharpest edge of the thesis: only three companies own meaningful nuclear fleets in deregulated markets that allow direct corporate deals — Constellation Energy, Vistra, and Talen Energy — and they're fenced in by legislation across eleven states. A legal monopoly on the one power source that can scale. Felix runs the names through his Winston app analysis: Constellation (Three Mile Island, $20–50/MWh contracts, 31% below its highs), Talen (a single-customer Microsoft story with weak fundamentals, ~20% off peak), and Vistra (a third below its high). Demand desperation keeps growing while the twenty-year contracts haven't yet hit reported numbers.

The picks-and-shovels alternative

For those who find the utilities unappealing, he offers the infrastructure angle: GE Vernova, the turbine and grid business. Turbine backlog has grown from 83 to 116 GW against roughly 20 GW of annual production — a six-year queue — plus a $42 billion grid backlog and a global transformer shortage. It's the classic picks-and-shovels play: whichever AI bets win, they all need electrons moved through equipment someone has to build.

Why efficiency makes it worse

One of the video's sharpest insights is counterintuitive: making AI more efficient won't fix the power crisis. The Jevons paradox says that cheaper compute summons an order of magnitude more demand — every efficiency gain gets swallowed by new use cases. That's why Big Tech's first rung (hardware efficiency) is a treadmill, and why the real race is over the second and third rungs: locking up power contracts and generation itself. The neuron-chip experiment is the tell. When the industry starts funding 20-watt biology to escape its own power bills, the constraint isn't a rumor — it's the central fact of the buildout.

The takeaway

The risks are stated plainly: data-center announcements are claims on power, not power; AI spending could slow; regulators could reverse course; and none of these names are cheap on fundamentals. But the core claim stands — the AI boom's bottleneck isn't chips or software, it's the grid, and the money is flowing to whoever controls the watts. His rule before any of it: size your positions and set your exits before you buy.