Glossary

Earnings before interest, tax, depreciation and amortisation, after the company's own adjustments. Many GPU owners lead with it, but it leaves out depreciation, which is their largest cost, so it flatters a business whose economics depend on how fast its chips lose value. Reading it beside cash capital spending and the stated useful life of the fleet gives a fairer picture. See GPU economic life and depreciation.

A datacentre built and run as a production line, with electricity and data as its inputs and tokens as its output. The term describes a design choice: where a cloud hall is built for many kinds of work, an AI factory is built for one, with liquid-cooled racks, a single network fabric and a power system matched to the GPUs it hosts. Because its output can be measured, it can be managed like a plant. See What is an AI factory?.

A bond backed by a pool of assets that pay predictable cash, such as leases or contract receivables, placed in a separate vehicle and issued in layers of seniority. The senior notes are paid first; the junior notes absorb losses first in return for a higher yield. For GPUs the notes must repay while the chips still earn, so the structures lean heavily on the customer contracts behind them. See GPU-backed financing.

An index value computed for a date before the index was published, or computed after its own date. The CTI Compute Cost Index's history before our own rate cards is back-calculated from the weekly provider medians of a public GPU price dataset, credited under its licence, and every such value is marked as back-calculated wherever it appears. See The CTI in plain words.

A GPU cloud that rents out dedicated servers or whole clusters with only a thin layer of software on top: no managed services, and the customer brings its own stack. Bare metal is faster to stand up and priced lower than a full-stack AI cloud, which wraps the same hardware in orchestration, storage and inference services. Many neoclouds run a blend of the two. See What is a neocloud?.

Compute Terminal's register of very large AI datacentre campuses that Epoch AI's register does not track. Rows are proposed automatically from company filings and the news, each with its source sentence, and enter the Capacity Index only after an analyst has checked and approved them, counting from the day of approval. See The CTI AI-Factory Capacity Index.

Capital expenditure measured as the cash actually paid for property and equipment in a period, read from the cash-flow statement, rather than guidance or announced budgets. Our capex figures for Microsoft, Alphabet, Amazon and Meta are cash capex from their filings:

Joining the pieces of an index so that a change in its make-up does not move its level. When a new chip enters the CTI Compute Cost composite, the new weights are linked on the day they take effect: the composite carries on from its old level, and only later price changes move it. Equity indices do the same thing through their divisor. See The CTI in plain words.

Optics built into the switch package itself, rather than plugged into its front panel as separate modules. Converting between electrical signals and light right beside the switch chip saves power and removes thousands of pluggable modules and the failures that come with them. The switch end of each link is moving to co-packaged optics first; the network cards in servers remain pluggable for now. See Networking and optics.

Renting space, power, cooling and connectivity in a datacentre owned by someone else. Retail colocation rents a cage or a few cabinets to many customers, who also pay for cross-connects to carriers and to one another; wholesale colocation leases whole halls to a few large tenants on long terms. The colocation provider owns the building, and the tenant owns the computers inside it. See What is a datacentre?.

Compute Terminal's family of economic indicators for the AI build-out. The CTI Compute Cost Index follows what a GPU-hour costs every business day, the CTI AI-Factory Capacity Index follows how much AI datacentre power is switched on every Monday, and the CTI AI-Factory Equity Index follows the listed companies behind them. Each has a public methodology, and each has been published since 11 November 2026. See The CTI in plain words.

Contracted backlog divided by debt: a quick test of how well a GPU borrower's signed revenue covers what it owes. It is crude, because backlog is gross revenue over many years while debt is principal, but it shows direction. Cover that falls quarter after quarter means debt is growing faster than signed business. See Take-or-pay compute contracts.

The revenue a company's signed contracts still have to pay over their remaining terms, reported in its filings as backlog or as remaining performance obligations. For a neocloud it is the clearest measure of committed demand, though it is gross revenue spread over many years. Lenders compare it with debt, and watch how much of it comes from a few customers. See Take-or-pay compute contracts.

Contracted power divided by active power, summed over the neoclouds that disclose both in their filings. A high ratio means a company has signed far more capacity than it has switched on: backlog to deliver, and capital to raise. Because each company files on its own dates, the ratio can move when one of them files, or joins the sum, without any change on the ground. See The CTI AI-Factory Capacity Index.

The pumps and heat exchangers at the heart of a liquid-cooled hall. A coolant distribution unit separates the clean loop that runs through the racks from the facility's water loop, holds the coolant above the dew point, balances the flow, and filters and monitors the fluid. Units range from small ones serving a single rack in a refit to large skids serving a whole row or hall. See Power and cooling.

The power available to the IT equipment itself, the servers, storage and network, as opposed to the site's total draw, which also feeds cooling, lighting and losses. Datacentre capacity, leases and build costs are quoted per megawatt of critical IT power, so it is the unit to use when turning a megawatt into a number of GPUs, or a lease into a price per GPU-hour. See What is a datacentre?.

Compute Terminal's record of transactions across AI infrastructure: deals, listings and sets of talks, each logged once with its date, sector, parties and source, and with a value in US dollars only where the source reported one. Compute Open shows the last thirty days with headline terms; subscribers see the full terms, the closing calendar and the IPO pipeline. See AI datacentre deals.

A loan agreed in full at the start but drawn in stages, as the borrower needs the money. GPU operators use them to fund clusters as the chips are delivered and installed, often through a special-purpose vehicle that holds the GPUs and the customer contract behind them. The loan is repaid from the customer's payments, and can be rated well above the operator when the customer is investment grade. See GPU-backed financing.

The cost of an asset spread over the years it is used, charged against profit each year although no cash leaves the business at the time. For a GPU owner depreciation is the largest cost, because the chips are paid for up front and replaced as new generations arrive. Whether the business makes money depends on whether the chips earn back their cost before they stop earning. See GPU economic life and depreciation.

The unit in which datacentre space is leased: the monthly rent for each kilowatt of critical IT power, usually for a powered, cooled shell without the GPUs. Because the landlord does not own the chips, a lease priced this way sits far below a GPU contract for the same megawatt. Converting between the two takes the number of GPUs per megawatt and the hours in a year. See From $/kW-month to $/GPU-hour.

The annual value of a compute contract for each megawatt of critical IT power it covers, usually quoted in millions of dollars. It lets contracts for different chips and sizes be compared, and it links GPU-hour prices to build costs: multiply a price per GPU-hour by the GPUs in a megawatt, by the hours in a year and by the share of hours billed. See From $/kW-month to $/GPU-hour.

The period over which an asset's earnings justify the cost of owning and running it, as opposed to the useful life stated in the accounts. For a GPU it ends not when the chip stops working but when newer generations push its rent too low to cover its costs. Our Economic Life Tracker reads an implied economic life from the trend in rental prices for each GPU generation. See GPU economic life and depreciation.

A site with a working grid connection and the electrical plant to take power today, as opposed to land whose connection is still in the queue. Energised sites are the scarcest input in AI infrastructure, because a GPU earns nothing until it sits in a powered, cooled hall. Companies that already held them, many of them former crypto miners, were the first to move into AI hosting. See What is a neocloud?.

How much of a company's business is AI infrastructure, as the CTI AI-Factory Equity Index classifies it. A pure play earns at least half of its revenue from AI infrastructure and counts at full weight; a significant name earns a smaller share above the methodology's floor and counts at half weight. The class is read from company filings in a fixed order, starting with the annual report. See The CTI AI-Factory Equity Index in plain words.

A lease that transfers most of the risks and rewards of owning an asset, so the tenant records it on its balance sheet as an asset and a liability. Hyperscalers acquire some datacentres and equipment this way, and those assets do not appear in cash capex; Microsoft's reported figure, for one, excludes them. Reading the lease notes beside capex shows more of the build-out. See Hyperscaler capex in 2026.

The part of a company's shares available to public investors, leaving out holdings such as those of insiders, governments or strategic owners. Equity indices weight companies by free-float market capitalisation, the share price times the shares times the float factor, so that a company with a large locked-up stake counts for what investors can actually hold. The CTI AI-Factory Equity Index takes its float factors from company filings. See The CTI AI-Factory Equity Index in plain words.

The earnings measure REITs report in place of net income: net income with property depreciation added back and gains or losses on property sales taken out. Adjusted funds from operations goes further and subtracts the recurring capital spending needed to keep the buildings competitive. Per share, and set against the dividend, it is the measure most investors in a datacentre REIT anchor on. See Datacentre REIT economics.

The share of a cluster's wall-clock time spent on useful work that does not have to be repeated: what is left after failures, restarts, checkpoints and gaps in scheduling. Two clusters with the same hardware can differ widely in goodput, and since customers pay for output, the difference is revenue. Buyers increasingly write delivered performance into their contracts. See GPU-hour economics.

One GPU used for one hour: the unit in which AI compute is priced and sold. Rental prices, contract values and the economics of a cluster can all be put per GPU-hour, so a megawatt lease, a multi-year contract and a marketplace listing can be compared on one scale. The price board publishes today's GPU-hour price for each chip and provider class. See GPU-hour economics.

The link between a datacentre and the electricity grid, with the substation that steps the grid's voltage down for the campus. In many markets the queue for a new large connection runs for years, because the utility must study the load and often build new lines and substations. Grid delivery, rather than chips or land, has become the binding constraint on how fast AI capacity is built. See What is a datacentre?.

One of the largest cloud providers, which run hundreds of services across global networks of datacentres and treat GPUs as one product line among many. On our price board the hyperscaler class is a census of four: Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure, read from the list prices on their own pricing pages. In capex, the term usually means the four largest spenders: Microsoft, Alphabet, Amazon and Meta. See What is a neocloud?.

The price board's cell for the four hyperscalers: Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure. Because there are only four, the board reports them all rather than a sample. The cell is published when all four list a price, and otherwise shows the average of those that do, with each one named. For the newest chips the census is often incomplete. See How to read the price board.

How far the hyperscalers' on-demand list price for a chip sits above the neocloud on-demand price for the same chip. The premium pays for the rest of the cloud, enterprise agreements, compliance and a global footprint. Large hyperscaler customers rarely pay list prices for long, so the premium overstates the gap in what customers actually pay, and it can only be measured once all four hyperscalers publish a price. See On-demand, reserved and spot: how GPU rental prices form.

The values an index's rules would have produced had the index existed before its launch. Every value of the CTI AI-Factory Equity Index before 11 November 2026 is a hypothetical back-test, and every place that shows one says so. A back-test is built with today's knowledge, including today's universe of companies, so it may look better than anything that could have been built at the time, and it says nothing reliable about the future. See The CTI AI-Factory Equity Index in plain words.

The committee that oversees the methodology of the Compute Terminal indices and decides discretionary matters, such as a capping that cannot be solved, a disputed corporate action or a change to the universe. Its members are Denis Rudev, its chair, and Gleb Petukhov, and every decision needs both. Its terms of reference are published on the independence page. See The CTI AI-Factory Equity Index in plain words.

The number an index's total market value is divided by to give its level. When an event changes the constituents' market value without any change in prices, such as a stock split, a spin-off, a delisting or a rebalance, the divisor is adjusted so that the level carries on unchanged. Only price moves, and in a total return version reinvested dividends, move the index. See The CTI AI-Factory Equity Index in plain words.

A price-board cell with fewer providers than the methodology's minimum for a published value. It shows those providers' trimmed mean, marked indicative with its sample size, or, where there are too few even for that, the latest weekly median from a credited public dataset. An indicative value is real information but not a market median, and should be quoted with its label. Indicative cells are open to Compute Open readers and every paid plan. See How to read the price board.

Running a trained model to answer a request: every prompt sent to a chatbot, every line of code suggested, every image generated. Inference runs all the time, close to its users and often inside a given country's borders, so it favours always-on capacity with a low cost per token. As the mix of work shifts from training towards inference, so does the demand for that kind of site. See What is an AI factory?.

A lossless network technology, now sold by NVIDIA, that is the incumbent fabric for the largest AI training runs. It never drops packets, routes around congestion and can perform part of the arithmetic of combining results inside its switches. Tuned Ethernet fabrics have closed much of the gap, and buyers choose between the two on ecosystem, price and time to performance. See Networking and optics.

In a datacentre, the direct links between customers in the same building: cross-connects to carriers, cloud providers and one another. The more customers a building holds, the more each connection is worth, so interconnection revenue tends to carry high margins and loyal customers. It is a large part of the economics of retail colocation, and much less of wholesale leasing to a single hyperscale tenant. See Datacentre REIT economics.

A credit rating in the upper tiers of the rating agencies' scales, signalling a low expected risk of default. It matters in GPU financing because loans secured on a take-or-pay contract with an investment-grade customer can themselves be rated investment grade, whatever the operator's own rating. Those loans are the cheapest layer of the capital stack, and the layer that moved least when interest rates rose in September 2026. See GPU-backed financing.

Carrying heat away from chips with liquid rather than air, through cold plates mounted on the processors or by immersing whole servers. The newest GPU racks draw far more power than air can cool, so AI halls are built for liquid from the start, with coolant distribution units beside the racks and pipework out to the plant. Cooling has become part of how many GPUs a building can hold. See Power and cooling.

The price a provider publishes on its own pricing page for capacity on demand, before any discount for a commitment or a negotiated contract. The price board's hyperscaler cell is a census of four list prices, from AWS, Azure, Google Cloud and OCI: for an H100 it reads

The price of spare or interruptible capacity offered on GPU marketplaces, including spot instances that the provider can reclaim at short notice. It usually sits below on-demand rates, because the capacity comes with fewer guarantees. The price board's marketplace class collects these prices, whoever quotes them: for an H100 it reads

The set of providers present on both of the days an index compares. The CTI Compute Cost Index measures each day's change only over that set, taking the trimmed mean of the same providers' prices on both days, so a provider joining or leaving does not move the index. When too few providers match, the day has no value of its own and the previous value is carried, for a limited time. See The CTI in plain words.

The share of a chip's theoretical peak arithmetic that becomes actual model mathematics. It is stricter than goodput, because it also discounts time spent waiting for memory or for other chips across the network, and code that does not use the hardware fully. Even the best training runs leave much of the silicon's peak unused, and closing that gap is what operators compete on. See GPU-hour economics.

A cloud provider built around one product: large, tightly networked pools of the latest AI accelerators, rented with as little friction as possible. Where a hyperscaler treats GPUs as one line item, a neocloud treats the GPU cluster as the whole company, from procuring the chips to designing the halls and selling the hours. The category grew out of the shortage of H100-class compute. See What is a neocloud?.

A version of an equity index that reinvests dividends, after the tax withheld on them, as well as following prices. The headline CTI AI-Factory Equity Index is a price return; its net total return version is published beside it for Plus and Pro subscribers. Over long periods the gap between the two versions is the dividend stream. See The CTI AI-Factory Equity Index in plain words.

NVIDIA's link between GPUs inside a server or a rack, which lets them read and write one another's memory directly, with no network packets in the way. On rack-scale systems NVLink joins a whole rack of GPUs over a copper backplane, so the most bandwidth-hungry traffic of training never touches the slower network between racks. See Networking and optics.

The hourly price for capacity used without a commitment, started and stopped at will and billed by the hour or the second. It is the most visible price in the market, and it usually sits above reserved and contract prices, because the provider carries the risk of idle hours. On the latest price board a neocloud H100 hour reads

The IT power energised and live at a datacentre, as opposed to power under construction, approved or only announced. The CTI AI-Factory Capacity Index sums it each Monday across the AI datacentres it tracks, and only a change in operational power moves the index. Capacity still being built shows up in the pipeline instead. See The CTI AI-Factory Capacity Index.

The module at each end of a fibre link that turns electrical signals into laser light and back. Every GPU in a large cluster drives several, across the training, front-end and storage networks, so transceivers take a noticeable share of a cluster's power and cost, and their failures create a steady maintenance load. Newer designs remove the power-hungry signal processor or move the optics into the switch. See Networking and optics.

How long a cluster's cash income takes to repay what it cost to build, before operating costs and interest unless stated otherwise. It depends mostly on the contract price per megawatt-year and on how much of the build a customer prepays. A short payback matters because a GPU has to earn back its cost before the next generation of chips reprices it. See GPU-hour economics.

A company that owns energised shells, buildings with power and cooling but no GPUs, and leases them to the operators that install the chips. Power landlords earn rent priced per kilowatt-month on long leases, so their economics are those of real estate: steady income, and no exposure to the GPUs going out of date, because the tenant owns the hardware. See What is a neocloud?.

The energy a whole datacentre uses divided by the energy delivered to its IT equipment. A value close to one means little is spent on cooling and other overhead, and liquid-cooled AI halls in cold climates come closest. PUE has a blind spot: electrical losses inside the racks count as IT load, so a good PUE can hide waste before the power reaches the chips. See Power and cooling.

A datacentre building with its grid connection, substation and core electrical plant in place, but without the computers, and often without the final fit-out for a particular tenant. Powered shells are leased per kilowatt-month to operators that install their own racks and GPUs. Because power is the slowest part of a new site to deliver, a shell that already has it can be fitted out in months rather than years. See What is a datacentre?.

Cash a customer pays upfront under a compute contract, before the capacity is delivered. Prepayments increasingly part-fund the build itself: they cut the debt an operator needs, shorten its payback and show that the customer is committed. Until the capacity is delivered they sit in the operator's accounts as a liability to the customer, not as revenue. See Take-or-pay compute contracts.

Capacity not yet energised, with each megawatt counted at the probability that it will be delivered, according to its stage: announced, approved or financed, or under construction. The CTI AI-Factory Capacity Index publishes it beside the headline, so that a large announcement counts for less than the same power already being built. It is a view of what is coming, not a forecast of when. See The CTI AI-Factory Capacity Index.

The groups the price board sorts providers into, because the same chip is sold on very different terms: neocloud on-demand, the hyperscalers' list prices, and marketplace and spot. Further cells cover all on-demand providers together, one-year and three-year reserved terms, and serverless pricing. Comparing a price only with others in its own class avoids most misreadings of what a GPU-hour costs. See How to read the price board.

The power one rack of servers draws, in kilowatts. It sets much of the rest of a hall's design: how the racks are cooled, how heavy the power distribution must be and how much floor a megawatt needs. A cloud hall built for air cooling runs at a small fraction of the density of a hall built for the newest GPU racks. See Power and cooling.

A company that owns income-producing property and, in return for paying out most of its taxable income as dividends, pays little or no corporate tax on it. Because it keeps little cash, a REIT funds its growth by issuing shares, borrowing, selling assets or bringing in partners, so its cost of capital matters as much as the demand for its buildings. See Datacentre REIT economics.

A reconstitution reviews which companies are in an index; a rebalance resets their weights. The CTI AI-Factory Equity Index reconstitutes in June and December, and reweights its existing constituents with fresh share counts and float in March and September, each change taking effect after the close on the third Friday of the month. Every change is announced in advance. See The CTI AI-Factory Equity Index in plain words.

Compute committed for a fixed term, usually one or three years, at a price below the on-demand rate in return for the commitment. Hyperscalers offer it as reservations, savings plans or committed-use discounts; neoclouds publish term rates, or sign larger blocks under take-or-pay contracts. The price board's one-year and three-year reserved cells are for Plus and Pro subscribers. See On-demand, reserved and spot: how GPU rental prices form.

What an asset is still worth at a point in its life, or at the end of a lease or loan. For a GPU it depends on the rent the chip can still command once newer generations are on the market. Lessors in operating leases carry residual-value risk, and our Economic Life Tracker shows the share of today's rent left after each of the coming years when rents are falling. See GPU economic life and depreciation.

Selling an asset a company already owns, such as a building or a fleet of equipment, and leasing it straight back. The seller keeps using the asset and frees the capital tied up in it for the next build, while the buyer earns the lease income. Sale-leasebacks recycle capital across the AI build-out, but they replace ownership with a fixed obligation to pay rent. See GPU-backed financing.

The number of providers behind a price-board cell, shown as n beside every value and counted before the tails are cut. It is the first thing to read: a value from many providers is a fair picture of the market, and a value from a few is a fair picture of those few. A cell keeps its providers and their range even when it has too few for a published value. See How to read the price board.

The two levels of a GPU cluster's network. Scale-up joins GPUs inside a server or a rack over fast copper links such as NVLink; scale-out joins racks into a cluster over a fibre network such as InfiniBand or Ethernet. Scale-up is far faster for each GPU, so cluster design keeps the chattiest traffic inside it and sends only the bulkier, less frequent exchanges across scale-out. See Networking and optics.

Optical fibre with a narrow core that carries light along a single path, now the standard in AI datacentres. Multimode fibre, with a wider core and many paths, smears pulses together as line rates rise, cutting its reach below the length of a row of racks. With single-mode, reach depends on the laser and the loss budget, and silicon photonics has brought the cost of its optics down. See Networking and optics.

A separate company set up to hold specific assets and the debt against them, kept legally apart from its parent so that the parent's creditors cannot reach them. In GPU financing the SPV owns the GPUs and the right to the payments under a take-or-pay contract, and the debt is repaid from those payments first. Lenders therefore underwrite the customer's promise to pay, with the GPUs only as a recovery floor. See GPU-backed financing.

Cloud capacity that the provider can reclaim at short notice, usually when a customer paying a higher price needs it. Spot prices sit below on-demand rates because the capacity comes with fewer guarantees, and they move with spare capacity rather than with a published rate card. Training jobs that save checkpoints often run on spot capacity, accepting interruptions in exchange for a lower price. See On-demand, reserved and spot: how GPU rental prices form.

A contract in which a customer commits to pay for a defined block of GPUs, or of megawatts, for several years whether or not it uses them. Take-or-pay contracts are the commercial core of the neocloud business: they turn demand into predictable revenue, and lenders finance clusters against them, looking first at the customer's ability to pay. See Take-or-pay compute contracts.

The unit that AI models read and write: a short word, part of a longer one, a fragment of code or a patch of an image. Every prompt and every answer is metered in tokens, and the industry prices its models per million of them. That makes the token the output of an AI factory, and the cost per token the measure of how well the factory runs. See What is an AI factory?.

Building a model by running it over very large amounts of data until it learns to predict well. Training comes in large, episodic runs that need the biggest coherent block of GPUs, power and cooling available, but it can sit far from its users. That is why training campuses go where power is cheap and plentiful, while inference moves closer to the people it serves. See What is an AI factory?.

An average taken after cutting the same share of values from the top and the bottom of a range. The price board uses one for each cell, so a single unusually high or low quote cannot move the value on its own; with only a few providers in a cell, nothing is cut. Our indices use the same trimmed mean to measure each day's change over the providers present on both days. See How to read the price board.

Equipment that conditions a datacentre's power and keeps the load running for the seconds between a grid failure and the start of the backup generators, using batteries or flywheels. In an AI hall it is one stage of a long conversion chain from the grid to the chip, and the next generation of electrical designs, built on direct current, aims to remove the central units altogether. See Power and cooling.

The number of years over which a company depreciates an asset in its accounts, stated in the notes and set by management's estimate. For servers and GPUs it is one of the most consequential estimates the largest owners make: lengthening it lowers annual depreciation and raises reported profit, with no change in cash. Our Economic Life Tracker sets each company's stated life against what rental prices imply. See GPU economic life and depreciation.

A development's stabilised operating income divided by what it cost to build. A datacentre owner creates value when it builds at a yield on cost comfortably above its cost of capital, and the spread between the two is the margin on development. Securing power, leasing the space before it is finished and keeping construction on time all protect that spread. See Datacentre REIT economics.