The CTI in plain words: the Compute Cost and Capacity indices

The Compute Terminal Index, or CTI, is a family of economic indicators for the AI build-out. Two of them measure the physical side of the market. The CTI Compute Cost Index follows what it costs to rent a GPU-hour, every business day. The CTI AI-Factory Capacity Index follows how much AI datacentre power is switched on, every Monday. Both are indexed to their level on 5 January 2026, and both have been published since 11 November 2026. A third member, the CTI AI-Factory Equity Index, measures the listed companies behind the build-out and has its own piece.

This piece explains in plain words how the first two are made. The live values are on the Compute Cost and Capacity pages, and the full rules are in the methodology. Every parameter named there is set in the configuration that the pipeline and the pages both read, so the published method and the calculation cannot drift apart.

Why an index, when there is already a price board

The price board answers a simple question each morning: what does an hour of this chip cost today, from this kind of provider? Its cells are trimmed means across whichever providers list the chip that day. That makes them good snapshots and poor time series. If a cheap provider stops listing an H100, the cell rises even though no price changed; if an expensive one joins, it rises again for the same reason.

An index has to separate those two things. The Compute Cost Index therefore does not chain together the board's daily values. It measures each day's change only across the providers present on both days, so a provider joining or leaving does not move it. What is left is a measure of price change for a like-for-like set of sellers.

The Compute Cost Index: what goes in

The index reads three cells of the board for each of four chips, the H100, H200, B200 and B300:

• Neocloud, the on-demand prices of GPU clouds whose main business is renting accelerators;
• Hyperscaler list, a census of the on-demand list prices of the four hyperscalers;
• Marketplace and spot, spare and interruptible capacity.

It reads a cell only when the cell is published, with enough providers behind it, or with all four hyperscalers reporting. Indicative and census cells are shown on the board but never enter the index. The board's other cells, the reserved terms, the serverless prices and the all-provider cell, are shown and never read.

Our own rate cards do not reach back to the base date, so the history before them comes from the GetDeploying GPU price dataset, published under CC BY 4.0: the weekly median across providers for each chip. Those values are back-calculated and marked as such wherever they appear.

The Compute Cost Index: how it is calculated

Each chip has its own index, and the headline is a weighted composite of them.

• The back-calculated history. For each chip, the weekly provider median from the dataset is set against its value in the base week. Each weekly value holds for the business days of its week.
• The link. On the first business day that our own neocloud cell for a chip has enough providers, the chip's index takes the back-calculated level that day. Our own price level never enters the index; only its changes from that day on do.
• The daily change. From the day after the link, each day's change is the ratio of two trimmed means taken over the same providers, those present both that day and on the last day with a value of its own. If too few providers match, the day has no value of its own.
• Carry-forward. A day without a value of its own carries the previous value for a limited number of business days, each one logged. Beyond that the value is empty and publication is held until the gap is reviewed.
• The composite. The chip indices are combined with fixed weights set in the methodology, with the H100 carrying the most. A chip that has no back-calculated history, as the B300 did not, starts its own index on its link day and joins the composite only after its neocloud cell has had enough providers for several days in a row. Each change of weights is chain-linked on the day it takes effect, so a chip joining does not move the headline.

The result answers one question: compared with the base date, how much more or less does the same basket of GPU-hours cost from the same kinds of seller?

The spreads

Two spreads are published beside the index for each chip. The hyperscaler premium is how far the hyperscalers' list price sits above the neocloud price, and the marketplace discount is how far marketplace and spot prices sit below it. Spreads read each day's cells directly, not the chain-linked index, so unlike the index they do move when a provider joins or leaves. A spread is withheld, with its reason, when either of its cells is missing; for the newest chips the hyperscaler census is often incomplete, so their premium is often withheld.

The Capacity Index in brief

The Capacity Index is built on the same principles: stored inputs, fixed rules and a value frozen when it is stored. Its subject is power rather than price. Each Monday it adds up the IT power energised at the AI datacentres it tracks, from Epoch AI's Frontier Data Centers register, published under CC BY 4.0, and from our own register of large campuses that Epoch does not cover. The index is that operational power set against its level on the base date. Beside it sit a probability-weighted pipeline of capacity not yet energised and a ratio of contracted to active power at the neoclouds that disclose both. Our piece on the Capacity Index explains each of them.

Checks before anything is published

Each index is computed only from stored inputs, and no run rewrites an earlier date: a run writes its own date and nothing else. A recompute check rebuilds every stored value from those inputs and reports any that differ, so every value can be traced back to the prices or sites behind it. A restatement is never automatic. It is a one-off correction approved by the owner and recorded value by value with its reason, and a correction to a figure already published goes on the corrections page. Rules that catch the unusual are logged as flags: a large daily move in the Compute Cost composite, a chip missing from it, a value carried too long, or a single Capacity site whose power jumps between weeks. A flag that needs review holds that day's or that week's publication until it is approved, so nothing is published with an open question against it. A value is marked back-calculated when it comes from the dataset's history, falls before the launch, or was computed after its own date.

What each plan sees

The CTI headline level and its thirty-day chart are free. The sub-indices for each chip, the spreads and the history beyond thirty days are in Standard for the last twelve months and in Plus in full, and Pro has them as CSV. For the Capacity Index the headline is free, Compute Open and Standard see the last seven days of history, Plus sees all of it, and Pro has the CSV.

How to read a move

A rise in the Compute Cost Index means the same kinds of seller are charging more for the same chips than they were on the comparison day. It does not mean that any one provider's price moved by that amount, and it is not a price anyone can rent at. A fall in a single chip's index while the composite holds steady usually means one generation is getting cheaper while the others are not. Read the chip indices and the spreads beside the headline, and the board for the providers behind them.

Sources

• Compute Terminal Index methodology, version 1.0-rc (2 November 2026); methodology.
• Compute Terminal price board; methodology.
• GetDeploying GPU price dataset, CC BY 4.0, for the history before our own rate cards.
• Epoch AI Frontier Data Centers, CC BY 4.0, and the Compute Terminal campus register, for the Capacity Index.