Corrections per day. Circle area is proportional to the count below 13; at that count or above, the circle size is fixed and the count is printed.
20
published errata: a published account of an error, its cause and the correction.
165
numerical corrections: a revised or withdrawn numerical value. 89 concern licensed records. Details are withheld under the licence; one daily mark links to the aggregate entry below.
100
name corrections: an organization or facility name checked against its own filing or website.
3
ownership-event corrections: a description of an ownership transaction corrected against a filing.
Each circle groups corrections of one kind made on the same day. Within each kind, vertical separation prevents overlaps. Select it to open the first corresponding entry below.
The register
Values are reproduced without rounding. Corrections to licensed records are grouped in one entry.
Paired re-derivation from the corrected chip count. Epoch's verbatim H100-equivalent for this row equals its chip count (2,816 = 2,816 × 1.0: Epoch converts the GH200 at 1.0 H100 per superchip), so the same 1.0 factor is held: 2,800 × 1.0 = 2,800. No convention change is smuggled in with the count change. Epoch AI's GPU-clusters catalogue (2025-07-21 pull) transcribes KAUST's 2024 pre-installation announcement — '704 nodes, each with quadruple Nvidia "Grace Hopper" Superchips … These 2,816 Superchips' — as a planned 2,816. KAUST's own release of 2025-11-17 states 2,800 GH200 accepted and transferred from HPE in October 2025, and the KSL system documentation states the GPU partition as 2,800 with 122.8 PFlop/s HPL. Same institution, seventeen months apart, the later figure post-acceptance; the later governs display. The 16-unit difference is not explained by either source. Substrate value preserved verbatim per DQP §3 + §16; contradiction recorded in source_conflicts per DQP §7.
Paired re-derivation from the corrected chip count: 2,800 × 989.5 TFLOP/s dense FP16 = 2,770.6 PFLOP/s. The per-superchip figure reproduces the substrate value exactly on the verbatim side (2,816 × 989.5 TFLOP/s = 2,786.432 PFLOP/s, the stored value), so the convention is held fixed across the correction; it is the H100-SXM5 dense figure of the canonical registry (src/lib/methodology/versions.ts GPU_MODELS.h100_sxm5), which Epoch applies to the GH200's Hopper GPU. Deterministic re-derivation, not a new estimate. Epoch AI's GPU-clusters catalogue (2025-07-21 pull) transcribes KAUST's 2024 pre-installation announcement — '704 nodes, each with quadruple Nvidia "Grace Hopper" Superchips … These 2,816 Superchips' — as a planned 2,816. KAUST's own release of 2025-11-17 states 2,800 GH200 accepted and transferred from HPE in October 2025, and the KSL system documentation states the GPU partition as 2,800 with 122.8 PFlop/s HPL. Same institution, seventeen months apart, the later figure post-acceptance; the later governs display. The 16-unit difference is not explained by either source. Substrate value preserved verbatim per DQP §3 + §16; contradiction recorded in source_conflicts per DQP §7.
Displayed chip count moves to KAUST's post-acceptance figure. Epoch AI's GPU-clusters catalogue (2025-07-21 pull) transcribes KAUST's 2024 pre-installation announcement — '704 nodes, each with quadruple Nvidia "Grace Hopper" Superchips … These 2,816 Superchips' — as a planned 2,816. KAUST's own release of 2025-11-17 states 2,800 GH200 accepted and transferred from HPE in October 2025, and the KSL system documentation states the GPU partition as 2,800 with 122.8 PFlop/s HPL. Same institution, seventeen months apart, the later figure post-acceptance; the later governs display. The 16-unit difference is not explained by either source. Substrate value preserved verbatim per DQP §3 + §16; contradiction recorded in source_conflicts per DQP §7.
Paired re-derivation across the three partitions. Epoch's verbatim 40 reproduces as 80 PFLOP/s ÷ 1.979 PFLOP/s (Epoch's CSV gives 40.42; the row stores the rounded 40) — Epoch's H100 reference is the 1,979 TFLOP/s figure — so the V100 term is kept at Epoch's 40 and the added partitions are converted at the same reference: H200 1,979 ÷ 1,979 = 1.0 per GPU (the dense ratio 989.5 ÷ 989.5 gives the same), A100 624 ÷ 1,979 = 0.315 per GPU (dense 312 ÷ 989.5 gives the same). 40 + 120 × 1.0 + 16 × 0.315 = 165.04, stored 165. The convention is held fixed on the verbatim side, and on the added side the ratio is the same under either convention. NCI's HPC-systems specification lists three GPU partitions — 30 × 4 NVIDIA H200 141GB, 2 × 8 NVIDIA A100 80GB, 160 × 4 NVIDIA Tesla V100 32GB — where Epoch AI's GPU-clusters catalogue carries the V100 partition alone (640, transcribing the 2024 specification). The NCI Annual Report 2025 dates the H200 expansion to 2025; the NCMAS 2024 guide (2023-11-09) lists the DGX A100 nodes as existing then. Substrate value preserved verbatim per DQP §3 + §16; partitions carried in gpu_inventory with per-entry provenance.
Paired re-derivation across the three partitions at the canonical registry's dense FP16 figures (src/lib/methodology/versions.ts GPU_MODELS): 640 × 0.125 PFLOP/s (V100 SXM2) = 80, which reproduces the substrate value exactly and so pins the convention; plus 120 × 0.9895 (H200 SXM, same compute as H100 SXM5) = 118.74; plus 16 × 0.312 (A100 80GB) = 4.992. Total 203.732 PFLOP/s. Deterministic re-derivation, not a new estimate. NCI's HPC-systems specification lists three GPU partitions — 30 × 4 NVIDIA H200 141GB, 2 × 8 NVIDIA A100 80GB, 160 × 4 NVIDIA Tesla V100 32GB — where Epoch AI's GPU-clusters catalogue carries the V100 partition alone (640, transcribing the 2024 specification). The NCI Annual Report 2025 dates the H200 expansion to 2025; the NCMAS 2024 guide (2023-11-09) lists the DGX A100 nodes as existing then. Substrate value preserved verbatim per DQP §3 + §16; partitions carried in gpu_inventory with per-entry provenance.
A G42 chip purchase authorization was misattributed to Stargate UAE Phase 1
What the site showed
An earlier correction replaced Epoch’s power-derived estimate of 100,000 chips with 35,000 GB300 chips, then derived 87,500 dense PFLOPS and 88,428 H100 equivalents. Those values appeared as Phase 1 capacity even though the cited authorization applied to G42 as a company.
Correction
The facility chip count and both derived performance values are withheld. Commerce’s November 19, 2025 statement authorizes purchases equivalent to up to 35,000 GB300 chips for G42; it does not specify how many were installed at this facility. G42’s October 16 construction update describes a first 200 MW phase within the 1 GW Stargate cluster and wider 5 GW campus, with delivery planned for 2026. These sources do not establish commissioning or an installed inventory.
Why it happened
The correction assigned G42’s company-wide purchase authorization to Stargate UAE Phase 1 without evidence of that allocation. References to the 200 MW phase supplied no chip inventory. A calculation based on assumed power shares was used to corroborate the count, although it could only test whether the assumed count fit those power assumptions.
Changes made
The numerical corrections layer withholds all three unsupported values and retains the original source record. The separate power correction for cumulative-campus double counting remains in place. Threshold estimates exclude this record when both corrected chip count and power are absent, and tests reject fallback to raw values after a null correction. A future chip-inventory correction requires evidence for the facility, hardware and quantity.
Commerce’s 19 November 2025 statement authorized G42 to purchase up to 35,000 GB300-equivalent chips but did not allocate them to Stargate UAE Phase 1. The former 35,000-chip correction and the compute figures derived from it have been withdrawn. Epoch’s original chip count was estimated from power with unconfirmed hardware, so it cannot establish a replacement inventory. G42’s 16 October 2025 update described the first 200 MW as under construction for planned delivery in 2026. The separate campus power correction is unchanged.
Changing chip-data coverage appeared as a shipment decline and a concentration trend
What the site showed
The upstream chip-flow chart ended with an apparent fall in cumulative shipments. Its final quarter contained only Nvidia and Google observations, all marked incomplete by the source; other designers disappeared into zero-filled series. Cambricon records with separate annual start dates were also mixed into the cumulative stack. The designer-concentration chart used the same changing population and described its early single-designer coverage as near-monopoly. Its incomplete marker appeared in a tooltip without withholding the plotted point.
Correction
The shipment and concentration figures now cover the same reported designers, with a consistent start date for each chip series. Quarters marked incomplete, or missing chips previously reported, are excluded and identified; gaps within the series remain visible. Growth is calculated from the displayed endpoint totals. Cambricon’s two annual observations are shown separately with their source windows. The figures cover the designers reported by the source; no raw source value was changed.
Why it happened
Both charts grouped records by end date without checking whether their start dates and covered populations were comparable. Missing rows became zero, and the source’s incomplete flag was ignored. The concentration chart used designer count to check coverage, so it missed partial periods with the usual number of designers and admitted annual observations into a cumulative series.
Changes made
Both charts now apply the same checks for source windows, incomplete quarters and missing records, and a quarter of unknown completeness is excluded. Screen and print use the same records, with excluded periods identified. The HHI band calculation is unchanged.
On the training-cost disclosure figure, Inflection-2 appeared on the rung for runs whose lab named only the hardware or the model size, despite retaining an explicitly reported FLOP value. Two CHIPS award records also lost their tax-credit disclosure qualifiers and other structured notes. Their failed-citation warnings remained in the database but did not reach the training-cost comparison, the facility pages or the sovereign embed.
Correction
The affected metadata was restored from retained ingestion records and the original ingestion code, after matching each record’s identity, numerical fields and citation. Historical failed-citation warnings remain attached to their original sources and check dates. No numerical value or source URL changed. The retained files are not a pre-overwrite database backup and cannot recover unknown database-only additions.
Why it happened
The check that records whether a source URL still resolves wrote its warning over the disclosure notes on some records, and three surfaces (the training-cost comparison, the facility pages and the sovereign embed) did not show the warnings at all.
Changes made
The check now writes its warning beside the notes, and only when the stored value and citation still match what it read. Warnings render with the individual source on the training-cost comparison, the facility pages (screen, Markdown and print) and the sovereign embed; a source held under licence shows no locator. BIS warning visibility is a separate correction.
Sovereign-AI gap allocations were published without the figures needed to reproduce them
What the site showed
The Inflation view assigned percentages of each programme’s announced-versus-disbursed gap to named causes, classified programmes as having or not having a gap, and summarised those allocations across the portfolio. Source links and confidence labels accompanied the allocations.
Correction
The archived manifest does not contain the announced, disbursed and cause-specific amounts needed to reproduce those percentages or the gap classifications. The quantitative matrix and its summaries are withdrawn. The separate programme records, government-budget comparisons and funding ledgers remain available; the archived allocations and their citations are retained for review. No replacement percentages have been estimated.
Why it happened
The review checked that the drafted percentages summed to 100 and carried citations, without checking the underlying amounts or whether those citations supported the allocations.
Changes made
The data loader withholds the archived allocations. Restoring them requires compatible amounts, units and dates, a reproducible calculation, and an independent reading of the cited sources. Automated checks verify that the supporting amounts are present and add up; editorial review must establish that the sources support them. The correction record and methodology retain the withdrawal notice; the former Inflation URL redirects to the programme register. The correction also covered the former Sovereign AI Brief, retired in September 2026.
Changing the stress-index weights recalculated the chart while downloads continued to use the original equal-weight history. With every weight set to zero, the historical composite reverted to that original series. Rounded score columns also prevented some downloaded composite values from being reproduced exactly. A fixed ten-per-cent input perturbation was presented as a confidence assessment without an uncertainty model.
Correction
Downloads now include the selected normalized weights, their raw sum and the component scores at full precision. These reproduce the displayed one-decimal composite. Setting every weight to zero removes the composite and its exports; the four component series remain available. The unsupported confidence assessment was removed. No confidence interval is calculated.
Why it happened
The chart and the download read different series; with every weight at zero a fallback restored the original equal-weight history; and the exported scores were rounded past the precision the composite needs.
Changes made
The chart and the download now read the same recalculated history and require a positive sum of weights; the export carries the inputs the composite reproduces from and names the source series.
Supply-chain records kept figures their new public citations did not support
What the site showed
Supply-chain records relabelled to cite public relationship evidence retained numerical attributes from an earlier import. Scores, annual values and price correlations could consequently appear beside a public citation that did not establish those quantities. A sole-source flag left at the importer’s default of no, and an empty list of alternative suppliers, could also read as verified statements about sole sourcing and alternatives. The values affected graph details, ranking and country aggregates.
Correction
The affected relationships and public citations remain available. Their inherited scores, shares, values, correlations, volumes, lead times, sole-source flags, alternatives and notes are withheld by both the application data loader and the public database view. Unknown fields remain empty, and raw records are preserved. Records from other sources retain their existing publication rules; their individual fields were outside this correction’s source review.
Why it happened
The import changed provenance labels and URLs without checking the accompanying attributes against the new citations. Graph details and aggregates then treated the public citation as support for the entire record. A public database view carried the same inherited attributes.
Changes made
The application loader excludes licensed-source rows and withholds unsupported attributes before combining or ranking records, including in country and bilateral totals. Cascade simulations identify their default values as model assumptions. On 6 September 2026 the public view received the same filters. Raw records were unchanged. A withheld field returns only after its value is verified against an independent source.
Core Scientific’s 26 February 2025 release gives approximately 260 MW of critical IT load for the Denton site. This cluster record repeats that site capacity. Totals use the 262 MW site record, with both figures available in the campus comparison. The cluster’s 100,000 accelerators remain in compute totals.
This is a second record of Google’s Papillion data center. Capacity totals use the 237 MW site record; the 250 MW figure remains in the campus source comparison.
The Prometheus cluster and the New Albany campus describe the same site at different scopes. Capacity totals use the campus record’s 631 MW; the cluster’s 1,281.28 MW estimate remains in the source comparison. Its 500,000 accelerators remain in compute totals.
The company writes its initialism as “BNY”; the stored name used “Bny”.
Primary source
BNY corporate site, page title "BNY | Global Financial Services" (read 2026-09-03). SEC EDGAR carries the registrant as "Bank of New York Mellon Corp" (CIK 0001390777). · https://www.bny.com/
The agency writes its initialism as “CSIRO”; the stored name used “Csiro”.
Primary source
CSIRO site, page title "Commonwealth Scientific and Industrial Research Organisation, Australian Government - CSIRO" (read 2026-09-03). · https://www.csiro.au/
