What It Costs to Train a Frontier AI Model

You now know what GPUs are and why they're priced like exotic cars. Time to run the most expensive experiment in software history: take thousands of them, run them flat-out for months, and produce a single AI model. This chapter is about that bill — what's actually on it, why it keeps growing, and the strange economics of what all that money buys.

The core of the bill: GPU-time

Training, mechanically, is simple to describe: show the model a huge slice of human text, have it predict the next token, nudge billions of parameters after every miss, and repeat until it's read more than any human could in a thousand lifetimes.

The industry measures this work in GPU-hours — one chip running for one hour — and frontier training runs consume them at a scale that's hard to hold in your head. Public reports about recent frontier-class models describe clusters of tens of thousands of GPUs running for weeks to months — tens of millions of GPU-hours for a single run. Multiply any plausible hourly cost for top-end chips by that, and you land where the reported estimates land: compute bills alone in the hundreds of millions of dollars for the largest models, up from single-digit millions just a few model generations ago.

Why does anyone sign off on a bill like that? Because of the most consequential discovery in modern AI.

Scaling laws: why bigger became a strategy

Around 2020, researchers established something remarkable: model quality improves predictably as you scale up compute, data, and parameters together. Not "sometimes, if you're lucky" — predictably, along smooth curves. These are the scaling laws, and they changed the economics of the whole field.

Scaling laws are also why the training bill compounds rather than settles: each generation's "predictably better" requires roughly an order of magnitude more compute than the last.

The rest of the bill

Compute headlines the invoice, but three other lines are far bigger than most people assume:

• Data. The era of freely scraping the whole internet is closing — publishers now license their archives for reported sums in the tens to hundreds of millions. Then the data must be cleaned, filtered, and deduplicated, which is its own engineering programme. Increasingly, labs also pay to create data: hiring experts to write worked solutions in maths, code, medicine, and law, because quality of data now matters as much as quantity.
• People. After pre-training comes the feedback phase — thousands of human raters comparing model answers so the model learns what good looks like, plus the elite research and engineering teams themselves, whose compensation packages have become industry legend. A frontier lab's payroll is a material fraction of its training economics.
• Dead ends. The headline run is the last one. Before it come months of smaller experiments — testing architectures, data mixes, and hyperparameters — plus runs that crash, diverge, or just disappoint. A meaningful share of every training budget buys knowledge about what not to train.

What all that money buys: a file

Here's the strangest fact in AI economics. Spend the better part of a billion dollars, and the deliverable is a single file of parameters — the trained model, the same billions of numbers a model is made of. That file:

• Cost a fortune to create — everything above.
• Costs nothing to copy. Duplicating it is duplicating any file. This is why one training run can serve the entire world, and why the real ongoing cost is running it, not making it.

Economists would call this a classic high fixed cost, near-zero marginal cost product — like a blockbuster film or a new drug. And the comparison holds one more uncomfortable truth.

The file goes stale

A factory built for nine figures produces goods for decades. A frontier model built for nine figures is typically surpassed within a year — often within months, sometimes by its own maker. Competitors ship, scaling laws keep operating, and yesterday's marvel becomes today's mid-tier option served at a discount.

So the training bill isn't a one-off capital expense; it's a subscription to the frontier. Stop paying and you don't keep your position — you fall off the leaderboard within a couple of product cycles. This, more than anything, explains the eye-watering fundraising rounds of the big labs: they aren't raising to train a model, but to afford the next several, each an order of magnitude costlier than the last.

Recap

• Frontier training consumes tens of millions of GPU-hours — thousands of chips for months — with reported compute costs in the hundreds of millions and rising each generation.
• Scaling laws made improvement predictable, which made spending rational, which made training budgets an arms race.
• Beyond compute, the bill includes data licensing and creation, human feedback, elite payroll, and failed experiments — each material on its own.
• The output is a file: astronomically expensive to make, free to copy — high fixed cost, near-zero marginal cost.
• The file goes stale fast, so frontier training is a repeating subscription, not a one-time purchase — which is why the labs keep raising ever-bigger rounds.

That's the mortgage. Now for the bill that never stops arriving: every message, from every user, metered token by token — why you pay per token.