Corrections

Public log of cases where Scrutica displayed a value later found to be wrong or misleading. Each entry records what the site showed, the corrected value, when the error was caught, the root cause, and what changed to prevent recurrence; the routine data-additions feed (new facilities, capacity updates) lives at /changelog.

Public record of cases where the site displayed a wrong or misleading value. Each entry states what the site showed, the corrected value, when we caught the error, why it happened, and what changed to prevent recurrence; a reader who saw the wrong number before the fix can find the full disclosure here. (The separate data-additions feed lives at /changelog.)

Pipeline failures detected· last run 2026-07-27 02:01 UTC

Record · What Changed (the routine feed of records added and updated)

One or more ingestion pipelines reported a genuine failure. The state is shown as-is rather than softened; the underlying repairs are in progress. The per-pipeline detail lives on What Changed.

2026-07-26

A cancelled €30bn fab was still listed as an announced project, and a UK AI campus named the wrong owner alongside a superseded chip count

What the site showed

Intel Fab 52 Magdeburg carried status "announced" with the description "Construction paused September 2024 amid Intel restructuring" and the claim that a €9.9 billion EU Chips Act subsidy had been approved. Intel cancelled the project on 2025-07-24, twelve months before this correction. Separately, Nscale Loughton listed CloudHQ as its owner, 45,000 NVIDIA GB200 accelerators, 112,500 PFLOP/s and 113,694 H100 equivalents, and recorded no relationship to Microsoft. The Loughton figures fed the UK column on the country rollups, the map, and every capacity surface that sums GPUs by geography.

What is actually right

Magdeburg is cancelled: Intel’s second-quarter 2025 earnings release (2025-07-24) states that Intel "will no longer move forward with planned projects in Germany and Poland." On the subsidy, the German Federal Government’s own statement about the June 2023 declaration of intent gives federal funding of "up to 9.9 billion euros" and adds that it "is subject to approval by the European Commission under state aid legislation" — pledged subject to approval, which is not the same as approved; Scrutica makes no claim about whether approval or disbursement followed, because no primary source for either was located. Loughton is owned by Nscale, which announced its purchase of the site on 2025-01-13, and Microsoft is recorded as a tenant: Nscale’s 2025-09-16 release with Microsoft states the campus "will play a vital role in providing Microsoft’s Azure services across the UK." The displayed accelerator figures are now 23,040 GB300, 57,600 PFLOP/s and 58,211 H100 equivalents, from that same release. Both chip figures are Nscale’s own: 45,000 GB200 was a January 2025 site-capacity ceiling ("can house up to"), 23,040 GB300 is a committed deployment nine months later. The earlier statement is preserved in the substrate and in a source-conflict record rather than deleted — a reader tracking announced-versus-deployed capacity needs both.

Why it happened

Two unrelated causes. The Magdeburg record was written from a September 2024 pause announcement and never revisited; nothing in the pipeline watches a facility record for a primary source that contradicts it, so the row aged into a false statement while looking maintained. The Loughton owner was manufactured by a matcher: Epoch’s GPU-clusters CSV leaves the Owner column empty for that row and reads "Cloud" in the Users column — a user-category descriptor, not a company — and the ingest fell back to it and resolved it through a prefix matcher, where the five-character word "cloud" cleared both length guards against the canonical name "cloudhq." Five further facilities in Sweden, France, Canada, India and one implied Mistral cluster carried the same invented owner. The chip count was not a Scrutica error: the pipeline faithfully transcribed the operator’s January figure and had no mechanism to notice the operator superseding it.

What is different now

The ingest no longer runs Users-column descriptors through the prefix matcher, and when such a descriptor matches no known organization the ingest omits the owner field entirely rather than writing a guess — so a refresh can no longer overwrite a curated owner with a category word. The resolver moved into a pure module (scripts/lib/epoch-org-resolver.ts) with regression tests, because the only prior way to exercise it was a hand-written copy, and a hand-written copy is how a matcher defect survives its own test suite. The five other misattributed records had their owners withdrawn to empty rather than replaced: no source names an owner for any of them, and an empty field is honest where a guess is not. Each carries a flag saying so. The status vocabulary had no value for a cancelled project — "decommissioned" asserts a facility that was built and retired, and this one never broke ground — so the schema gained "cancelled" the same day and Magdeburg is the first record to carry it. Where an accelerator breakdown and a corrected total now disagree, the facility page says which figure supersedes which instead of printing both silently.

Affected pages

/facilities /map /country /compare /capacity /companies

2026-07-1560b37a6d, e0c12b74

SEC-rederived ownership and supply edges carried citations that named both endpoints without asserting the relationship; four edges asserted wrong endpoints outright

What the site showed

About 3,167 edges re-derived during the June 2026 licensing sweep cited SEC documents selected on a "both endpoint names appear in the document" test. For flat Exhibit-21 subsidiary lists that test is unsound: a flat list proves every named entity is a subsidiary of the filing registrant and asserts nothing about relationships between co-listed entities, so a sub-to-sub edge citing one displayed tier-1 SEC provenance the document does not contain. Three edges additionally presented AMETEK as the parent of FARO Technologies subsidiaries (Lanmark Controls, Advanced Technical Solutions Beijing, SiteScape) — AMETEK appears in the cited FARO filings only as a named competitor — and one supply edge presented Coulter Pharmaceutical as a counterparty where the cited 1998 10-K discusses the unrelated Coulter Corporation, a surname collision.

What is actually right

A full-population deterministic audit (CIK-to-filer identity resolution, then document-structure classification) followed by per-document adjudication of all 220 third-party-filer edges: 127 by a ten-agent wave returning verbatim supporting quotes, 93 by the structural classifier with an orchestrator review pass. Outcome: 4 edges deleted (wrong endpoints; the correct FARO-parented edges already existed), 8 repointed to documents that assert the specific relationship, and 168 provenance-limitation flags applied (citation_co_membership_only, citation_document_mismatch, citation_indirect_support, relationship_nature_mismatch). Flagged edges remain in the graph — the underlying relationships are vendor-asserted and frequently true; each flag states exactly what its citation does and does not evidence. Pre-write snapshots: planning/audits/citation-apply-snapshot-2026-07-15.json and citation-artifacts-apply-snapshot-2026-07-15.json.

Why it happened

The June 2026 re-derivation attached citations by name co-occurrence. A document that names both endpoints does not necessarily assert the relationship between them — flat Exhibit-21 lists never do — and name-presence also admitted non-parties mentioned in passing: a competitor named in a filing became a parent; a surname collision became a supply counterparty.

What is different now

scripts/lib/ex21_hierarchy.py now classifies every Exhibit-21 as flat, indented, tiered-column, or explicit-text (10 tests) and the Item-21 parser emits registrant-to-subsidiary edges only from flat lists; sibling edges require hierarchy-encoding structure for the specific pair. The classifier doubles as a reusable audit predicate over the full deployed set. The mangled vendor name "REGENETP INC" is corrected to PolarityTE, Inc. through the editorial corrections layer, which preserves the verbatim source value.

Affected pages

/companies /ownership /supply-chain

2026-07-14

Chain-of-title walker elected an inverted ownership edge and concluded a false "dispersed float" ultimate owner for Harman International

What the site showed

The company page for Harman International Industries presented "Harman Hungary Financing Kft" — a Harman financing subsidiary — as the documented parent (HIGH confidence, tier 1, citing Harman’s own 2016 SEC Exhibit 21), walked the chain through it, and concluded ultimate beneficial ownership was dispersed public float. Samsung Electronics America (SEC merger exhibit) and Samsung (GLEIF) appeared only as alternates. Samsung has wholly owned Harman since March 2017.

What is actually right

Samsung Electronics America, Inc. is Harman’s documented parent (SEC EDGAR merger exhibit, accession 0001193125-17-078968), with Samsung the GLEIF-recorded apex. The inverted edge was endpoint-swapped on 2026-07-14 (snapshot: planning/audits/harman-inversion-snapshot-2026-07-14.json) — an Exhibit 21 lists subsidiaries OF THE FILER (CIK 800459 = Harman), so it cannot evidence the filer’s parent.

Why it happened

Two stacked defects. (1) Data: the edge was ingested as a correct PARENT-typed row ("Harman International is Harman Hungary’s parent"); the June 2026 direction-normalization (PR #31, which corrected 3,557 reversed edges) re-typed this row to SUBSIDIARY without swapping its endpoints, inverting its meaning — a sibling failure mode of the class that fix targeted. (2) Design: when two DIFFERENT documents claimed different parents at equal confidence and authority tier with no as-of dating, the walker’s sort was stable-but-arbitrary (row order) and it walked the elected branch to a UBO conclusion, while page copy stated the walker "does not pick a winner." 3,502 entities in the substrate carry more than one distinct HIGH-confidence parent claim — the election defect was structural, not a one-row problem. Caught by an external reviewer, 2026-07-14.

What is different now

The walker now STOPS at contested hops: when distinct parents tie on confidence and authority tier and dated recency cannot arbitrate (a missing as-of date is unknown, not old), it renders the competing claims side-by-side with each claim’s onward chain and draws no ultimate-owner conclusion (terminal: contested / terminus_contested, both walkers). Same-document co-parent structures (founder vehicles, split holdings) are exempted by document-identity check and keep the existing walk-with-branches behavior. Regression tests cover the contested stop, the dated-recency arbitration path, and the co-parent exemption. The re-derived edge population is under a document-type-aware direction audit (an Exhibit-21 filer must be the parent side; merger documents are legitimately filed by the acquired side) — further corrections from that audit will be logged here.

Affected pages

/companies/org-harman-international-industries /ownership

2026-06-309fa2c450, 09b22bde

3,959 ownership edges were walked backwards for weeks — subsidiaries rendered as parents across chain-of-title, visibility tiers, and closure walks

What the site showed

The historical oce-ingest-b ingest wrote PARENT-typed edges with the opposite direction convention from the schema’s, so every consumer that walked them showed inverted ancestry — e.g., the ownership closure presented TSMC Arizona as the parent of TSMC. The inverted edges fed the chain-of-title panels, the compute-visibility tier classifier, and facility ownership walks from the ingest date until the fix.

What is actually right

3,557 edges were direction-corrected per-edge on 2026-06-30, and a follow-up root-caused the ingest convention and resolved the remaining 402 residue rows to zero, with integrity guards (migration 20260630000000_ownership_edge_integrity.sql) asserting the convention going forward.

Why it happened

The oce-ingest-b ingest encoded "A is parent of B" with source/target reversed relative to the schema convention the walkers assume. Downstream walkers were convention-correct, so the error surfaced as inverted chains rather than crashes. The class was caught during the June 2026 audit campaign; a residue instance survived even the fix (one row re-typed without an endpoint swap — see the Harman entry above, caught by external review on 2026-07-14).

What is different now

Integrity guards land with the migration (direction-convention assertions + tests that fail on reversed PARENT edges). The contested-parentage stop (Harman entry) adds a behavioral backstop: where residual bad edges tie with good ones, the walker now stops and shows the disagreement instead of electing a branch. TRANSPARENCY NOTE: this entry was added on 2026-07-14 — two weeks after the fix — after an external reviewer observed that the largest correction event in the site’s history had no public log entry. The omission itself violated this page’s stated purpose; logging it late, with that admission, is the honest remainder.

Affected pages

/companies /ownership /compute-leverage/visibility /facilities

2026-05-19wave14-fix-2026-05-19

Stargate UAE Phase 1 GPU count was inflated 2.86× from a power-derived Epoch estimate, contradicting the actual 35,000 GB300 export-control authorisation

What the site showed

The Facility Intelligence Card for Stargate UAE Phase 1 (fac-epoch-gc-stargate-uae-phase-1) rendered 100,000 NVIDIA GB300 GPUs / 250,000 PFLOPS dense FP16 / 8,338 H100 equivalents alongside the 200 MW Phase 1 power capacity. The same row inflated the UAE column on /compare, /sovereign-ai, /capacity/thresholds, and the country-level GPU rollups.

What is actually right

35,000 NVIDIA GB300 Blackwell Ultra GPUs / 87,500 PFLOPS dense FP16 / 88,428 H100 dense equivalents, anchored to the US Commerce Department's November 2025 export-control authorisation for G42. Two verified primary sources contain this figure verbatim alongside "200 MW" Phase 1 context: The National 2025-12-05 ("The Mubadala-backed company can purchase the equivalent of up to 35,000 of Nvidia's Blackwell chips (GB300s), the Commerce Department said.") and DCD 2026-01-23 ("the export licenses allowing each country's so-called 'AI champion' — which, for the UAE, is G42 — to purchase up to 35,000 worth of Nvidia GB300 systems or their equivalent."). G42's own October 2025 PR Newswire release independently confirms the "first 200MW of the 1GW mega-scale infrastructure" topology.

Why it happened

Epoch AI's GPU-clusters CSV emitted "Total number of AI chips" = 100,000 as a power-derived inference, flagged verbatim in the CSV's Note column ("Chip amount roughly estimated based on power capacity. Assuming for now that they're using GB300s, but this hasn't been officially confirmed"). Scrutica's ingest passed the value through with is_estimated = true but no second-source check. The physical sanity bound also fails: 200 MW / 100,000 GB300 = 2.00 kW/GPU is infeasible at GB300 TDP 1.4 kW with chip-share-of-power 0.49 (per-chip envelope = 0.98 kW < TDP); 200 MW / 35,000 GB300 = 5.71 kW/GPU is plausible. The Epoch CSV estimate predates the November 2025 Commerce Dept disclosure that anchored the actual chip ceiling. A separate, second inconsistency lived inside the same old row: 8,338 H100 equivalents alongside 250,000 PFLOPS is a ~30× internal mismatch. The 8,338 derived from Epoch’s SPARSE Max-OP/s ratio while every other H100e figure on the platform uses the dense FP16 convention (src/lib/methodology/versions.ts) — the F-30 methodological-inconsistency finding. The corrected 88,428 H100e follows the dense convention; the full derivation is in the numerical-corrections overlay entry for h100_equivalent_gpus on this page.

What is different now

A new editorial overlay table, editorial_facility_numerical_corrections, allows numeric column values to be corrected at display time without modifying the substrate (CLAUDE.md §16; mirrors the type/name correction tables already shipped). The Stargate UAE Phase 1 correction landed as 3 paired INSERTs (total_gpu_count, peak_pflops_fp16, h100_equivalent_gpus). The substrate verbatim row remains unchanged so Epoch CSV refreshes do not re-introduce the inflated value. The ingest pipeline (scripts/ingest-epoch-gpu-clusters.ts) gained a gate that flags any Epoch CSV row whose Note column matches the power-derived inference pattern with a chip_count_power_derived_estimate data-quality flag. CORRECTION TO THIS ENTRY (2026-07-14, caught by external review): that gate had never executed against production — the pipeline’s last wet run predated the gate, so the ~13 peer facilities (GB200/B200 rows matching the same power-derived pattern) carried no visible flag while this entry claimed they did. The 13 flags were minted directly on 2026-07-14 (flag ids f28-peer-power-derived-*) and now render on each facility page; their chip counts remain uncorrected by design — no primary-source-converging replacement exists, and substituting one would violate the no-fabrication principle. The gpu-clusters pipeline’s wet-run cadence is under review separately since Epoch deprecated the upstream dataset in favor of its AI Data Centers hub.

Affected pages

/facilities /map /sovereign-ai /capacity/thresholds /economics/cost-index /compare /country /concentration /cascade /validation

2026-05-04e5bfd89, 567fdb5

Country profile pages and geography queries returned zero results after an ISO2 schema migration

What the site showed

The query engine returned 0 rows for geography queries (e.g., "facilities in Japan"), and country profile pages rendered empty until the access layer was updated post-migration. Geographic queries returned no results despite facility records being present in the database.

What is actually right

The underlying data remained intact, and the routes now correctly render the full facility counts and query results per country. Verified via test queries against the primary database demonstrating that geographic linkage persisted while the access layer alone had failed.

Why it happened

A prior migration altered the canonical country column shape from full names ("United States") to ISO2 codes ("US"). Two accessors — the NL query engine tool-executor and the /api/query route — were not updated during the migration and continued filtering the new ISO2 columns using full country names, producing a silent zero-row intersection.

What is different now

Accessors now route through a canonical async countryToIso2() wrapper (src/lib/data/country-iso2.ts) that ensures query parameters and extracted tool entities match the database column shape. The accessors also apply strict sanitization, returning zero rows for unrecognized input rather than throwing 500 errors.

Affected pages

/query /country

2026-05-045d29515, cfbb39e

Chain stub counts rendered as 0 across 4,529 facility rows due to a stale precompute

What the site showed

The Allied Coordination Gap Analyzer (now /compute-leverage/coordination) rendered a chain stub count of 0 for 4,529 facilities until the precompute was re-run.

What is actually right

774 of the 4,529 rows correctly possess non-zero chain stub counts. The correct distribution was verified by a parity check matching live state against the intended precompute logic.

Why it happened

A database migration added the chain_stub_count column with DEFAULT 0 NOT NULL, but the precompute script was never re-run after the column was introduced. The live substrate carried the default zero values for all 4,529 rows, which the application faithfully rendered.

What is different now

The precompute script was re-run to populate the column correctly. A persistent parity check script was wired into the pipeline to assert chain_stub_count values match real-time hop counting, ensuring any future stale precompute state is caught immediately.

Affected pages

/compute-leverage/coordination

2026-04-013b9a2cc

H100-PCIe and H100-NVL memory specifications did not match NVIDIA's per-SKU datasheets

What the site showed

H100-PCIe was listed with the memory type of the H100-SXM baseline. H100-NVL's memory capacity was shown at the SXM value rather than its own 94 GB per GPU.

What is actually right

H100-PCIe uses HBM2e (the PCIe form factor predated HBM3 availability on that SKU). H100-NVL has 94 GB of HBM3 memory per GPU. Both figures are now sourced from NVIDIA's per-SKU product page rather than inherited from the SXM entry.

Why it happened

The GPU spec table was originally populated per chip family (H100) rather than per SKU variant (H100-SXM, H100-PCIe, H100-NVL). Form-factor variants diverge from the family baseline on memory type and capacity, and the single-entry approach hid those divergences.

What is different now

GPU spec entries in lib/data/chip-classifications.ts are per-SKU rather than per-family. Each entry requires a source URL pointing to NVIDIA's specific datasheet for that SKU. The change applies to all subsequent additions (R200, B300, TPU v6e, etc.).

Affected pages

/capacity /methodology /facilities

2026-03-30

Cascade scenarios, supply-chain graphs, and BIS matches were truncated by a silent PostgREST row cap

What the site showed

Cascade simulations appeared to terminate earlier than the full supply-chain topology warranted. The supply-chain graph displayed a subset of bilateral relationships. BIS cross-reference counts read lower than the underlying data actually contained. The UI showed no indication that results were truncated.

What is actually right

The supply-chain graph holds the full 21,037 bilateral edges (majority from a licensed supply-chain database, with additional coverage from a licensed corporate-ownership database and CSET) — the figure cited site-wide. Cascade propagation reaches deeper downstream tiers. BIS matching returns more candidate matches.

Why it happened

PostgREST — Supabase's auto-generated REST API — caps query results at 1,000 rows by default and returns no warning or metadata flag indicating truncation has occurred. Any Scrutica query against a table larger than 1K rows was silently returning a partial result set, which the code then treated as complete.

What is different now

A site-wide fetchAllRows() helper (exported from lib/supabase/server-user.ts) automatically paginates any query expected to return >1K rows. Data-access modules that query the supply-chain, BIS, and cascade tables now route through this helper, with the cap documented inline so it is caught on review rather than re-discovered.

Affected pages

/cascade /supply-chain /export-controls /map /chokepoint-anatomy

Numerical corrections (editorial overlay)

Per-facility numeric column corrections that overlay the substrate at display time. The substrate row is preserved verbatim per DQP §6 (temporal provenance); the corrected value is rendered everywhere the column is consumed. Each entry cites primary sources verified at curation time.

2026-07-27Tier 1 (primary)

Nscale Loughton · total_gpu_count

Substrate verbatim

45,000

Editorial overlay

23,040

Reason

Epoch AI’s GPU-clusters CSV transcribes Nscale’s January 2025 announcement, which stated the Loughton site “can house up to 45,000 of the latest Nvidia GB200 GPUs” — a site-capacity ceiling. Nscale’s 2025-09-16 release with Microsoft states the campus “will initially house 23,040 NVIDIA GB300 GPUs delivered in Q1 of 2027”, a committed deployment of a different chip. Both statements are the operator’s own, nine months apart; the later and more specific one governs. Power is unchanged — both releases give 50 MW scalable to 90 MW, and the substrate holds the 90 MW ceiling. Substrate value preserved verbatim per DQP §3 + §16; contradiction recorded in source_conflicts per DQP §7.

Primary sources

2026-07-27Tier 1 (primary)

Nscale Loughton · peak_pflops_fp16

Substrate verbatim

112,500

Editorial overlay

57,600

Reason

Paired re-derivation from the corrected chip count: 23,040 GB300 × 2,500 TFLOP/s dense FP16 = 57,600 PFLOP/s. The per-GPU figure is the canonical registry’s (src/lib/methodology/versions.ts GPU_MODELS.gb300, dense, no 2:4 sparsity), and it reproduces the substrate value exactly on the verbatim side (45,000 × 2,500 TFLOP/s = 112,500 PFLOP/s), so the convention is held fixed across the correction. Deterministic re-derivation, not a new estimate. Epoch AI’s GPU-clusters CSV transcribes Nscale’s January 2025 announcement, which stated the Loughton site “can house up to 45,000 of the latest Nvidia GB200 GPUs” — a site-capacity ceiling. Nscale’s 2025-09-16 release with Microsoft states the campus “will initially house 23,040 NVIDIA GB300 GPUs delivered in Q1 of 2027”, a committed deployment of a different chip. Both statements are the operator’s own, nine months apart; the later and more specific one governs. Power is unchanged — both releases give 50 MW scalable to 90 MW, and the substrate holds the 90 MW ceiling. Substrate value preserved verbatim per DQP §3 + §16; contradiction recorded in source_conflicts per DQP §7.

Primary sources

2026-07-27Tier 1 (primary)

Nscale Loughton · h100_equivalent_gpus

Substrate verbatim

113,694

Editorial overlay

58,211

Reason

Paired re-derivation: 23,040 × (2,500 / 989.5) = 58,211.2 H100-SXM5 dense-FP16 equivalents, rounded as the substrate rounds. The 989.5 TFLOP/s denominator is the platform’s dense H100 convention (src/lib/methodology/versions.ts GPU_MODELS.h100_sxm5) and is the same denominator Epoch used for the verbatim value — 45,000 × 2,500 / 989.5 = 113,693.78, stored rounded to 113,694 — so no convention change is smuggled in with the count change. Epoch AI’s GPU-clusters CSV transcribes Nscale’s January 2025 announcement, which stated the Loughton site “can house up to 45,000 of the latest Nvidia GB200 GPUs” — a site-capacity ceiling. Nscale’s 2025-09-16 release with Microsoft states the campus “will initially house 23,040 NVIDIA GB300 GPUs delivered in Q1 of 2027”, a committed deployment of a different chip. Both statements are the operator’s own, nine months apart; the later and more specific one governs. Power is unchanged — both releases give 50 MW scalable to 90 MW, and the substrate holds the 90 MW ceiling. Substrate value preserved verbatim per DQP §3 + §16; contradiction recorded in source_conflicts per DQP §7.

Primary sources

2026-07-20Tier 3 (press)

HUMAIN Riyadh Phase 1 · power_capacity_mw

Substrate verbatim

100

Editorial overlay

Withheld (DQP §5)

Reason

HUMAIN Riyadh Phase 1 (100 MW) is a press-tier per-campus split of the SAME NVIDIA HUMAIN AI-factories program whose primary source discloses ONLY the 500 MW 5-year aggregate, already carried by fac-epoch-gc-humain-saudi-arabia-phase-2. Firm 200-MW double-count (Riyadh+Dammam) against phase-2's 500 (humain_dup_map, wrds-mw-semantics-2026-07-18.md §4). Withheld (corrected_value=null) so the tier-1 aggregate remains the single authoritative HUMAIN capacity. Row's own notes state the primary announcement discloses only the 500 MW aggregate; the 100 MW campus figures are press-tier (Al Arabiya/DCD). | verdict_class=DUP_CROSS_SUBSTRATE; delta_mw=-100. Evidence (same source document): NVIDIA release 2025-05-13 (tier 1): 'up to 500 megawatts ... over the next five years' for the HUMAIN AI-factories program; the per-campus 100 MW figures are press-tier and not an incremental primary disclosure. [humain-rollup-2026-07-19]

Primary sources

https://nvidianews.nvidia.com/news/humain-and-nvidia-announce-partnership-to-build-ai-factories-of-the-future-in-saudi-arabia

2026-07-20Tier 3 (press)

HUMAIN Dammam Phase 1 · power_capacity_mw

Substrate verbatim

100

Editorial overlay

Withheld (DQP §5)

Reason

HUMAIN Dammam Phase 1 (100 MW) is the second press-tier per-campus split of the SAME 500 MW NVIDIA HUMAIN program aggregate carried by fac-epoch-gc-humain-saudi-arabia-phase-2 (same referent as the Riyadh row). Withheld (corrected_value=null) to remove the 200-MW double-count while keeping the tier-1 500 MW aggregate as the single authoritative HUMAIN capacity. | verdict_class=DUP_CROSS_SUBSTRATE; delta_mw=-100. Evidence (same source document): NVIDIA release 2025-05-13 (tier 1): 'up to 500 megawatts ... over the next five years'; the Riyadh and Dammam 100 MW splits together equal a 200 MW subset already inside the 500 MW aggregate. [humain-rollup-2026-07-19]

Primary sources

https://nvidianews.nvidia.com/news/humain-and-nvidia-announce-partnership-to-build-ai-factories-of-the-future-in-saudi-arabia

2026-07-17Tier 2 (research)

Anthropic-Amazon New Carlisle · power_capacity_mw

Substrate verbatim

910

Editorial overlay

Withheld (DQP §5)

Reason

Campus dedup (Task #28 / F6-2): this row is the Epoch current-deployment estimate for the New Carlisle / Project Rainier campus, whose canonical power figure is elected on fac-aws-new-carlisle (2,200 MW full build-out, press+operator sourced). Withheld from aggregates to prevent triple-counting the same campus power envelope; the 910 MW figure is retained and surfaced, cited, in the campus source-disagreement panel (facility_campus_memberships). NOT a claim the 910 MW is wrong — it is a different scope (current deployment vs full build-out) of the same campus.

Primary sources

2026-07-17Tier 2 (research)

Project Rainier · power_capacity_mw

Substrate verbatim

350

Editorial overlay

Withheld (DQP §5)

Reason

Campus dedup (Task #28 / F6-2): this row is the SemiAnalysis-sourced named "Project Rainier" cluster (350 MW / 400k GPU) within the New Carlisle campus, whose canonical power figure is elected on fac-aws-new-carlisle (2,200 MW). power_capacity_mw withheld to prevent triple-counting; GPU/PFLOP columns NOT corrected (this row is the sole source of the campus compute figures and keeps counting). The 350 MW figure is retained and surfaced, cited, in the campus panel.

Primary sources

2026-07-10Tier 2 (research)

NVIDIA SATURN V Phase 1 · power_capacity_mw

Substrate verbatim

0.591

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (NVIDIA SATURN V Phase 1) is superseded by NVIDIA SATURN V Phase 3 (fac-epoch-gc-nvidia-saturn-v-phase-3), whose Power Capacity (MW)=3.704064 is a CUMULATIVE running total already inclusive of this row's 0.5912816 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

JAMSTEC ZettaScaler-2.0 Gyoukou · power_capacity_mw

Substrate verbatim

0.316

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (JAMSTEC ZettaScaler-2.0 Gyoukou) is superseded by JAMSTEC ZettaScaler-2.2 Gyoukou (fac-epoch-gc-jamstec-zettascaler-2-2-gyoukou), whose Power Capacity (MW)=2.5996 is a CUMULATIVE running total already inclusive of this row's 0.3158 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

CSCS Piz Daint Phase 1 · power_capacity_mw

Substrate verbatim

1.812

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (CSCS Piz Daint Phase 1) is superseded by CSCS Piz Daint Phase 2 (fac-epoch-gc-cscs-piz-daint-phase-2), whose Power Capacity (MW)=4.3744 is a CUMULATIVE running total already inclusive of this row's 1.811992 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Eni HPC4 Phase 1 · power_capacity_mw

Substrate verbatim

2.482

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Eni HPC4 Phase 1) is superseded by Eni HPC4 Phase 2 (fac-epoch-gc-eni-hpc4-phase-2), whose Power Capacity (MW)=1.98653 is a CUMULATIVE running total already inclusive of this row's 2.4824 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Petrobras Fênix Phase 1 · power_capacity_mw

Substrate verbatim

0.523

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Petrobras Fênix Phase 1) is superseded by Petrobras Fênix Phase 2 (fac-epoch-gc-petrobras-f-nix-phase-2), whose Power Capacity (MW)=0.6981 is a CUMULATIVE running total already inclusive of this row's 0.5226 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Jean Zay Supercomputer Phase 1 · power_capacity_mw

Substrate verbatim

0.695

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Jean Zay Supercomputer Phase 1) is superseded by Jean Zay Supercomputer Phase 4 (fac-epoch-gc-jean-zay-supercomputer-phase-4), whose Power Capacity (MW)=2.41672704 is a CUMULATIVE running total already inclusive of this row's 0.69542928 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

NSCC ASPIRE 2A Phase 1 · power_capacity_mw

Substrate verbatim

0.218

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (NSCC ASPIRE 2A Phase 1) is superseded by NSCC ASPIRE 2A Phase 2 (fac-epoch-gc-nscc-aspire-2a-phase-2), whose Power Capacity (MW)=0.2998195199999999 is a CUMULATIVE running total already inclusive of this row's 0.21805056 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

SURF Snellius Phase 1 · power_capacity_mw

Substrate verbatim

0.247

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (SURF Snellius Phase 1) is superseded by SURF Snellius Phase 3 (fac-epoch-gc-surf-snellius-phase-3), whose Power Capacity (MW)=0.7370854400000001 is a CUMULATIVE running total already inclusive of this row's 0.24740352 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Princeton Della Phase 1 · power_capacity_mw

Substrate verbatim

0.262

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Princeton Della Phase 1) is superseded by Princeton Della Phase 2 (fac-epoch-gc-princeton-della-phase-2), whose Power Capacity (MW)=0.6800102400000001 is a CUMULATIVE running total already inclusive of this row's 0.26225472 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

PARAM Siddhi-AI Phase 1 · power_capacity_mw

Substrate verbatim

0.294

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (PARAM Siddhi-AI Phase 1) is superseded by AIRAWAT-PSAI Phase 2 (fac-epoch-gc-airawat-psai-phase-2), whose Power Capacity (MW)=0.5444275199999999 is a CUMULATIVE running total already inclusive of this row's 0.2935296 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Jean Zay Supercomputer Phase 3 · power_capacity_mw

Substrate verbatim

2.077

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Jean Zay Supercomputer Phase 3) is superseded by Jean Zay Supercomputer Phase 4 (fac-epoch-gc-jean-zay-supercomputer-phase-4), whose Power Capacity (MW)=2.41672704 is a CUMULATIVE running total already inclusive of this row's 2.07659088 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

NSC Berzelius Phase 1 · power_capacity_mw

Substrate verbatim

0.412

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (NSC Berzelius Phase 1) is superseded by NSC Berzelius Phase 2 (fac-epoch-gc-nsc-berzelius-phase-2), whose Power Capacity (MW)=0.62409984 is a CUMULATIVE running total already inclusive of this row's 0.4123391999999999 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Jean Zay Supercomputer Phase 2 · power_capacity_mw

Substrate verbatim

1.766

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Jean Zay Supercomputer Phase 2) is superseded by Jean Zay Supercomputer Phase 4 (fac-epoch-gc-jean-zay-supercomputer-phase-4), whose Power Capacity (MW)=2.41672704 is a CUMULATIVE running total already inclusive of this row's 1.7664192 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Lawrence Livermore NL Lassen Phase 1 · power_capacity_mw

Substrate verbatim

1.882

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Lawrence Livermore NL Lassen Phase 1) is superseded by Lawrence Livermore NL Lassen Phase 2 (fac-epoch-gc-lawrence-livermore-nl-lassen-phase-2), whose Power Capacity (MW)=2.0756736 is a CUMULATIVE running total already inclusive of this row's 1.88225856 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

AIST ABCI 1.0 · power_capacity_mw

Substrate verbatim

2.994

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (AIST ABCI 1.0) is superseded by AIST ABCI 3.0 (fac-epoch-gc-aist-abci-3-0), whose Power Capacity (MW)=8.587779200000002 is a CUMULATIVE running total already inclusive of this row's 2.99400192 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

AIST ABCI 2.0 · power_capacity_mw

Substrate verbatim

2.819

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (AIST ABCI 2.0) is superseded by AIST ABCI 3.0 (fac-epoch-gc-aist-abci-3-0), whose Power Capacity (MW)=8.587779200000002 is a CUMULATIVE running total already inclusive of this row's 2.8185 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Gcore data center Phase 1 · power_capacity_mw

Substrate verbatim

0.457

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Gcore data center Phase 1) is superseded by Gcore data center Phase 2 (fac-epoch-gc-gcore-data-center-phase-2), whose Power Capacity (MW)=1.9019 is a CUMULATIVE running total already inclusive of this row's 0.4566016 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

NVIDIA SATURN V Phase 2 · power_capacity_mw

Substrate verbatim

3.632

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (NVIDIA SATURN V Phase 2) is superseded by NVIDIA SATURN V Phase 3 (fac-epoch-gc-nvidia-saturn-v-phase-3), whose Power Capacity (MW)=3.704064 is a CUMULATIVE running total already inclusive of this row's 3.6324288 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Lawrence Livermore NL El Capitan Phase 1 · power_capacity_mw

Substrate verbatim

0.793

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Lawrence Livermore NL El Capitan Phase 1) is superseded by Lawrence Livermore NL El Capitan Phase 2 (fac-epoch-gc-lawrence-livermore-nl-el-capitan-phase-2), whose Power Capacity (MW)=35 is a CUMULATIVE running total already inclusive of this row's 0.793182208 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

NVIDIA Selene Phase 1 · power_capacity_mw

Substrate verbatim

2.408

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (NVIDIA Selene Phase 1) is superseded by NVIDIA Selene Phase 2 (fac-epoch-gc-nvidia-selene-phase-2), whose Power Capacity (MW)=4.7391 is a CUMULATIVE running total already inclusive of this row's 2.4075 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Voltage Park Texas Phase 1 · power_capacity_mw

Substrate verbatim

1.427

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Voltage Park Texas Phase 1) is superseded by Voltage Park Texas Phase 2 (fac-epoch-gc-voltage-park-texas-phase-2), whose Power Capacity (MW)=5.7289232000000005 is a CUMULATIVE running total already inclusive of this row's 1.42688 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

NVIDIA Israel-1 Phase 1 · power_capacity_mw

Substrate verbatim

1.487

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (NVIDIA Israel-1 Phase 1) is superseded by NVIDIA Israel-1 Phase 2 (fac-epoch-gc-nvidia-israel-1-phase-2), whose Power Capacity (MW)=2.9222502400000003 is a CUMULATIVE running total already inclusive of this row's 1.48721664 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Sustainable Metal Cloud Singapore Phase 1 · power_capacity_mw

Substrate verbatim

1.461

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Sustainable Metal Cloud Singapore Phase 1) is superseded by Sustainable Metal Cloud Singapore Phase 2 (fac-epoch-gc-sustainable-metal-cloud-singapore-phase-2), whose Power Capacity (MW)=7.007000000000001 is a CUMULATIVE running total already inclusive of this row's 1.4611251200000002 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

NVIDIA Eos Phase 1 · power_capacity_mw

Substrate verbatim

1.487

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (NVIDIA Eos Phase 1) is superseded by NVIDIA Eos Phase 2 (fac-epoch-gc-nvidia-eos-phase-2), whose Power Capacity (MW)=6.69247488 is a CUMULATIVE running total already inclusive of this row's 1.48721664 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Horizon Compute Baobab Phase 1 · power_capacity_mw

Substrate verbatim

1.461

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Horizon Compute Baobab Phase 1) is superseded by Horizon Compute Baobab Phase 2 (fac-epoch-gc-horizon-compute-baobab-phase-2), whose Power Capacity (MW)=2.8700672000000003 is a CUMULATIVE running total already inclusive of this row's 1.4611251200000002 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

TensorWave MI300X Cluster 1 Phase 1 · power_capacity_mw

Substrate verbatim

1.529

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (TensorWave MI300X Cluster 1 Phase 1) is superseded by TensorWave MI300X Cluster 1 Phase 2 (fac-epoch-gc-tensorwave-mi300x-cluster-1-phase-2), whose Power Capacity (MW)=15.015000000000002 is a CUMULATIVE running total already inclusive of this row's 1.5288000000000002 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

HUMAIN Saudi Arabia/NVIDIA Phase 1 · power_capacity_mw

Substrate verbatim

50.45

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (HUMAIN Saudi Arabia/NVIDIA Phase 1) is superseded by HUMAIN Saudi Arabia Phase 2 (fac-epoch-gc-humain-saudi-arabia-phase-2), whose Power Capacity (MW)=500 is a CUMULATIVE running total already inclusive of this row's 50.45040000000001 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Tesla A100 Cluster Phase 1 · power_capacity_mw

Substrate verbatim

4.948

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Tesla A100 Cluster Phase 1) is superseded by Tesla A100 Cluster Phase 2 (fac-epoch-gc-tesla-a100-cluster-phase-2), whose Power Capacity (MW)=6.2689536 is a CUMULATIVE running total already inclusive of this row's 4.9480704 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Meta Research SuperCluster (RSC-1) Phase 1 · power_capacity_mw

Substrate verbatim

5.179

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Meta Research SuperCluster (RSC-1) Phase 1) is superseded by Meta Research SuperCluster (RSC-1) Phase 2 (fac-epoch-gc-meta-research-supercluster-rsc-1-phase-2), whose Power Capacity (MW)=13.278719999999998 is a CUMULATIVE running total already inclusive of this row's 5.1787008 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Sakura's H100s Phase 1 · power_capacity_mw

Substrate verbatim

2.877

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Sakura's H100s Phase 1) is superseded by Sakura's B200s Phase 2 (fac-epoch-gc-sakura-s-b200s-phase-2), whose Power Capacity (MW)=16.416400000000003 is a CUMULATIVE running total already inclusive of this row's 2.87659008 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Sesterce Synapse Phase 1 · power_capacity_mw

Substrate verbatim

2.922

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Sesterce Synapse Phase 1) is superseded by Sesterce Synapse Phase 2 (fac-epoch-gc-sesterce-synapse-phase-2), whose Power Capacity (MW)=5.7401344000000005 is a CUMULATIVE running total already inclusive of this row's 2.9222502400000003 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Sesterce H100s Phase 1 · power_capacity_mw

Substrate verbatim

2.922

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Sesterce H100s Phase 1) is superseded by Sesterce H100s Phase 2 (fac-epoch-gc-sesterce-h100s-phase-2), whose Power Capacity (MW)=5.844500480000001 is a CUMULATIVE running total already inclusive of this row's 2.9222502400000003 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

NFDG Andromeda Phase 1 · power_capacity_mw

Substrate verbatim

3.648

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (NFDG Andromeda Phase 1) is superseded by NFDG Andromeda Phase 2 (fac-epoch-gc-nfdg-andromeda-phase-2), whose Power Capacity (MW)=5.808624640000001 is a CUMULATIVE running total already inclusive of this row's 3.64832832 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Nebius Kansas City Phase 1 · power_capacity_mw

Substrate verbatim

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Nebius Kansas City Phase 1) is superseded by Nebius Kansas City Phase 2 (fac-epoch-gc-nebius-kansas-city-phase-2), whose Power Capacity (MW)=40 is a CUMULATIVE running total already inclusive of this row's 5 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Telangana Yotta H1 Hyderbad AI City Cluster Phase 1 · power_capacity_mw

Substrate verbatim

5.606

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Telangana Yotta H1 Hyderbad AI City Cluster Phase 1) is superseded by Telangana Yotta Hyderbad AI City Cluster Phase 2 (fac-epoch-gc-telangana-yotta-hyderbad-ai-city-cluster-phase-2), whose Power Capacity (MW)=50 is a CUMULATIVE running total already inclusive of this row's 5.6056 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Yotta Shakti Cloud NM1 Phase 1 · power_capacity_mw

Substrate verbatim

5.845

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Yotta Shakti Cloud NM1 Phase 1) is superseded by Yotta Shakti Cloud NM1 Phase 2 (fac-epoch-gc-yotta-shakti-cloud-nm1-phase-2), whose Power Capacity (MW)=22.9605376 is a CUMULATIVE running total already inclusive of this row's 5.844500480000001 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Alps Supercomputer Phase 1 · power_capacity_mw

Substrate verbatim

8.683

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Alps Supercomputer Phase 1) is superseded by Alps Supercomputer Phase 2 (fac-epoch-gc-alps-supercomputer-phase-2), whose Power Capacity (MW)=11.9088 is a CUMULATIVE running total already inclusive of this row's 8.6825 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Nebius 8k Finland Phase 1 · power_capacity_mw

Substrate verbatim

11.415

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Nebius 8k Finland Phase 1) is superseded by Nebius Finland Phase 2 (fac-epoch-gc-nebius-finland-phase-2), whose Power Capacity (MW)=84.084 is a CUMULATIVE running total already inclusive of this row's 11.41504 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Stargate UAE Phase 1 · power_capacity_mw

Substrate verbatim

200

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Stargate UAE Phase 1) is superseded by Abu Dhabi UAE/USA 5GW Campus Phase 2 (fac-epoch-gc-abu-dhabi-uae-usa-5gw-campus-phase-2), whose Power Capacity (MW)=5000 is a CUMULATIVE running total already inclusive of this row's 200 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

NVIDIA CoreWeave Eos-DFW Phase 1 · power_capacity_mw

Substrate verbatim

15.616

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (NVIDIA CoreWeave Eos-DFW Phase 1) is superseded by NVIDIA CoreWeave Eos-DFW Rumored Phase 2 (fac-epoch-gc-nvidia-coreweave-eos-dfw-rumored-phase-2), whose Power Capacity (MW)=44.8448 is a CUMULATIVE running total already inclusive of this row's 15.61577472 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Tesla Cortex Phase 2 · power_capacity_mw

Substrate verbatim

70.07

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Tesla Cortex Phase 2) is superseded by Tesla Cortex Phase 3 (fac-epoch-gc-tesla-cortex-phase-3), whose Power Capacity (MW)=140.14 is a CUMULATIVE running total already inclusive of this row's 70.07 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Tesla Cortex Phase 1 · power_capacity_mw

Substrate verbatim

71.344

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Tesla Cortex Phase 1) is superseded by Tesla Cortex Phase 3 (fac-epoch-gc-tesla-cortex-phase-3), whose Power Capacity (MW)=140.14 is a CUMULATIVE running total already inclusive of this row's 71.344 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

xAI Colossus Memphis Phase 1 · power_capacity_mw

Substrate verbatim

150

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (xAI Colossus Memphis Phase 1) is superseded by xAI Colossus Memphis Phase 3 (fac-epoch-gc-xai-colossus-memphis-phase-3), whose Power Capacity (MW)=352.352 is a CUMULATIVE running total already inclusive of this row's 150 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Applied Digital CoreWeave Ellendale Phase 1 · power_capacity_mw

Substrate verbatim

100

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Applied Digital CoreWeave Ellendale Phase 1) is superseded by Applied Digital Ellendale Possible Phase 3 (fac-epoch-gc-applied-digital-ellendale-possible-phase-3), whose Power Capacity (MW)=400 is a CUMULATIVE running total already inclusive of this row's 100 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

xAI Colossus Memphis Phase 2 · power_capacity_mw

Substrate verbatim

280.28

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (xAI Colossus Memphis Phase 2) is superseded by xAI Colossus Memphis Phase 3 (fac-epoch-gc-xai-colossus-memphis-phase-3), whose Power Capacity (MW)=352.352 is a CUMULATIVE running total already inclusive of this row's 280.28 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

OpenAI Stargate Abilene Oracle OCI Supercluster Phase 1 · power_capacity_mw

Substrate verbatim

240.24

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (OpenAI Stargate Abilene Oracle OCI Supercluster Phase 1) is superseded by OpenAI Stargate Abilene Oracle OCI Supercluster Phase 3 (fac-epoch-gc-openai-stargate-abilene-oracle-oci-supercluster-phase-3), whose Power Capacity (MW)=2200 is a CUMULATIVE running total already inclusive of this row's 240.24000000000004 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

OpenAI Stargate Abilene Oracle OCI Supercluster Phase 2 · power_capacity_mw

Substrate verbatim

800.803

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (OpenAI Stargate Abilene Oracle OCI Supercluster Phase 2) is superseded by OpenAI Stargate Abilene Oracle OCI Supercluster Phase 3 (fac-epoch-gc-openai-stargate-abilene-oracle-oci-supercluster-phase-3), whose Power Capacity (MW)=2200 is a CUMULATIVE running total already inclusive of this row's 800.8028028000001 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

xAI Colossus 2 Memphis Phase 1 · power_capacity_mw

Substrate verbatim

264.264

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (xAI Colossus 2 Memphis Phase 1) is superseded by xAI Colossus 2 Memphis Phase 2 (fac-epoch-gc-xai-colossus-2-memphis-phase-2), whose Power Capacity (MW)=1409.408 is a CUMULATIVE running total already inclusive of this row's 264.264 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Applied Digital CoreWeave Ellendale Phase 2 · power_capacity_mw

Substrate verbatim

250

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Applied Digital CoreWeave Ellendale Phase 2) is superseded by Applied Digital Ellendale Possible Phase 3 (fac-epoch-gc-applied-digital-ellendale-possible-phase-3), whose Power Capacity (MW)=400 is a CUMULATIVE running total already inclusive of this row's 250 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

DataVolt Neom 1.5 GW Phase 1 · power_capacity_mw

Substrate verbatim

300

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (DataVolt Neom 1.5 GW Phase 1) is superseded by DataVolt Neom 1.5 GW Phase 2 (fac-epoch-gc-datavolt-neom-1-5-gw-phase-2), whose Power Capacity (MW)=1500 is a CUMULATIVE running total already inclusive of this row's 300 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-07-10Tier 2 (research)

Stargate UAE Phase 2 · power_capacity_mw

Substrate verbatim

1,000

Editorial overlay

Withheld (DQP §5)

Reason

Epoch gpu_clusters.csv 'Builds Upon' chain: this row (Stargate UAE Phase 2) is superseded by Abu Dhabi UAE/USA 5GW Campus Phase 2 (fac-epoch-gc-abu-dhabi-uae-usa-5gw-campus-phase-2), whose Power Capacity (MW)=5000 is a CUMULATIVE running total already inclusive of this row's 1000 MW (verified against CSV Quote text describing sequential building phases at the same site, e.g. Applied Digital Ellendale 100+150+150=400 MW). Withheld to prevent double-counting in operator/country MW rollups (audit finding: Epoch Builds Upon supersession double-count, 2026-07-10). Substrate facilities.power_capacity_mw is unchanged; only display/aggregation is corrected.

Primary sources

2026-05-19Tier 3 (press)

Stargate UAE Phase 1 · total_gpu_count

Substrate verbatim

100,000

Editorial overlay

35,000

Reason

Epoch GPU-clusters CSV "Total number of AI chips" column emits 100,000 as power-derived inference. CSV Note column verbatim: "Chip amount roughly estimated based on power capacity. Assuming for now that they're using GB300s, but this hasn't been officially confirmed." Two verified primary sources cite the US Commerce Department's November 2025 export-control authorisation of 35,000 GB300-equivalent chips for G42: The National 2025-12-05 verbatim "the US authorised the export of Nvidia's powerful AI chips to G42. The Mubadala-backed company can purchase the equivalent of up to 35,000 of Nvidia's Blackwell chips (GB300s), the Commerce Department said"; DCD 2026-01-23 verbatim "the export licenses allowing each country's so-called 'AI champion' — which, for the UAE, is G42 — to purchase up to 35,000 worth of Nvidia GB300 systems or their equivalent." Physical sanity bound confirms: 200 MW / 100,000 GB300 = 2.00 kW/GPU is infeasible at GB300 TDP 1.4 kW with chip-share-of-power 0.49 (per-chip envelope = 0.98 kW < TDP); 200 MW / 35,000 GB300 = 5.71 kW/GPU is plausible (per-chip envelope = 2.80 kW > TDP, leaves headroom for Grace CPU, NVLink, facility overhead). Per CLAUDE.md DQP §1 (no fabricated data) + §6 (temporal provenance), corrected to operator-direct convergence anchored at the export-control ceiling.

Primary sources

2026-05-19Tier 3 (press)

Stargate UAE Phase 1 · peak_pflops_fp16

Substrate verbatim

250,000

Editorial overlay

87,500

Reason

Paired correction to total_gpu_count: Epoch CSV peak_pflops_fp16 = 250,000 PFLOPS = 100,000 GB300 × 2.5 PFLOPS dense FP16/BF16 per GPU (per src/lib/methodology/versions.ts GPU_MODELS gb200/dense FP16 = 2,500 TFLOPS — GB300 Blackwell Ultra carries the same dense Tensor throughput per NVIDIA spec). Corrected 35,000 × 2.5 = 87,500 PFLOPS dense FP16. Per CLAUDE.md §16 + DQP §1 + §6, this is a deterministic re-derivation from the corrected chip count + published dense methodology, not a new estimate. Source convergence identical to the total_gpu_count correction.

Primary sources

2026-05-19Tier 3 (press)

Stargate UAE Phase 1 · h100_equivalent_gpus

Substrate verbatim

8,338

Editorial overlay

88,428

Reason

Paired correction to total_gpu_count: Epoch CSV "H100 equivalents" column = 8,337.5 derives from Epoch's Max-OP/s sparse ratio (1.65e+19 ÷ 1.979e+15 H100 sparse FP16) — a sparse-based methodology that diverges from the dense convention published in src/lib/methodology/versions.ts:81-95 and applied across other read sites (per F-30 methodological-inconsistency finding in end-number-spot-check-wave12-2026-05-19.md). Corrected value applies the dense convention consistently: 35,000 × 2,500 / 989.5 = 88,428 H100 dense equivalents (GB300 dense FP16 = 2,500 TFLOPS, H100 dense FP16 = 989.5 TFLOPS). This re-anchors the rendered value to the same dense convention that the FLOP Capacity Engine and /cost-index quote, eliminating cross-site methodology mismatch. Source convergence identical to the total_gpu_count correction.

Primary sources

Licensed-database records

89 capacity corrections apply to facility records derived from a licensed commercial database, held under academic subscription and not redistributed (net -16,923.4 MW across the affected records) · applied 2026-07-18. Per-row detail — the record identifiers and the source excerpts behind each correction — is withheld under the database license, consistent with how those facility records are treated elsewhere on the site. The corrected values are reflected in every capacity total the site reports.

Numerical corrections (editorial overlay)

2026-07-27Tier 1 (primary)

Nscale Loughton · total_gpu_count

Substrate verbatim

45,000

Editorial overlay

23,040

Reason

Primary sources

2026-07-27Tier 1 (primary)

Nscale Loughton · peak_pflops_fp16

Substrate verbatim

112,500

Editorial overlay

57,600

Reason

Primary sources

2026-07-27Tier 1 (primary)

Nscale Loughton · h100_equivalent_gpus

Substrate verbatim

113,694

Editorial overlay

58,211

Reason

Primary sources

2026-07-20Tier 3 (press)

HUMAIN Riyadh Phase 1 · power_capacity_mw

Substrate verbatim

100

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

https://nvidianews.nvidia.com/news/humain-and-nvidia-announce-partnership-to-build-ai-factories-of-the-future-in-saudi-arabia

2026-07-20Tier 3 (press)

HUMAIN Dammam Phase 1 · power_capacity_mw

Substrate verbatim

100

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

https://nvidianews.nvidia.com/news/humain-and-nvidia-announce-partnership-to-build-ai-factories-of-the-future-in-saudi-arabia

2026-07-17Tier 2 (research)

Anthropic-Amazon New Carlisle · power_capacity_mw

Substrate verbatim

910

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-17Tier 2 (research)

Project Rainier · power_capacity_mw

Substrate verbatim

350

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

NVIDIA SATURN V Phase 1 · power_capacity_mw

Substrate verbatim

0.591

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

JAMSTEC ZettaScaler-2.0 Gyoukou · power_capacity_mw

Substrate verbatim

0.316

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

CSCS Piz Daint Phase 1 · power_capacity_mw

Substrate verbatim

1.812

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Eni HPC4 Phase 1 · power_capacity_mw

Substrate verbatim

2.482

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Petrobras Fênix Phase 1 · power_capacity_mw

Substrate verbatim

0.523

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Jean Zay Supercomputer Phase 1 · power_capacity_mw

Substrate verbatim

0.695

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

NSCC ASPIRE 2A Phase 1 · power_capacity_mw

Substrate verbatim

0.218

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

SURF Snellius Phase 1 · power_capacity_mw

Substrate verbatim

0.247

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Princeton Della Phase 1 · power_capacity_mw

Substrate verbatim

0.262

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

PARAM Siddhi-AI Phase 1 · power_capacity_mw

Substrate verbatim

0.294

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Jean Zay Supercomputer Phase 3 · power_capacity_mw

Substrate verbatim

2.077

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

NSC Berzelius Phase 1 · power_capacity_mw

Substrate verbatim

0.412

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Jean Zay Supercomputer Phase 2 · power_capacity_mw

Substrate verbatim

1.766

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Lawrence Livermore NL Lassen Phase 1 · power_capacity_mw

Substrate verbatim

1.882

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

AIST ABCI 1.0 · power_capacity_mw

Substrate verbatim

2.994

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

AIST ABCI 2.0 · power_capacity_mw

Substrate verbatim

2.819

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Gcore data center Phase 1 · power_capacity_mw

Substrate verbatim

0.457

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

NVIDIA SATURN V Phase 2 · power_capacity_mw

Substrate verbatim

3.632

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Lawrence Livermore NL El Capitan Phase 1 · power_capacity_mw

Substrate verbatim

0.793

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

NVIDIA Selene Phase 1 · power_capacity_mw

Substrate verbatim

2.408

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Voltage Park Texas Phase 1 · power_capacity_mw

Substrate verbatim

1.427

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

NVIDIA Israel-1 Phase 1 · power_capacity_mw

Substrate verbatim

1.487

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Sustainable Metal Cloud Singapore Phase 1 · power_capacity_mw

Substrate verbatim

1.461

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

NVIDIA Eos Phase 1 · power_capacity_mw

Substrate verbatim

1.487

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Horizon Compute Baobab Phase 1 · power_capacity_mw

Substrate verbatim

1.461

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

TensorWave MI300X Cluster 1 Phase 1 · power_capacity_mw

Substrate verbatim

1.529

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

HUMAIN Saudi Arabia/NVIDIA Phase 1 · power_capacity_mw

Substrate verbatim

50.45

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Tesla A100 Cluster Phase 1 · power_capacity_mw

Substrate verbatim

4.948

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Meta Research SuperCluster (RSC-1) Phase 1 · power_capacity_mw

Substrate verbatim

5.179

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Sakura's H100s Phase 1 · power_capacity_mw

Substrate verbatim

2.877

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Sesterce Synapse Phase 1 · power_capacity_mw

Substrate verbatim

2.922

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Sesterce H100s Phase 1 · power_capacity_mw

Substrate verbatim

2.922

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

NFDG Andromeda Phase 1 · power_capacity_mw

Substrate verbatim

3.648

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Nebius Kansas City Phase 1 · power_capacity_mw

Substrate verbatim

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Telangana Yotta H1 Hyderbad AI City Cluster Phase 1 · power_capacity_mw

Substrate verbatim

5.606

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Yotta Shakti Cloud NM1 Phase 1 · power_capacity_mw

Substrate verbatim

5.845

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Alps Supercomputer Phase 1 · power_capacity_mw

Substrate verbatim

8.683

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Nebius 8k Finland Phase 1 · power_capacity_mw

Substrate verbatim

11.415

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Stargate UAE Phase 1 · power_capacity_mw

Substrate verbatim

200

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

NVIDIA CoreWeave Eos-DFW Phase 1 · power_capacity_mw

Substrate verbatim

15.616

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Tesla Cortex Phase 2 · power_capacity_mw

Substrate verbatim

70.07

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Tesla Cortex Phase 1 · power_capacity_mw

Substrate verbatim

71.344

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

xAI Colossus Memphis Phase 1 · power_capacity_mw

Substrate verbatim

150

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Applied Digital CoreWeave Ellendale Phase 1 · power_capacity_mw

Substrate verbatim

100

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

xAI Colossus Memphis Phase 2 · power_capacity_mw

Substrate verbatim

280.28

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

OpenAI Stargate Abilene Oracle OCI Supercluster Phase 1 · power_capacity_mw

Substrate verbatim

240.24

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

OpenAI Stargate Abilene Oracle OCI Supercluster Phase 2 · power_capacity_mw

Substrate verbatim

800.803

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

xAI Colossus 2 Memphis Phase 1 · power_capacity_mw

Substrate verbatim

264.264

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Applied Digital CoreWeave Ellendale Phase 2 · power_capacity_mw

Substrate verbatim

250

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

DataVolt Neom 1.5 GW Phase 1 · power_capacity_mw

Substrate verbatim

300

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-07-10Tier 2 (research)

Stargate UAE Phase 2 · power_capacity_mw

Substrate verbatim

1,000

Editorial overlay

Withheld (DQP §5)

Reason

Primary sources

2026-05-19Tier 3 (press)

Stargate UAE Phase 1 · total_gpu_count

Substrate verbatim

100,000

Editorial overlay

35,000

Reason

Primary sources

2026-05-19Tier 3 (press)

Stargate UAE Phase 1 · peak_pflops_fp16

Substrate verbatim

250,000

Editorial overlay

87,500

Reason

Primary sources

2026-05-19Tier 3 (press)

Stargate UAE Phase 1 · h100_equivalent_gpus

Substrate verbatim

8,338

Editorial overlay

88,428

Reason

Primary sources

Licensed-database records