Draft, not yet released. In stewardship review. Not indexed until the first approved release.

AI compute is physical infrastructure built on a few chips and a few places

Part I describes systems and their couplings. It ranks no country or operator and assesses no authority.

7 min read2 figures1 table12 notesNR-XTECH-SEMICONDUCTORS-GLOBAL-2026-BASELINE

Contents
  1. ·Edition home
  2. ·Executive summary
  3. 1AI compute is physical infrastructure built on a few chips and a few places
    1. 1.1 The questions this part answers
    2. 1.2 The argument
    3. 1.3 Twenty couplings between five systems
  4. 2Key findings
  5. 3Systems and sectors at baseline
  6. 4Risks and capability
  7. 5Compute build-out meets local power, water and packaging limits
  8. 6Readings from the other desks
  9. 7The 2027 agenda
  10. 8Systems in depth
  11. AMethodology
  12. BData appendix
  13. CReferences and record
  14. ·Complete edition, for print

1.1The questions this part answers

Key questions

  • Where is semiconductor and compute risk concentrated?
  • Which couplings carry semiconductor and compute shocks into other systems?
  • Which semiconductor and compute capabilities hold under compound stress?
  • What must 2027 evidence measure first?

In this part

  • Narrative
  • Headline finding
  • Lead figure
  • Five footprint and governance chapters
  • Twelve-domain reading
  • Sector screens
  • Dependency matrix
  • Risk objects
  • Capability ladder
  • Scenario
  • Three lenses
  • Portfolio

1.2The argument

AI compute is usually discussed as software. The 2026 evidence reads it as a physical build: chips from a handful of fabrication clusters, machines from specialised equipment, memory and packaging, and sites that need electricity and cooling189.

Demand grows fast. Leading AI supercomputers doubled in performance about every nine months from 2019 to 2025, while their power needs doubled about every year1. Epoch AI estimates that training compute grows about four times a year and that power binds first3. Earlier work finds a doubling time of about six months in the deep learning era2.

Supply follows money. Chip sales rose 25.6% in 20258, equipment billings reached USD 135.1 billion9, and governments added the CHIPS and Science Act and the EU Chips Act4412. Export controls shape who can buy advanced computing items13.

The footprint is the least measured part. US data centres used 176 TWh in 2023 and may use 325 to 580 TWh in 202815. Water use of fabrication plants is reported by single companies, not on a common basis23. Peer-reviewed work shows large gaps between models, sites and hardware67.

Figure 1.1How fast do compute measures double? Doubling time, months

Two sources report the speed of growth. The AI supercomputer dataset shows performance doubling about every nine months, while power and cost double about every year1. An earlier study finds training compute doubling about every six months in the deep learning era and about every twenty months before it2. These are historical fits, not forecasts.

ItemBar, scaled to the largest valueValue
AI supercomputer performance, 2019 to 20259
AI supercomputer power need and hardware cost12
Training compute, deep learning era6
Training compute, before deep learning20

Source: as cited in the note to this figure. Year 2025.

Figure ID FIG-XTECH-SEMICONDUCTORS-GLOBAL-2026-BASELINE-01-003

9months

AI supercomputer performance doubling1

12months

power and cost doubling, about a year1

6months

training compute, deep learning era2

20months

training compute, earlier era2

1.3Twenty couplings between five systems

The semiconductor and compute-centred matrix types the twenty couplings between the five core systems, read for their semiconductor and compute content. Strengths are Nexus estimates (ESTIMATED) from the cited evidence and analogue events; they set the graph that 2027 evidence will test.

Figure 1.2Couplings between the five systems

Each line is one coupling from one system to another, drawn by the strength the edition assigns. Strengths are Nexus estimates from the cited evidence and analogue events.

  • Strong
  • Moderate
  • Weak

Source: this edition’s coupling matrix (Table 1.1); status Estimated.

Table 1.1Coupling matrix: from the row system to the column system
From, toWATENEFOOHLTBIO
WAT Water·MODERATE
water for power and cooling		WEAK
no documented link		WEAK
no documented link		WEAK
no documented link
ENE EnergyMODERATE
cooling and fab water demand		·WEAK
no documented link		WEAK
no documented link		WEAK
no documented link
FOO FoodWEAK
no documented link		WEAK
no documented link		·WEAK
no documented link		WEAK
no documented link
HLT HealthWEAK
workforce		WEAK
workforce		WEAK
no documented link		·WEAK
no documented link
BIO BiodiversityWEAK
no documented link		WEAK
no documented link		WEAK
no documented link		WEAK
no documented link		·
Strength: strong, moderate or weak; a Nexus estimate (status Estimated) from cited evidence and analogue events. The scale is defined in Annex A.

Transmission paths

  1. 01

    AIC → ENE → INF

    Compute demand raises electricity demand and constrains where sites can be built.

    Reported US data centre estimates15

  2. 02

    TRD → INF

    Export controls change who can buy advanced chips and equipment.

    Reported US guidance of May 202613

  3. 03

    FIN → INF

    Sales and subsidies fund plants and equipment.

    Reported sales and CHIPS Act figures844

  4. 04

    WAT → INF

    Water supply constrains plant siting.

    Inferred one company water report23

  5. 05

    GOV → INF

    Subsidy programmes set where capacity is built.

    Reported EU and US programmes4412

Notes

Numbers are the edition’s reference numbers, so a note reads the same in every chapter. The full list, with every source and its licence, is in Annex C. Annex C, references

  1. 1Pilz, K.F., Sanders, J., Rahman, R. and Heim, L., Trends in AI Supercomputers, arXiv:2504.16026, submitted 22 April 2025 (abstract page: 500 systems from 2019 to 2025; performance doubling about every nine months; cost and power doubling about every year; one March 2025 system with about 200,000 chips, USD 7 billion and 300 MW; about 75% of performance in the United States). arXiv, Ithaca. Observation year 2025. Accessed 2026-11-03.back
  2. 8Semiconductor Industry Association (ed.), Global Annual Semiconductor Sales Increase 25.6% to $791.7 Billion in 2025 (release; WSTS data; logic USD 301.9 billion; memory USD 223.1 billion; regional growth; about USD 1 trillion projected for 2026). Semiconductor Industry Association, Washington. Observation year 2025. Accessed 2026-11-03.back
  3. 9Evertiq (ed.), Semiconductor equipment billings reached $135B in 2025, 9 April 2026, a trade report of a SEMI release (USD 135.06 billion in 2025 against USD 117.14 billion in 2024; regional table); the SEMI page itself returned an access error, so this is a secondary report and weak evidence. Evertiq, Munich. Observation year 2025. Accessed 2026-11-03.back
  4. 3Sevilla, J. and others, Can AI scaling continue through 2030?, Epoch AI, 20 August 2024 (first 100,000 characters read: about 4 times a year; 1 to 5 GW campuses; about 2e29 FLOP feasible by 2030; power binds first). Epoch AI, San Francisco. Observation year 2024. Accessed 2026-11-03.back
  5. 2Sevilla, J., Heim, L., Ho, A., Besiroglu, T., Hobbhahn, M. and Villalobos, P., Compute Trends Across Three Eras of Machine Learning, arXiv:2202.05924, 2022 (abstract page: doubling about every 20 months before 2010 and about every 6 months in the deep learning era). arXiv, Ithaca. Observation year 2022. Accessed 2026-11-03.back
  6. 44United States Congress, CHIPS and Science Act, Public Law 117-167, enacted 9 August 2022 (text of section 102 on Congress.gov, first 100,000 characters read: the funding lines for sections 9902 and 9906 of Public Law 116-283 by fiscal year 2022 to 2026 sum to USD 39 billion for manufacturing incentives and USD 11 billion for research, USD 50 billion in all; the totals are sums of the listed amounts, not stated in the text). Observation year 2022. Accessed 2026-11-03.back
  7. 12European Commission (ed.), European Chips Act policy page (entered into force 21 September 2023; three pillars; more than EUR 33 billion across 19 State aid decisions; second Chips Act proposed June 2026). Directorate-General for Communications Networks, Content and Technology, Brussels. Observation year 2026. Accessed 2026-11-03.back
  8. 13US Department of Commerce, Bureau of Industry and Security (ed.), home page and news items (May 2026 guidance on licences for advanced computing items to Country Group D:5 entities; 7 April 2026 extension of the authorised IC designer timeline to 31 December 2026; Section 232 investigation of semiconductor imports listed). BIS, Washington. Observation year 2026. Accessed 2026-11-03.back
  9. 15US Department of Energy (ed.), DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers, 20 December 2024 (176 TWh in 2023, about 4.4%; 325 to 580 TWh, about 6.7% to 12%, in 2028). US Department of Energy, Washington. Observation year 2023. Accessed 2026-11-03.back
  10. 23AZFamily (ed.), TSMC Arizona expansion could make it Phoenix’s largest water customer, 22 July 2026 (total planned investment USD 265 billion; about 5,300 acre-feet a year for the first fab; 65% reclaimed; goal 90%); local news reporting company figures, weak evidence. Gray Media, Phoenix. Observation year 2026. Accessed 2026-11-03.back
  11. 6Luccioni, A.S., Jernite, Y. and Strubell, E., Power Hungry Processing: Watts Driving the Cost of AI Deployment?, ACM FAccT 2024, arXiv:2311.16863 (abstract page: multi-purpose generative models orders of magnitude more costly per 1,000 inferences than task-specific models). ACM, New York. Observation year 2024. Accessed 2026-11-03.back
  12. 7Patterson, D. and others, Carbon Emissions and Large Neural Network Training, arXiv:2104.10350, April 2021 (abstract page: sparse models under one tenth of the energy; location effect 5 to 10 times; choice of model, datacentre and processor 100 to 1000 times). arXiv, Ithaca. Observation year 2021. Accessed 2026-11-03.back

Cite this chapter

Valadbaygi, S. (Ed.). (2026). Chapter 1: AI compute is physical infrastructure built on a few chips and a few places. In Exponential Technology Report 2026: Advanced Semiconductors and AI Compute (Baseline Edition 2026, Report No. NR-XTECH-SEMICONDUCTORS-GLOBAL-2026-BASELINE) [Draft under review]. The Global Centre for Risk and Innovation (GCRI).

Every citation form of the whole report, BibTeX and RIS included, is on the edition home. Cite this report

ReportsContribute

Send evidence or a correction on Advanced Semiconductors and AI Compute.

Reports is the Consortium’s quarterly publication, written by six desks. This page sits with the Global Centre for Risk and Innovation (GCRI) Systems Evidence desk. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work. Sent to the wrong desk, it is forwarded, not refused.

GCRI Systems Evidence deskExponential Technology

Never send personal data, sensitive infrastructure detail or anything you have no right to share. A contribution is reviewed before use; sending it confers no authorship or editorial rights, and sponsorship never buys a finding.

CC BY 4.0 Free to read. Text and figures: CC BY 4.0. Data: CC0 1.0. Authors keep copyright. Published by GCRI Canada.

The Global Centre for Risk and Innovation

Science · evidence · measurement

Not for profit

Our articles, maps and twins are free to read and free to republish. Readers and partners keep it that way.

Partner with us

Who funds us

GCRI is a not-for-profit institution. To ask who funds a piece of work, use the funding questions form.

Independence

Editors and the expert reviewers named on each article decide what we publish. Funders and partners do not review or approve articles before publication.

Corrections policy

Licence

Articles: CC BY 4.0. Figures, maps, twins and data: CC BY 4.0. Other uses: ask through the permissions form.

Privacy choices