In this article
Today we're talking about the money generative artificial intelligence moves, and it's worth looking at the zeros slowly because almost everyone confuses them, journalists included. The figure that circulates in Spanish-press headlines —"training GPT-5 costs so many millions"— is always just one piece of the puzzle, and almost never the most expensive. How much does it really cost to make one of the big models? The answer shifts the debate from "what does the model do?" to "who can afford to do it?", which is the question the sector would prefer we didn't ask. My take: if you don't understand the zeros, you don't understand why there are only five or six serious competitors in the world.
I've spent two years getting hooked on the technical notes of Epoch AI, a British organisation devoted precisely to estimating what the labs don't publish. Once you get used to their breakdowns, the headlines become legible. Without those breakdowns, the figure of "a billion" or "five hundred million" blends in with a footballer's salary and turns abstract. The intent of this article is to break that abstraction.
The chips, the floor of the bill
A frontier model of 2025 is trained on tens of thousands of Nvidia GPUs working in parallel for weeks or months. The reference GPU for the GPT-4 to GPT-4.5 generation of models was the H100. The next generation —already in full deployment— uses H200 and B200/B300, successively more powerful and more expensive.
The nominal price of an H100 runs from 25,000 to 31,000 dollars per unit, depending on the integrator and the sales format. The most informative thing, though, isn't the purchase price but the hourly rental price in the cloud. There the numbers tell the story clearly. At the start of 2024 an H100 rented for around 8 dollars an hour; by the end of 2025 it had dropped to the 1.50-to-3-dollar range; in March 2026 the one-year contracts were rising again, according to the SemiAnalysis index, to about 2.35 dollars an hour. The swing is brutal and explains why training-cost figures published on different dates are hard to compare.
To get a sense of the order of magnitude, training GPT-3 in 2020 consumed around 1,300 megawatt-hours of electricity over thousands of earlier GPUs. Training GPT-4 in 2022-2023 consumed, according to Epoch AI's most conservative public estimates, around 100 million dollars in compute for the main run alone. The figure rose to several hundred million for the immediately following generation. And this is where the GPT-5 case gets interesting.
What Epoch AI says about GPT-5
In August 2025, Epoch AI published an analysis titled Why GPT-5 used less training compute than GPT-4.5 —available at epoch.ai/gradient-updates— that broke a very widespread mental schema. The intuitive hypothesis was that each generation doubles or triples the compute. But the Epoch team documented that GPT-5 had used less training compute than GPT-4.5, its immediate predecessor. The improvement came not from raw size but from algorithmic efficiency.
That doesn't mean GPT-5 was cheap. It means the figure for training the final model isn't the one that best describes total investment. The operating accounts leaked and reconstructed by several sources point to OpenAI having spent in 2024 around 3 billion dollars on model-training compute —including all the experiments prior to the definitive model— plus 1 billion more in research compute, adding up to roughly 5 billion dollars of R&D spending on compute alone. Epoch's projection for 2025 puts that figure near 9 billion.
The difference between "the cost of the final run" and "total R&D compute spending" is on the order of ten times. When a journalist writes that GPT-5 cost 500 million, they're reporting roughly the first figure. And they're omitting everything else.
What doesn't make the headline
There are seven line items the headline omits and that are worth having in front of you.
Prior research. Before the final run there are dozens or hundreds of failed pretraining experiments. Each costs between hundreds of thousands and tens of millions. The sum is probably greater than the definitive model. It's the most important piece and the one hardest to see from outside.
Salaries. A frontier lab employs between five hundred and fifteen hundred engineers and scientists. The average US salary for a senior AI researcher in 2025 runs from 800,000 to 1.5 million dollars a year in total compensation, according to the public ranges of OpenAI, Anthropic, Google DeepMind and Meta AI offers. Multiply and look.
Data. The licences OpenAI signed with News Corp, Axel Springer, the Financial Times, Le Monde, the Associated Press and other publishers add up, according to the combined estimates of Press Gazette and Reuters, to figures that summed already approach a billion dollars in multi-year commitments. That's only what's public. There are other line items in human-annotated data for reinforcement learning with human feedback, in safety tests commissioned to contractors, and in synthetic data generated with the lab's own compute.
Inference. Once the model is trained, it has to be served. ChatGPT serves, according to OpenAI's most recent public figures, around 700 million weekly active users in 2025. Each query to a big model consumes electricity and GPU time. SemiAnalysis and other analysts estimate that the annual inference cost for a company like OpenAI already equals or exceeds the annual training cost.
Safety and evaluation. The red-team teams —humans who try to break the model before it's released— consume their own budget. Automatic evaluations against dozens of benchmarks add compute. External audits, the agreements with safety institutes —AISI in the UK, AISI in the US—, the regulatory lawyers. All of that weighs.
Infrastructure. Building a data centre with the electrical density required for clusters of a hundred thousand GPUs requires civil works, dedicated power generation, liquid cooling, low-latency networks, proprietary orchestration software. Microsoft announced in 2024 an 80-billion-dollar investment in data centres during fiscal year 2025. The figure has been revised upward since then.
Amortised prior capital. GPUs aren't bought for one model: they're bought for five years of use. When the training cost is reported it's calculated by multiplying GPU-hours consumed by the theoretical hourly price, ignoring the real amortisation of the asset and the financial costs of the immobilised capital.
Any figure that doesn't include these seven line items is an incomplete figure. Probably fine for headlines. No good for understanding the sector.
Inference, the hidden cost
It's worth pausing on inference because it's the piece that most changes the public debate. When someone says "the free model," they're describing a situation that isn't free for anyone.
Every time a ChatGPT user sends a message, OpenAI's servers process that message through a model of hundreds of billions of parameters, generate an answer, and that consumes electricity and hardware depreciation. The marginal cost per query is very small —cents of a cent— but multiplied by hundreds of millions of users and dozens of queries per day per user, it gives astronomical figures.
The public estimates of SemiAnalysis and other analysts put OpenAI's annual inference cost in 2025 in the range of several billion dollars, a figure comparable to the annual training cost. Anthropic and Google are in the same order of magnitude, adjusted to their user base. This means the free model of the marketing is, in reality, the model subsidised by capital from Microsoft, Amazon, Google and the investors who still trust that one day the economics will add up.
Who can pay the entry fee
As of today, the actors that have trained a frontier model in the last twelve months are half a dozen. OpenAI with Microsoft. Anthropic with Amazon and Google. Google with DeepMind. Meta with its own AI division. DeepSeek with the financial backing of the quant fund High-Flyer and, presumably, with Chinese state support. Alibaba with Qwen. xAI with Elon Musk's personal funding. Mistral, in a clearly lower division but with a niche of its own.
That list is so short not through bad luck, but because the sum of the seven line items above exceeds ten billion dollars a year in sustained spending, and only the big hyperscalers —Microsoft, Amazon, Google— and the few competitors able to pay comparable deals have that muscle. The barrier to entry isn't set by the algorithm. It's set by the bank account.
The DeepSeek blow and what it changes
In January 2025, DeepSeek published its R1 model with the claim that the final run of its predecessor V3 had cost 5.6 million dollars in compute. The figure lit up social media —and, let's recall, generated the biggest stock-market drop in Nvidia's history on 27 January 2025. The figure was counted with honest but partial criteria: it was the cost of the final run in GPU-hours multiplied by the theoretical hourly price of an H800, excluding everything else —prior experimentation, amortised infrastructure, personnel, data. The real total cost of the project was surely on the order of hundreds of millions, funded by the founder's own quant fund, High-Flyer, run by Liang Wenfeng.
The 5.6-million figure, deceptive headline though it was, had a true core. DeepSeek's algorithmic efficiency —more aggressive mixture of experts, training with less but better-curated data, reinforcement learning with outcome verification— demonstrated that you could train a competitive model with an order of magnitude less compute. If that efficiency curve continues, the barrier to entry comes down too. And if the barrier comes down, the sector stops being six actors and becomes fifteen or twenty.
We don't yet know whether that curve will hold. But the effect on the debate is already felt. Investors now look less at pure training figures and more at compute efficiency per benchmark point. The US Bureau of Industry and Security, in charge of chip-export controls to China, is reviewing whether the controls produced the opposite of the intended effect: forcing the Chinese labs to innovate precisely in efficiency, and then export that efficiency through public papers.
The political question
This is personal opinion, but it's held up by the data I've been listing. The cost structure we've just seen turns generative AI into an inevitably oligopolistic sector. Not through conspiracy, not through regulatory bad faith, but through economic physics: there's a minimum capital threshold to take part, and that threshold sits at ten billion dollars a year sustained over several years in a row.
This has consequences worth naming without drama. The first is that national regulation has limited capacity: if the six or seven actors able to train frontier are almost all in the United States and China, what European regulators decide about AI affects use more than production. The second is that the debate about democratising AI through open-weight models is only operational if training costs come down. If they don't come down, the openness is rhetorical: the weights are open but no one can retrain them. The third is that venture capital and the hyperscalers have fused their interests, which is financial news as much as technological.
What I'd say to the Spanish press and to the public conversation: every time someone writes that "training model X cost so many millions," the immediate question should be "which line items were included in that figure?" Probably the answer is "only the final run." And that turns the data into a decorative piece, not useful information.
The latest firm data available, gathered by Epoch AI in its analysis Training compute costs are doubling every eight months for the largest AI models published in May 2025, shows that the training cost of leading models has multiplied between 2 and 3 times a year over the last eight years. If that curve holds, the top models of 2027 will cost more than a billion dollars in their final run alone, and the actors able to do it will be four or five, not ten.
Definitions
Final-run cost: spending on compute for the training session of the definitive model, calculated as GPU-hours multiplied by the hardware's theoretical hourly price. It's the figure usually cited in headlines. It excludes everything else.
Compute amortisation: compute allocated to general research and to experiments prior to the final model. In OpenAI's operating accounts it represented, in 2024, a line item on the order of a billion dollars.
Pretraining experiment: an exploratory training run that tests an architectural hypothesis before the final model. They tend to be smaller-scale but add up, in aggregate, to more compute than the definitive model.
Inference: the compute consumed by each query to an already-trained model. For mass services like ChatGPT, its aggregate annual cost equals or exceeds that of training.
References
Epoch AI, How much does it cost to train frontier AI models? (blog, 2024, ongoing update). Public estimates of training cost per model and historical trend.
Epoch AI, Why GPT-5 used less training compute than GPT-4.5 (but GPT-6 probably won't) (gradient-updates, August 2025). Analysis of the specific case and of algorithmic efficiency.
Epoch AI, Training compute costs are doubling every eight months for the largest AI models (data-insights, May 2025). The cost curve and projection to 2027.
Epoch AI, Most of OpenAI's 2024 compute went to experiments (data-insights, 2025). Breakdown of OpenAI's reconstructed operating accounts.
SemiAnalysis, H100 vs GB200 NVL72 Training Benchmarks (newsletter, 2024-2025). Data on hardware, prices and operating profitability.
SemiAnalysis, The Great GPU Shortage – Rental Capacity (newsletter, March 2026). The H100 hourly-price index and the dynamics of the rental market.
Stanford HAI, Artificial Intelligence Index Report 2026 (April 2025). The chapter on compute and investment dynamics in frontier models.
Press Gazette, coverage of OpenAI's licence agreements with publishers (2024-2025). Source of the figures on the deals with News Corp, Axel Springer, FT and others.
To go deeper
Daron Acemoglu & Simon Johnson, Power and Progress (PublicAffairs, 2023). A historical framework on how general-purpose technologies concentrate benefits and why the direction isn't neutral.
Mariana Mazzucato, The Entrepreneurial State (Anthem Press, 2013). An analysis of the role of public capital in technological R&D, useful for understanding why the hyperscalers have occupied the space that in other sectors the State occupies.
Vili Lehdonvirta, Cloud Empires (MIT Press, 2022). How the cloud oligopoly that today pays the AI bill was built.
Cade Metz, Genius Makers (Dutton, 2021). A narrative history of the sector, useful for understanding how six labs ended up setting the pace.
You might also like
- Epoch AI says the data runs out in 2028
- What the AI Index Report 2026 says that the Spanish press doesn't tell you
- DeepSeek and the end of the American monopoly
- Nvidia is worth more than the economy of Spain
Comments0
No comments yet.
Leave a comment