Facilities written 2026-08
Capacity
Compute capacity is estimated site by site, then read against the thresholds in statute.
Compute capacity is estimated site by site, then read against the thresholds in statute.
The Act attaches its systemic-risk obligations at 10²⁵ FLOP of cumulative training compute. Read that as a quantity of physical capacity and it becomes a question about buildings: how long would each site take to accumulate it? Of 499 facilities that disclose a chip count or a power figure, 178 would get there inside a month of sustained operation, 250 inside a quarter and 332 inside a year; the fastest needs under a day. None of that is a compliance finding: the obligation attaches to a model placed on a market, and no site trains one model with its whole floor. It is the shape of the population the line is drawn across.
Plate ITime to 10²⁵ FLOP499 sites · complete
Plate I. One vertex per facility (every estimable site, none sampled out), ranked left to right by how many days of 24-hour sustained operation its estimated daily FLOP budget would need to accumulate 10²⁵ FLOP.
The vertical axis is that duration on a log scale, floored at one day and capped at ten years: durations beyond a decade add no policy information and letting the tail set the scale would flatten everything under it.
The first three bands are NESTED (a site clearing in twelve days is counted in all three), so they do not sum to the population. A site’s estimate is the chip-count path where a chip count is disclosed and the power path otherwise.
The population is every facility whose record has at least one of those two inputs: 499 sites, of which those clearing within a year are spread across 29 countries. That is 499 of 4,573 facilities Scrutica tracks: 512 disclose a chip count or a power figure, one of which is a chip-manufacturing site (logic and memory fabs, packaging), excluded because their megawatts power chip production, which the power path would read as training compute. The rest disclose neither figure.
A site absent from this curve is absent because of what its record does not say; that is a fact about disclosure, and it says nothing about capacity. The curve is therefore a floor on how many sites could clear the line, and the ladder above states how far short of a census that floor is.
The Estimator view keeps the fabs, because a fab’s chip count and power figure can still be checked against each other; that is the whole difference between the two denominators.
A duration here is a statement of capability; no legal finding follows from it. It assumes the whole site training one model, which no operator does, so every figure is an upper bound on speed and a lower bound on time. All values are estimates.
The same estimate, read against the lines that carry legal consequence (the EU AI Act’s 10²⁵ trigger, China’s ~10²⁴ registration threshold, the rescinded US 10²⁶), with the verbatim statutory anchors, a reverse calculator, and the Article 52(6) register reconstruction.
Per site, per regime
Plate I holds one regime fixed. Here every regime is selectable (the active EU trigger, China’s registration threshold, the rescinded US order, the vetoed California bill, an academic proposal), against every facility, with the residual shown beside the EU AI Office’s ±30% measurement tolerance from the July 2025 GPAI Guidelines. Estimates use dense FP16 specs and no 2:4 structured sparsity. The reverse calculator solves the other direction: what a site would need, in chips or in days, to reach a selected line.
Facility-side: which sites have the physical compute to put a model across the EU AI Act 10²⁵ FLOP threshold (and the China CAC, US EO 14110 lines).
FACILITIES ≤ 30 DAYS (10²⁵ FLOP)
FACILITIES ≤ 90 DAYS
FACILITIES ≤ 1 YEAR
MODELS ABOVE EU
MODELS ABOVE US
EU AI Act: Systemic Risk GPAI
Cumulative training compute, measured in total FLOP; inference throughput is out of scope. Article 51(2): "A general-purpose AI model shall be presumed to have high impact capabilities ... when the cumulative amount of computation used for its training measured in floating point operations is greate...
Regulation (EU) 2024/1689, Article 51(2) + Annex XIII · Official Journal of the European Union · Tier 1
Threshold-Capable Facilities by Country (within 1 year, 10²⁵ FLOP)
⚠ = facility is in a jurisdiction where EU AI Act: Systemic Risk GPAI applies
| Facility | Country | Daily FLOP | Days to 10²⁵ FLOP | Estimation Path | Regime Exposure | Expand details |
|---|---|---|---|---|---|---|
| (†) | AE | <1 | Power: 5000 MW | Not subject | ||
| US | <1 | Hardware: 700,000 GPUs (gb200) | Not subject | |||
| US | <1 | Power: 2260 MW | Not subject | |||
| US | <1 | Power: 2200 MW | Not subject | |||
| FR | <1 | Hardware: 500,000 GPUs (gb200) | SUBJECT | |||
| US | <1 | Power: 2000 MW | Not subject | |||
| IN | <1 | Hardware: 450,000 GPUs (gb200) | Not subject | |||
| US | <1 | Power: 1800 MW | Not subject | |||
| SA | <1 | Power: 1500 MW | Not subject | |||
| US | <1 | Power: 1400 MW | Not subject | |||
| US | <1 | Power: 1400 MW | Not subject | |||
| US | <1 | Power: 1200 MW | Not subject | |||
| US | <1 | Hardware: 300,000 GPUs (b200) | Not subject | |||
| US | <1 | Power: 1000 MW | Not subject | |||
| US | <1 | Power: 1000 MW | Not subject | |||
| US | <1 | Hardware: 530,000 GPUs (H100 assumed) | Not subject | |||
| US | <1 | Power: 750 MW | Not subject | |||
| US | <1 | Hardware: 180,000 GPUs (gb200) | Not subject | |||
| US | <1 | Power: 636 MW | Not subject | |||
| US | <1 | Power: 631 MW | Not subject | |||
| US | <1 | Hardware: 150,000 GPUs (gb200) | MAY APPLY | |||
| US | <1 | Hardware: 150,000 GPUs (gb200) | Not subject | |||
| FR | <1 | Hardware: 150,000 GPUs (gb200) | SUBJECT | |||
| FR | <1 | Hardware: 150,000 GPUs (gb200) | SUBJECT | |||
| US | <1 | Power: 600 MW | Not subject | |||
| US | ~1 | Power: 510 MW | Not subject | |||
| SA | ~1 | Power: 500 MW | Not subject | |||
| IT | ~1 | Power: 500 MW | SUBJECT | |||
| US | ~1 | Power: 500 MW | Not subject | |||
| FR | ~1 | Hardware: 120,000 GPUs (gb200) | SUBJECT | |||
| US | ~1 | Hardware: 131,072 GPUs (b200) | Not subject | |||
| US | ~1 | Power: 467 MW | Not subject | |||
| US | ~1 | Hardware: 110,000 GPUs (gb200) | Not subject | |||
| US | ~1 | Power: 450 MW | Not subject | |||
| US | ~2 | Power: 369 MW | Not subject | |||
| US | ~2 | Power: 340 MW | Not subject | |||
| US | ~2 | Power: 339 MW | Not subject | |||
| US | ~2 | Hardware: 202,224 GPUs (H100 assumed) | Not subject | |||
| US | ~2 | Power: 303 MW | Not subject | |||
| US | ~2 | Power: 300 MW | Not subject | |||
| US | ~2 | Power: 300 MW | Not subject | |||
| US | ~2 | Power: 300 MW | Not subject | |||
| US | ~2 | Power: 284 MW | Not subject | |||
| US | ~2 | Power: 262 MW | Not subject | |||
| US | ~2 | Power: 250 MW | Not subject | |||
| US | ~2 | Power: 250 MW | Not subject | |||
| US | ~2 | Power: 250 MW | Not subject | |||
| US | ~2 | Power: 250 MW | Not subject | |||
| US | ~2 | Power: 250 MW | Not subject | |||
| KR | ~2 | Hardware: 60,000 GPUs (gb200) | Not subject |
Estimates assume continuous operation at full training capacity using dense FP16 TFLOP/s (no sparsity). Default MFU: 40%. Default interconnect efficiency: 85%. Hardware-based estimates assume H100 SXM5 where GPU model is unspecified. Power-based estimates use PUE 1.15 (operator fleet disclosures: Microsoft FY25, AWS, xAI, Stargate, Google, Meta) and 49% GPU share of IT load (Scrutica-editorial midpoint, calibrated against SemiAnalysis's 700 W / 1,275 W per-GPU server decomposition; see Calibration tab). Residuals from this calibration fall inside the EU AI Office's ±30% measurement tolerance for cumulative training FLOP under the July 2025 GPAI Guidelines. Methodology version: 1.3.6.
(†) Marked rows separate a published envelope from operational capacity; hover the marker for the per-facility caveat. The row's daily-FLOP value is the published-envelope ceiling, and the caveat states the Phase-1 operational capacity where the two materially differ.
Article 52(6)
The Act requires the Commission to maintain a public list of models with systemic risk. This is a reconstruction of what that list would contain, assembled from provider disclosures and compute estimates, with each entry’s basis stated and its uncertainty shown.
Eleven Commission surfaces, checked on 21 July 2026, have no published list of general-purpose AI models with systemic risk. Article 52(6) requires one. It is the Commission's duty and the only public surfacing of which models cross the line drawn by Article 51(2), the presumption of high-impact capabilities above 10²⁵ FLOP of cumulative training compute; notification, the providers' side, runs to the Commission in private by the Act's design. The table fills the gap with a reconstruction from Epoch AI's public compute estimates, every model whose estimate clears the presumption line, against the columns a register would settle and this cannot. Nothing in this table asserts that any provider missed an obligation.
| Model | Provider | Training compute (Epoch est.) | Epoch publication date | Art 111(3) cohort | Code of Practice | Public notification record |
|---|---|---|---|---|---|---|
| Grok 4 | xAI | 5 × 10²⁶Speculative | 9 Jul 2025 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Signatory (Safety and Security chapter only) | Not found in the enumerated sources as of 21 Jul 2026† |
| GPT-4.5 | OpenAI | 3.8 × 10²⁶Likely | 27 Feb 2025 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026† |
| Grok 3 | xAI | 3.5 × 10²⁶Likely | 17 Feb 2025 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Signatory (Safety and Security chapter only) | Not found in the enumerated sources as of 21 Jul 2026† |
| GPT-5 | OpenAI | 6.6 × 10²⁵Speculative | 7 Aug 2025 | After 2 Aug 2025 · obligations apply from placement | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026 |
| Llama 4 Behemoth (preview) | Meta AI | 5.184 × 10²⁵Likely | 5 Apr 2025 | Unreleased per Epoch · placement trigger not publicly established‡ | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026 |
| Gemini 1.0 Ultra | Google DeepMind | 5 × 10²⁵Speculative | 6 Dec 2023 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Signatory, full Code (listed as Google) | Not found in the enumerated sources as of 21 Jul 2026† |
| Llama Nemotron Ultra 253B | NVIDIA | 3.911 × 10²⁵Likely | 18 Mar 2025 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
| Composer 2.5 | Cursor | 3.87 × 10²⁵Likely | 18 May 2026 | After 2 Aug 2025 · obligations apply from placement | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026 |
| Llama 3.1-405B | Meta AI | 3.8 × 10²⁵Confident | 23 Jul 2024 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
| Claude 3.7 Sonnet | Anthropic | 3.35 × 10²⁵Likely | 24 Feb 2025 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026† |
| Grok-2 | xAI | 2.96 × 10²⁵Confident | 13 Aug 2024 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Signatory (Safety and Security chapter only) | Not found in the enumerated sources as of 21 Jul 2026† |
| Claude 3.5 Sonnet | Anthropic | 2.7 × 10²⁵Speculative | 20 Jun 2024 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026† |
| Doubao-pro | ByteDance | 2.505 × 10²⁵Likely | 28 Oct 2024 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
| Composer 2 | Cursor | 2.32 × 10²⁵Confident | 19 Mar 2026 | After 2 Aug 2025 · obligations apply from placement | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026 |
| GPT-4 (Jun 2023) | OpenAI | 2.1 × 10²⁵Likely | 13 Jun 2023 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026† |
| GPT-4 (Mar 2023) | OpenAI | 2.1 × 10²⁵Likely | 15 Mar 2023 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Signatory (full Code) | Not found in the enumerated sources as of 21 Jul 2026† |
| Nemotron-4 340B | NVIDIA | 1.8 × 10²⁵Confident | 14 Jun 2024 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
| Qwen3-Max | Alibaba | 1.512 × 10²⁵Speculative | 5 Sep 2025 | After 2 Aug 2025 · obligations apply from placement | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026 |
| Pangu Ultra | Huawei | 1.069 × 10²⁵Confident | 10 Apr 2025 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
| Aramco Metabrain AI | Saudi Aramco | 1.05 × 10²⁵Likely | 4 Mar 2024 | Unreleased per Epoch · placement trigger not publicly established‡ | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026 |
| Inflection-2 | Inflection AI | 1.001 × 10²⁵Confident | 22 Nov 2023 | Pre 2 Aug 2025 · compliance due by 2 Aug 2027 | Not on the signatory list | Not found in the enumerated sources as of 21 Jul 2026† |
† For models on the market before 2 August 2025 the notification clock is genuinely unsettled. Article 111(3) defers “the obligations laid down in this Regulation” for this cohort to 2 August 2027, and it does so without carving the two-week notification duty out of that deferral; no published Commission guidance has fixed whether the duty was already live, and neither has any national court. The Commission's own FAQ states that providers of models placed on the market before 2 August 2025 “must comply with the AI Act obligations by 2 August 2027”. A blank cell here records that ambiguity, nothing about a provider.
‡ Article 52(1) can run before release (“...or it becomes known that it will be met”); unreleased models above the line are retained for that reason, with placement status shown as Epoch records it.
Enforcement runs on its own clock. The Commission's GPAI enforcement powers, information requests and evaluations, with Article 101 fines behind them, switch on from 2 August 2026 under Article 113; the cohort column above tracks when obligations fall due, which is a different question.
The provider files first, and privately. Once a model meets the Article 51(1), point (a) condition, or it becomes known that it will, the provider “shall notify the Commission without delay and in any event within two weeks”; the presumption that pulls a model into that condition is Article 51(2), the 10²⁵ FLOP line. Notifications travel through the Commission's EU SEND platform (“EU SEND ensures the confidentiality, integrity, and authenticity of the information shared”) or the AI Office's dedicated mailbox, under Article 78 confidentiality. The public's side of the bargain is Article 52(6): “The Commission shall ensure that a list of general-purpose AI models with systemic risk is published and shall keep that list up to date, without prejudice to the need to observe and protect intellectual property rights and confidential business information or trade secrets in accordance with Union and national law.” Provider-side privacy is the Act's design; the register is the single public surfacing the Act provides.
The companion instruments arrived on schedule. GPAI Guidelines came on 18 July 2025, the Code of Practice on 10 July 2025, the training-content summary template on 24 July 2025, and the serious-incident reporting template followed. The Article 52(6) register did not. Chapter V has applied since 2 August 2025 (Article 113), the publication duty has no internal deadline, and eleven months on none of the eleven Commission surfaces in the log below has the list or a link to it. No designation under the Commission's own-initiative power (Article 51(1), point (b)) turned up in the public record either. What circulates instead is third-party reconstruction from Epoch AI's compute estimates, which is what this table is, and says so.
A row appears when Epoch's point estimate of cumulative training compute clears 10²⁵ FLOP. These are estimates, and Epoch labels each by confidence: “Confident” means within a factor of 3, “Likely” within a factor of 10, and “Speculative” within a factor of 30; the July 2025 GPAI Guidelines add a separate ±30% measurement tolerance on cumulative FLOP. Each value is copied exactly from the public database, keeps Epoch's per-row vintage, and shows here at four significant figures. Crossing that line does not place a model on the Union market: placement under Article 3 is a legal determination the public record does not settle per model, and two of these rows are unreleased per Epoch.
There is also a known direction to the undercount. Epoch estimates training compute, while Recital 111 directs that the cumulative total behind the Article 51(2) presumption count compute more broadly, “such as pre-training, synthetic data generation and fine-tuning”. A reconstruction built on the narrower measure marks a floor for the population the presumption reaches.
Nor does a row say anything about a provider's compliance. It is not a claim that the provider owed or missed a notification: Article 111(3) provides that “Providers of general-purpose AI models that have been placed on the market before 2 August 2025 shall take the necessary steps in order to comply with the obligations laid down in this Regulation by 2 August 2027”. And it is not a claim that the presumption has survived, since Article 52(2) invites a provider to rebut it and any such rebuttal is as private as the notification itself. What a published register would resolve per model, this reconstruction cannot, which is the gap it documents.
Scope of the claim: no evidence of a published list after these checks, on this date; that is not a proof of non-existence, and the check is re-run before any republication of this table. Open-web searches on 2026-07-21 (re-running the 2026-06-09 pass) for a Commission publication of the list, and for provider self-disclosures of Article 52(1) notifications, surfaced neither; only statutory-text mirrors and third-party reconstructions (Epoch-derived rosters) appear.
| Source | Result | Page vintage | Method |
|---|---|---|---|
| Commission AI regulatory framework policy page | No register, no link to one | Last update 7 July 2026 | direct fetch |
| Guidelines for providers of general-purpose AI models | No register; describes notification duty only | Last update 28 April 2026 | direct fetch |
| Commission FAQ on GPAI provider obligations | No register; states the 2 August 2027 rule for pre-2 August 2025 models | Last update 11 November 2025 | direct fetch |
| AI Act Service Desk, Article 52 (Commission-operated) | Quotes the 52(6) duty verbatim; links no list | No page date; footer "Official version of 13 June 2024" (statute text vintage) | direct fetch |
| Commission factpage on GPAI obligations | No register; confirms the private notification mailbox (EU-AIOFFICEGPAI-SR-PROVIDERS@ec.europa.eu) | Last update 1 August 2025 | direct fetch |
| Commission news item on the GPAI Guidelines | No register | Published 18 July 2025, updated 31 July 2025 | direct fetch |
| GPAI Code of Practice policy page | No register, no link to one | Last update 25 June 2026 | direct fetch |
| Code of Practice contents + signatory list | Signatory list present; no model register | Last update 23 April 2026 | direct fetch |
| Commission Q&A on GPAI models in the AI Act | No register; references Article 111(3) special rules | Last update 9 September 2025 | direct fetch |
| European AI Office policy page | No register; classification work described, no published list | Last update 17 July 2026 | direct fetch |
| Site-scoped searches across Commission AI Office surfaces | Zero register hits; only obligation descriptions and templates | n/a | recorded search |
Sources
Regulation (EU) 2024/1689, Articles 51, 52, 78, 111(3), 113 (tier 1) · eur-lex.europa.eu (CELEX 32024R1689)
AI Act Service Desk, Articles 51 and 52 pages (Commission-operated; Article 52(1) and 52(6) wording verified 21 July 2026) · ai-act-service-desk.ec.europa.eu/en/ai-act/article-52
European Commission, Guidelines for providers of general-purpose AI models (EU SEND platform and notification mechanics) · digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers
European Commission, FAQ on obligations for GPAI providers (the 2 August 2027 sentence; page updated 11 November 2025) · digital-strategy.ec.europa.eu/en/faqs/guidelines-obligations-general-purpose-ai-providers
European Commission, GPAI Code of Practice contents and signatory list (page updated 23 April 2026; checked 21 July 2026) · digital-strategy.ec.europa.eu/en/policies/contents-code-gpai
Epoch AI, “Data on AI Models” (tier 2; CC-BY; database accessed 21 July 2026; per-row vintage in the export) · epoch.ai/data/ai-models
What a FLOP count misses
A compute threshold governs the training run. It does not govern what is drawn out of a model afterwards (scaffolding, tool use, inference-time search, fine-tuning), and that gap is what this addendum documents.
Article 51(2) presumes systemic risk above 10²⁵ FLOP; the rebuttal mechanism (and downstream deployment decisions) leans on capability evaluations, and the capability evaluations the public can read are floors. The UK AI Security Institute's 13 April 2026 evaluation of Claude Mythos Preview ran a simple agent scaffold held constant for cross-model comparability. On 13 May 2026 AISI put the caveat on the page in its own words: “the cap, alongside our use of a simple agent scaffold, artificially lowers success rates and understates what models can do with more tokens and stronger scaffolds.”
Two weeks after Mythos shipped, Inspect (AISI's own evaluation framework) added deepagent() (subagent delegation, persistent memory, structured planning) in v0.3.213 (27 April 2026); the published cyber methodology has not yet adopted it. At identical FLOP cost a model can clear a deployment threshold under a single-agent eval that it would fail under a multi-agent eval. Compute keeps accumulating while capability evaluations report floors, and policy cites the floors; no public evaluator states a ceiling.
Primary sources (authority tier 1)
AISI · “Our evaluation of Claude Mythos Preview's cyber capabilities” (13 Apr 2026): aisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities
AISI · “How fast is autonomous AI cyber capability advancing?” (13 May 2026; floor-estimate admission verbatim): aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capability-advancing
UK Government BEIS · Inspect framework CHANGELOG, v0.3.213 (27 Apr 2026; deepagent() + subagent delegation): github.com/UKGovernmentBEIS/inspect_ai/blob/main/CHANGELOG.md
European Parliament & Council · Regulation (EU) 2024/1689, Article 51(2) (10²⁵ FLOP presumption of systemic risk): eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689
Parameters
Every figure on this page (the reach curve, the days-to-threshold column, the reverse calculator) is estimated from the same five constants the Estimator exposes as sliders, held once in the methodology registry. Each card shows the default, its adjustable range, the paths it touches and the source it was calibrated against. Move them on the Estimator and the arithmetic behind these durations moves with them.
Fraction of peak throughput achievable across multi-node interconnect. 0.85 for NVLink, 0.7 for InfiniBand/Ethernet backend. Ethernet surpassed InfiniBand in AI back-end network market share in 2025 (Dell'Oro Group); UEC 1.0 spec released June 2025. Meta validated Ethernet RoCE at 24K-GPU scale for LLaMA 3.
Fraction of peak FLOPs achieved during training. Epoch AI (Sevilla et al, 2022) recommends 0.30 for LLMs, 0.40 for other networks. PaLM 540B (Chowdhery et al, 2022, Table 3) reports MFU 0.462, the highest published MFU figure we can verify against a primary source; the related HFU (hardware FLOP utilization, including rematerialization) is 0.578, a distinct metric that must not be read as the upper endpoint of an MFU range. LLaMA 3 405B (Meta 2024) achieves 380 TFLOP/s/GPU sustained on 16K H100 = 0.384 MFU. SemiAnalysis (Patel et al, "100,000 H100 Clusters", 2024-06-17) reports FP16 MFU of 0.40 on trillion-parameter training runs across the 100K-H100 reference cluster; the 0.40 default coincides with this figure and is recalibrated above Epoch’s 2022 LLM recommendation to reflect 2023–2025 frontier-lab achievements. Bounds 0.20–0.50 cover poorly-parallelized small-model runs through the published PaLM ceiling with a small headroom margin. We have not located any public training run with a measured MFU exceeding 0.50; if one surfaces, the upper bound widens with citation.
Ratio of total facility power to IT equipment power. Lower is more efficient. The 1.15 default reflects the empirical median across publicly-disclosed AI training fleets (Microsoft FY25 1.16; AWS 2024 1.15; xAI Memphis 1.18; Stargate Abilene 1.12 with heat recovery). Google fleet (1.09) and Meta (1.08) are below this; APAC and hot-climate facilities are above (Microsoft APAC 1.28). No operator publishes AI-pod-specific PUE separately, so the default is grounded in fleet averages.
Fraction of total IT power drawn by accelerator silicon (GPU TDP only), excluding in-server overhead and external infrastructure. SemiAnalysis (Patel et al, June 2024) is the primary source for the per-GPU server decomposition: 700 W GPU + ~575 W in-server overhead (CPUs, NICs, PSUs) = 1,275 W per-GPU server, implying an in-server GPU share of ~0.55. The cluster-level share (~0.47) and the 0.49 default that splits the two readings are Scrutica-editorial, extending the published per-server decomposition to the cluster level. Ranges allow for compute-dominant Blackwell/NVL72 racks (upper) and mixed-workload clusters (lower).
GPUs as a fraction of total data center capital expenditure. SemiAnalysis estimates 40–50%.
Every value here is adjustable on the Capacity Estimator sliders and traces to src/lib/methodology/versions.ts. The full derivation chain, with each formula and its calibration, is documented at /methodology. Estimation bounds are parameter-sensitivity ranges (not statistical confidence intervals).
The Estimator shows how any one of these durations is produced: the three back-solves behind a site’s daily budget, and how far they agree.