InsightsAI & Technology

The Economics of Training Large AI Models

Idrak Insights TeamJuly 27, 202613 min read

Hardware · Staff · Energy

Key takeaways

  • Compute-cost estimates put GPT-4's training at around $78 million and Google's Gemini Ultra at around $191 million, and Anthropic's CEO has said current frontier runs cost between $100 million and $1 billion.
  • Hardware isn't the only major line item. Epoch AI's detailed cost research found R&D staff costs, including equity, make up 29-49% of a frontier model's total development cost, nearly as much as hardware itself.
  • Training costs have grown at roughly 2.4x per year since 2016. If that continues, the largest training runs are projected to exceed $1 billion by 2027.
  • DeepSeek's widely reported $5.6 million figure for DeepSeek V3 excluded infrastructure, experimentation, and failed runs, and doesn't represent an apples-to-apples comparison with fully-loaded frontier model costs.
  • A single company, Nvidia, supplies more than 60% of the world's AI compute, meaning the economics of training frontier models are also, in large part, the economics of one hardware supply chain.

Introduction

Headlines throw around AI training cost figures constantly, $78 million here, "over $1 billion" there, often without saying what's actually included in the number. That inconsistency matters, because two very different measurements get flattened into the same casual claim: the cost of raw compute for a single training run, and the fully-loaded cost of developing a model, including the people who built it. This guide breaks down what's actually being spent, where the money goes, how fast it's rising, and why understanding this matters even if your organization never trains a model of its own.

What "Training Cost" Actually Means

Most headline figures for AI training costs describe compute alone: the cost of renting or amortizing the GPU time needed for the final training run. That's a real number, but it's not the whole picture. The most rigorous public research on this topic, a detailed cost model built by Epoch AI, in collaboration with the data behind Stanford's AI Index, breaks full model development costs into three categories: hardware (the chips, servers, and networking equipment), R&D staff costs (including salaries and equity compensation), and energy consumption. Reported figures that only cite "compute cost" are typically describing just the hardware slice of a much larger total.

Where the Money Actually Goes: A Cost Breakdown

Where AI training money actually goes

  • Hardware 57%

    Chips, servers, and interconnect. Epoch AI's published range is 47-67%.

  • R&D Staff 39%

    Salaries and equity compensation. Published range is 29-49%.

  • Energy 4%

    Electricity for the training run itself. Published range is 2-6%.

Percentages are midpoint estimates from Epoch AI's published ranges for frontier models such as GPT-4 and Gemini Ultra. Actual splits vary by model and organization.

The staff figure tends to surprise people. Coverage of AI costs fixates on GPUs and data centers, but paying the senior researchers, infrastructure engineers, and safety teams capable of building a frontier model is nearly as expensive as the hardware itself, sometimes more, when equity compensation is factored in.

How Fast Costs Are Rising

Training cost by model

Estimated training cost, millions of USD

GPT-4$78M

Compute cost estimate, Stanford AI Index / Epoch AI.

Meta Llama 3.1 405B$170M

Estimate compiled from Stanford AI Index / Epoch AI data.

Google Gemini Ultra$191M

Estimate, Stanford AI Index / Epoch AI.

DeepSeek V3$5.6M

Self-reported and not directly comparable, see the caveat below.

Epoch AI's research found that the amortized hardware and energy cost for the final training run of frontier models has grown at roughly 2.4x per year since 2016 (90% confidence interval: 2.0x to 2.9x). Extrapolated forward, that trajectory puts the largest training runs on track to exceed $1 billion by 2027. Anthropic CEO Dario Amodei has said publicly that current frontier developers are likely spending close to $1 billion on a single training run, with runs of up to $10 billion possible within the next two years.

The Outlier That Made Headlines: DeepSeek

DeepSeek's claimed $5.6 million training cost for DeepSeek V3 generated enormous attention as apparent proof that frontier-level AI could be built far more cheaply than assumed. The figure deserves real context before being treated as comparable to the numbers above. Reporting from TechCrunch noted the figure explicitly excluded infrastructure costs, experimentation runs, and failed training attempts, all of which are included in the Epoch AI and Stanford estimates for other models. OpenAI also reportedly investigated whether DeepSeek used distillation, training a new model using outputs from an existing one, a meaningfully cheaper approach than training entirely from scratch. None of this means DeepSeek's achievement wasn't real, but the $5.6 million figure and the $78-191 million figures above aren't measuring the same thing.

Who Actually Supplies the Compute

Who supplies AI compute

  • Nvidia 60%

    Over 60% of total AI compute.

  • All other providers 40%

    Under 40% combined. Google and Amazon supply most of this remainder, with Huawei holding a small but growing share.

Source: Stanford HAI, 2026 AI Index Report.

This concentration matters economically as much as technically: when one supplier controls the majority of the hardware behind an entire industry's training runs, that supplier's pricing, production capacity, and export policies directly shape what frontier AI development costs, and who can afford to compete at all.

The Cost Nobody Puts on the Invoice: Carbon

Estimated training emissions by model

Estimated tons of CO2-equivalent

GPT-45,184 t

Estimated tons of CO2-equivalent.

Meta Llama 3.1 405B8,930 t

Estimated tons of CO2-equivalent.

xAI Grok 472,000 t

Estimated at over 72,000 tons of CO2-equivalent.

Source: Stanford HAI, 2026 AI Index Report. Stanford notes these figures are estimates and should be interpreted with caution.

None of the dollar-cost estimates above include this. The scale of the increase, roughly 14 times GPT-4's estimated footprint for the most recent frontier training run Stanford tracked, is a reminder that the economics of AI training and its environmental cost are rising together, not separately.

Where the Numbers Are Headed

Growth trajectory at a glance

01

2.4x

per-year growth in amortized training costs since 2016

Epoch AI

02

$1B+

projected cost of the largest training runs by 2027, if current trends hold

Epoch AI

03

$100M-$1B

estimated cost of current frontier training runs

Dario Amodei, Anthropic

04

Up to $10B

potential training run costs within two years

Dario Amodei, Anthropic

Why This Matters Beyond AI Labs

Almost no business will ever train a frontier model from scratch, and given the numbers above, that's the economically rational choice, not a limitation. The overwhelming majority of organizations are, and should be, buyers of AI capability rather than builders of foundation models: accessing frontier intelligence through APIs and subscriptions rather than shouldering nine-figure training budgets themselves. Understanding this cost structure still matters, though. It explains why enterprise AI pricing looks the way it does, why only a handful of organizations can meaningfully compete at the frontier, and why the compute supply chain, dominated by a single chip supplier, has become a genuine strategic consideration for any company building a long-term AI roadmap.

Sources

  • Cottier, Rahman, Fattorini, Maslej, Besiroglu & Owen (Epoch AI), "The Rising Costs of Training Frontier AI Models" (arxiv.org, 2405.21015)
  • Epoch AI, "How Much Does It Cost to Train Frontier AI Models?" (epoch.ai)
  • Stanford HAI, The 2026 AI Index Report, including the Research and Development chapter (hai.stanford.edu/ai-index)

This article reflects publicly available research as of July 2026. AI training cost and emissions estimates are inherently approximate, model developers rarely disclose exact figures, and analysts caution these numbers should be interpreted as informed estimates rather than precise accounting.

Frequently Asked Questions

How much does it actually cost to train a frontier AI model?+

Estimates for compute costs alone place GPT-4 at around $78 million and Google's Gemini Ultra at around $191 million, according to Stanford's AI Index and Epoch AI research. Anthropic CEO Dario Amodei has said current frontier training runs cost between $100 million and $1 billion.

What's actually included in AI training cost estimates?+

Epoch AI's research breaks the full cost of developing a frontier model into three categories: hardware, including chips, servers, and interconnect (47-67% of total cost); R&D staff costs including equity compensation (29-49%); and energy consumption (2-6%). Estimates citing only compute cost typically describe just the hardware slice.

Why did DeepSeek's training cost look so much lower than other models?+

DeepSeek's widely reported $5.6 million figure for DeepSeek V3 excluded infrastructure, experimentation, and failed training runs, according to TechCrunch's reporting. OpenAI also reportedly investigated whether DeepSeek used distillation from existing models, a fundamentally cheaper approach than training from scratch.

How fast are AI training costs rising?+

Epoch AI's research found the amortized hardware and energy cost for the final training run of frontier models has grown at roughly 2.4 times per year since 2016. If that trend continues, the largest training runs are projected to exceed $1 billion by 2027.

Who actually supplies the computing power behind AI training?+

According to Stanford's 2026 AI Index Report, Nvidia accounts for over 60% of total AI compute, with Google and Amazon supplying much of the remainder and Huawei holding a small but growing share.

Does the cost of training an AI model include its environmental impact?+

Not in the dollar figures typically reported. Stanford's 2026 AI Index estimated that training a model like xAI's Grok 4 generated over 72,000 tons of carbon-equivalent emissions, compared to an estimated 5,184 tons for GPT-4, a cost that doesn't appear on any training budget.

Advisory

Deciding where AI investment actually belongs in your roadmap? We can help.

Talk to our team