Three separate costs are bundled under "the cost of AI": building a frontier model, paying to use a model, and supplying the electricity for the data centers that run both. Epoch AI estimates that the hardware and energy cost of the final training run of frontier models has grown about 2.4 times per year since 2016, while the price of buying a fixed level of capability has collapsed, according to Stanford's 2025 AI Index. Both trends are real, because they measure different things.
Key takeaways
- Training costs for the largest models rose roughly 2.4 times per year from 2016 on, according to Epoch AI's 2024 estimate, which projected runs above USD 1 billion by 2027.
- The price of reaching a fixed capability level fell by a factor of several hundred for GPT-3.5-level performance in under two years.
- A falling unit price does not guarantee a falling bill, because newer models can cost more per query than older ones.
- The International Energy Agency projects that electricity demand from data centers will more than double by 2030, and its totals cover far more than AI.
- Canada's public compute spending is CAD 2 billion over five years, announced in Budget 2024.
Three costs that are often mixed together
A figure about "AI costs" can refer to a developer's spending, a customer's invoice, or a grid operator's load forecast. The table separates them. It is a reading aid built from the sources listed below, not a ranking.
| Cost | Who pays it | Latest measured trend | Source |
|---|---|---|---|
| Training a frontier model | Model developers | Amortized hardware and energy cost of the final run up about 2.4 times per year since 2016 | Epoch AI, 2024 paper (revised February 2025) |
| Using a model at a fixed capability level | Businesses and consumers | Price for GPT-3.5-level performance fell by a factor above 280 from late 2022 to late 2024 | Stanford HAI, AI Index 2025 |
| Powering data centers | Operators, utilities, and ultimately ratepayers | Demand projected to more than double between 2024 and 2030 | International Energy Agency, 2025 |
What it costs to train a frontier model
The most cited academic estimate comes from Epoch AI researchers Ben Cottier, Robi Rahman, Loredana Fattorini, Nestor Maslej, Tamay Besiroglu, and David Owen, in The rising costs of training frontier AI models. First posted on May 31, 2024 and revised on February 7, 2025, it estimates the amortized cost of hardware and energy for the final training run. The growth rate is about 2.4 times per year since 2016, and the authors state that the largest runs would pass USD 1 billion by 2027 if the trend continued. That is a projection from a trend, not an observed bill.
The same work breaks total development cost for models such as GPT-4 and Gemini Ultra into hardware at 47 to 67 percent, research and engineering staff at 29 to 49 percent, and energy at 2 to 6 percent. Two cautions apply. The headline growth figure covers only the final run, while the breakdown also counts supporting experiments and staff time. Estimates for hardware that developers own, such as Google's TPUs, rest on estimated production costs and are more uncertain than estimates based on actual prices. Renting the same compute from a cloud provider gives figures about twice as high on average.
Why the price of using a model keeps falling
Stanford's Institute for Human-Centered AI tracks what customers pay. In the 2025 AI Index, the inference cost of a system performing at GPT-3.5 level fell by a factor above 280 from November 2022 to October 2024, a span of about two years. The report also points to yearly gains in hardware cost and energy efficiency as part of the explanation.
The comparison holds capability fixed. A buyer who wants the newest class of model faces a different price. A business that moves to a more capable model, or that sends far more requests because each one is cheap, can see total spending rise even as the price per unit falls. For that reason, cost reports are most useful when they state the task, the model, the volume, and the date.
The electricity bill
The third cost is power. The International Energy Agency's executive summary on energy and AI (2025) projects that electricity use by data centers will grow to more than twice its 2024 level by 2030, starting from a low single-digit share of global supply. The category covers every data center workload, from email to video streaming, so it should not be read as the footprint of chatbots alone. The exact terawatt-hour totals are in the report itself. How much of any increase reaches household electricity prices depends on local grids and tariffs, and none of the sources checked for this article settles that question for Canada or the United States.
Public money for compute in Canada
Costs are not carried only by companies. The federal government's Canadian Sovereign AI Compute Strategy commits CAD 2 billion over five years starting in 2024-25, announced in Budget 2024. Up to CAD 700 million goes to the AI Compute Challenge for domestic data-center capacity, up to CAD 1 billion to public supercomputing infrastructure, and up to CAD 300 million to the AI Compute Access Fund, which helps Canadian businesses and researchers buy compute. Budgets of this kind are commitments, so actual disbursement should be checked in departmental reports.
Questions to ask when reading a cost claim
- Which cost is measured: training, usage fees, staff, or electricity?
- Is the figure an estimate, a price list, or an audited expense?
- Does it cover the final run only, or experiments and failed attempts too?
- Which model, task, and date does a price comparison use?
- Is the currency stated, and is it adjusted for inflation?
- Does the headline number describe one company or the whole industry?
FAQ
Is AI getting cheaper or more expensive?
Both, depending on the measure. Reaching a fixed level of performance became far cheaper from 2022 to 2024, while the cost of training the very largest models kept rising.
How much did GPT-4 cost to train?
No official figure from the developer was found. Epoch AI gives only a percentage breakdown of development cost for models in that class, so any single dollar number for GPT-4 should be treated as an outside estimate.
Does AI use a large share of the world's electricity?
Not by itself. The IEA counts every kind of data center, including those serving many non-AI services, and puts their combined draw at a small fraction of world electricity supply.
Will training runs really cost USD 1 billion?
The estimate was a 2024 projection that assumed the growth trend would continue. Whether it happens depends on developer decisions that the paper does not predict.
Related reading
- How AI spending affects business growth
- Where companies report AI savings
- AI and sustainability
- AI's wider effects on society
More in AI Tools and Industry.
Last verified: October 2026
Sources
- The rising costs of training frontier AI models, Cottier et al., Epoch AI, arXiv:2405.21015, 2024 (revised 2025)
- How much does it cost to train frontier AI models?, Epoch AI, 2024 (updated January 13, 2025)
- Artificial Intelligence Index Report 2025, Stanford Institute for Human-Centered AI, 2025
- Energy and AI, International Energy Agency, 2025
- Canadian Sovereign AI Compute Strategy, Innovation, Science and Economic Development Canada, 2024-2026