AI is getting cheaper. The things it runs on are getting scarcer.
When something becomes abundant, the money moves to whatever is still scarce. This page tracks both sides: the falling price of machine intelligence, and the prices of the physical inputs it needs. Then it checks which scarce inputs the funds have already noticed.
The same capability costs a thousand times less than in 2023, and the fall is speeding up
Each line follows one fixed level of capability and shows the cheapest model that reached it, priced per million tokens. Capability is measured on Epoch AI's Capabilities Index, which combines many benchmarks into one score.
Price is a blend of 3 input tokens to 1 output token. Prices to early 2025 are launch prices from Epoch AI; later ones are today's cheapest listed price on OpenRouter, which for open models is a third-party host. The GPT-4 line has not fallen since August 2025, which suggests a floor near the cost of serving a small model. The best models today (Claude Opus 5.5 at $8, Gemini 4 Argon at $4) cost less than GPT-4 did at launch while scoring 41 points higher.
Each unit got cheaper, and total spending went up twelvefold
When a useful thing gets cheaper, people tend to use so much more of it that total spending rises. That is happening here. The five biggest buyers of AI capacity now spend more on buildings and equipment in one quarter than they did in all of 2019, two and a half times over.
Cash paid for property and equipment each calendar quarter, from each company's 10-Q and 10-K as filed in SEC XBRL: Microsoft, Alphabet, Amazon, Meta and Oracle. Oracle's fiscal quarters end a month earlier and are placed in the nearest calendar quarter. Finance leases are not included, so the true figure is higher.
If AI keeps its trend, US AI power has to reach about 119 GW by the end of 2028. The grid can support about 61
Installed AI compute grew 2.6 times in the last year, and frontier training compute has grown 4 to 5 times a year for a decade. Chips get about 1.4 times more efficient each year, so the power AI needs grows about 1.9 times a year. Following that line forward shows what the physical world has to supply, and where it cannot.
History is the power of AI chips sold since 2022 (Epoch AI) times 1.75 for servers, networking and cooling, times the US share of 77.5%. The trend band and the supply band each come from 20,000 runs over the ranges in the assumptions table below; lines are medians and bands run from the 10th to the 90th percentile. Supply counts the gas, batteries, solar, wind and reactor restarts already planned in EIA's generator inventory, minus retirements and other load growth. The dots are the 77 US campuses Epoch AI tracks, which are the biggest AI sites but not all AI capacity.
Assumptions behind the odds
When the grid is the limit, the money has gone to ways around it
If the grid can carry only about half the power AI wants, the scarce thing is power that does not wait in the grid queue. Over the last year the market paid for exactly that. Makers of turbines, fuel cells and engines that data centers install on site rose. Owners of existing plants fell, because what they can charge is limited by a capped capacity market, cheaper Texas power and new state tariffs for data centers.
The whole chain: where it binds, and whether the funds have noticed
Each row is one link in the chain an AI data center needs, in the order it binds. The odds come from the scaling model above: the share of runs in which that link falls short by the end of 2027 and 2028 if AI keeps its trend. The price number is that link's price change over the last year, against the year before. The second is the share of all 13F filers holding at least one of its listed companies, June 2025 to June 2026. Tap a row for the detail.
| Link | Binds on the trend | Price, last 12 months | Share of funds holding | Reading |
|---|
Measured, inferred and guessed
Measured
- Prices of transformers, switchgear, turbines and cable (BLS producer prices), copper and uranium (IMF), and consumer electricity and computers (BLS).
- Big Tech capital spending, from SEC filings.
- How many 13F filers hold each company, from SEC's quarterly data sets.
- The cheapest price for a fixed capability score, from Epoch AI scores and listed prices.
- AI chips sold and their power draw (Epoch AI), and the power plants planned and under construction (EIA).
Inferred
- That price rises mean scarcity. They also reflect metals, tariffs and general inflation.
- That AI is the cause. Transformer prices rose 61% from 2019 to 2022, before ChatGPT, from grid upgrades and supply chains.
- Memory scarcity, which is read from consumer computer prices because no free memory-price series exists.
Guessed
- The ranges in the scaling model: how much grid power each chip watt needs, how much new AI load lands on gas, copper and workers per megawatt, and how fast supply grows. Each is a range rather than a single number, and the odds are only as good as those ranges.
- That the trend holds at all. The odds say which link breaks first if it does, not whether it will. A binding link slows the trend rather than stopping AI.
Not shown yet
- Whether "scarce but uncrowded" stocks go on to beat the market. Most such rules fail a real backtest, so this needs one before it is trusted. The data for it exists: price series back decades and 13F filings back to 2013.
- Electrician wages, transformer lead times and grid connection queues as live series. No current free source was reachable, so they enter the model as fixed figures from the reports listed below.
Researched and built by Claude (Anthropic). Every number links to its source, the odds are model estimates, and nothing here is investment advice.