Every time you type a question into an AI chatbot, you are standing at the top of one of the most expensive supply chains ever built. Below your $20-a-month subscription — or your free account — sits a stack of costs that runs through data centers the size of small towns, chips that sell for more than a car, electricity demand rivaling entire countries, and capital spending that now exceeds what the world invests in oil and gas production.

This article follows the money through five layers: the hardware, the training, the running costs, the developer pricing, and finally the end user. Every figure below comes from company filings, official pricing pages, or research organizations that specialize in tracking AI economics — and where a number is an estimate rather than a disclosed fact, it is labeled as one.

Layer 1: The hardware — a $725 billion year

The foundation of the AI economy is physical: chips, servers, networking, buildings, and power. The scale of spending here is genuinely without precedent in corporate history.

The four largest cloud companies — Amazon, Microsoft, Alphabet (Google), and Meta — plan to spend roughly $725 billion combined on capital expenditure in 2026, up about 77% from approximately $410 billion in 2025, according to figures compiled by the Financial Times from company earnings reports. Amazon leads at roughly $200 billion, Microsoft has guided to about $190 billion for the calendar year, Alphabet to $175–185 billion, and Meta to $115–135 billion. The overwhelming majority of this money goes to AI data centers, GPUs, custom chips, and the power infrastructure to run them.

To put that in perspective, the International Energy Agency noted in its April 2026 report that the capital expenditure of just five large technology companies exceeded $400 billion in 2025 — more than global investment in oil and natural gas production. Goldman Sachs now projects a combined $5.3 trillion in capex from the four biggest hyperscalers between fiscal 2025 and 2030.

Most of that money flows toward one company. NVIDIA, which makes the GPUs that train and run most AI models, reported $215.9 billion in revenue for its fiscal year 2026 (which ended in January 2026), up 65% from the year before. In its most recent quarter alone (ended April 26, 2026), NVIDIA reported $75.2 billion in data center revenue — up 92% year over year — at gross margins around 75%. Analyst estimates put the production cost of a high-end Blackwell GPU at roughly $3,000–3,500, against selling prices that can exceed $30,000 for top configurations. Those figures are estimates, not NVIDIA disclosures, but the company's published ~75% gross margin tells the same story: the chip layer is currently the most profitable part of the entire AI stack.

The buildings matter too. Research group Epoch AI estimates that a typical AI data center with 1 gigawatt of IT power costs roughly $38 billion in up-front capital expenditure, and that gigawatt-scale facilities take about two years to build. The Stargate project — a joint venture between OpenAI, SoftBank, and Oracle announced in early 2025 — plans around $500 billion of AI data center infrastructure in the United States, the largest private infrastructure commitment in AI history.

Layer 2: Training — from $40 million to $1 billion per model

Training a frontier AI model means running tens of thousands of chips continuously for weeks or months. Companies almost never publish exact training costs, so the best numbers come from independent researchers — chiefly Epoch AI, whose cost models underpin the Stanford AI Index.

Epoch AI's peer-reviewed analysis estimated the amortized hardware and energy cost of GPT-4's final training run at about $40 million, and Google's Gemini Ultra at about $30 million. Using an alternative method based on cloud rental prices — the approach used in the Stanford AI Index — the estimates are roughly twice as high: about $78 million for GPT-4 and $191 million for Gemini Ultra. The difference between methods is itself a useful lesson: "how much did it cost to train model X" has no single true answer, only well-documented estimates.

What is not in dispute is the direction. Epoch AI found that the amortized cost of training the most compute-intensive models has grown at roughly 2.4x per year since 2016, and projected that if the trend continues, the largest training runs will cost more than $1 billion by 2027. Industry estimates place the 2026 generation of frontier training runs in the hundreds of millions of dollars. Anthropic CEO Dario Amodei predicted this trajectory publicly back in 2024, describing training runs approaching a billion dollars.

It is also worth knowing what the money actually buys. According to Epoch AI's breakdown, hardware (chips, servers, interconnect) accounts for roughly 47–67% of a model's total development cost, research staff for 29–49%, and — perhaps surprisingly — energy for only 2–6%. The final training run is also just a fraction of the total: most of the budget goes to the experiments and failed runs that come before it.

Layer 3: Inference — the electricity bill of the AI age

Training happens once; inference — actually answering queries — happens billions of times a day, and it is where AI meets the power grid.

The IEA's benchmark data: global data centers consumed about 415 terawatt-hours of electricity in 2024, roughly 1.5% of world electricity consumption. In 2025, data center electricity demand grew 17% — far outpacing the 3% growth in overall global electricity demand — and consumption by AI-focused data centers specifically surged about 50%. The IEA's base case projects total data center consumption roughly doubling to around 945 TWh by 2030 (slightly more than Japan's entire current electricity consumption), with AI-specific demand tripling. The United States accounts for about 45% of global data center electricity use, and by 2030 is projected to consume more electricity for data centers than for producing aluminium, steel, cement, and chemicals combined.

There is, however, a powerful counter-trend that alarmist coverage often skips. The IEA reports that energy efficiency per AI task has been improving at a pace unprecedented in the history of energy technology — energy use per task has been dropping by at least an order of magnitude annually in recent years. A simple text query today typically consumes less electricity than running a television for the same amount of time. Total demand keeps rising anyway, because usage is growing even faster than efficiency improves — more users, and more energy-hungry applications like AI agents.

Layer 4: What developers pay — the token economy

This is the one layer where prices are fully public, because AI companies publish them. Businesses that build on AI models pay per "token" — roughly three-quarters of a word — with separate rates for input (what you send) and output (what the model generates), priced per million tokens (MTok).

Published rates as of early July 2026:

Two things stand out. First, the spread: the cheapest capable model costs less than 2% of the most expensive flagship per token, which is why serious AI products route easy tasks to cheap models and reserve premium models for hard ones. Second, the direction: Epoch AI's tracking shows the cost of running a model at a fixed level of performance has been collapsing — falling by between 9x and 900x per year depending on the capability level. Yesterday's frontier intelligence rapidly becomes today's commodity.

Standard discounts amplify this: batch (non-urgent) processing typically costs 50% less, and prompt caching can cut input costs by up to 90%. Menlo Ventures estimates enterprise spending on generative AI reached $37 billion in 2025, up from $11.5 billion in 2024.

Layer 5: What you pay — and why it's so cheap

At the top of the stack sits the consumer, and the pricing here is strikingly uniform: free tiers with usage limits, a standard tier at about $20 per month (ChatGPT Plus, Claude Pro, Gemini's paid tier), and premium tiers at $100–200 per month for heavy users.

Twenty dollars a month is a remarkable price for something running on billion-dollar infrastructure — and that is the point. Consumer AI is priced for market share, not margin. OpenAI reported more than 900 million weekly ChatGPT users as of early 2026, with roughly 50 million paid consumer subscribers by reported figures — meaning the overwhelming majority of users pay nothing at all. The falling cost of inference (that 95%+ decline since GPT-4's 2023 launch, per industry reporting) is what makes the free tier economically survivable; reported gross margins on subscription tiers have turned positive even as companies lose money overall.

Does the money add up?

Here is the tension that defines AI economics in mid-2026. On one side: roughly $725 billion in hyperscaler capex this year alone, trillions projected through 2030. On the other: the actual revenues of AI companies, which are growing explosively but remain an order of magnitude smaller.

OpenAI reached approximately $25 billion in annualized revenue by early 2026 — about $2 billion per month, up from $3.7 billion in all of 2024 — yet is reportedly projected to lose around $14 billion in 2026, according to internal documents cited by The Information. Anthropic announced in April 2026 that it had passed a $30 billion annualized run rate, up from about $1 billion at the start of 2025, with roughly 80% of revenue coming from business customers. (Estimates of both companies' revenues vary between research firms; the figures above are the companies' own reported numbers.) Menlo Ventures' $37 billion estimate for all enterprise generative AI spending in 2025 sits against $410 billion of hyperscaler capex that same year.

Whether this is a rational build-out ahead of demand — the way railways and fiber-optic cables were eventually filled — or an overshoot that ends in write-downs is the single most debated question in technology and finance right now. The revenue growth rates are real and historically unprecedented. So is the gap. Both facts can be true at once, and honest analysis holds them together rather than picking a side.

What is certain is the shape of the machine: money flows from hundreds of millions of users paying $0–20 a month, through businesses paying dollars per million tokens, to AI labs spending hundreds of millions per model, to cloud giants spending hundreds of billions on data centers, and finally to the chipmakers — where, for now, most of the profit in artificial intelligence actually lands.

Disclosure note for readers: revenue run rates, training cost figures, and GPU production costs are estimates or annualized snapshots unless drawn directly from audited filings, and different research firms report different numbers. Where sources conflicted, this article used the companies' own disclosed figures and noted the disagreement.