← AI capability and resilience

AI capability & resilience · Compute · UK · 2025–2030

How much AI compute does Britain have, and what limits it?

Public AI compute was 21 AI ExaFLOPS in 2025, with a target of 420 by 2030. That is a public programme open to researchers and smaller firms by application — by its own account a small share of all UK compute. Growing it needs chips from abroad, power and grid connections.

Evidence last reviewed 3 October 2026

What exists, and who can use it

AI-01 · Capability ledger

Announced, installed and accessible are different stages

Facility or programmeStageOperationalPublished capacity (own unit)Who can applyAccess termsEvidence date
Isambard-AI, University of Bristol (AIRR)BIST/UKRI, AI Research Resource (AIRR): advanced supercomputers for the UKAccessible nowFully running from July 20255,448 Nvidia GH200 superchips (hardware count, not a performance measure)UK micro, small and medium organisations (Rapid Access, up to 20,000 GPU hours); academia, industry and other UK organisations (Gateway, up to 10,000 GPU hours)Allocated by application through the AIRR portal; both routes open with no closing date. Charging terms not stated on the pages checked.28 September 2026
Zenith, University of Cambridge (AIRR)BIST/UKRI, AI Research Resource (AIRR): advanced supercomputers for the UKAccessible nowListed as an AIRR access route; a £36m upgrade was announced in January 2026AMD-based system; published capacity not stated in a comparable unitAs for Isambard-AI (same AIRR routes)As for Isambard-AI. The earlier Dawn system is now described in the past tense.28 September 2026
New AIRR supercomputer, EdinburghDSIT, UK Compute RoadmapAnnouncedNot yet operationalNot yet specifiedNot yet publishedUp to £750m announcedUK Compute Roadmap; UK AI Hardware Plan (June 2026)
AIRR programme as a wholeDSIT, UK Compute RoadmapTarget21 AI ExaFLOPS in 2025; 420 targeted by 2030AI ExaFLOPS (the precision basis is not stated, so not comparable with vendor chip figures)Eligible UK researchers and organisationsPart of a Compute Roadmap commitment of up to £2bn to 2030, of which over £1bn is AIRRNo official update on delivered capacity against the target found
Commercial cloud (Microsoft, Amazon Web Services, Google and others)CMA, Cloud services market investigation: final decisionAccessible nowOperatingNot published for the UKAny customer able to payCommercial contracts. The CMA found egress fees and software licensing make switching harder; fewer than 1% of customers switch each year.CMA final decision, 31 July 2025

Location is not control

Who owns the equipment, who allocates time on it, and who controls the software and licences can all differ from where a facility sits. For the public AIRR systems, ownership of the equipment is not set out in the pages checked. Ownership of commercial data centres varies by site and is not compiled in any official source. No official source measures how much UK compute is under UK operational control.

Funding commitments overlap — never add them up

Three funding figures appear across AIRR-related sources and must never be added together: the overall Compute Roadmap commitment (up to £2bn to 2030), the AIRR-specific component within it (over £1bn — a PART of the £2bn, not additional to it), and the Edinburgh heterogeneous supercomputer (up to £750m) — whose relationship to the £1bn+ AIRR component is not independently confirmed. Separately, the UK AI Hardware Plan's own £1.1bn+ headline combines targeted public AND private investment, and its own listed category figures (Innovation £120m, Skills £80m, Procurement £750m, Investment funds) do not cleanly reconcile to that £1.1bn total either — the plan's own text does not resolve this arithmetic, and this page does not attempt to on its behalf.

Who is responsible now

The Compute Roadmap and AI Hardware Plan were published by the Department for Science, Innovation and Technology. After the July 2026 machinery-of-government changes, the AI Research Resource and compute sit with the Department for Business, Innovation, Science and Trade (BIST, the renamed Department for Business and Trade), while AI strategy and public-sector AI adoption moved to the Cabinet Office.

Power: what data centres use, and what electricity costs

How much electricity do data centres use?

Data centres’ grid electricity use rose from 1.3 to 4.5 TWh between 2020 and 2024

DESNZ estimate of electricity consumed from the grid by data centres in Great Britain, terawatt-hours a year, 2020 and 2024. Years between were not published.

Great Britain · 2020 and 2024 only
TWh a year0.02.04.02020n/a2021n/a2022n/a202320244.5
On DESNZ’s estimate, data centres used about 3.5 times as much grid electricity in 2024 as in 2020. That is still a small share of national demand, but it is concentrated in a few places on the grid, which is why connection dates matter more than the national total.

Great Britain · 2020 and 2024 only · Official estimate from meter-data analysis; data centres are not yet separately identified in DESNZ’s main collections

Sources, definitions and data — Data centres’ grid electricity use rose from 1.3 to 4.5 TWh between 2020 and 2024
Geography:
Great Britain — Data centres identified in electricity meter data
Measure:
TWh a year
Period:
2020 and 2024 only
Status:
Official estimate from meter-data analysis; data centres are not yet separately identified in DESNZ’s main collections
Calculation:
Quoted from DESNZ commentary; not a DUKES table cell. Missing years are shown as gaps, not zeros.
Compare with:
2020
Not included / breaks:
Not all AI: data centres also run ordinary cloud, web and business services. Excludes on-site generation. Not comparable with NESO’s scenario figure in the next chart, which uses a different definition.

Source: DESNZ, Digest of UK Energy Statistics (DUKES) 2026, Chapter 5, Chapter 5 commentary, p.2 (released 30 July 2026). Chart AI-03.

View the data as a table
YearTWh
20201.3
2021not published
2022not published
2023not published
20244.5

How much might data-centre demand grow?

In one NESO scenario, data-centre demand reaches about 50 TWh by 2040

Great Britain electricity demand from data centres under NESO’s Holistic Transition scenario, TWh a year, 2024–2040. Dashed line: scenario projection.

Great Britain · 2024 (estimate) to 2040 (scenario)
TWh a year02040602024202820322036204049.7 TWh
NESO’s Holistic Transition pathway has data-centre demand rising from 7.6 TWh in 2024 to 22.8 TWh in 2030 and 49.7 TWh in 2040. It is a planning scenario, not a prediction, and it assumes the connections and generation it needs are built. Whether that demand is met in Britain or served from data centres abroad depends partly on grid connections and power prices.

Great Britain · 2024 (estimate) to 2040 (scenario) · Scenario projection by the system operator — not a forecast of one certain outcome

Sources, definitions and data — In one NESO scenario, data-centre demand reaches about 50 TWh by 2040
Geography:
Great Britain — Data centres, as defined in NESO’s Future Energy Scenarios
Measure:
TWh a year
Period:
2024 (estimate) to 2040 (scenario)
Status:
Scenario projection by the system operator — not a forecast of one certain outcome
Calculation:
As published by NESO for the named scenario. Points are the years NESO publishes; the line between them is a guide, not data.
Compare with:
2024 starting point, 7.6 TWh
Not included / breaks:
One of four NESO scenarios; the others give different paths. NESO’s 2024 starting figure uses a different definition from DESNZ’s meter-based estimate and is higher. Already included in total demand, not additional to it.

Source: NESO, Future Energy Scenarios 2025: Data Workbook, Holistic Transition, data centre demand (released 28 November 2025 (version 6)). Chart AI-04.

View the data as a table
YearTWhType
20247.6estimate
20259.3scenario
203022.8scenario
203537.2scenario
204049.7scenario

What do medium-sized non-domestic users pay for electricity?

The UK had the highest price of these seven countries in 2025 H2

Non-domestic electricity price for users of 2,000–19,999 MWh a year, including taxes and levies but not VAT, pence per kWh, 2025 H2.

UK and six EU countries · 2025 H2
2025 H2
  • United Kingdom25.1p
  • Ireland19.1p
  • Italy17.0p
  • Germany16.7p
  • Netherlands15.8p
  • France10.6p
  • Spain10.4p
The UK’s price for medium-sized non-domestic users was well above Germany’s and France’s in the same band and period. For very large industrial users, government attributes much of the UK’s higher price to policy costs carried on bills (see Powering Britain). For compute, power price is one factor among several — connection dates, land and chips also decide where data centres are built.

UK and six EU countries · 2025 H2 · Official statistics

Sources, definitions and data — The UK had the highest price of these seven countries in 2025 H2
Geography:
UK and six EU countries — Non-domestic users in the Medium consumption band
Measure:
pence per kWh
Period:
2025 H2
Status:
Official statistics
Calculation:
As published by DESNZ (converted to pence by DESNZ).
Compare with:
Six EU countries in the same band and period
Not included / breaks:
A large data centre uses far more than this band and usually buys power on negotiated contracts, so this is an indicator of the price environment, not a data-centre tariff. Household price caps are a different measure.

Source: DESNZ, International non-domestic energy prices, QEP Table 5.4.2 (Medium band, incl. taxes excl. VAT). Chart AI-05.

View the data as a table
Country2025 H2, p/kWh
United Kingdom25.08
Ireland19.05
Italy16.97
Germany16.74
Netherlands15.78
France10.56
Spain10.38

Grid connections are covered in Britain’s grid queue, and industrial prices for very large users in Powering Britain.

Competing readings

Build at home. More domestic compute gives UK researchers and firms capacity that cannot be withdrawn by a foreign provider, and keeps sensitive work under UK law. This is the case for AIRR and the Hardware Plan.

Buy access, secure the terms. Most organisations will always use commercial cloud; resilience may come more cheaply from switching rights, contracts and oversight of critical providers than from building rival capacity.

The binding constraint may be elsewhere. Grid connection dates, power prices, chip supply and skills limit how fast any compute can be used productively. No source checked for this page establishes which binds most.

Options are compared on the main AI page.

Sources and methods (6)

Method notes

  • Each facility keeps its own published unit; accelerator counts and ExaFLOPS figures are never summed or converted.
  • DESNZ’s consumption estimate and NESO’s scenario use different definitions and are shown in separate charts.