Cost & Economics

D1 · StoryFor young children and total beginners

Guess which costs more: building the AI, or answering one question?

Making an AI costs a lot of money. Builders pay for special chips, big buildings full of computers, and lots of power.

Then each answer costs a little too. The computers must work every time someone asks a question.

So builders look for ways to do the same job with less. One trick is to save work they did before, so they need not do it twice.

Like: Building a bakery costs a lot once; every loaf after that costs a little.

D2 · PictureFor curious beginners

Costs come in two shapes. Build cost is paid once per model: chips, data, staff and the training run. Run cost is paid on every request: the computing behind each answer, usually priced per million tokens.

Training the most compute-hungry models has grown about 2.4 times more expensive each year since 2016. Chips and staff are the biggest items; energy is only 2 to 6 percent.

Run cost grows with use and with how a question is answered. Anthropic measured its own agents using about 4 times as many tokens as a chat, and its multi-agent research system about 15 times.

Much optimisation work is cost work: batching many requests together, caching repeated text (on one major platform, cached input is billed at a tenth of the normal price or less), and smaller or compressed models.

Like: Building a bakery costs a lot once; every loaf after that costs a little.
Where it breaks: A loaf costs about the same every time. An AI answer can cost a hundred times more or less depending on its length, the model, and whether an agent loops many times to produce it.

D5 · FrontierFor experts

Per-token prices make AI look like a metered utility, but the bill depends on how each answer is produced and who pays for the build.

Reading the headline. The capital-spending figure mixes AI with everything else the four companies build, and filings don't split it by layer.

Training costs. If costs keep rising 2.4 times a year, the largest runs would pass a billion dollars by 2027, leaving frontier development to the best-funded organisations (Epoch AI, 2024).

Cost moves to every answer. Recent capability gains come largely from post-training and extra computing at the time of use, which is paid per request (International AI Safety Report, 2026).

Agents multiply tokens. Anthropic measured its own agents using about 4 times the tokens of a chat and its multi-agent research system about 15 times; token use alone explained 80% of performance differences on one browsing test (Anthropic, 2025).

Energy. A small share of training cost but a growing share of electricity: data-centre use is projected to more than double to around 945 TWh by 2030 (IEA, 2025).

On every layer

• Layer 9 · Applications & Interfaces: Every request costs computing, so an AI app's bill grows with its use; in 2023 many AI start-ups were spending over 80% of the money they raised on compute. (Navigating the High Cost of AI Compute (Andreessen Horowitz))
• Layer 8 · Agents: Agents loop many times per task: Anthropic measured its own agents using about 4 times the tokens of a chat, and its multi-agent research system about 15 times. (How we built our multi-agent research system (Anthropic))
• Layer 7 · Orchestration & Tooling: Long, repeated instructions and documents are a big part of the bill; prompt caching bills repeated input at a tenth of the normal price or less. (Prompt caching (Claude API documentation))
• Layer 6 · Inference & Optimization: Better memory management lets a server batch more requests at once: PagedAttention raised throughput 2 to 4 times at the same latency, lowering the cost of each token. (Efficient Memory Management for Large Language Model Serving with PagedAttention)
• Layer 5 · Foundation Models: The cost of training the most compute-hungry models has grown about 2.4 times a year since 2016. (The rising costs of training frontier AI models (Epoch AI))
• Layer 4 · Data: Human labels are slow and costly to collect; training from AI feedback against a written list of principles is one way to need far fewer of them. (Constitutional AI: Harmlessness from AI Feedback)
• Layer 3 · Cloud & Infrastructure: Clouds rent accelerators by the hour or in reserved blocks of days or weeks, turning a large upfront purchase into a running cost. (Amazon EC2 User Guide: Capacity Blocks for ML)
• Layer 2 · Compute Hardware: Accelerator chips and staff are the two biggest costs of training a frontier model, each reaching tens of millions of dollars. (The rising costs of training frontier AI models (Epoch AI))
• Layer 1 · Physical & Energy: Electricity is a small part of what a frontier training run costs, about 2 to 6 percent; chips and staff cost far more. (The rising costs of training frontier AI models (Epoch AI))

Examples as of 2026-10-30

• Big Tech capital spending: Microsoft, Alphabet, Amazon and Meta
• Per-token API price lists from OpenAI, Anthropic and Google
• Hourly GPU rental from hyperscalers and neoclouds
• The Stargate data-centre programme (OpenAI, Oracle, SoftBank)

Sources

1. The rising costs of training frontier AI models (Epoch AI) (2024)
2. How we built our multi-agent research system (Anthropic) (2025-06-13)
3. Prompt caching (Claude API documentation) (2026)
4. International AI Safety Report 2026 (2026-02-03)
5. IEA: Energy and AI, executive summary (2025)
6. Navigating the High Cost of AI Compute (Andreessen Horowitz) (2023-04-27)