AI Lattice / Corrections
Corrections
Changes to published facts are logged here with a date and a reason. Earlier versioned snapshots keep their original contents.
Since publication
v0.2.1 (October 30, 2026) corrected nine items on the money page after an independent audit of the v0.2.0 figures, v0.2.2 restated the big-four cash gap, v0.2.3 revised the headline verdict after a review of the argument, v0.3.0 revised it again on the results of the stress test, and v0.3.1 narrowed it after a review of the stress test; each is logged below. v0.1.1 added a missed development, Meta's Muse agent app, to the default-assistant chokepoint and the timeline; see the release note. v0.2.0 added the money layer; see the release note.
Correction log
- v0.5.9 Two chip-backed loans are now credited to the lenders their sources name instead of the category 'Private credit funds': the up to $35 billion special-purpose company for Anthropic to Apollo with Blackstone (flow F136, link L155), and the up to $12.5 billion GPU loan for xAI to Apollo with Diameter Capital (flow F137, link L156). Amounts, dates and sources are unchanged; each lender's share is not disclosed, so amounts are not split. Diameter Capital is added to the map. Blackstone and Diameter each get a link to the same flow, and flows.csv gains a co_funders column. Where a source names the funder, the map credits that funder and any co-lender it names, not a generic category.
- v0.5.9 Two entries leave the map because no sourced link, money flow, power record, permission or front door connects them to anything else on it: 'Private credit funds', whose two loans now sit with the named lenders (its sector-wide figures stay on the money page), and Mitsubishi Heavy (Power), which stays named in the gas turbine chokepoint's sourced facts. Their pages redirect to the money page and the chokepoint. Sourced links were added for three companies that had none: GE Vernova (gas turbines for Crusoe's AI data centers), Eaton (800-volt power design with Nvidia) and ByteDance (a reported plan to buy Huawei Ascend chips). No company stays on the map with zero connections. Each release is now checked for this, and the build stops if one is found.
- v0.5.6 SoftBank on the money page, stress test exposure table and data files: the $10 billion loan secured on SoftBank's OpenAI shares is now shown as drawn in August 2026, citing SoftBank's financial report for the six months to September. The borrowing chart shows $69 billion, now labeled as historical borrowing including refinancing. The $40 billion bridge loan is now labeled as the facility limit, dated March 27, 2026, cites SoftBank's execution notice and financial report, and is marked closed. The loan was described as being arranged after SoftBank reported it drawn. The bridge loan record carried a March 6 date and cited an April press story about a different loan; the $69 billion was not labeled as including money later refinanced.
- v0.5.6 Oracle answer and money page: the single 'about $72 billion' partner financing row is now four rows by site: Texas, a $22.75 billion construction loan reported closed; Wisconsin, a $14.75 billion loan facility; New Mexico, about $18 billion of bank loans reported secured; Michigan, about $16 billion of bonds and equity closed in April 2026. Oracle's new-contract statement again includes 'or similar structures'. The single total mixed loan facilities, loans reported secured and a bond-and-equity package of different dates, and could be read as a funded total or as Oracle debt. It is neither. The earlier wording also dropped part of Oracle's statement.
- v0.5.6 Cloud backlogs answer: Anthropic's Microsoft and Google contracts now use the minimums in its public S-1 (at least $44.9 billion and about $114.5 billion), and Amazon's $578 billion backlog cites Amazon's 10-Q. The page no longer gives a current share range or calls our count a floor; it says about half as of The Information's May 2026 report and shows contracts and backlogs side by side. Anthropic's suppliers outside the four clouds again include smaller GPU rental firms, Nscale and Lambda, citing the S-1. The page used $31.4 billion and $111.1 billion from reporting on Anthropic's confidential draft, and a secondary source for Amazon. Gross multiyear contract values divided by a single date's remaining backlog cannot set an upper or lower bound on the share. The S-1 lists Nscale and Lambda.
- v0.5.6 OpenAI answer, key findings and signals: the October 30 report is now described as the FT reporting that OpenAI reduced its fundraising target by $10 billion, to about $20 billion, at a reported $1.25 trillion valuation. Wording that investors reduced or withdrew commitments was removed, as were details of the first close and a UAE investor. The FT reports on the target and on revenue now cite Reuters' accounts; the November wire still cites Bloomberg. The Reuters account of the FT report supports a reduced target, not investor withdrawals, and the removed details came from a copy we could not read in full.
- v0.5.6 Nvidia answer: the July 2026 reports are now described as Bloomberg's account of a $750 billion-plus web of AI financing linking Nvidia, OpenAI, CoreWeave, Oracle and Microsoft, and The Wall Street Journal's report of talks on a guarantee of up to $250 billion and on chip purchases expected to cost about $350 billion. The guarantee is now described with the two triggers in Nvidia's filing: OpenAI's insolvency resulting in a lease default, or its failure to make lease payments, and pays the shortfall between a guaranteed minimum lease value and what a replacement lease or sale recovers. The page called the $750 billion a total of Nvidia's own financing talks and the $350 billion chip financing. The source describes the first as a web of financing across several companies and the second as the expected cost of the chips. Nvidia's filing names two specific triggers; the page had given a looser list including OpenAI failing to perform.
- v0.5.6 Chip packaging and memory answer: Samsung is now described as expected to supply HBM4 for AMD's MI455X; CXMT's HBM3E yield as reported at 20 to 30 percent; Samsung's HBM4 profit lift as reported by Korean press; Micron's $100 billion as minimum contracted revenue under its strategic customer agreements; the CoWoS shortfall as a supply-demand gap of about 20 percent; and the 52 to 78 week lead times as trade reports from late 2025. The same changes were made on the map. The sources do not call Samsung the sole supplier, give a 20 to 30 percent yield range, present the Samsung profit link as an inference, and attribute the lead times to trade reports that TSMC has not confirmed.
- v0.5.6 Oracle answer and money page: the April 2026 financing of the Michigan site is now described as a $16.3 billion package, mostly bonds, of which Pimco bought about $10 billion. The page called the whole $16.3 billion a debt package anchored by Pimco. It included about $2 billion of equity, and Pimco bought about $10 billion of a bond tranche of about $14 billion.
- v0.5.6 SoftBank answer: the $45 billion is now described as SoftBank's total investment gain on OpenAI, from valuation; the loan-to-value policy is now stated as below 25 percent in normal times with a 35 percent upper limit, citing SoftBank's 2026 report; the bridge-loan row now says the $30 billion drawn was fully repaid by September 15, 2026 and the $10 billion undrawn was canceled effective September 30, citing SoftBank's financial report for the six months to September; the loan-to-value ratio is now given at both June 30 (13.0 percent) and September 30 (15.9 percent). The cited report gives the $45 billion gain but does not call it unrealized. The 25 percent figure is a normal-times policy, not a ceiling. The page cited only the September 9 prepayment notice. SoftBank's November 10 financial report confirms the full repayment and the cancellation, and its November 10 results give the September 30 ratio. The money page's chart counted the full $40 billion limit as borrowed; it now counts the $30 billion drawn.
- v0.5.6 OpenAI answer: the reported November wire and the possible $20 billion round size now cite Bloomberg's own report, with Reuters' separate report of the close. Microsoft's June backlog figure on the cloud backlogs answer now cites its source. Map: OpenAI's entry now says it closed a $122 billion round of committed capital in March 2026, $110 billion of it announced in February. The Reuters copy previously cited does not mention a wire or a $20 billion total. The OpenAI entry said it raised $110 billion, while its March round was $122 billion of committed capital, not all documented as paid in.
- v0.5.5 The $11 billion of OpenAI's current round is now described everywhere as a reported wire based on unnamed sources, not a confirmed payment. The model still counts it; without it, the 2028 need is about $92 billion on plan and $118 billion with front-loaded spending. Bloomberg reported the wire on November 6, citing people familiar with the matter, and Reuters also reported the close that day, citing its own source. OpenAI has not confirmed it. Several places on the site stated it as money raised.
- v0.5.5 Power page and record: Oracle's New Mexico force-majeure notice is now attributed to press reports of September 24, 2026, which said it cited potential delays in securing power, with Oracle's response that the notices did not establish a delay and that the project remained on its planned schedule. The statement that the gas pipeline is delayed was removed from the site and event records. The power page now keeps Project Jupiter apart from projects already delayed: it is a schedule risk, Bloomberg reported the notice could let Oracle defer payments, and Reuters reported that Oracle cannot terminate the lease. The notice is not published. The earlier wording stated it as fact, dated the event September 27 from a later write-up, and cited a site directory page that predates the notice.
- v0.5.5 OpenAI's new-money need by the end of 2028 changed from about $40 to 95 billion to about $81 billion on its own plan and $107 billion with front-loaded spending; the half-growth case by 2030 from about $140 billion to about $184 billion; the stall case from about $290 billion to about $331 billion. The headline, key figures, key findings, questions and the OpenAI answer page now use these figures. The stress test treated OpenAI's $122 billion March round as fully paid. OpenAI announced it as committed capital. The model now counts only the parts with a dated payment record: Amazon's $50 billion (Amazon's 10-Q) and SoftBank's three $10 billion tranches (SoftBank's releases), $80 billion in all. Nvidia's $30 billion and about $12 billion from others have no documented payment and are switched off; the sensitivity table shows the need if they have been paid. Page figures now come from one model output, rounded once: the previous headline said $40 billion where the OpenAI answer table said $39 billion.
- v0.5.5 OpenAI's share of Oracle's backlog removed from the headline answer, key figures, questions, company summary and answer pages. Oracle is now described by its own filing figures: $664 billion of contracted backlog and $288 billion of signed data-center leases not yet started. Oracle does not disclose customer shares. The 45 percent was our arithmetic on a press-reported contract, and the 'roughly half' estimate comes from a trade write-up that names no source. The source trail keeps both the reported contract and the trade estimate, labeled as outside estimates. The statement that Oracle's non-OpenAI backlog more than doubled, which rests on the same unsourced split, was removed.
- v0.5.4 OpenAI's new-money need on its own plan changed from about $60 to 105 billion by the end of 2028 to about $40 to 95 billion; the half-growth case by 2030 from about $320 billion to about $140 billion; the stall case from about $470 billion to about $290 billion. The headline, key findings, questions and the OpenAI answer page now use these figures. An independent audit found the stress test treated 80% of planned compute as uncuttable in every year, but much of the $688 billion of signed contracts falls after 2030. The floor is now each year's payments on the signed contracts, spread over their reported terms (about 64% of plan over 2026 to 2030). 2026 burn now stays at the February plan's figure, since OpenAI burned $3.7 billion in the first quarter. The model also counts the $11 billion of OpenAI's current round wired by November 6. The old method is kept as a sensitivity row.
- v0.5.4 Supplier money put into the labs and neoclouds changed from about $93 billion firm (up to $242 billion) to about $122 billion firm (up to $236 billion). Amazon's own 10-Q shows its second $35 billion in OpenAI was paid by the third quarter, so it is now firm. Nvidia's $6.3 billion obligation to buy unsold CoreWeave capacity is a purchase commitment, not an investment, and no longer counts toward the investment total; it stays in the loops table.
- v0.5.4 OpenAI's share of Oracle's backlog changed from 'roughly half, about $330 billion' to 'its reported $300 billion contract, about 45% (our arithmetic)', with press estimates of roughly half noted. The $330 billion was half of the backlog, taken from a trade write-up that gives no source. Oracle does not report the split.
- v0.5.4 The verdict no longer says the big four pay for 2026 data centers 'most of it from their own cash' without qualification; it now says they spent about $29 billion more than their businesses produced in July to September and are covering the gap with bonds and leases. The earlier wording contradicted the page's own figures.
- v0.5.4 Cloud backlogs answer: Google Cloud's backlog updated from $514 billion (June) to $560.8 billion (September 30); total from about $2.43 trillion to $2.48 trillion; the labs' share from about 39% to 38%. Microsoft's $677 billion now cites its 10-Q. Claims that OpenAI and Anthropic were about half of AWS's backlog and over 80% of its first-quarter increase were removed. Alphabet's 10-Q for the third quarter was filed October 29. The AWS-specific claims are not in the cited source, which reports about half across the four clouds combined.
- v0.5.4 Oracle answer: the list of other customers (AMD, Meta, Nvidia, TikTok, xAI) was removed, and Blue Owl's December 2025 refusal to fund the Michigan site now notes the $16.3 billion financing that closed in April 2026. Neither cited source names those customers. The Blue Owl line was out of date.
- v0.5.4 Nvidia answer: the reported talks on a $250 billion OpenAI data-center guarantee are now marked as overtaken by the signed $105 billion guarantee on the same Ohio campus; the guarantee terms are restated from the filing (set portions of rent and power, ending 20 years after each lease starts or earlier on a satisfactory OpenAI credit rating); first-half gains on stakes are now stated as before tax, against after-tax net income; Nvidia's filed $25 billion of unpaid equity commitments added. The July talks and the August guarantee concern the same SB Energy campus. 'The leases run 20 years' and 'about 20% of net income' did not match the filing.
- v0.5.4 OpenAI answer: the end-2025 cash figure ($40 billion) now cites its source; the $50 billion and $70 billion revenue reports are explained rather than called conflicting; the November 6 close of about $11 billion added. The cash figure was uncited on this page. The FT explained the gap between the two revenue figures on October 8.
- v0.5.4 Ledger: the Amazon, Nvidia and SoftBank parts of OpenAI's round are marked as included in the $122 billion total; OpenAI's record now shows the $250 billion Azure commitment and aligns its Cerebras note with the ledger; SoftBank's $4.6 billion gap between its two rows and its reported $64.6 billion total is noted. Adding OpenAI's funding rows double-counted about $110 billion, and two records disagreed with the ledger.
- v0.5.4 Packaging and memory answer: the statement that Micron can meet only 50 to 67% of key customers' demand is now attributed to a JPMorgan analyst. It is the analyst's reading after meeting Micron's management, not Micron's statement.
- v0.5.4 Power page: a note under the site table now says how many sites report computing load, total site power, or neither. Site totals add different measures, which can differ by a fifth to a half.
- v0.4.2 Power page: firm chip and cloud deals cut from about 14 GW to about 9 GW; verdict reworded to 'developers whose financing depends on on-time delivery are most exposed, while well-funded buyers with secured power or flexible grid connections have more options'. AMD's SEC filing shows OpenAI's binding commitment covers only the first 1 GW of its up-to-6 GW deal. A final review also found the earlier verdict predicted the order of failures more firmly than the record supports.
- v0.4.2 Power page: Google Cheyenne's computing load shown as not disclosed; deal totals no longer set against running site capacity; four October 2026 policy and financing events added; Homer City's first-power date shown as disputed; its first gas turbine set on its foundation (developer, October 20, 2026) added. 1.8 GW was Crusoe's figure for its five-building design; Google's redesign has four buildings and no published size.
- v0.4.1 Power page verdict changed to 'Power is the main brake on US AI through 2028. Much of the build-out will be late; buyers with cash and their own power will adapt, and the failures will start with developers whose financing assumes on-time delivery.' An independent review found 'power, not money' and 'slip, not collapse' stronger than the record supports: the 32 GW gap is Morgan Stanley's scenario, not a measured ceiling, and delay can turn into cancellation when payments continue and contracts allow customers to leave. The claim that delay pushes OpenAI's funding test out by a year or two was removed; it was not modeled.
- v0.4.1 Power page: removed the ratios comparing gigawatt-sized deals with the selected sites; firm deal total cut from about 16 GW to about 14 GW. The ratios mixed chip and cloud deals, computing load and grid draw, and global deals with a selection of campuses. AMD's deal with Anthropic is up to 2 GW, not firm.
- v0.4.1 Power record corrected: Stargate New Mexico changed its planned power supply to up to 2.45 GW of Bloom fuel cells, not 2.2 GW of gas turbines; Calvert Cliffs split into 190 MW of new nuclear capacity and 500 MW of existing output; Crane received a FERC waiver in June 2026 to transfer 760 MW of grid-connection rights; Stargate Norway's capacity went to Microsoft; Beacon Point's reported tenant is Nvidia, with Lambda reselling about 350 MW to Anthropic; Nscale Monarch shown as Anthropic's 460 MW share; Homer City's gas plant is being built; Crusoe's 'paused' Cheyenne project is the same campus as Google's Project Tembo and is now one row. Each was checked against the company or regulator's own announcement after an independent review.
- v0.4.1 Power record expanded: OpenAI's 4.25 GW PORTS-Pike campus in Ohio, Microsoft's 2 GW Pecos campus, Google's Cheyenne campus and three running Google campuses, Amazon's GW Ranch and Homer City campuses, Constellation's nuclear deals with Google, the Black Hills supply for Google, Crusoe's 900 MW plant for Microsoft at Abilene, and a dated section on regulatory and federal financing changes. The site list is a selection of large AI campuses and is now labeled that way; its total is no longer used as an industry-wide comparison.
- v0.3.3 Google DeepMind (Gemini) added as a model lab on the map; the money page now says which frontier labs depend on outside money. Google appeared only as a cloud company, which hid that it is also a frontier lab. The page had described the frontier as OpenAI and Anthropic; Google, Meta and xAI also train frontier models but are funded by their parent companies.
- v0.3.2 SoftBank's bridge loan removed as an open risk; the exposure table now shows the bonds and loans that replaced it. SoftBank's own September 9, 2026 announcement shows it repaid the remaining $25.9B of the bridge loan on September 15. The page had still listed refinancing the bridge loan as something to watch.
- v0.3.1 Headline verdict changed to 'Not a house of cards. The demand is real, but the riskiest part rests on one company: OpenAI needs $60 to 105 billion of new money by 2028 just to stay on plan.' An independent review found 'the whole structure rests on one wall' overstated what the stress test shows: it tests two labs, not the whole industry, and a funding need is not a funding failure.
- v0.3.1 Stress test: OpenAI's stall-case need raised from about $370B to about $470B by 2030, and the half-growth case from about $250B to $320B; Anthropic's harshest case now leaves about $18B at end-2028, not $35B. Version 0.3.0 let all of OpenAI's and Anthropic's non-compute spending fall with revenue. Half is now treated as fixed. Sensitivity rows for compute floors, borrowing costs and the old method were added.
- v0.3.1 Anthropic's 2026 free cash flow in the stress table and stress.csv corrected to the second half only. The table was labeled second half but showed the full year; the model's cash totals were already right.
- v0.3.1 Exposure table relabeled as gross exposure with no order of losses; SoftBank's bridge loan shown as up to $30B; Nvidia's guarantee described as paying only the shortfall after recovery. The previous wording implied an order of who loses first that the record does not support. $10B of the $40B bridge loan was cancelled undrawn.
- v0.3.0 Headline verdict changed to 'Not a house of cards. The demand is real, but the whole structure rests on one wall: OpenAI's ability to keep raising money.' A year-by-year stress test was added. v0.2.3 said the verdict was not stress-tested. The stress test finds Anthropic can largely meet its commitments from cash on hand, needing its IPO money only in the harshest case, while OpenAI needs new outside money by 2028 even on its own plan, so the risk sits on one company's fundraising.
- v0.2.3 Headline verdict changed to 'Not a house of cards, but not stress-tested either. The demand is real; the financing is narrow and increasingly borrowed.' An independent review of the argument found the old headline treated the big four's survival as proof the whole system is safe, and read concentrated contracts as narrow demand. The page has no year-by-year payment test yet, so the headline now says so.
- v0.2.3 The $657B of Anthropic compute commitments is now attributed to Fortune's tally, not to the S-1 itself. Anthropic's S-1 reports $265.6B of fixed or minimum commitments as of June 30, 2026 and leaves the total blank. $657B is Fortune's sum of the deals named in the filing. The page had presented it as the filing's own number.
- v0.2.3 '56% of US businesses' now reads '56% of businesses on Ramp's payment platform'. Ramp's sample is about 85,000 of its own customers, not a national survey.
- v0.2.3 'Has no buyer' changed to 'has no signed buyer today', with the end-customer numbers behind the two labs added. Two concentrated buyers is not the same as narrow demand; OpenAI and Anthropic report tens of millions of paying users and customers between them.
- v0.2.2 Big-four cash gap for July to September 2026 changed from $27B to $29B, with the $47B gap excluding Microsoft added. Amazon's and Meta's finance leases are now included in capex, from their own releases. Microsoft's finance leases are still missing from public sources we could open, so the true gap is larger.
- v0.2.1 Headline figure for lab compute commitments changed from $1.3T to $1.2T firm ($1.35T including cancellable contracts). The $1.3T added OpenAI's firm contracts to Anthropic's $657B S-1 total, which includes cancellable deals. Firm-only is the right comparison.
- v0.2.1 Supplier money into labs and neoclouds raised from about $85B to about $93B. Amazon's first $8B in Anthropic (2023 to 2024) was missing from the ledger; added as flow amazon to anthropic.
- v0.2.1 OpenAI's share of Microsoft's backlog now dated: 45% as of December 2025. Microsoft's backlog has since grown, and Microsoft says the growth came from customers outside the big labs, so the current share is probably lower. The figure was shown without a date.
- v0.2.1 'Two to three times a year' revenue growth changed to 'roughly doubles every year through 2028'. The labs' own projections (OpenAI about $36B in 2026 to $350B in 2030; Anthropic about $190 to 200B in 2028) imply roughly 1.8x a year, not 2 to 3x.
- v0.2.1 'Roughly forty times the next buyer' changed to 'roughly twenty times'. The next buyer after the two labs is Meta at $36B firm; $688B divided by $36B is about 19, not 40.
- v0.2.1 SpaceX bar on the 2026 capital spending chart changed from a half-year figure to a full-year estimate. The chart compares full-year numbers; the SpaceX row showed July to December only.
- v0.2.1 SoftBank's borrowing shown as $69B firm plus a $10B loan being arranged, not $79B. The $10B loan secured on OpenAI shares was reported as being arranged, not completed.
- v0.2.1 CoreWeave's $3.5B April 2026 convertible restored to the ledger as its own row. It had been merged with the September convertible and cited to the wrong press release, then cut in review. It is supported by CoreWeave's April 10, 2026 release.
- v0.2.1 Amazon's 2026 capex and negative free cash flow now cite Amazon's own earnings release. They had been cited to a secondary write-up; the primary release confirms $220B expected cash capex and -$42.4B of trailing free cash flow.
Before publication
Before release, an independent reviewer checks every changed figure and claim in this update against its cited source.
Report an error
Found something wrong? Contact the editor through mikeye.com with the link ID or page.