Why the Same AI Model Carries a $40M or $100M Price Tag — and Both Are True
AI training costs are real but slippery — the same model can carry a $40 million or $100 million price tag, both accurate, because one counts only the final compute run and the other includes staff and failed experiments that add up to nearly half the total bill.
What are the documented costs of AI model training at major labs, and how do publicly cited figures compare to verifiable evidence from financial filings, leaks, and independent analysis?
- 1Independent research and executive statements broadly agree on the scale of frontier model costs once you account for what each figure measures: a final training run for GPT-4 was estimated at roughly $40 million, while Sam Altman said total GPT-4 training exceeded $100 million — figures that reconcile if the higher number includes research and development overhead.
- 2The single biggest driver of confusion is definitional: staff and R&D overhead account for between 29% and 49% of a frontier model's total amortized cost, so a 'final run' figure and a 'total development' figure for the same model can differ two- to threefold without either being wrong.
- 3AI training costs have grown at about 2.4 times per year since 2016, making today's hundred-million-dollar figures the expected result of a steep, documented trend rather than outliers.
- 4Companies disclose selectively: Alphabet's filings confirm it does not break out DeepMind or Gemini training costs at all, folding AI spending into company-wide capital expenditure that nearly doubled to $25.2 billion in the first half of 2024.
When a tech CEO says training their AI model cost a hundred million dollars, the number sounds precise — but the closer you look, the more it dissolves. This investigation traced what the big labs actually said, held those statements against financial filings and independent technical analysis, and found a puzzle: the numbers are real, but they almost never mean what a casual reader thinks.
Sam Altman said GPT-4 training exceeded $100 million. Independent researchers pegged the final training run at roughly $40 million. Both figures are accurate — they're just counting different things. The $40 million captures only the last successful run, the raw compute hours on the chips. The $100 million includes the staff, the failed experiments, the overhead. For frontier models, that overhead runs between 29% and 49% of the total bill, which is why the same model can wear tags that differ two- or threefold without either being dishonest.
The confirmed evidence clusters in the tens to low hundreds of millions per model. Costs are climbing steeply — about 2.4 times every year since 2016 — making today's hundred-million-dollar figures the expected output of a documented curve, not shocking outliers. But the most dramatic claims, including leaked figures putting OpenAI's training spend at $25 billion in 2025, trace back to second-hand reporting that could not be independently verified here. And no major lab has published an audited, model-specific cost breakdown — the single document that would resolve most of the debate simply does not exist in the public record. The gap between spending that is clearly enormous and figures that cannot be checked is precisely where hype, valuation and policy all get decided.
The Full Investigation
7 sections · 11 min read
Confirmed facts and attributed reporting read normally; only contested, unverified, or speculative sentences are highlighted. Hover any sentence for its grade and sources.
The question behind the headline number
When a tech chief executive names a price for training an AI model, the figure lands like a fact. But peer at it and the ground shifts. The same model can carry a $40 million price tag from one source and a $100 million tag from another — and both can be accurate, because they are counting different things.
This report traces what the big labs — OpenAI, Google, Anthropic and Meta — have actually said about training costs, and holds those statements against three other kinds of evidence: official financial filings, leaked internal documents, and independent technical analysis. The stakes are practical. These numbers shape how investors value companies, how regulators think about competition, and how the public understands whether frontier AI is a game only the richest firms can play.
The cast is small and familiar. OpenAI's Sam Altman and Anthropic's Dario Amodei have both spoken publicly about costs. Meta's Mark Zuckerberg has disclosed sweeping infrastructure plans. On the other side sit independent analysts — chiefly Epoch AI and Stanford's Human-Centered AI institute — who rebuild cost estimates from hardware and compute data. The tension between these two camps is the story.
What the labs have actually said about their training costs
Start with the clearest voice in the room. At an April 2023 event at MIT, Sam Altman was asked whether training GPT-4 had cost $100 million. His answer, widely reported: it was more than that. Stanford's AI Index and other independent sources record the same indication — GPT-4 training exceeded $100 million. That is about as solid as public statements get: a named executive, on the record, corroborated across sources.
Dario Amodei of Anthropic has been the industry's most vocal forecaster. He told CNBC in April 2024 that current models cost roughly $100 million to build, and predicted the next year's models would run to about $1 billion. By July 2024, Stanford's researchers reported him saying $1 billion training runs were already underway. Two independent accounts landing on the same billion-dollar scale for 2024-2025 — they line up cleanly. Amodei also described Anthropic's own Claude 3.5 Sonnet as costing 'a few $10M's to train' and said it was not built using any larger or more expensive model — a pointed claim of efficiency, though it rests on his word alone.
Meta's disclosures came through Zuckerberg. On the company's Q2 2024 earnings call in late July 2024, he said training its next model, Llama 4, would need ten times the compute of Llama 3. Meta also procured a cluster of more than 100,000 Nvidia graphics chips for model development, according to reporting on the company. These are statements about scale and trajectory more than tidy per-model price tags — a pattern that runs through the whole industry.
What is missing is as telling as what is present. No lab in this dossier published an audited, model-specific training cost. The figures are soundbites, forecasts, and infrastructure boasts — attributable and mostly confirmed, but never itemized.
2019-07
- Microsoft invested $1 billion in OpenAI
2023-04
- Sam Altman stated at MIT event that GPT-4 training cost more than $100 million
2024-04
- Dario Amodei told CNBC current AI models cost ~$100M to develop and predicted next year's models would cost ~$1 billion
2024
- Meta reported Q2 2024 capital expenditures rose 33% to $8.5 billion from $6.4 billion year-earlier
2024-07
- Dario Amodei stated that $1 billion training runs were already underway
- Microsoft's Form 10-K described OpenAI as both strategic partner and competitor
2024-07-23
- Alphabet filed SEC Form 10-Q reporting capex of $25.2B for H1 2024 (up from $13.2B H1 2023)
2024-08-01
- Mark Zuckerberg stated on Meta Q2 earnings call that Llama 4 training would require 10x more compute than Llama 3
2025-02-07
- Epoch AI and Stanford published study estimating AI training costs grew 2.4x annually since 2016
Open: What did the original Reuters and SemiAnalysis analyses estimate for adding ChatGPT-like features to Google Search, and why do those figures differ ($6 billion versus $3 billion) [C-001][C-002]?; Is there any independent corroboration for Meta's reported 100,000-GPU cluster beyond the single outlet that reported it [C-006]?
Filings confirm the spending — but hide the model-level detail
If executives speak in soundbites, their companies' legal filings should speak in hard numbers. They do — but not the numbers a curious reader wants. The filings confirm enormous spending. They just refuse to tie it to any single model.
Alphabet's official quarterly filing with US regulators shows the point plainly. Its capital expenditures — spending on data centres, servers and networks — hit $25.2 billion in the first half of 2024, up from $13.2 billion a year earlier. That is a jump of about 91% year-over-year, which our own check of the figures confirms. Its research and development expenses reached $11,860 million in Q2 2024. Yet the same filing states that Alphabet does not break out DeepMind or Gemini training costs separately; AI development is folded into company-wide activity. The detail simply is not there to extract.
Meta's numbers tell a parallel story. Its capital expenditures rose nearly 33% to $8.5 billion in a single quarter of 2024, up from $6.4 billion a year earlier — an increase our arithmetic confirms as about 32.8%. These are real, verifiable, primary-source figures. But they cover all the infrastructure a company uses, not the cost of training one model. Treating a capex line as a training bill would badly overstate the case.
Then there is the material that would fill the gap — if it could be independently verified. A Wall Street Journal report, based on confidential financial documents, reportedly showed OpenAI expecting to spend four to five times more on training than Anthropic each year for roughly five years — neither OpenAI nor Anthropic has confirmed these numbers, and this investigation could not verify them independently. A separate account of leaked WSJ documents put Anthropic's training costs at $4.1 billion against $13 to $25 billion for OpenAI — a three-to-six-fold gap — but again, neither company has confirmed the figures and they could not be independently verified here. The most striking figure of all: leaked investor documents reportedly projected OpenAI's 2025 training spend at $25 billion and $121 billion by 2028 — but once more, OpenAI has not confirmed this and the investigation could not verify it. Here is the catch. Every one of these leaked figures reached this investigation second-hand — through a Reddit thread and a Substack newsletter citing the WSJ, never the original Journal reporting itself. Microsoft's own filing, which described OpenAI as both a strategic partner and a competitor, adds context but no cost breakdown. The most dramatic evidence, in short, is also the least checkable.
Open: Do the original Wall Street Journal articles specify whether the $25 billion and $4.1 billion figures represent single-model training, annual aggregate R&D, or total capital investment [C-025][C-030][C-031]?; Can Microsoft's $1 billion (2019) and $10 billion (post-ChatGPT) investments in OpenAI be confirmed beyond the single aggregator that reported them here [C-029]?
How independent analysts rebuild the numbers from scratch
Where executives assert and filings obscure, a handful of independent researchers try to reconstruct costs from the ground up — counting chips, hours and watts. Their work is the closest thing to a check on the industry's own figures, and it tells a subtly different story.
The anchor is a February 2025 study from Epoch AI and Stanford. It found that the amortized cost of training the most compute-hungry models has grown about 2.4 times every year since 2016, with a statistical range of 2.0 to 2.9 times. That single trend does a lot of quiet work: it makes today's hundred-million-dollar models look like the predictable output of a steep curve rather than shocking outliers. The same study put GPT-4's final training run at roughly $40 million and Google's Gemini Ultra at around $30 million — both notably below the executive-quoted figures.
The methods behind these estimates are transparent, which is what makes them useful. Epoch AI estimates GPT-4's training compute at 2.1e25 floating-point operations — a raw measure of calculation — based on the hardware used and how long training ran. The specialist firm SemiAnalysis independently landed at about 2.15e25, a difference under 3%. When two independent teams converge that tightly, it is strong evidence the underlying compute figure is solid. Meta's own documentation offers another cross-check: it recorded 30.8 million GPU-hours to train Llama 3.1 405B, which works out to $61 to $92 million at typical cloud rental rates of $2 to $3 per hour — arithmetic that holds up on inspection.
Here the estimates diverge in an instructive way. Epoch AI pegged Llama 3.1's training at $170 million, while Meta's documented GPU-hours imply the lower $61-92 million range. That two-to-threefold gap is not a mistake. The $170 million figure includes amortized R&D overhead — staff, failed experiments, infrastructure — while the GPU-hour figure captures only the final run's raw compute. The same fork explains the range for Gemini Ultra, where a compute-cost basis and a final-run basis differ severalfold. Read carefully, the independent numbers do not contradict the companies. They measure a different slice of the same pie.
Open: Is there any independent estimate of Gemini 1.0 Ultra's training compute to corroborate Epoch AI's single figure of 5.0e25 floating-point operations [C-021]?; Can SemiAnalysis's detailed 100,000-GPU cluster costs — more than $4 billion in server capital and $123.9 million a year in electricity — be corroborated by any other source [C-033]?
Competing explanations for the gap between stated and verified costs
So why does the same model wear a $40 million tag in one place and top $100 million in another? Four explanations compete, and the evidence sorts them into stronger and weaker.
One explanation holds that public figures systematically undercount by leaving out overhead. On this reading, companies quote only the final successful run and omit the failed experiments, staff, and infrastructure behind it — making stated costs two to five times lower than the true total. The evidence here is genuinely strong. Staff costs alone run 29% to 49% of the amortized total; Meta's raw GPU-hour figure of $61-92 million sits well below Epoch's all-in $170 million estimate; and the reported but unverified finding that most research compute goes to unreleased work all point the same way, though that last claim rests on a single source and could not be independently confirmed.
A second reading is less accusatory: the spread reflects honest methodological differences, not misrepresentation. GPT-4 ranges from $40 million to over $100 million, Gemini Ultra from $30 million to $192 million, Llama 3.1 from $61 million to $170 million — but each figure openly states its scope, whether final-run, compute-only, or fully amortized. The dossier did not contain direct access to the original Wall Street Journal reporting or primary leaked documents that might show same-model-same-scope inconsistencies, so the inconsistency question is tested against available evidence only, not exhaustively. No case of a company quoting different numbers for the same model in the same accounting sense was found in the material available here. That absence supports the innocent reading.
A third, more troubling explanation says leaked documents reveal multi-year spending five to twenty times higher than public single-run figures suggest — OpenAI's alleged $25 billion in 2025 against publicly discussed model costs in the hundreds of millions. This one stays merely plausible, not proven, for two reasons. The leaked figures could not be independently verified here. And the reported but unverified finding that most compute goes to experimentation means a huge annual budget can coexist honestly with modest per-model figures — turning apparent deception back into a question of scope.
A fourth explanation looks forward: costs are climbing exponentially and will hit $5 to $100 billion within a few years. The documented 2.4x annual growth rate supports the trend's shape, and both Amodei's forecasts and the leaked OpenAI projections point upward. But these are predictions by nature. Only the actual disclosed costs of models released in 2026 through 2028 will show whether the curve holds or bends.
What the evidence forces us to conclude
Pull the threads together and a measured verdict emerges — neither the industry's reassurance nor its critics' alarm survives fully intact.
The numbers companies cite are real in the narrow sense that matters most: they are attributable, mostly confirmed, and consistent with independent technical analysis once scope is accounted for. Altman's over-$100 million for GPT-4 and Epoch's $40 million final-run estimate are not a contradiction to be exposed; they are two honest measurements of different things, bridged by the documented fact that overhead adds up to nearly half the total. The steep, verified 2.4x annual cost growth makes the whole landscape coherent.
Where the industry's critics are on firmer ground is disclosure. No lab in this dossier published an audited, model-specific cost. Alphabet's confirmation that it does not break out Gemini or DeepMind costs at all is the clearest example: the granularity needed for genuine verification simply does not exist in public filings. That is a real transparency gap, even if it is not proof of dishonesty.
The most explosive claims remain unproven rather than debunked. The leaked figures suggesting OpenAI spends tens of billions annually would, if verified, reframe the entire debate — but they reached this investigation only through second-hand aggregators, and the reported but unverified finding that most compute goes to experimentation offers an innocent explanation for why annual budgets could legitimately dwarf single-model costs. On a speculative basis, and labelled as such, the trajectory looks genuinely steep: independent researchers project the largest training runs could exceed $1 billion by early 2027, and Amodei has floated eventual figures reaching $100 billion. Those are forecasts, not receipts. What the evidence forces is narrower and duller than either camp would like: the numbers are mostly real, the definitions are slippery, and the disclosure is thin.
Why it matters
How much it costs to train an AI model is not a trivia question. If the true figure is a few tens of millions, frontier AI is within reach of many players and the market stays competitive. If it is tens of billions a year, the field belongs to a handful of giants — with consequences for prices, jobs, and who controls a transformative technology. The evidence shows costs climbing about 2.4 times a year and companies pouring tens of billions into infrastructure. But because no lab discloses an audited per-model cost, investors, regulators and the public are left reading executive soundbites and unverified leaks. That gap between spending that is clearly enormous and figures that cannot be checked is precisely where hype, valuation and policy all get decided.
- No major lab has published an audited, model-specific training cost with a documented breakdown of final-run compute versus total amortized development — the single document that would resolve most of the debate does not exist in the public record.
- Whether frontier training costs will actually reach the $5-100 billion range forecast for 2026-2028 depends on disclosed figures for models not yet released, which no current evidence can confirm or refute.