Corrections

The Disclosure view of the Companies register listed 68 companies linked to tracked facilities, each with the reason its LEI record gives for reporting no consolidating parent, a link to that record and the company’s sites and capacity; its totals counted 491 declaring organisations. Each of the 68 company records linked to its row. For 24 of the 68 the LEI, and so the filing, belonged to another legal entity: University of Oxford showed the filing of Swiss Friends of Oxford University (Zug, Switzerland), Together AI that of TOGETHER (Perros-Guirec, France), Crusoe that of CRUSOE LIMITED (Gibraltar, lapsed since 9 April 2022), Orange that of a Marseille company named ORANGE whose registration is retired, and Princeton Digital Group that of another company in its group. Expedient’s LEI could not be tied to Expedient either way. The register labelled filings in the non-public class “Declines to name its parent”. The participation page sent, with each site it lists, whether the site’s operator and holders had filed such a reason, though it printed none of it. The Huawei Technologies and Microsoft dossiers described their non-consolidating declaration as control by an entity that does not consolidate the company’s accounts, and the visibility methodology said organisations are never matched by name alone and that corporate form is never stripped as noise.

The Concentration Monitor printed supplier shares, a Herfindahl and an editorial composite for four rows whose own source lines said the cited sources do not verify the shares. Advanced 2.5D Packaging (CoWoS-class): TSMC 65%, Amkor 15%, ASE Group (SPIL) 12% and Samsung 8%, a Herfindahl of 4,658 and a composite of 5,543 (Elevated), labelled Stable; its source line kept the shares as editorial advanced-packaging estimates. EML Laser Chips (800G+ Optical Interconnect): Lumentum 55%, Coherent (incl. Source Photonics) 30% and others 15%, a Herfindahl of 4,150 and a composite of 9,172 (Critical), labelled Stable; its source line said the product announcement it cites does not verify the shares. EUV Photomasks & Pellicles: Toppan 30%, Dai Nippon Printing (DNP) 25%, Photronics 20%, HOYA 15% and others 10%, a Herfindahl of 2,250 and a composite of 2,565 (Moderate), labelled Stable; its source line called the shares editorial estimates without a verified EUV-specific market breakdown. Wafer Inspection & Metrology Equipment: KLA Corporation 60%, Applied Materials 15%, Hitachi High-Tech 10%, Onto Innovation 8% and others 7%, a Herfindahl of 4,038 and a composite of 4,179 (Elevated), labelled Stable; its source line said the filing it cites does not verify the market breakdown.

The United Arab Emirates’ country page and its profile in the Country Access register closed the note on approved recipients with “G42 and Core42’s authorization expires April 6, 2027 absent subsequent BIS notice.” The G42 dossier printed the same note in its export-control posture row.

The Concentration Monitor printed supplier shares, a Herfindahl and an editorial composite for three rows whose shares no cited source gave. Chiplet Interconnects (UCIe): TSMC 55%, Intel 20%, Samsung 10% and others 15%, a Herfindahl of 3,750 and a composite of 5,063 (Elevated), labelled Concentrating. The shares estimated advanced-packaging capacity by a method the row did not disclose, and none of the sources cited here gives supplier shares for UCIe interconnects. US Data Center Grid Interconnection: PJM Interconnection 40%, ERCOT 25%, MISO 15% and other grid operators 20%, a Herfindahl of 2,850 and a composite of 3,078 (Moderate), labelled Concentrating. The percentages had no source or definition, and the page ranked their Herfindahl among company Herfindahls under the heading Supplier Concentration; grid operators are not suppliers a developer can choose between at one site. Semiconductor-Grade Neon Gas: Ukraine (Ingas, Cryoin) 55%, China 25%, Japanese purifiers 8%, Korean purifiers 4% and others 8%, a Herfindahl of 3,794 and a composite of 3,547 (Moderate), labelled Stable, with Ukraine as the top supplier. The card’s own source, a USITC briefing of April 2022, gave Ukraine roughly half of world supply and reported that Cryoin and Ingas had stopped production in March 2022; no source was cited for the other shares.

Five rows printed a trajectory that the evidence on their own cards did not support. Wafer Inspection & Metrology Equipment read Concentrating (Herfindahl 4,038, composite 4,179) over a filing that, the card said, does not validate a change in the 60% KLA share. EML Laser Chips (800G+ Optical Interconnect) read Concentrating (Herfindahl 4,150, composite 9,172) over a prototype demonstration that measures neither laser-chip production nor a shortfall. EUV Photomasks & Pellicles read Concentrating (Herfindahl 2,250, composite 2,565) over a pellicle maker’s production and a photomask development target, neither of which describes a change in the photomask shares scored. Advanced 2.5D Packaging (CoWoS-class) read Deconcentrating (Herfindahl 4,658, composite 5,543) over TSMC’s plan for two US packaging sites, which gives no production date and would add capacity to the largest supplier. ABF Dielectric Film read Deconcentrating (Herfindahl 9,050, composite 10,317), assessed 19 May 2026 on evidence that cites a glass-substrate announcement of 24 July 2026, which reports no commercial product shipment. The page’s lead read “6 chokepoints are concentrating, 6 stable and 2 diversifying”.

Compute estimates multiplied dense peak throughput by a utilization default of 0.40 and an interconnect default of 0.85, so at the defaults a site was estimated to sustain 0.34 of its peak. The 0.40 is SemiAnalysis’s reported model FLOP utilization for trillion-parameter training runs on H100s. Model FLOP utilization is achieved training throughput over peak, so it already includes the time lost to communication between accelerators, and the 0.85 took that loss off a second time. On the records of 26 September, the Thresholds page counted 19 of 184 estimable sites that could accumulate 10²⁵ FLOP within a month, 48 within a quarter and 90 within a year, across 21 countries, with 94 taking longer than a year and the fastest 1.25 days.

864 facility records read “Hyperscale DC” on their own pages and wherever the site listed them: country ledgers, the facility register, site search, the Query Builder and the MCP server. No source any of them cites states a class. 718 are OpenStreetMap sites imported by their data-centre tags, 96 are sites in the IM3 Open Source Data Center Atlas and 21 are sites from an OpenStreetMap extract, each recorded only as a data centre; 29 are interconnection-queue requests in New York and Texas. Two of those requests are energy storage by their own sources: Giga Texas Data Center (ERCOT request 25INR0688, battery storage) and Port Jeff -Mt. Sinai Energy Storage (NYISO request 0821, withdrawn). The register’s “Hyperscale DC” filter and the Query Builder’s facility search returned these records under that class, and with them the 455 OpenStreetMap records corrected on 31 May, whose pages already read “Class not stated”: 1,319 records in all. The 864 were counted in the database on 25 September 2026: records stored as Hyperscale DC, from these four sources, with no class correction on record.

From 17 July 2026 until this correction the Sovereign AI Program Index download was labelled v2026.3, released 17 July 2026. Four different CSV files were served under that one label in turn: SHA-256 19a19e3f91fad40a2978b5b733f338eaa13f774f40906be003bf8133868dec56 (17 July, 34 programmes), 38c2dbda0a3acc41e571eab2fdd1a56c6b45faaa59dea982d3ae9ed230a81f13 (14 August, 36), 6be7b5ddc8b04ae01aaf36aabf9961045e051391a657f3488dcc9c5ca698a68f (14 August, 36) and cd296f9b045355c44b4696817e45fe61dd7af396c9f1c3dc456a2e4f8225304b (8 September, 36). From 21 September the methodology file served with the last of them said the v2026.3 CSV, its checksum and its release date were unchanged.

The Sovereign AI dashboard rated 35 of its 36 programmes high for TSMC dependency, and 35 on BIS jurisdictional reach (27 low, 3 medium, 3 high, 2 none). Neither rating had a published rubric, and no record cited a source for its rating. The export-control ratings contradicted the records: the UAE was rated low beside G42’s announcement of US approval for advanced-chip exports, while the Philippines and Thailand were rated high with no export-control source. China, the Czech Republic and Poland were rated high for NVIDIA dependency and Ireland medium, although none of the four records names NVIDIA hardware or lists a source about NVIDIA.

Spain’s only priced component read “EUR 1.5B National AI Strategy 2024 | Government | $1.62B”, yet its programme page said “No government-only dollar amount on record”, and its government-funding field was empty on the dashboard and in the dataset; its country page gave the same $1.62B as a programme amount.

The visibility sheet read that 46.9% of operating capacity has no controlling jurisdiction identified: 7,761.0 MW of 16,555.8 MW hosted. Google’s data centres were counted in that share, because their AI-hardware owner was recorded on a second Google record separate from Alphabet Inc.’s, and that record had no parent in the ownership graph. The same split sent 78 recorded training runs and deployments, and 20 facility relationships, to the second record, with 54 more runs, two Together AI hardware-owner links, one IREN supply-chain link and one IREN name alias on second records of Meta, Together AI, IREN and Alphabet. The U.S. Department of Energy’s record read “Us Department Of Energy”, gave no headquarters country, typed the department as a power provider and cited an Epoch AI file that does not name it.

Fifteen country pages whose country holds no facility records (Afghanistan, Bolivia, Cyprus, Iraq, Jamaica, Jordan, Kyrgyzstan, Madagascar, Malta, Myanmar, Nicaragua, North Korea, South Sudan, Syria and Uzbekistan) printed “0.0% with capacity data or a site description” in their facility-detail figure. Its screen-reader label gave the same share of none (“0 of 0 facility records held for Afghanistan have capacity data or a site description (0%)”), and the power figure’s label beside it read “Operating plant is 0% here”.

The United Arab Emirates page listed 5.0 GW of power under construction beside 3 MW operating and 160 MW announced, the countries register counted the same 5,000 MW in its under-construction stage, and Stargate UAE’s record called the figure not operating capacity because the site is under construction. The 5,000 MW is the campus’s announced envelope: the campus note on the record says the envelope should not be read as capacity already under construction, and puts the anchor project at about 1 GW with a first cluster of about 200 MW.

The training-evidence register gave Oracle Corporation about 170 MW of operating power. Oracle’s Abilene campus is filed as a current deployment of 590 MW, a review of which admits it as operating plant inside the 1,200 MW build-out, three phase records and the build-out itself, and the register keeps one record per site. The deployment’s figure was carried onto the kept record only when the deployment was the record the kept one replaced; at Abilene the phase records replaced it first, so the 590 MW was dropped.

Alphabet’s dossier counted 101 linked facilities as sites, including 88 operating and 9 announced, while Alphabet’s company record counted 96: the dossier included five second records of sites Alphabet already holds (Google’s Lincoln, Omaha and Papillion campuses in Nebraska, The Dalles in Oregon and Waltham Cross in England). The company records made the opposite error for a site a company holds only through a second record. G42’s record listed 15 facilities and not Stargate UAE, because G42’s record of that site is a second record of the campus whose canonical record is Oracle’s; the Stargate OpenAI record listed none.

The passage on ownership documentation over time read that across 3 snapshots from 23 to 25 September 2026 the operating-capacity share classified as Tier 1 had not moved, from 73.8% to 73.4%, with the data-gap share rising from 5.9% to 6.5%, and that the snapshots contained 4,266 and 4,266 facilities respectively, so the share moved with the set of records held. The 23 September snapshot counts Epoch’s records at IT power and the later ones at facility power: operating capacity went from 13,536 MW to 16,642 MW overnight with no tier’s facility count changing. The register table also printed the United Kingdom’s capacity controlled from elsewhere as 0.6 MW, where the country’s attribution figure read 0.5 MW.

Huawei’s and G42’s dossiers read “Searched US SEC filings (EDGAR full-text search), Aug 19, 2026: nothing shows ownership, operation or a capacity contract.” Huawei’s search account said “Huawei Cloud” returned 44 filings from 19 registrants, and that the two data-centre operators among them, GDS Holdings and VNET Group, identified no lease, hosting agreement or other facility-control arrangement with Huawei. G42’s said that neither NVIDIA’s filing nor Cerebras’s established facility ownership, operation or contracted capacity. DeepSeek’s gave a press report as “a self-built site of about 1 GW in Ulanqab, Inner Mongolia, and leased capacity from unnamed providers”.

The opening line of six company dossiers stated a universal or a count wider than the search behind it. Anthropic’s read “Anthropic holds all its documented compute by contract, from 5 companies”; Microsoft’s “Microsoft Corporation runs its own sites and contracts capacity from 5 companies”; Alphabet’s “Alphabet runs its own sites; no contracted capacity is documented”; G42’s “No primary document shows how G42 holds its compute”, with the same sentence on DeepSeek’s dossier and, beneath its Entity List line, on Huawei’s. The searches had read other companies’ disclosures, and for G42 and Huawei only US SEC filings; none had read the subject’s own filings, announcements or papers. Anthropic’s own announcements show what that left out: it announced four of the recorded arrangements itself, and two more that the record did not hold. Its dossier also read, under “What cannot be seen”, “All 7 disclosures of these arrangements come from Anthropic’s counterparties”, every control plate printed an empty band as “No contractual capacity documented”, “No operating arrangement documented” or “No site-specific title documented”, and missing fields read as facts about the sources: “Anthropic’s share not stated” beside Colossus 1, which Anthropic’s own post of 6 May 2026 answers, “No contract start dates were disclosed”, “3 of 5 contracts state no capacity” and “No SEC filings under its own name”, which a 2004 Schedule 13D naming Huawei Technologies as a reporting person contradicts.

The Anthropic dossier listed Amazon’s April 2026 agreement under “Announced” as “up to 5.0 GW; current and future Trainium generations; no delivery dates”, and its note read “Amazon also announced an immediate $5 billion equity investment and up to $20 billion tied to milestones”. The note on TeraWulf’s August 2026 presentation said its approximately $19 billion contracted-revenue figure “supplements the earlier lease announcement”.

The Anthropic dossier held seven disclosures, all by its counterparties, and said under “What cannot be seen”: “All 7 disclosures on record come from Anthropic’s counterparties; none is Anthropic’s own”. Anthropic had announced seven of its compute arrangements itself, and this site’s chokepoints page already cited one of those announcements, of 6 May 2026, for Colossus 1. Three of the seven were missing from the record altogether, not the two an earlier notice gave: $30 billion of Azure compute capacity from Microsoft, up to one gigawatt on NVIDIA systems from counterparties Anthropic does not name, and a $50 billion investment in data centres built with Fluidstack. The receipt for TeraWulf’s August 2026 document called it a “furnished earnings presentation” that “repeats the approximately $19 billion of contracted revenue”, and the Project Rainier line dated Amazon’s release June 2025.

The G42 dossier said Supplement no. 8 to part 740 approves G42 and Core42 to receive specified advanced chips without a license and that, absent further notice from BIS, the approval expires on April 6, 2027, citing “91 FR 43034, Jul 10, 2026”. It described the United Arab Emirates as “treated as a US ally under US chip-export rules”.

The notice of 23 September on planned capacity counted as present capacity said that operating capacity is 16.6 GW on 322 records, that a further 56.2 GW is recorded as announced, under construction, build-out beyond operating plant or earlier phases, and that the worldwide operating share every country page prints is 31.6%. The first two figures count every record, those from licensed sources included, and over them 16.6 GW is 22.9% of 72.8 GW; the 31.6% is taken over the publishable records alone, 52.6 GW. The company disclosure register gave the same 56.2 GW without saying that 20.2 GW of it sits on licensed records the country pages leave out, and the countries register printed 32% operating over the records placed in a country, 52.5 GW, leaving out 125 MW on 112 records whose country does not resolve. No page named the records it counted.

Since 31 May, 455 facility records from the OpenStreetMap corpus have read “Class not stated” on their own pages: the ingest stamped all 1,173 of its records Hyperscale DC, and no source states a class for these 455. Country pages printed the stamped class instead, on at least 315 rows across 77 country pages, and the Query Builder, the site search and the MCP server returned it too.

630 of the 5,485 public supply-chain records describe partnerships, which name neither company as the other’s supplier. Company records listed each one under a company’s suppliers or its customers, according to the column it happened to be stored in; the supply-chain graph and the map drew them as supplier-to-customer links; and the Query Builder and the MCP server’s supply-chain tool returned them as upstream or downstream relationships. NVIDIA’s record, headed “Upstream Suppliers · 116 records”, opened with AAEON Tech, ABB, Adlink, Advantech, AEye and Altair: 91 of the 118 records it read were partnerships, Barclays among them, whose cited filing is a share-repurchase agreement.

From 8 May, the Colossus 1 record, its Markdown copy and the Query Builder described one development, credited to xAI’s post: Anthropic leasing all of Colossus 1’s compute as the facility’s operating tenant, xAI keeping ownership of the facility, and the price and term undisclosed. The neocloud chart called the arrangement a whole-cluster lease. The chokepoints page gave Anthropic’s figure as “220,000 NVIDIA GPUs”.

The home page counted 142 designation-to-company matches, 102 of them through ownership chains. The Entity List cross-reference read 138 of 3,420 active designations as matched (4.0%), with 142 matches, 139 of them involving organisations with supply-chain records, 95 designations Confirmed and 43 Screening leads, and 102 chain matches between 101 designations and 10 organisations, 58 at high confidence and 44 at medium. Among them, 52 separately designated Huawei companies were matched to HiSilicon, 35 China Electronics Technology Group institutes and subsidiaries to their own parent group, four Pakistani companies with “Systems” in their names to Systems (Pakistan), AccoTest Technology and a hair-accessories maker to Anhui Conch Cement, and Xinjiang Police College to COL. Company records listed export-control designations reaching 78 organisations through affiliates, among them 92 on each of Rusal, EuroSibEnergo and Vostsibugol and 53 on HiSilicon. The methodology gave 699 of 3,435 designations (20.35%) an anchor in the ownership graph, and 101 (2.94%) a path to a compute-relevant organisation. The Impact Pre-Scorer measured its projections against 51 entities reached through corporate relationships in the Huawei 2019 family and 2 in SMIC 2020, and its removal case for CETC-7 and CETC-30 reached 13 closure entities.

The procurement column of the Sovereign AI dashboard gave the United States $40.4B, recorded across 2,298 of 2,471 records, beside a green dot labelled “High confidence (≥0.9)” and authority tier 1. The figure admitted any record the relevance classifier scored 0.7 or above, whatever it had rated the record’s relevance: 2,412 of the 2,471 were records it rated unrelated to AI ($39.9B), and 2,087 of those ($39.4B) were rejections by a keyword pre-filter, which records a score of 0.95 without a model reading the record. Grants.gov funding opportunities, advertised budgets set before any award, were counted as committed spending (948 records, $39.5B). One State Department opportunity, stored once for each of its four postings, was counted four times at $2,147,483,647, the largest value its source field accepts. The dot showed the classifier’s mean score for its own calls, and the tier graded the source documents; neither measured whether the figure was right.

The coordination view classed the 208 facilities it attributes to Microsoft, Equinix, Nvidia and ASML as OPAQUE, each chain described as having “no confirmed public-filing or government endpoint”, although all four companies file with the SEC: the records for Microsoft Fairwater Atlanta, Equinix LD10, NVIDIA Helios and ASML Veldhoven each labelled the chain “Unresolved ownership chain” beside a tier 1 visibility reading whose stated basis is the company’s SEC identifier. The OPAQUE total read 3,669 facilities and 4,894,253 operating H100-equivalent GPUs; NONE read 509 facilities and 7,722,834.

The Query Builder’s procurement search could return three records drawn from EU public procurement notices (Tenders Electronic Daily), and its allocation rollup three totals built from them: an open tender in Lithuania and contract awards in Switzerland and Poland, each rollup total reading $0. None described a real procurement. Each was filed under a number from inside a notice instead of the notice’s own publication number, so its source link led to no published notice; each named the contracting authority as the contract’s winner; each carried a three-letter country code where every other record carries two; and each was marked unrelated to AI although no notice had been read by the classifier: a filter rejected the notices’ product codes because they lacked a check digit, and the rejection was stored as the classification.

The regional-billings panel on Trade Impact compares WSTS billings in the twelve months after an action with a counterfactual, and the WSTS series on the page ends in January 2026. Selecting any of the five US actions dated after it, from the Super Micro indictment of 19 March 2026 to the UAE’s move to A:5 on 10 July 2026, printed a net gap of +$0M and, for each of the four regions, “+$0M vs counterfactual · +0.0%”: months the series does not hold were read as zero. For the ten actions from March 2025 to January 2026 whose following year the series holds only in part (1 to 11 months), the panel summed the months it had and described the sum as covering the twelve post-event months, and the per-action table printed seven of those sums beside full-year sums with no month count.

The Iran threat scenario’s summary, as the server sent it, read “14 location records fall within the published range envelopes”. The 14 were the supplemental cloud-region records alone: the page requested the facility records only after it reached the browser, so readers whose browser ran it saw the full count a moment later, while search engines, readers without scripts and any load whose facility request failed kept 14, and a failed request was not mentioned.

The Query Builder’s tool register described the BIS licensing tool as “2,587 rows covering CY2018-2022”, each carrying approval, denial and return-without-action counts, a licence value, and a friction rate and tier, with a start year defaulting to 2018 and a friction filter to use for “majority denied”. It described the trade tool as reading five sources, and that tool’s year and partner groupings, described as annual totals, added every source’s rows and every direction together, though the sources measure the same trade.

Every corridor on Trade Impact added together the rows of every source that reports it, though each of those sources measures the same trade. Korea → China for the October 2022 controls read $101.6B → $82.9B a year (−18.4%): CEPII BACI’s $60.4B plus UN Comtrade’s $41.3B, a figure neither source reports. China’s IC imports from the United States read $30.0B → $17.8B (−40.7%), China GACC’s two-way trade added to UN Comtrade’s imports. The reporter totals each verdict is judged against, and the seller baskets on Plate IX, were summed the same way (Korea’s exports at about $229B a year; the Korean, Japanese and German baskets each drew “Other Asia, nes” twice, once per source’s code). The discovered corridors added sources as well, counted GACC’s two-way figures as exports (China → Taiwan, $174.8B), and listed one flow twice (CEPII BACI’s United States → China beside China’s own imports from the United States; Korea → Other Asia, nes under two codes). The December 2024 action was compared with 2025, a year neither UN Comtrade nor CEPII BACI reports, so its before-years added three sources and its after-year held one (Japan → China, −83.6%).

The cross-reference register joined ten Entity List designations to organisations they do not name: Schneider GmbH to Schneider Electric SE as a confirmed match, and nine others as screening leads, among them China Communications Construction Company to China Construction Bank, the University of Science and Technology of China and Nanjing University of Science and Technology to the University of Electronic Science and Technology of China, and Harbin Engineering University to the Harbin Institute of Technology. The change feed listed all ten as matches added.

On at least two requests (15 August and 23 September 2026) a batch of organisation reads inside the Regulatory Impact Pre-Scorer timed out, and the page showed the simulation built from the reads that succeeded: the organisations in that batch lost their names and headquarters countries, so the report’s country breakdown left them out. Which entity was simulated, and the figures those two requests showed, were not recorded.

The plate on /cascade ranked every organization by its count of recorded downstream relationships within three links and described each count as how far “a failure here reaches”, naming Siemens EDA and TSMC the widest at 191 organizations each, most of that count running through partnership records.

The About page gave 16,755 distinct supplier–customer relationships and the methodology page 16,755 supplier–customer–product keys, counting as separate relationships records that differed only in a product label the site withholds, while the cascade engine held 16,346 (earlier on 23 September the pages gave 14,831 against 14,422). The methodology page also said partnership records were in that count; the cascade leaves them out. The API index described the supply-chain graph endpoint as “Full supply-chain graph (20,798 edges)”, the number of stored records, licensed ones included.

The cascade propagated disruptions along every relationship on record as a present dependency, including those whose only evidence is SEC disclosures for fiscal periods that ended five or more years before the latest filing wave: 1,367 of 13,233 supplier–customer pairs in the full corpus and 991 of 1,938 in the public graph drawn on /cascade. ASYST Technologies, last disclosed in fiscal 2002 and bankrupt since 2009, was among them.

The cascade engine treated every stored supply-chain record as a supplier–customer dependency, including 3,337 that record a partnership: 1,008 licensed relationships of the partner kind and 2,329 whose product line reads “… partnership”. Disruptions travelled along them from one partner to the other, and /cascade read 14,422 relationships among 4,950 organizations (16,346 among 5,374 once the 1,924 licensed relationships added on 23 September reached it).

Marvell's record gave revenue of $21.1B and capex of $0.3B: Cognizant Technology Solutions' figures, read from Cognizant's filer number. Equinix's gave revenue of $6.0B, its 2020 figure, and capex of $14.9M, which matches no line in its filings, under a 2024 annual-report citation. Iron Mountain's gave revenue of $4.5B, its 2021 figure, under the same kind of citation. No company's revenue or capex stated the fiscal year it covered.

From 14 August 2026 the worldwide series on the Compute Visibility Index was built from the first 1,000 of the snapshot rows (the table held 38,717 on 23 September). It read “Across 4 snapshots between 2026-04-25 and 2026-04-29, the capacity share classified as Tier 1 has risen, from 39.6% to 43.8%”, over 4,529 and 3,669 facilities, with the data-gap share at 12.4% and 13.1%. The last of those snapshots was cut part-way: it holds 4,529 facilities, of which the page read 3,669.

Capacity totals added the power and accelerator figures of records in every status, and the operating figures that did read status counted two things a status does not settle: a campus build-out figure filed on an operating record, and an earlier phase of a cluster whose operating later phase already counts its chips. The United Arab Emirates page gave 5.2 GW of tracked capacity, 5,000 MW of it one campus under construction, and 91,888 accelerators, 85,000 of them announced; the countries register ranked it second. The United States page gave 16.3 GW operating, counting the New Carlisle and Abilene build-out figures (2,200 MW and 1,200 MW) as operating plant, and 6,987,464 accelerators of every status. The Compute Visibility headline read 67.7% of tracked capacity as having no controlling jurisdiction identified, over 72.3 GW across 4,057 sites; the disclosure register placed its declaring organisations at 11.7 GW of 72.8 GW; the training-evidence register led with Meta at 4,625 MW; the participation page gave the United States 16,330 MW on 152 sites; the neocloud chart gave CoreWeave 1,158,959 H100-equivalents and Applied Digital 846,387, none of Applied Digital’s on an operating site; the threshold plot ranked 201 facilities, 97 reaching 10²⁵ FLOP within a year; the coordination view summed H100-equivalents of every status; the impact pre-scorer summed power of every status (8,032 MW with Oracle as the seed); and Alphabet’s dossier counted 736 MW announced on 3 records, one of them a second record of Google’s Papillion campus whose figure the corrections withhold.

The coordination view walked each facility’s ownership chain from whichever organisation its record named as operator, owner or hardware owner, whether or not a source established that role. It read 104 facilities as STRONG (a chain reaching a non-aligned or restricted jurisdiction), 28 as MILD, 780 as NONE and 3,354 as OPAQUE.

The attribution chart on the Compute Visibility Index classed a facility’s control from whichever organisation its record named as operator, owner or AI-hardware owner, whether or not any source established that role, and walked every recorded ownership link above it. It read 43.6% of 72.3 GW as attributable to a controlling jurisdiction (34.8% to the host country, 8.8% elsewhere), 32.9 GW (45.5%) as an unresolved chain and 7.9 GW (10.9%) as no organisation on record.

From 10 July 2026 the 52 facility records drawn from Epoch AI’s data-centre table gave Epoch’s IT power as their power capacity, while records from every other source give facility power, cooling and other overhead included. Country and operator totals added the two: the United States page read 13.8 GW operating of 36.5 GW tracked, and its capacity chart put Colossus 2 at 1,276 MW beside a record reading 946 MW.

Country pages printed a programme’s government funding under the word “announced”: the United States read $30.0M announced, against the programme’s ~$2.6B headline announcement.

Two disruption-scenario rationales gave Synopsys and Cadence about 75% of the EDA market while the scenario descriptions gave 62%, and the coordination analysis gave an unsourced 80% or more for the three largest vendors.

The coordination analysis cited TrendForce’s second-quarter 2025 foundry figures but gave TSMC 71% and Samsung Foundry both 6.8% and 8%.

The Czech Republic’s accelerator total counted the Karolina supercomputer’s 576 GPUs twice, once from IT4I’s hosting record and once from a catalogue’s deployment record: 1,152 GPUs. EuroHPC JU’s organisation totals carried the same 576 twice.

Both company records gave the February 2019 transaction a single value of $135.0M, a figure the seller’s filings do not state.

640 GPUs: the V100 partition. Once the count was corrected to 776, the compute estimate applied the V100 rate to all 776 GPUs, and the record’s catalogue figures gave 203.73 PFLOP/s and 165 H100-equivalent by assuming the SXM form of the H200.

The compute estimate treated the facility’s 64 Cerebras CS-2 systems as 64 NVIDIA H100 GPUs, because no rate was recorded for Cerebras hardware.

2,816 GH200 superchips, from the planned configuration. Once the count was corrected to 2,800, the record’s compute figures applied the NVIDIA H100 SXM rate to each superchip: 2,770.6 PFLOP/s and 2,800 H100-equivalent.

G42’s company-wide authorization to purchase up to 35,000 GB300-equivalent chips was assigned to Stargate UAE Phase 1.

Incomplete quarters and annual observations were included alongside cumulative shipment series. Changes in source coverage appeared as changes in shipments and concentration.

Changing the weights updated the chart, but exports retained the equal-weight series.

Some relationship records included numerical values and sourcing flags that their public citations did not establish.

The former funding-gap breakdown attributed shares of the difference between announced and disbursed funding to particular causes.

The chart included unsupported relationships and weights that did not consistently describe the connections drawn.

The MI350X record used sparsity-assisted figures for dense performance.

The compute-leverage analysis described advanced packaging as outside export controls.

The Loughton record named CloudHQ as owner and used a 45,000-chip site-capacity ceiling as the deployment count.

The facility retained the IO Datacenter name from its source map.

The PCIe and NVL records reused memory specifications from the SXM variant.