Plain words
Glossary
29 terms you meet when looking for compute, from MW versus MVA to capacity blocks.
These entries say what a word means. They are written in our own words from general knowledge and carry no numbers about the world, so each is marked Definition, not a sourced statistic. Where this site holds sourced facts about a term, the entry links to them. A word can be used differently by different providers, so ask what a figure covers before relying on it.
Jump to: Accelerator · Bare metal · Behind the meter · Capacity block · Colocation · CoWoS · Data residency · Egress · Facility load · GPU cloud (neocloud) · H100-equivalent · HBM (high-bandwidth memory) · Hyperscaler · InfiniBand · Interconnection queue · IT load · Liquid cooling · MW and MVA · On-demand · Powered shell · PPA (power purchase agreement) · PUE (power usage effectiveness) · Redundancy (N+1) · Reserved · SLA (service level agreement) · Sovereign cloud · Spot · Tier levels (Uptime Institute) · Training and inference
Accelerator
A chip built to do the heavy arithmetic of AI and scientific work much faster than a general-purpose processor. GPUs are the best known kind. TPUs and other custom chips are also accelerators.
See: Chip and accelerator entries in the supply chain
Definition, not a sourced statistic
Bare metal
Renting a whole physical server with no virtualisation layer shared with other customers, so the software has direct access to the hardware.
See: Provider directory
Definition, not a sourced statistic
Behind the meter
Generation or storage on the customer's side of the electricity meter, serving the site directly rather than feeding the public grid first. A site can run behind the meter fully, or partly alongside a grid connection.
See: Bloom Energy; Caterpillar and Hunt Energy agreement
Definition, not a sourced statistic
Capacity block
A fixed amount of accelerator capacity booked for a fixed window in advance, for example a set number of GPUs for a set period starting on a set date. You pay for the booking and get the capacity at the start, instead of competing for it on demand. Providers use their own names and rules for this shape of contract.
See: Guide: capacity blocks and the other contract shapes
Definition, not a sourced statistic
Colocation
Renting space, power and cooling in someone else's data centre for your own servers. The operator runs the building; you own and run the machines.
See: Equinix; Digital Realty
Definition, not a sourced statistic
CoWoS
TSMC's name for a packaging method that places an accelerator chip and its memory stacks side by side on one carrier so they work as a single unit. The name is a brand, and other makers use other methods.
See: TSMC 3DFabric and CoWoS packaging; Advanced packaging capacity
Definition, not a sourced statistic
Data residency
A requirement or promise about the country or region where data is stored and processed. It is different from who can legally compel access to it.
See: Policy register
Definition, not a sourced statistic
Egress
Data leaving a provider's network, for example copying results out or moving to another provider. Some providers charge for it, which can make leaving harder than arriving.
See: Guide: questions to ask a provider
Definition, not a sourced statistic
Facility load
The power drawn by the whole building: the IT load plus cooling, power conversion, lighting and other overheads. Facility load is always at least the IT load. Adding IT and facility figures together double counts, so this site shows them as separate metrics.
See: Records that keep IT and facility load apart
Definition, not a sourced statistic
GPU cloud (neocloud)
A cloud provider that focuses on accelerator capacity for AI work rather than a broad service catalogue. "Neocloud" is an industry nickname for the same idea. The terms describe a type of business, not a size or a quality.
Definition, not a sourced statistic
H100-equivalent
A way to compare different accelerator chips on one scale by expressing a fleet as the number of NVIDIA H100 chips that would give the same throughput. It depends on the benchmark and precision chosen, and on the person doing the conversion, so it is an estimate and should be read with its source.
See: Frontier AI data centres, where Epoch AI gives H100-equivalents
Definition, not a sourced statistic
HBM (high-bandwidth memory)
Memory chips stacked vertically and mounted right beside an accelerator so data moves between them very fast. It is a separate part from the accelerator, bought from memory makers.
See: High-bandwidth memory supply; SK hynix
Definition, not a sourced statistic
Hyperscaler
A very large general-purpose cloud provider that runs its own global network of data centres and sells a wide range of services beyond raw compute.
See: Amazon Web Services (EC2); Microsoft Azure
Definition, not a sourced statistic
InfiniBand
A networking standard built for very low delay between machines in a cluster, widely used to link accelerators during training. Ethernet-based designs are an alternative approach to the same job.
See: NVIDIA Quantum InfiniBand; Ultra Ethernet Consortium
Definition, not a sourced statistic
Interconnection queue
The waiting list kept by a grid operator for new generators, batteries and large loads that want to connect to the grid. Each request is studied for its effect on the network before it is allowed to connect. A place in the queue is not a connection.
See: Ofgem data centre connection reforms (GB); Ireland CRU data centre connection policy
Definition, not a sourced statistic
IT load
The electrical power drawn by the computing equipment itself: servers, storage and network gear. It leaves out cooling, lighting and power conversion losses. Figures marked "IT" or "critical" usually mean this.
See: Records that show IT load as its own metric
Definition, not a sourced statistic
Liquid cooling
Removing heat with a liquid instead of only air. Direct-to-chip systems run liquid through plates on the hottest chips; immersion systems submerge the hardware in a non-conductive liquid. Dense accelerator racks are the usual reason to use it.
Definition, not a sourced statistic
MW and MVA
Both measure electrical power. MW (megawatts) is the real power that does work and becomes heat. MVA (megavolt-amperes) is apparent power, the total the equipment and cables must carry. The two are linked by the power factor: MW equals MVA multiplied by the power factor, so MW is never more than MVA. A site quoted in MVA and one quoted in MW are not comparable until you know the power factor and what the figure covers.
See: Data centre records, with each capacity figure labelled by metric
Definition, not a sourced statistic
On-demand
Compute you start when you want and stop when you want, billed for the time used, with no commitment. Availability is not guaranteed at the moment you ask.
Definition, not a sourced statistic
Powered shell
A data centre building delivered with its structure, grid connection and base power, but without the fit-out inside (cooling, racks, internal electrical distribution). The tenant or a partner completes the fit-out.
See: Data centre records, with stage shown for each site
Definition, not a sourced statistic
PPA (power purchase agreement)
A long-term contract to buy electricity from a named generator at an agreed price or formula. Buyers use it to secure supply or to match their use with renewable generation. It is a financial and supply contract, not a grid connection.
See: Power entries in the supply chain
Definition, not a sourced statistic
PUE (power usage effectiveness)
Facility energy divided by IT equipment energy over the same period. A value of 1.0 would mean every unit of energy reaches the computing equipment and none goes to overheads; higher values mean more overhead. It depends on how it is measured (annual average or best moment, design target or measured), so a PUE with no period and method is a weak figure.
See: Data sources that publish a PUE field
Definition, not a sourced statistic
Redundancy (N+1)
Having spare equipment so one failure does not stop service. N is the number of units needed to carry the load; N+1 adds one spare, 2N duplicates the whole system.
See: Guernsey, where an operator lists N+1 generators
Definition, not a sourced statistic
Reserved
Capacity or a billing rate committed for a longer term, such as months or years, in return for certainty of supply, a different rate, or both. Reserved can mean a billing discount on usage or a physical set-aside of machines, and providers differ, so ask which.
Definition, not a sourced statistic
SLA (service level agreement)
The written promise about service: what is measured (for example availability), how, and what the customer receives if it is missed. Read what is excluded as carefully as what is promised.
See: Guide: questions to ask a provider
Definition, not a sourced statistic
Sovereign cloud
A cloud service designed so that data and operations stay under the legal control of a chosen country or bloc. What counts as sovereign is set by contract, law or certification and differs between providers, so ask what is actually guaranteed (location, operator nationality, access by foreign authorities).
See: STACKIT; Core42; Policy register
Definition, not a sourced statistic
Spot
Spare capacity offered on terms that let the provider or host take it back, often at short notice. Spot work must be able to stop and resume. Some providers call the same idea interruptible or preemptible.
Definition, not a sourced statistic
Tier levels (Uptime Institute)
A classification, from Tier I to Tier IV, of how much redundancy a data centre's design has so that maintenance or a failure need not stop the load. Higher tiers mean more redundancy in the design. A tier can apply to a design, a built facility or an operation, and "Tier III standard" in marketing is not the same as being certified, so check which is claimed.
See: Isle of Man, where operators describe Tier III facilities
Definition, not a sourced statistic
Training and inference
Training is the process of building a model by adjusting it over a large amount of data, which usually needs many accelerators working together for a long time. Inference is using a finished model to answer requests, which can run on much smaller slices of hardware but must respond quickly.
See: Providers by type
Definition, not a sourced statistic
Checked on 3 October 2026 and due for review by 3 April 2027. A definition that looks wrong or misleading? Tell us.