The CTI AI-Factory Capacity Index

The CTI AI-Factory Capacity Index measures how much AI datacentre power is switched on. Every Monday it adds up the IT power energised at the AI datacentres it tracks and sets the total against its level on 5 January 2026. Beside it, it publishes how much more capacity is on the way, weighted by how likely it is to arrive, and how far the neoclouds' signed contracts run ahead of the power they have live. It has been published since 11 November 2026.

Where the Compute Cost Index follows the price of a GPU-hour, this index follows the supply that sets it. The live values are on the Capacity page, and the full rules are in the methodology. This piece explains them in plain words.

Why power, not chips

A GPU earns nothing until it sits in a powered, cooled hall, and through this cycle the grid connection, not the chip, has been the slowest part to deliver. Power is also the unit the industry itself uses: campuses, leases and build costs are all quoted in megawatts of critical IT power. Counting energised megawatts therefore measures the capacity that can actually run, rather than the chips that have been ordered or the campuses that have been announced. Our piece on power and cooling explains why power became the constraint.

What goes in

The index has three inputs, each stored with the date it was read.

• Epoch AI's Frontier Data Centers register, published under CC BY 4.0: a register of the largest AI datacentre sites, with a timeline of each site's IT power. It is fetched every week and versioned, so the register as it stood on any Monday can be read again.
• Our own campus register, for very large campuses that Epoch does not track. Each row comes from a company's filing or a press report and is published only after an analyst has checked and approved it.
• Company disclosures of active and contracted power, read from the SEC filings of the neoclouds that report both, with the sentence each figure came from.

How the campus register is filled

The register is proposed automatically and published by hand. Each morning a job reads the filings and releases of the companies in our coverage, and the daily news for developers that do not file, and proposes any sentence that names a single campus with a power figure in the range the register covers. A company's own statement is graded as the stronger evidence, a press report as the weaker.

Most of the work is in what the rules leave out. Company-wide totals, land options, letters of intent, power plants built to serve a campus, bitcoin-mining capacity, tables of figures, and sites Epoch already tracks under another name are all skipped, and each skipped sentence is kept with the rule that skipped it, so the rules can be audited. Sentences that give only a potential figure are flagged for the analyst. When a press report's link cannot be opened by readers, the analyst replaces it with the company's own release or the publisher's article before the row can be approved. Approval also refuses any row outside the register's size range.

Stages and probabilities

Not all capacity is equal. For each Epoch site, power dated on or before the Monday counts as energised, and power dated later counts as under construction. A register row carries its own stage, from the wording of its source: energised, under construction, approved or financed, or announced. A row is counted only from the day it was approved, and never when Epoch tracks a site of the same name.

Each stage carries a probability that the capacity will be delivered, rising from announced to energised, with energised power counted in full. The probabilities are fixed in the methodology, so the same stage always counts the same way.

The series

The index publishes a small family of series each week.

• Operational power: the energised IT power across all tracked sites.
• The index: operational power set against its level on the base date.
• Pipeline power: the capacity not yet energised, each megawatt weighted by the probability of its stage.
• By region and by operator: operational power split across the US, Europe, the Middle East, Asia and other regions, and across hyperscalers, AI labs, neoclouds, colocation providers and developers, and others.
• The neocloud contracted-to-active ratio: contracted power divided by active power, summed over the neoclouds that disclose both, each figure from that company's latest filing on or before the Monday.

Only energised power moves the index. A campus that is announced, approved or under construction moves the pipeline, not the headline.

What changes a week's value

Each Monday's values are read once, from the Epoch register and our own register as they stand that morning, and then frozen. That has two consequences worth knowing when you read the series.

• Revisions land as steps. When Epoch revises a site, the revision is logged and applies from the next Monday. Earlier weeks are not restated, so a revision shows up as a step in that week's change rather than as a rewritten history.
• Approvals land as steps. A register row counts from the day it was approved and is not back-filled, so an approval appears as a step in pipeline power. Unless the row is energised, it does not move the index at all.

The ratio has a similar trait. Because the companies in it report on different dates, the ratio can move when a company files, or when one joins the sum, even if nothing changed on the ground.

Checks before publication

A single site whose operational power moves by an unusually large amount between two stored weeks is flagged for review, and an open or rejected flag holds that Monday's publication. Each run also compares the newest Epoch vintage with the last stored Monday and logs what Epoch revised. A disclosure is kept only when the filing states the figure once, unambiguously, in a sentence about current power; forward-looking statements are ignored, and anything ambiguous is left out and logged. A Monday held for review is published on the first run after its flag is approved.

How to read it

A rise in the index means more AI datacentre power was switched on at the sites we track. It does not count capacity on order, and it says nothing about how fully that power is used. Read it with the pipeline beside it: a growing pipeline with a flat index means capacity is being planned faster than it is being delivered. The contracted-to-active ratio says how much signed demand the neoclouds have yet to serve from live power.

For the sites themselves, the Global AI Datacentre Monitor maps them one by one.

What each plan sees

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.

Sources

• Compute Terminal Index methodology, version 1.0-rc (2 November 2026), section on the Capacity Index; methodology.
• Epoch AI Frontier Data Centers, CC BY 4.0.
• Compute Terminal campus register, analyst-approved rows only.
• Company filings on Forms 10-K, 10-Q, 6-K and 8-K, for active and contracted power.