Mbinu
How the edition was made: methods, the model card, the evidence cut and its age, the acceptance record and the glossary.
Contents
- ·Edition home
- ·Executive summary
- 1HPC is power, cooling and policy before it is speed
- 2Key findings
- 3Systems and sectors at baseline
- 4Risks and capability
- 5Power and siting limits meet fast-growing demand
- 6Readings from the other desks
- 7The 2027 agenda
- 8Systems in depth
- AMbinu
- BData appendix
- CReferences and record
- ·Complete edition, for print
A.1Mbinu
The global baseline uses published values and four simple derivations. Each is shown in full so that any reader can reproduce it.
| No. | Derivation | Calculation | Result |
|---|---|---|---|
| (1) | Combined power of the five largest systems, June 2026 | 42,220 + 29,685 + 24,607 + 38,698 + 15,794 kW | 151,004 kW, about 151 MW |
| (2) | Growth of combined list performance in six months | 18.73 ÷ 14.99 | about 1.25 times |
| (3) | Share of listed systems that use accelerators | 276 ÷ 500 | 55.2% |
| (4) | Ratio of El Capitan HPL-MxP to Linpack | 16.7 ÷ 1.809 | about 9.2 times |
| Strength | Definition |
|---|---|
| STRONG | a disruption in the source system has historically changed the target system’s service or output within one season, in more than one region |
| MODERATE | an effect is documented but buffered or slower |
| WEAK | an effect is plausible but not documented at global scale |
| Label | Rule |
|---|---|
| Reported | value as issued by its source, with year |
| Inferred | reasoned from cited evidence; never a measurement |
| Imekadiriwa | Nexus analytical value with stated method |
| Hali tarajiwa | conditional on stated assumptions |
| Expert input | structured editorial judgement awaiting evidence |
A.2Model card, HPC System Model v0
- Model ID
- MDL-XT07-HPC · v0.1
- Madhumuni
- represent HPC and advanced computing as infrastructure coupled to electricity, cooling, chips, public funding, workforce and security policy
- Familia
- graph-based system of systems
- Inputs
- the cited sources of this edition; no proprietary data
- Mbinu
- expert-specified coupling strengths (STRONG · MODERATE · WEAK) from assessments and analogue events
- Validation
- none yet, first validation against federation evidence in 2027
- Kutokuwa na uhakika
- qualitative; strengths are hypotheses, not measurements
- Domain of validity
- global scale; not for asset-, site- or country-level use
- Reproducibility
- strength table and sources published in this edition
typed in v0
documented within one season
documented, buffered
plausible, undocumented
A.3The baseline evidence cut
Every Nexus edition is published from a frozen evidence cut. The global baseline cut is documentary: it pins the exact edition of every source cited. Federation cuts begin with the first national Foundry cuts in 2027.
- Cut ID
- DC-XTECH-HPC-GLOBAL-2026-B0
- Type
- baseline documentary cut
- Resolution policy
- pinned · as of 30 September 2026
- Sources pinned
- 43 cited sources, each at the edition and access date in the references
- Model runs
- none at baseline, analytical values are stated with their method
- Digest
- computed at freeze and recorded in the publication record
- Supersedes
- nothing, first cut of the series
- Superseded by
- the 2027 HPC annual cut
Lineage
- 01
source edition
- 02
pinned citation
- 03
finding
- 04
page
- 05
publication record
| Element | Baseline 2026 | 2027 global annual |
|---|---|---|
| Vyanzo | cited global publications, pinned by edition | national, regional and sector cuts through one graph |
| Analysis | stated methods and editorial estimates | typed analysis specifications |
| Models | HPC System Model v0 specification | model runs with receipts |
| Matokeo ya uchunguzi | twelve baseline findings | findings re-bound to federation evidence |
| Mapitio | five passes scheduled | five passes on the exact digest |
A.4Evidence age and coverage
A 2026 global edition rests on evidence observed in many different years. This page shows each time base, so that no single year label hides older evidence.
| Indicator | Imeonekana | Released | Cadence | Next release |
|---|---|---|---|---|
| Top-system results | 2026 | June 2026 | periodic | November 2026 list |
| Efficiency results | 2026 | June 2026 | periodic | November 2026 list |
| US data centre energy | 2023 | December 2024 | one-off | next national report |
| Export rule timeline | 2026 | April and May 2026 | per event | further guidance |
| EuroHPC budget | 2021 to 2027 | page read 2026-11-03 | continuous | regulation changes |
| Footprint papers | 2021 to 2024 | preprints and papers | one-off | new studies |
| MPI and OpenMP | 2024 to 2026 | June 2025 · July 2026 | periodic | OpenMP 6.1 |
| FugakuNEXT plan | 2025 | August 2025 | one-off | detailed design |
| Mfumo | Coverage | Thinnest dimension |
|---|---|---|
| Mifumo | GOOD | unlisted systems |
| Efficiency | PARTIAL | facility-level figures |
| Nishati | PARTIAL | common basis for facilities |
| Utawala | PARTIAL | text of rules and guidance |
| Uwekezaji | PARTIAL | public funding outside the US and EU |
| Supply | THIN | accelerator and memory supply |
| Skills | THIN | workforce numbers |
| Maji | THIN | cooling-water use |
| Nyanja | Source that would fill it | Blocker |
|---|---|---|
| Cooling-water use of large computing facilities | operators, national laboratories, regulators under the EU scheme | not retrieved; no primary figure read |
| Facility energy and demand on a common basis for HPC and for commercial AI workloads | national statistics, operators, IEA | US figures read in the Berkeley Lab report29 and global figures in the IEA report28; no split between HPC and commercial AI |
| Data centre electricity by country outside the US | IEA Energy and AI28, national statistics | global figures read in the report; national series not bound |
| Lawrence Berkeley report text, including water use | LBNL 2024 United States Data Center Energy Usage Report29 | report read; its water estimates are not yet used in this edition |
| Power draw and cost of JUPITER from primary pages | EuroHPC, Jülich Supercomputing Centre | power and budget are not stated on the page read9; list value only1 |
| LineShine architecture and independent description | operator and peer-reviewed or institutional papers | list pages and one trade analysis30; no operator description |
| Specifications and schedules of DOE Solstice, Equinox, Lux and Discovery | Argonne and Oak Ridge system pages; DOE announcement31 | the DOE announcement gives 100,000 and 10,000 Blackwell GPUs for Solstice and Equinox and 2026 delivery for Equinox, without performance figures31; Oak Ridge pages name Lux and Discovery without specifications827 |
| El Capitan system description from the laboratory | Lawrence Livermore National Laboratory32 | news item read: 1.742 exaflops on HPL and 2.79 exaflops peak in November 202432; current values from the list1 |
| Text of the 2023 US advanced computing rule | Federal Register33 | Federal Register record and abstract read; full rule text not analysed |
| EU delegated regulations on data centre reporting and rating from the primary text | EUR-Lex3435 | the 2024 regulation was read34; the 2026 rating proposal is known from a law firm note35 |
| Supply of accelerators, high-bandwidth memory and fabrication capacity for HPC | industry and trade statistics | not retrieved; no figure used |
| HPC workforce size and skills gaps | national statistics, OECD, industry surveys | not retrieved; capability rung stays LISTED |
| Survey data on data centre efficiency and density | Uptime Institute global survey42 | report read; weighted average PUE of 1.54 in 2025; no figure used in the findings |
| Public funding totals for HPC outside the US and EU | national strategies and ministries | not consolidated; UK plan gives no budget figure23 |
| Results of the November 2026 TOP500 list | TOP500 news and lists6 | not yet published at page read; the news page shows June 2026 as latest |
| Capability rungs beyond list monitoring, planning and funding | agencies, operators, laboratories | not reviewed |
A.5Evidence cards, two anchor sources
Published by the TOP500 project at ISC 2026 on 23 June 2026, 67th edition. Periodic. Role: binds INF-001 and the dashboard power figures. Coverage: systems that submit results.
EVD-XT07-001
TOP500 list of June 20261
Published by the TOP500 project on the same date. Periodic. Role: binds ENE-004. Coverage: listed systems that report power.
EVD-XT07-002
Green500 list of June 20263
| Dimension | EVD-XT07-001 | EVD-XT07-002 |
|---|---|---|
| Completeness | submitted systems only | listed systems that report power |
| Timeliness | twice a year | twice a year |
| Revision | new list replaces old | new list replaces old |
| Method transparency | benchmark method published | power method stated by the project |
| External validation | results reported by submitters | results reported by submitters |
Evidence cards to follow in 2027
A.6Publication acceptance record
Editorial self-check at evidence close. Every applicable item carries a disposition; the remaining items are review, approval and deposit.
| Group | Area | Items | Pass | N/A | Pending |
|---|---|---|---|---|---|
| D1 | Identity and scope | 8 | 8 | 0 | 0 |
| D2 | Source and data | 12 | 9 | 1 | 2 |
| D3 | Analysis and models | 12 | 3 | 8 | 1 |
| D4 | Findings and claims | 9 | 8 | 0 | 1 |
| D5 | Figures, tables, maps | 10 | 6 | 2 | 2 |
| D6 | Editorial and citations | 10 | 8 | 0 | 2 |
| D7 | Review and approval | 13 | 0 | 0 | 13 |
| D8 | Artefact and design | 12 | 7 | 0 | 5 |
| D9 | Registry and release | 9 | 0 | 0 | 9 |
Publication gate
- Conditions met
- 14 of 23 at evidence close
- Blocking problems
- none recorded
- Open conditions
- reviews, approval of the exact digest, deposit, registry pointer
Not applicable at baseline
- Model runs
- first runs in Q1 2027
- Ramani
- plates in Q2 2027
- Waivers
- none
A.7Status words
| Hali | What it means |
|---|---|
| Imeonekana | Measured directly and published by the source. |
| Reported | Published statistics of an official or institutional source. |
| Declared | Stated by the body concerned; not independently measured. |
| Imekadiriwa | A Nexus estimate from cited evidence, with its method in Annex A. |
| Modelled | Produced by a stated model from cited inputs. |
| Inferred | Reasoned from cited evidence; not measured. |
| Forecast | A projection by the cited source. |
| Simulated | A simulation result under stated assumptions. |
| Hali tarajiwa | A conditional statement under stated assumptions; never a prediction. |
| Expert input | Expert judgement, recorded with its basis. |
| Candidate | Proposed for a later edition; not yet evidence. |
| Evidence-quality levels (EQL 1 to 5) grade the source behind a value; they are not a score of the subject. | |
A.8How the edition is made, released and corrected
Every edition is built from a sealed evidence cut, reviewed under three keys, signed and deposited before any site shows it as published; corrections go to Nexus Standards as change proposals and feed the next cut.
| Step | What happens | Chanzo |
|---|---|---|
| 01 Evidence arrives from the pillars | Foundry seals observations and indicator series, Universe hands over after-action and scenario findings, and contributions are screened at every gateway, including this site’s. | |
| 02 The cut is sealed | On the cut date the evidence is pinned under one identifier, so anyone can rebuild the edition from exactly what it was built on. | |
| 03 The edition is assembled | The edition is drafted from its template and content pack. Agents may draft; they never publish and never decide. | |
| 04 The keys turn | At the publisher of record the truth key comes first: every value is checked against its source or stated method before the meaning and capital keys turn. No edition moves on until every key has been turned by a named steward. | |
| 05 The gate opens | A named approver releases one exact digest of the edition; any change after this point is a new version, never a silent edit. | |
| 06 Signed and deposited | The PDF, the web edition and the data bundle are signed together with a manifest and an RO-Crate description, and deposited on Zenodo with a concept and a version DOI. | |
| 07 Published on its desk’s site | Only then does the edition show as published on its desk’s Reports site, and the Collection lists it as a part. | |
| 08 Corrections go to Standards | Method and correction findings go to Nexus Standards as change proposals; citations, corrections and new contributions feed the next evidence cut. |
A teaching model of the process, not a forecast and not a record of any run; dates and counts on this page come from the programme record. Sources: Reports governance at GCRI Canada (2026); Zenodo, repository of record for Reports deposits (2026).
Lifecycle, one state at a time: Draft, Computed, Evidence bound, Reviewed, Approved for release, Published, Corrected, Superseded. This edition is a draft.
A.9Glossary and abbreviations
| Istilahi | Meaning |
|---|---|
| Baseline Edition | the inaugural annual edition that fixes the reference state for the series |
| WEFH+B | water, energy, food, health and biodiversity, the national nexus core |
| Muunganiko | a typed dependency through which one system transmits to another |
| Protection gap | the share of losses not covered by insurance |
| Capability ladder | LISTED → EXISTS → ADEQUATE → TESTED → DEMONSTRATED |
| EQL | evidence-quality level, 1–5 |
| Evidence cut | a frozen resolution of exact sources and versions |
| NNC · NLC · NWG | National Nexus Consortium · National Leadership Council · National Working Groups |
| Abbreviation | Expansion |
|---|---|
| BIS | Bureau of Industry and Security |
| CPU | central processing unit |
| DOE | US Department of Energy |
| ECP | Exascale Computing Project |
| EF | exaflops |
| GPU | graphics processing unit |
| HBM | high-bandwidth memory |
| HPC | high-performance computing |
| HPCG | High Performance Conjugate Gradients benchmark |
| HPL-MxP | mixed-precision High Performance Linpack benchmark |
| PUE | power usage effectiveness |
| TWh | terawatt-hours |
A.10Technical annex
| Item | Annex | Contents | Hali |
|---|---|---|---|
| A.1 | Indicator definitions | the global indicator set, with units and sources | IN THIS EDITION |
| A.2 | Coupling strengths | the twenty-cell matrix and its rationale | IN THIS EDITION |
| A.3 | Scenario specification | SCN-XT07-PWR-01 assumptions and signposts | IN THIS EDITION |
| A.4 | Capability placements | rung evidence for fifteen functions | IN THIS EDITION |
| A.5 | Model diagnostics | first runs and validation | 2027 |
| A.6 | Sensitivity analyses | influence of coupling strengths | 2027 |
| A.7 | Cluster notes | cards for the major exascale and data centre clusters | 2027 |
| A.8 | Review statements | pass-by-pass review record | AT RELEASE |
A.1 · indicator definition: Linpack result (Rmax)
- Construct
- sustained speed of a dense linear solve at the largest size the system can run
- Unit
- floating-point operations per second, shown here in exaflops
- Mbinu
- the Linpack benchmark run by the submitter and reported to the list
- Chanzo
- TOP500 list, June 20261
- Kutokuwa na uhakika
- reported by submitters; not independently repeated here
- Known limits
- one benchmark; dense algebra only; not a measure of useful scientific output
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
- 29Shehabi, A., Smith, S.J., Hubbard, A., Newkirk, A., Lei, N., Siddik, M.A.B., Holecek, B., Koomey, J., Masanet, E. and Sartor, D., 2024 United States Data Center Energy Usage Report, LBNL-2001637 (December 2024; report PDF read; 176 TWh in 2023, 4.4% of US electricity; about 325 to 580 TWh in 2028, 6.7% to 12.0%; direct water estimates included). Lawrence Berkeley National Laboratory, Berkeley. Observation year 2023. Accessed 2026-11-05.back
- 28IEA (ed.), Energy and AI, World Energy Outlook Special Report (April 2025; report PDF read; around 415 TWh for data centres in 2024, about 1.5% of world electricity; around 945 TWh in 2030 in the Base Case). International Energy Agency, Paris. Observation year 2024. Accessed 2026-11-05.back
- 9EuroHPC Joint Undertaking (ed.), JUPITER launching Europe’s exascale era (news item, 5 September 2025; about 24,000 GH200 superchips; warm-water cooling; funding shares; power and budget not stated). EuroHPC JU, Luxembourg. Observation year 2025. Accessed 2026-11-03.back
- 1TOP500 (ed.), TOP500 list, June 2026, 67th edition (list page; top systems, Rmax, Rpeak, cores and power). TOP500 project, Hamburg. Observation year 2026. Accessed 2026-11-03.back
- 30Chips and Cheese (ed.), Cozma, G., TOP500 at ISC’26: We have a New Number 1 (25 June 2026; LineShine described as the first Chinese submission to the TOP500 in nine years; other Chinese exascale systems not listed); trade analysis, weak evidence. Chips and Cheese. Observation year 2026. Accessed 2026-11-05.back
- 31US Department of Energy, Office of Science (ed.), Energy Department Announces New Partnership with NVIDIA and Oracle to Build Largest DOE AI Supercomputer (bulletin, 28 October 2025; Solstice with 100,000 NVIDIA Blackwell GPUs; Equinox with 10,000 Blackwell GPUs, expected to be delivered in 2026; no performance figures stated). US Department of Energy, Washington. Observation year 2025. Accessed 2026-11-05.back
- 8Oak Ridge Leadership Computing Facility (ed.), compute systems (index page; Lux and Discovery labelled as introduced; no specifications). Oak Ridge National Laboratory, Oak Ridge. Observation year 2026. Accessed 2026-11-03.back
- 27Oak Ridge National Laboratory (ed.), The Genesis Mission (page; Lux and Discovery listed with one-line descriptions only). Oak Ridge National Laboratory, Oak Ridge. Observation year 2026. Accessed 2026-11-03.back
- 32Lawrence Livermore National Laboratory (ed.), Lawrence Livermore National Laboratory’s El Capitan verified as world’s fastest supercomputer (news item, 18 November 2024; 1.742 exaflops on HPL; 2.79 exaflops peak; more than 11,000 compute nodes). Lawrence Livermore National Laboratory, Livermore. Observation year 2024. Accessed 2026-11-05.back
- 33US Department of Commerce Bureau of Industry and Security (ed.), Implementation of Additional Export Controls: Certain Advanced Computing Items; Supercomputer and Semiconductor End Use; Updates and Corrections (interim final rule, 88 FR 73458, Federal Register document 2023-23055, 25 October 2023, effective 17 November 2023; Federal Register record and abstract read). Office of the Federal Register, Washington. Observation year 2023. Accessed 2026-11-05.back
- 34European Commission (ed.), Commission Delegated Regulation (EU) 2024/1364 of 14 March 2024 on the first phase of the establishment of a common Union rating scheme for data centres (OJ L series, 17 May 2024; text read via the Publications Office; operators with installed IT power demand of at least 500 kW report by 15 September 2024, then by 15 May 2025 and every year thereafter). Publications Office of the European Union, Luxembourg. Observation year 2024. Accessed 2026-11-05.back
- 35Latham & Watkins (ed.), European Commission Proposes Mandatory Sustainability Label for Data Centres (client note, 25 September 2026; Commission proposal of 21 September 2026 for a delegated regulation; A to G ratings for power and water usage effectiveness; labels issued annually from 15 August 2027; two-month scrutiny by Parliament and Council; data centres of 500 kW or above); law firm summary, the Commission text was not opened. Latham & Watkins. Observation year 2026. Accessed 2026-11-05.back
- 42Donnellan, D., Lawrence, A., Bizo, D., Rogers, O., Judge, P., Davis, J., Smolaks, M. and Weinschenk, R., Uptime Institute Global Data Center Survey 2025, UII Keynote Report 180 (July 2025; report PDF read; more than 800 owners and operators surveyed; weighted average annual PUE of 1.54 in 2025). Uptime Institute, New York. Observation year 2025. Accessed 2026-11-05.back
- 23UK Government (ed.), AI Opportunities Action Plan (page; expand the AI Research Resource by at least 20 times by 2030; non-AI HPC and exascale flagged as unresolved; no budget figure). Department for Science, Innovation and Technology, London. Observation year 2025. Accessed 2026-11-03.back
- 6POST-CLOSE, context only: TOP500 (ed.), TOP500 news index (headlines and dates of the lists of November 2024, June 2025, November 2025 and June 2026; no November 2026 list at page read). TOP500 project, Hamburg. Observation year 2026. Accessed 2026-11-03.back
- 3TOP500 (ed.), Green500 list, June 2026 (list page; top five by gigaflops per watt). TOP500 project, Hamburg. Observation year 2026. Accessed 2026-11-03.back