What it really costs to train a frontier model

What does it actually cost to train a frontier model? Nobody outside the labs knows precisely, and the numbers that circulate are usually one component reported as if it were the total.

The headline figure you read is almost always the final training run. That’s the cheapest honest way to count it, and it leaves out most of the money.

The cost to train a frontier model everyone quotes

When a lab says a model cost some amount to train, they usually mean the compute for the run that produced the released weights. Chips multiplied by hours multiplied by a rate.

It’s a real number and it’s rising fast. Epoch AI reports frontier training compute growing 5x per year since 2020, roughly ten-thousandfold in total for the leading models.

But it’s the equivalent of costing a film by the price of the final shooting day.

What the headline leaves out

Cost	Usually in the headline?	Rough weight
Final training run compute	Yes	The quoted figure
Failed and abandoned runs	No	Often larger than the final run
Research experiments at small scale	No	Continuous, all year
Data licensing and acquisition	No	Rising sharply
Human annotation for tuning	No	Meaningful, ongoing
Salaries	No	Frequently the biggest line
Idle and reserved capacity	No	Paid whether used or not

Failed runs deserve the most attention, because they’re invisible and enormous. A frontier training run can diverge weeks in, and when it does the money is simply gone.

The published cost is the price of the run that worked. The bill includes every run that didn’t.

Salaries are the line nobody models

Frontier research talent is among the most expensive labour in the world, and the compensation packages that get reported in hiring stories are not outliers.

How much that talent is worth showed up starkly this month when four senior Google researchers left together to found a company, a move that knocked 4% off Alphabet. Markets priced a handful of people at billions.

Spread that across a team over a multi-year development cycle and salaries can rival or exceed the compute line, which is why “cost to train” is a slippery phrase.

Data stopped being free

Early models were trained on scraped text at essentially the cost of bandwidth. That era is closing.

Villalobos and colleagues projected that models “will be trained on datasets roughly equal in size to the available stock of public human text data between 2026 and 2032“. As the free supply tightens, licensed corpora become a purchase rather than a download.

Chinchilla made that worse arithmetically. DeepMind found that “for every doubling of model size the number of training tokens should also be doubled”, per Training Compute-Optimal Large Language Models, so data demand scales with the model rather than staying flat.

Overtraining multiplies the bill deliberately

Here’s a decision that raises training cost on purpose, and it’s the right call.

Because inference dominates lifetime cost, labs train smaller models far past the compute-optimal point so they’re cheaper to serve. Meta’s Llama 3 family, built around a 405B dense flagship plus much smaller siblings, follows exactly that logic.

So a chunk of modern training spend is buying lower serving costs later. It’s capital expenditure to reduce operating expenditure, and counting it purely as a training cost misreads what it’s for.

The infrastructure is now financed, not bought

The largest structural change is that the compute underneath all this increasingly sits on somebody else’s balance sheet.

Six private capital firms signed non-binding agreements toward up to $500 billion of AI data centre financing, with Nvidia guaranteeing collateral value. Jensen Huang told CNBC his chips are an “investable asset”.

That converts a capital cost into a financing cost, which changes who bears it and when. It also means the true cost of training now includes an interest rate.

Training is not where the money ends up

Here’s the part that reframes the whole question. For a widely used model, lifetime inference spend exceeds training spend, often by a lot.

Training is a single bill. Inference is a bill that arrives with every request, forever, and it scales with success rather than with ambition.

You can see labs pricing against that reality. Grok 4.6 arrived at $6 per million output tokens against $30 for a comparable model, which is a serving-cost decision showing up on a price list.

Reasoning models push the balance further toward inference, since they spend compute at request time. So “what did it cost to train” is becoming a less interesting question than “what does it cost to run”, and the two are managed as one budget.

What it costs everyone else

Frontier figures are close to useless if you’re deciding whether to train something yourself, and the good news is that the relevant numbers are much smaller.

Fine-tuning an open model with parameter-efficient methods runs on rented hardware for tens of dollars, not millions, because you’re adjusting a small adapter rather than retraining a network.

Continued pretraining on domain text sits in the middle, meaningful but not exotic. And full pretraining from scratch is almost never the right answer below frontier scale, since an open-weight model plus fine-tuning gets you further for a fraction of the spend.

That gap is why Meta’s decision to publish a 30B model under Apache 2.0 matters commercially. It moves the starting line for everyone who wasn’t going to spend nine figures anyway.

Why costs might fall rather than rise

The counter-case is real and it’s mostly about efficiency compounding.

Epoch AI reports pre-training compute efficiency improving around 3x per year, so a given capability level costs progressively less to reach. Architecture helps directly: mixture-of-experts buys capacity without proportional compute.

DeepSeek’s R1 pushed a different lever, getting reasoning from reinforcement learning alone, “obviating the need for human-labeled reasoning trajectories”. That removes an expensive annotation step entirely.

Put together, matching last year’s frontier gets cheaper every year even as reaching the new frontier gets more expensive. Both statements are true and people talk past each other by picking one.

Why nobody publishes the real number

It’s worth asking why this is opaque, because the reasons are commercial rather than technical.

A low figure signals efficiency to investors and makes competitors look wasteful. A high figure signals a moat, since it implies nobody underfunded can follow. Labs pick whichever story serves them, and both are defensible from the same underlying ledger.

There’s a competitive reason too. Detailed cost breakdowns reveal training recipes, hardware utilisation and data strategy, which is exactly what rivals want.

And most frontier labs are private or embedded inside larger companies, so nothing forces disclosure. Alphabet and Microsoft report AI spending inside enormous capital expenditure lines that reveal little about any single model.

What a sensible estimate looks like

If you want a defensible way to think about it rather than a number to repeat, use this shape.

Take the quoted training compute cost as a floor. Assume failed runs and research experiments add a comparable amount again. Add a research team’s annual compensation for the development period. Add data licensing, which is now material rather than negligible.

The result is usually several times the headline. Anyone quoting a single tidy figure for a frontier model is quoting one line of a much longer invoice.

The same caution applies to comparisons between labs. Two companies can report costs computed on entirely different bases, one counting only the successful run at internal transfer prices and the other including amortised infrastructure, and the resulting figures aren’t comparable in any useful sense.

Cloud accounting muddies it further. A lab inside a hyperscaler pays internal rates for compute that may bear little relation to what an external customer would be charged for the same hardware.

What would change this

Meaningful disclosure would, and there’s a plausible route to it. If AI infrastructure is financed by institutional capital, lenders will eventually want audited numbers, and audited numbers leak into the public record.

A sustained break in Epoch’s 5x compute growth would also change the picture, because it would mean the industry had decided the next frontier run wasn’t worth the money.

Compute access is also a policy variable rather than a purely commercial one, and export controls moved twice in five months in ways that change what a training run costs depending on where you sit.

Where that spending ends up is the other half of the question, which our piece on which layer of the stack keeps the money works through.

Until then, treat every published training cost as a marketing number. It’s chosen by the company, it’s unaudited, and it describes the component that makes the achievement look most efficient. Our piece on how model claims get framed covers the same instinct applied to capability.