hawkes.aiKey ideas
- The one idea — Electricity — generation, transmission, cooling — increasingly caps AI's growth more than chip supply does; compute strategy is now energy strategy.
- How it is shown — A growing GPU farm hitting a power-meter ceiling; grid queues in years; labs shopping for reactors.
- The trap to avoid — Modeling AI's future on chip supply alone — the grid moves at permit speed, and it's now the long pole.
AI companies aren't just buying chips anymore — they're shopping for nuclear reactors. Because increasingly the chips exist and the electricity doesn't.
The one idea
Electricity — generation, transmission, cooling — increasingly caps AI's growth more than chip supply does; compute strategy is now energy strategy.
For years the story was chip scarcity. That constraint is inverting: increasingly the chips exist and the electricity doesn't. Datacenters wait in line not for GPUs, but for gigawatts. The act closes on the wall the whole industry is now negotiating with: power. Measure the appetite. Datacenters already drink about four percent of US electricity; credible projections roughly double that share within years — driven overwhelmingly by AI. This stopped being an IT line and became a line on national energy forecasts. Utilities now plan around training runs. Why can't the grid just say yes? A gigawatt campus asks for a power plant's worth of electricity — and plants, transmission, and interconnection move at permit speed: years of queues and approvals.
How it works — the demo
A growing GPU farm hitting a power-meter ceiling; grid queues in years; labs shopping for reactors.
Silicon iterates in two-year generations; grids in decade-scale projects. The clock mismatch is the bottleneck's anatomy. The industry's response tells you it's real: AI companies are signing nuclear deals — restarting shuttered reactors, backing small modular designs, contracting dedicated solar and storage — and siting campuses where electrons, not engineers, are plentiful. Software companies have become energy procurers. When a sector starts buying power plants, believe its constraint. This re-prices act three's scaling story. The curves still want compute — but compute's denominator is now watts, and every efficiency trick this act taught is really capability per watt. The frontier's pace is now, in meaningful part, an energy question. That's not a metaphor. It's the meter.
The trap to avoid
Modeling AI's future on chip supply alone — the grid moves at permit speed, and it's now the long pole.
Why it matters — and what’s next
The trap for anyone forecasting this industry: modeling the future on chip supply alone. The long pole moved — to generation, transmission, permits: none respond to venture capital on venture timelines. This act's parting discipline is the one it opened with: find the binding constraint. It was bandwidth inside the chip. It's the grid outside the building. That completes the physical machine: multiply, chip, rack, warehouse, grid straining to feed it — you know what AI runs on, end to end. Act five asks the next question: given this expensive machine, how do you SHAPE a model to spend it wisely? Dense brains, committees of experts, and the architecture zoo. Next act: the model zoo opens.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
For years the story was chip scarcity. That constraint is inverting: increasingly the chips exist and the electricity doesn't. Datacenters wait in line not for GPUs, but for gigawatts. The act closes on the wall the whole industry is now negotiating with: power.
Measure the appetite. Datacenters already drink about four percent of US electricity; credible projections roughly double that share within years — driven overwhelmingly by AI. This stopped being an IT line and became a line on national energy forecasts. Utilities now plan around training runs.
Why can't the grid just say yes? A gigawatt campus asks for a power plant's worth of electricity — and plants, transmission, and interconnection move at permit speed: years of queues and approvals. Silicon iterates in two-year generations; grids in decade-scale projects. The clock mismatch is the bottleneck's anatomy.
The industry's response tells you it's real: AI companies are signing nuclear deals — restarting shuttered reactors, backing small modular designs, contracting dedicated solar and storage — and siting campuses where electrons, not engineers, are plentiful. Software companies have become energy procurers. When a sector starts buying power plants, believe its constraint.
This re-prices act three's scaling story. The curves still want compute — but compute's denominator is now watts, and every efficiency trick this act taught is really capability per watt. The frontier's pace is now, in meaningful part, an energy question. That's not a metaphor. It's the meter.
The trap for anyone forecasting this industry: modeling the future on chip supply alone. The long pole moved — to generation, transmission, permits: none respond to venture capital on venture timelines. This act's parting discipline is the one it opened with: find the binding constraint. It was bandwidth inside the chip. It's the grid outside the building.
That completes the physical machine: multiply, chip, rack, warehouse, grid straining to feed it — you know what AI runs on, end to end. Act five asks the next question: given this expensive machine, how do you SHAPE a model to spend it wisely? Dense brains, committees of experts, and the architecture zoo. Next act: the model zoo opens.