hawkes.aiKey ideas
- The one idea — Frontier pretraining compute runs tens to hundreds of millions of dollars — thousands of chips, months of burn, real failure risk — and market structure follows the physics.
- How it is shown — The bill itemized: the chip fleet, the electricity, the months; a loss-spike scare mid-run; the jumbo-jet comparison.
- The trap to avoid — Confusing training cost with serving cost — the run is a capital expense paid once; inference is the perpetual bill (and small players CAN fine-tune and serve).
- What it sets up — We need a unit for effort itself.
Some frontier training runs get quietly abandoned — months of compute and millions of dollars written off. The AI price tag has a risk column outsiders never price.
The one idea
Frontier pretraining compute runs tens to hundreds of millions of dollars — thousands of chips, months of burn, real failure risk — and market structure follows the physics.
Training one frontier model can cost more than a jumbo jet — and unlike the jet, there's no guarantee it flies. Tonight: the bill, itemized — and what it did to the industry's shape. Line one: the fleet. Frontier runs occupy thousands — sometimes tens of thousands — of Earth's most advanced chips, each priced like a luxury car, all reserved for the same job at the same time for months. Before a single token burns, the hardware bill alone explains most of the meter. Line two: the burn. Megawatts, continuously, for months — small-town draw, industrial rates, no nights off. Electricity is the quieter line-item that never stops accruing — a whole episode waits in the hardware act. Line three: risk — the item outsiders never price. Runs crash.
How it works — the demo
The bill itemized: the chip fleet, the electricity, the months; a loss-spike scare mid-run; the jumbo-jet comparison.
Curves spike; weeks roll back to checkpoints. Hardware fails daily at fleet scale. Some runs are quietly abandoned — months and millions written off. A frontier run is a jet-sized lottery ticket bought by teams who mostly, not always, win. The totals, from public estimates: GPT-4-class training compute landed somewhere between forty and a hundred-plus million dollars. The current frontier tier runs from a hundred million toward a billion. Market structure follows: the list of organizations that can sign that check and absorb a failure is a handful. The oligopoly wasn't plotted — it was priced. The trap: letting the jet-sized number stand for all AI economics. Training is capital, paid once per generation.
The trap to avoid
Confusing training cost with serving cost — the run is a capital expense paid once; inference is the perpetual bill (and small players CAN fine-tune and serve).
Why it matters — and what’s next
We need a unit for effort itself.
Serving — answering your prompts — is the perpetual bill, priced per token forever, and it's where most of the industry's ongoing spend actually lives. And the jet-price applies to pretraining from scratch: fine-tuning an open model costs luxury-car money, not fleet money. That door — episode eighty-five — is wide open. But dollars are a translation. Chip prices shift, electricity varies by geography, and the honest quantity underneath is simpler: the raw count of arithmetic operations performed. The industry runs its real accounting — and now, its regulations — in that unit. Next: the FLOP. Two minutes, and you'll read AI's power structure in its native currency.
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
Training one frontier model can cost more than a jumbo jet — and unlike the jet, there's no guarantee it flies. Tonight: the bill, itemized — and what it did to the industry's shape.
Line one: the fleet. Frontier runs occupy thousands — sometimes tens of thousands — of Earth's most advanced chips, each priced like a luxury car, all reserved for the same job at the same time for months. Before a single token burns, the hardware bill alone explains most of the meter.
Line two: the burn. Megawatts, continuously, for months — small-town draw, industrial rates, no nights off. Electricity is the quieter line-item that never stops accruing — a whole episode waits in the hardware act.
Line three: risk — the item outsiders never price. Runs crash. Curves spike; weeks roll back to checkpoints. Hardware fails daily at fleet scale. Some runs are quietly abandoned — months and millions written off. A frontier run is a jet-sized lottery ticket bought by teams who mostly, not always, win.
The totals, from public estimates: GPT-4-class training compute landed somewhere between forty and a hundred-plus million dollars. The current frontier tier runs from a hundred million toward a billion. Market structure follows: the list of organizations that can sign that check and absorb a failure is a handful. The oligopoly wasn't plotted — it was priced.
The trap: letting the jet-sized number stand for all AI economics. Training is capital, paid once per generation. Serving — answering your prompts — is the perpetual bill, priced per token forever, and it's where most of the industry's ongoing spend actually lives. And the jet-price applies to pretraining from scratch: fine-tuning an open model costs luxury-car money, not fleet money. That door — episode eighty-five — is wide open.
But dollars are a translation. Chip prices shift, electricity varies by geography, and the honest quantity underneath is simpler: the raw count of arithmetic operations performed. The industry runs its real accounting — and now, its regulations — in that unit. Next: the FLOP. Two minutes, and you'll read AI's power structure in its native currency.