The record shows that the AI build-out has passed the fibre build-out in spending relative to US GDP and repeats its supplier-funds-customer financing, while its largest asset class (accelerators) loses value within a few years, so any cheap inheritance from an AI repricing will lie mostly in grid connections, sites and buildings, and open-weight models, and the rules that would decide who acquires them are, as of 7 November 2026, proposals.
Overview
Between about 1996 and 2001, new US and European carriers borrowed heavily to bury long-haul optical fibre. Equipment makers lent their customers money to buy the gear. A forecast that internet traffic doubled every 100 days, which Kerry Coffman and Andrew Odlyzko of AT&T Labs wrote “appear to trace back to statements of John Sidgmore of MCI WorldCom’s UUNet or his colleagues,” justified the spending. Traffic was in fact doubling about once a year. When credit closed in 2001, dozens of carriers failed, and an often-quoted FDIC estimate from June 2001 held that only 2 to 5 per cent of installed fibre was lit [TBV, traced to a secondary citation of the FDIC]. The fibre cable lasted. Over the following decade it was bought cheaply, lit with new equipment, and carried broadband, video, cloud computing and, under secret legal authority, state collection of internet traffic.
Since 2023, five US companies (Microsoft, Alphabet, Amazon, Meta and Oracle) and the frontier labs have committed spending on data centres that Torsten Slok, Apollo’s chief economist, put at 1.4 per cent of US GDP in 2025, writing that “the consensus expects hyperscaler capex to run at roughly 3% of GDP every year from 2027 to 2029,” against a telecom peak that “topped out at 1.2% of GDP in 2000” [IC]. Financing has shifted from cash flow toward debt during 2026. In the first week of November 2026, the yield on CoreWeave’s 9.625 per cent bond due 2032 climbed above 16 per cent, according to Deutsche Bank data cited by CNBC on 6 November; Bloomberg reported on 5 November that ING was selling a dozen data centre loans worth about US$500 million, including a US$25 million slice backing a Stargate site in New Mexico; and CNBC reported that credit default swaps on Nvidia debt were up 76 per cent in a month, to roughly 139 basis points [IC].
The difference that matters most for inheritance is physical life. A fibre strand laid in 1999 still carries traffic. An AI accelerator bought in 2024 is depreciated over five or six years in hyperscaler filings. Michael Burry of Scion Asset Management, who disclosed bets against Nvidia, argued in November 2025 that its economic life is two to three years and that “understating depreciation by extending useful life of assets artificially boosts earnings - one of the more common frauds of the modern era” [IC for the filings; the critique is an interested party’s view]. What a repricing leaves behind that someone else could buy cheaply is therefore mostly the slow layers: interconnection agreements, substations, buildings, cooling plant, water permits, and trained open-weight models.
The brief tests one organising claim: that infrastructure built in a speculative installation period is repriced when the speculation ends, and the uses it serves afterward depend on who acquires it cheaply and under which institutions. The claim draws on Carlota Perez’s account of technological revolutions and on the wiki’s S-curve thesis. For fibre it holds well, with gaps in the record on shares. For AI the evidence so far supports the first half (financing stress and repricing in credit markets) and is silent on the second, because no distressed sale of AI capacity has yet occurred. The early November 2026 snapshot records the week the repricing began.
Right of reply
This page describes companies, governments, agencies and people only through their filings, court records, published statements and recorded actions. None has been invited to comment. Corrections and responses will be recorded on this page. On this site, [MS] marks a claim that rests on the account of the party the claim is about or on the wiki’s own project documents.
Conflict of interest
This brief was drafted by Claude, an AI model made by Anthropic, a frontier AI company whose finances, government contracts and investors fall inside the subject of this page. The Provenance section sets out the conflict and how it was handled.
The fibre build-out and what it left
Scale. US telecom-sector capital expenditure peaked at roughly US$120 billion in 2000, about 1.2 per cent of US GDP (Meridian Report, using sector capex over 2000 GDP; Apollo uses the same 1.2 per cent peak) [IC]. A 7GC analysis gives an inflation-adjusted peak of about US$213 billion and more than US$500 billion spent from 1996 to 2000 [MS, investor analysis]. Telecom equipment spending grew about 17 per cent a year from 1995 to 2000, then fell 24 per cent in 2001 [TBV, secondary]. In February 2001, America’s Network reported “600,000 miles of new inter-city fiber is on the way,” quoted by the Federal Reserve Bank of Richmond’s 2003 review, which found “massive overcapacity in long-haul fiber” and noted that “the glut was exacerbated by the dramatic advances in technology for increasing the capacity of each strand of fiber” [IC].
How much was dark. Three figures circulate. An FDIC assessment from June 2001 that 2 to 5 per cent of installed fibre was lit is cited in commentary; a Wall Street Journal figure of 2.7 per cent of fibre in use in 2002 is cited by Fabricated Knowledge; and investor Gavin Baker is quoted saying “at the 2000 peak, 97% of America’s fiber was dark” [all TBV; the FDIC and WSJ originals were not reached]. The Richmond Fed’s point about wavelength capacity matters for reading them. Dense wavelength division multiplexing multiplied what each lit strand could carry, so much of the glut was unused wavelengths on lit fibre, and a strand count overstates idleness in one sense and understates it in another.
Forecasts. Coffman and Odlyzko showed that “one of the great inspirations for the Internet bubble, the myth of ‘Internet traffic doubling every 100 days,’ was false”; their data showed traffic “approximately doubling each year” [IC]. Odlyzko traced the claim to UUNET and WorldCom officials, naming Bernie Ebbers and John Sidgmore among those who repeated it as late as 2002, and found it “was often a clincher for new venture business plans.” He also recorded that doubling every 100 days “was approximately true in 1995 and 1996” [IC, Odlyzko’s papers]. By 2009 he reported wireline traffic growth had slowed to “a range of 50-60% per year” [IC]. The forecast had the three features H3 tests: one interested source, extrapolation from an early growth rate, and repetition without independent measurement.
Financing. By fiscal 2000 Lucent “had entered into agreements with customers to provide up to $8.1 billion in credit or loan” and took bad-debt provisions of US$2.2 billion in 2001 and US$1.3 billion in 2002, according to William Lazonick and Edward March’s business history of Lucent [IC]. Lucent offered WinStar Communications up to US$2 billion and wrote off US$700 million in bad debts after WinStar’s bankruptcy in April 2001 [IC]. Nortel reported US$1.6 billion of vendor financing on its books at the end of 2000 and US$2.1 billion of total customer financing at the end of fiscal 2001 in its 10-K, as reported by The Globe and Mail [IC]. A finance professor’s count of SEC filings, reported by Newsweek, found bad loans rose from 2.6 to 60 per cent of Lucent’s customer loans and from 25.5 to 80 per cent of Nortel’s between the end of 2000 and the end of 2001 [TBV]. A McKinsey estimate of about US$25.6 billion of vendor financing across nine suppliers at the end of 2000, and a Cisco figure of about US$2.4 billion, appear in secondary sources only [TBV]. The role of analysts, including Jack Grubman, and the 2003 Global Analyst Research Settlement were outside what the research run could document [TBV].
Accounting. WorldCom disclosed in June 2002 that it had improperly booked about US$3.8 billion of line costs as capital spending, and the total later grew (CNN Money; Light Reading) [IC]. Global Crossing and Qwest booked revenue from capacity swaps under indefeasible rights of use (IRUs); the court and SEC records weren’t reached, so the swaps carry [TBV].
The crash. Telecom stocks lost on the order of US$2 trillion of market value from 2000 to 2002, by widely cited estimates [TBV]. 360networks skipped a debt payment in early June 2001 and filed for bankruptcy on 28 June 2001 [TBV, single secondary source]. More than 60 telecom companies failed between 2001 and 2003 [TBV]. Global Crossing filed for Chapter 11 on 28 January 2002, in the Southern District of New York and in Bermuda, and WorldCom filed on 21 July 2002, listing about US$107 billion of assets, the largest US bankruptcy to that date (eWeek; CNN Money; Light Reading) [IC]. Sale prices of networks in bankruptcy as a share of build cost were not found in a primary record.
Who bought, and what the fibre carried. Commentary describes Level 3 acquiring distressed fibre assets “at prices that would have been inconceivable during the mania” [TBV]. Reports of Google buying dark fibre from about 2005, the acquisitions by research and education networks such as National LambdaRail, municipal networks and the state laws restricting them, the AT&T Folsom Street facility disclosed in 2006, and the NSA Upstream programme disclosed in 2013 all bear on H5, and their primary records were not reached [TBV]. The comparison with Africa (SAT-3/WASC wholesale prices before 2009; SEACOM and EASSy from 2009) is also [TBV] here.
Institutions. Commentary attributes the speed of the build-out partly to the Telecommunications Act of 1996, under which cities had to grant rights of way; the Seattle Times reported in 2002 that three carriers trenched the same boulevard in Bellevue [TBV, secondary]. The open-access and unbundling rules of the 1996 Act, their later rollback, universal service funds and net neutrality decided the terms under which cheap capacity was sold. No primary evidence was obtained on how those rules affected shares of fibre acquired, so H8’s fibre half rests on structural inference.
The AI build-out to date
Scale. Apollo’s chief economist Torsten Slok put hyperscaler capex at 0.3 per cent of US GDP in 2019 and 1.4 per cent in 2025, with consensus at about 3 per cent a year from 2027 to 2029, “more than twice the peak of the telecom and fiber buildout of the late 1990s, which topped out at 1.2% of GDP in 2000” [IC]. In February 2026 Apollo put 2026 hyperscaler capex at about US$646 billion, about 2 per cent of US GDP [IC]. Company guidance tracked by later analysts rose to about US$780 to 800 billion for the five largest, with Amazon at US$220 billion and Alphabet up to US$205 billion [TBV, analyst compilations]. Meridian Report cautions that the ratio is “mechanical”: the numerator is global company spending and the denominator is US GDP, while the telecom figure was US-sector spending [MS, analyst caveat]. Goldman Sachs, using a narrower definition in late 2025, put AI capex at 0.8 per cent of GDP and said hyperscaler capex would need to reach US$700 billion in 2026 to match the telecom peak [IC]. On either definition, 2026 guidance passes that mark.
Financing. In September 2025 Nvidia signed a letter of intent to invest up to US$100 billion in OpenAI, released as each gigawatt of Nvidia systems was deployed, against at least 10 gigawatts [IC]. The investment that closed was US$30 billion, in OpenAI’s February 2026 round, and Nvidia’s chief executive later indicated the larger figure was unlikely [TBV, two secondary sources agree]. AMD granted OpenAI warrants for up to 160 million AMD shares, about 10 per cent of the company, vesting with deployment of 6 gigawatts of AMD chips [IC]. OpenAI agreed to buy about US$300 billion of compute from Oracle over about five years from 2027; Oracle’s remaining performance obligations rose to US$455 billion and later to US$664 billion [TBV for the later figure]. Reported OpenAI commitments also include US$250 billion to Microsoft Azure and US$22.4 billion to CoreWeave; figures for Amazon Web Services conflict between US$38 billion and US$138 billion across sources, and this brief does not choose between them [TBV]. Analyst compilations total about US$1.15 trillion of OpenAI commitments for 2025 to 2035 [TBV]. A Korean semiconductor analysis described “four distinct circular structures: supplier equity stakes, reverse equity paid to the customer (the AMD warrant), supplier backstops of unsold capacity, and equity-free concentration (the Oracle contract)” [MS, analyst framing].
Debt carries a growing share. Goldman Sachs credit strategists said 27 per cent of hyperscaler capex was debt-financed in 2025, expected about one-third in 2026 and 35 per cent in 2027, and noted that “17 out of the 23 sort of data center JV deals are now trading wide to originated yields” [IC]. Bank financings named in reporting include Société Générale’s US$7.1 billion for the first Stargate site, US$38 billion from MUFG and JPMorgan for two Oracle projects, and US$18 billion from SMBC for an Oracle project in New Mexico [TBV, reported by The Information via an aggregator]. The same report said CleanSpark, a bitcoin miner building data centres for Meta, sold US$2.3 billion of high-yield bonds at 98.5 cents with a 7.875 per cent coupon [TBV].
Accounting. Amazon extended the useful life of its servers from three to four years in 2020, four to five in 2022, and five to six in 2024, the last change reducing 2024 depreciation and amortisation by US$3.2 billion; effective 1 January 2025 it shortened a subset back to five years, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning” [IC, Amazon 10-K filings as quoted by several analyses]. Sources quote different effects for the 2025 change (an anticipated US$0.7 billion reduction in operating income, and a US$1.4 billion increase in depreciation in the FY2025 10-K); the gap may reflect estimate and outcome and was not resolved. Microsoft’s filing for the year to 30 June 2026 gives servers and network equipment lives of 2 to 6 years [TBV, secondary reading of the 10-K]. Michael Burry, who disclosed bets against Nvidia and Palantir, argued in November 2025 that hyperscalers overstate accelerator lives, which he put at two to three years, and estimated US$176 billion of understated depreciation over 2026 to 2028 [IC that he said it; the claim itself is contested]. An accountant quoted by CNBC replied that lives must be justified to auditors with engineering data [IC]. This brief reports no wrongdoing. It reports that the life assumption moves earnings by billions of dollars and that the companies have moved it in both directions.
Demand evidence. Allianz Research in March 2026 found “a ~46% growth gap between investment and” revenue [MS, insurer research]. 9fin’s review of third-quarter 2026 results was titled “Big tech borrows to bridge AI spending gap as capex outruns cash flow” and found that “demand for AI is real and growing” [IC]. Amazon reported negative free cash flow of US$42.4 billion over twelve months [TBV, secondary]. A July 2026 McKinsey survey found 11 per cent of 750 respondents said their reorganisation achieved “reinvention” [TBV, secondary]. Palantir’s chief executive told shareholders on 2 November 2026 that “the Sovereign AI movement is imposing the most dramatic dichotomy between haves and have nots that I have witnessed in 30 years of business” [IC, SEC 8-K]. The Bank of England’s November 2026 Monetary Policy Report, recorded in the wiki’s snapshot, estimates that AI has raised UK productivity by about 0.5 per cent over three years, most of it from building and deploying AI [IC]. Nothing found in the research run overturns the snapshot’s reading that evidence of productivity gains is thin.
The November 2026 repricing. The wiki’s early November snapshot records a Democratic sweep of Congress, voters approving local limits on data centres in nearly every measure on the ballot, a one-session fall of more than 5.5 per cent in the Nasdaq Composite on 5 November (about US$2.5 trillion of value), and a Federal Reserve that has turned toward raising rates [IC, as tagged on that page]. Ballotpedia’s count, published on 6 November, is 40 of 41 local measures favouring limits on 3 November [IC]. Since then:
- On 6 November, CNBC reported, citing Deutsche Bank, that the yield on CoreWeave’s 9.625 per cent bond due 2032 had climbed above 16 per cent, that credit default swaps on Nvidia were up 76 per cent in a month to roughly 139 basis points, Broadcom’s up 49 per cent and Amazon’s up 44 per cent, and quoted Jefferies strategist Christopher Wood: “The most plausible way the AI capex arms race will end is that the credit card will be taken away” [IC].
- CNBC reported the market “grappling with the prospect that at least some data center spending will be delayed or scrapped under incoming Democratic governors in Ohio, Texas, and Georgia” [IC]. National Economic Council Director Kevin Hassett said he expects data centre growth to continue as forecast [IC].
- Bank of America estimated AI-related credit issuance rising 31 per cent to US$685 billion in 2027 and flagged “downside to these forecasts” from “a potential slowdown in AI infrastructure projects on the back of the Blue Wave” [IC].
- ING sought bids, due 19 November with a US$10 million minimum, for about US$500 million across 12 data centre loans, including a US$25 million slice backing the Stargate site in New Mexico, according to Bloomberg on 5 November as reported by CNBC [IC].
- 9fin reported on 3 November that “AI-linked debt markets held up better than equities,” with one portfolio manager calling the outlook for credit “largely manageable” [IC].
- A Bloomberg and Sightline estimate reported in May 2026 put about half of planned 2026 US data centre capacity as delayed or cancelled [TBV, secondary].
- Kevin Warsh, chair of the Federal Reserve, cited tech debt issuance as one factor in higher government yields [TBV, secondary].
No distressed sale of an AI data centre, and no bankruptcy of a frontier lab or neocloud, was found as of 7 November 2026.
What survives a repricing, layer by layer
| Layer | Book life in filings | Resale and survival evidence | Grade |
|---|---|---|---|
| Sites and building shells | 25 to 40 years (hyperscaler policies as summarised by analysts) | Shells and land retain alternative use; ING’s loan sale shows a secondary market for the debt | [TBV] for lives; [IC] for loan sale |
| Grid interconnections and power contracts | Not separately disclosed in sources reached | Morgan Stanley puts a cumulative 2026 to 2028 US shortfall near 57 GW gross, so an approved connection is scarce [TBV]; contracts can be cancelled, and governors’ decisions may remove planned capacity [IC] | [TBV] |
| Cooling plant and water rights | Not found | Local votes against data centres (wiki snapshot) show permits are contested [IC] | [IC] |
| Accelerators | 5 to 6 years in hyperscaler books; 4 to 5 at neoclouds (Nebius 4, Lambda 5) | Burry estimates 2 to 3 years economic life; one-year H100 rental pricing is cited as the first glut signal [TBV] | [IC] lives; [TBV] resale |
| Networking | Booked with servers | Fibre links between campuses outlast the servers | [TBV] |
| Trained open-weight models | Not capitalised in sources reached | Apertus, InkubaLM and others are downloadable and survive their makers’ finances | [IC] |
| Skilled staff | Not an asset on the books | No data found | none |
The record supports a two-speed residue. Microsoft disclosed that about two-thirds of a recent quarter’s capex went to short-lived assets [TBV, secondary], which, if confirmed, puts most of the dollar value in the layer that does not wait.
Fibre and AI parallels
The same table is available as CSV.
| dimension | fiber_record | ai_record | parallel | note | grade | sources |
|---|---|---|---|---|---|---|
| Scale | US telecom capex peak about US$120bn, 1.2% of GDP, 2000 | Hyperscaler capex 1.4% of GDP 2025; about US$646bn (about 2%) guided for 2026; consensus about 3% for 2027 to 2029 | Strong | AI exceeds the telecom peak on the mechanical ratio; numerators differ | [IC] | Apollo 2026; Meridian Report 2026; Goldman Sachs 2025 |
| Financing | Lucent up to US$8.1bn customer credit; Nortel US$2.1bn customer financing end FY2001 | Nvidia LOI up to US$100bn in OpenAI (US$30bn closed); AMD warrants for 160m shares; about one-third of 2026 capex debt-financed | Strong | Supplier funds customer in both; AI borrowers include the most profitable firms | [IC] | Lazonick and March; Globe and Mail 2002; Goldman Sachs 2026 |
| Accounting | WorldCom line-cost capitalisation, US$3.8bn disclosed June 2002; IRU swaps | Server lives extended to 6 years 2020 to 2024; Amazon cut a subset to 5 in 2025 | Partial | Fibre cases were fraud findings; the AI case is an estimate dispute with no finding of wrongdoing | [TBV] | Amazon 10-K; CNBC 2025 |
| Forecasts | ”Doubling every 100 days” from UUNET and WorldCom; actual about yearly doubling | Multi-trillion capex forecasts from banks, consultancies and the companies; revenue gap about 46% (Allianz) | Partial | Many interested forecasters; usage is measured; revenue lags | [TBV] | Coffman and Odlyzko 2000; Allianz 2026 |
| Analysts and media | Grubman and the 2003 settlement | Analyst estimates revised upward each quarter in 2025 | Weak | No regulator finding on AI research conflicts found | [TBV] | Goldman Sachs 2025 |
| Physical life | Fibre strands usable for decades | Accelerators booked 5 to 6 years; argued 2 to 3 economically | Breaks | The largest AI asset class does not wait to be lit | [IC] | Amazon 10-K; CNBC 2025 |
| Crash trigger | Credit closed to carriers in 2001; recession from March 2001 | Credit spreads widened, CDS up, ING loan sale, elections and Fed tightening, November 2026 | Partial | Trigger rhymes in credit; equity fall reported; no defaults yet | [IC] | Richmond Fed 2003; CNBC 6 Nov 2026 |
| Buyers of distressed assets | Level 3, platforms, later Google | None yet; ING loan buyers sought among private credit, asset managers, insurers | Too thin | No distressed AI asset sale on record | [TBV] | CNBC 6 Nov 2026 |
| Surveillance | Backbone tapping (Room 641A 2006; Upstream 2013) | Frontier models on classified networks; Maven program of record; Pentagon designation of Anthropic over surveillance limits | Partial | State as user is documented in both; legal form differs | [TBV] | Breaking Defense 2025; CNBC 25 Sep 2026 |
| Public and community access | NRENs, municipal networks, later restricted by state laws | EU AI Factories (€10bn), Apertus, NAIRR pilot, proposals for public cloud from stranded assets | Partial | Public compute exists at small scale; acquisition rules proposed only | [TBV] | Vanderbilt 2026; EU InvestAI 2025 |
| Institutions | 1996 Act open access and rights of way; later rollback | No open-access rule for compute found; procurement terms used to shape use | Weak | Open access has no compute equivalent on record | [TBV] | Search of record in the research run |
The two strongest rows are scale and financing, both graded [IC]. The row that breaks is physical life. The buyers row cannot be graded for AI because the event has not occurred.
AI for defence, security and surveillance
Frontier labs and the Pentagon. In July 2025 the Chief Digital and Artificial Intelligence Office awarded Anthropic, Google, OpenAI and xAI prototype agreements, each with a US$200 million ceiling, “to develop agentic AI workflows across a variety of mission areas,” and made generative models available inside platforms including the Maven Smart System [IC]. Anthropic said its work built on deployment at Lawrence Livermore National Laboratory, integration with Palantir on classified networks, and its “Claude Gov” models [IC]. xAI and OpenAI launched “for Government” units [IC].
The Anthropic dispute is the clearest court record of a state deciding the terms of AI use through procurement. Anthropic declined to remove contract clauses prohibiting mass domestic surveillance and fully autonomous weapons; the Pentagon designated it a supply chain risk in late February and early March 2026 and stated that “the military will not allow a vendor to insert itself into the chain of command by restricting the lawful use of a critical capability” [IC, NPR]. Hours later OpenAI announced a deal to replace Anthropic in classified environments [IC, NPR]. A federal judge enjoined the designation on 26 March 2026; on 27 August District Judge Rita Lin set aside the designation made under 10 U.S.C. § 3252, ruling it “unlawful retaliation” in violation of the First Amendment [IC]. On 25 September the D.C. Circuit, 2 to 1, upheld a parallel designation made under the Federal Acquisition Supply Chain Security Act (41 U.S.C. § 4713), writing that the test turns “on what Anthropic does, not why Anthropic does it”; the dissent wrote that the law does not treat “a contractor’s honest and upfront enforcement of restrictions” as a supply chain risk [IC]. The panel stayed its decision to give Anthropic time to seek rehearing (Taft Law; CNBC) [IC]. Right of reply: Anthropic said it disagreed with the appellate decision; Defense Secretary Pete Hegseth posted “Confirmed: @AnthropicAI = Supply Chain Risk” [IC]. A tracker reports that in May 2026 Pentagon agreements expanded to Nvidia, Microsoft, AWS and Reflection AI for classified networks [TBV].
Defence-first firms. Palantir reported third-quarter 2026 revenue of US$2.313 billion, up 96 per cent; US government revenue of US$953 million, up 96 per cent; and US commercial revenue of US$984 million, which “surpassed U.S. government revenue for the first time ever” [IC, SEC 8-K]. It raised full-year guidance to US$8.530 to 8.534 billion [IC]. A tracker citing Reuters reports that a memo of 9 March 2026 made the Maven Smart System a program of record [TBV]; contract ceilings of about US$10 billion for the Army enterprise agreement and US$795 million or more for Maven appear in secondary compilations [TBV]. Palantir’s ImmigrationOS work for US immigration enforcement, Anduril, Shield AI, Scale AI and Helsing could not be documented from primary records and appear in the register as [TBV]. The wiki’s Surveillance Capitalism page treats Palantir as part of the data-fusion complex [MS]. Right of reply: Palantir’s published statements describe its products as giving customers “maximal control over their operations, data, and decisions” [IC, SEC 8-K]; a tracker reports Palantir “distanced itself from AI-driven targeting deaths” [TBV].
Targeting in recent conflicts. Reporting by +972 Magazine and others on AI-assisted targeting systems used in Gaza, the Israel Defense Forces’ published responses, and reported changes to Microsoft cloud services for an Israeli military unit could not be retrieved in the research run. They are listed in the register as [TBV] with no figures. This brief gives no verdict on the conduct of any armed force.
Border, migration and policing. Biometric border systems, Clearview AI and its regulatory penalties, and automated licence plate reader networks are in scope and were not reached [TBV]. Pope Leo XIV’s address of 30 October 2026 names the border control database as a decision that must not be delegated: “When a person is placed on a target list or entered into a border control database, there is a danger that he or she will be reduced to a mere data point” [IC].
Commercial Earth observation: Planet Labs. In the quarter to 31 July 2026, Planet reported record revenue of US$116.1 million, up 58 per cent [IC, SEC 8-K]. Defence and intelligence revenue roughly doubled to about US$81 million, about 70 per cent of the total, up from 57 per cent a year earlier [IC, earnings call and press]. The quarter included an US$8 million award from the National Geospatial-Intelligence Agency for the Global Monitoring Service, about US$14 million of Pelican satellite delivery revenue from the Swedish Army, and a six-figure renewal with a hyperscaler using Pelican imagery “to track data center and semiconductor facility construction” [IC for NGA; TBV for the hyperscaler]. In the prior quarter NGA extended a US$21.9 million maritime surveillance contract [IC, SEC 8-K]. In the same period Planet shipped its second Tanager satellite for methane detection, the platform behind Carbon Mapper’s public data [IC]. The wiki’s Planet Labs brief records founder voting control, a public benefit charter, and Planet AI app terms under which queries are kept out of training and reports belong to Planet [MS]. This is dual use inside one firm. The record shows defence revenue growing faster than commercial (about 15 per cent) and civil government (about 5 per cent) revenue; it does not show what Planet intends.
International governance. The CCW Group of Governmental Experts adopted its final report (CCW/GGE.1/2026/3) on 4 September 2026, containing “a set of elements of an instrument, without prejudging its nature” [IC, UN Geneva]. Human Rights Watch called the final report “significantly watered down” and reported “a record number of states now supporting negotiations on legally binding rules” [IC]. A Cloud Security Alliance note says US and Russian delegations removed language requiring human review of AI-selected targets and counts 76 states backing negotiations; a Stop Killer Robots briefing counts 130 states supporting a legally binding instrument [TBV; the counts measure different things and conflict]. The UN Secretary-General and the ICRC issued a joint appeal on 25 August 2026 [IC]. The Seventh Review Conference meets at the Palais des Nations from 16 to 20 November 2026 under Swiss Ambassador Julien Thöni [IC]. It has not met as of 7 November, so there is no outcome to report.
AI for place-based and prosocial outcomes
A prosocial use serves a public good anywhere. A place-based use serves a specific place and is governed, at least in part, by the people of that place. The register keeps the labels separate.
Science as a public good. The AlphaFold Protein Structure Database, run by EMBL-EBI with Google DeepMind, launched in July 2021 with about 360,000 structures and “has since grown to a staggering 200 million protein structures, from over one million organisms” [IC]. Its Nobel Prize in Chemistry was announced on 9 October 2024 [IC]. A secondary source reports 261,552,403 models as of 24 September 2026 [TBV]. ECMWF began operational forecasts with the deterministic Artificial Intelligence Forecasting System on 25 February 2025, and the ensemble version on 1 July 2025; a 15-day forecast runs “in minutes” on ECMWF’s computing facility [IC, ECMWF Newsletter 186]. Google DeepMind’s GraphCast and WeatherNext models are [TBV] here.
Environmental monitoring. Carbon Mapper’s Tanager-1, built by Planet with a NASA JPL spectrometer, launched on 16 August 2024 and released its first methane detections on 10 October 2024; the coalition reported “over $130 million in philanthropic funding,” and RMI states that funding “to date has been entirely philanthropic” [IC]. Climate TRACE, Global Fishing Watch, Global Forest Watch, the Allen Institute’s Skylight and EarthRanger, and Rainforest Connection appear in the register with [TBV] grades.
Agriculture. Digital Green’s Farmer.Chat is funded by the Gates Foundation under grant INV-047346; a September 2024 paper reported “15,000+ users” and “over 300,000 unique queries in the past year,” and its chief executive later said “some 250,000 farmers use the app” [IC for the paper; MS for the later count]. It runs on OpenAI’s GPT-4 with retrieval [IC]. A Gates and Google commitment of US$100 million for smallholder AI tools is reported secondhand [TBV].
Language and data sovereignty. Te Hiku Media’s Papa Reo platform received NZ$13 million over seven years from 2020 from New Zealand’s Strategic Science Investment Fund [IC]. Its te reo Māori speech recognition reaches 92 per cent accuracy, trained with NVIDIA NeMo on A100 GPUs [MS]. Its Kaitiakitanga License holds that “data is not owned but is cared for under the principle of kaitiakitanga and any benefit derived from data flows to the source of the data” [MS]. Lelapa AI raised a US$2.5 million seed round in February 2023 and released InkubaLM in August 2024, a 0.4 billion parameter model covering Swahili, Yoruba, isiXhosa, Hausa and isiZulu, openly on Hugging Face [IC for the model; TBV for funding]. Masakhane and the CARE Principles for Indigenous Data Governance are [TBV] in the register for their AI applications.
Public compute. Switzerland’s Apertus, released on 2 September 2025 by EPFL, ETH Zurich and CSCS in 8 billion and 70 billion parameter sizes, was trained on 15 trillion tokens on the Alps supercomputer and released with open weights, architecture and training data “under a permissive open-source license” [IC]. Reported public funding is about CHF 20 million plus more than 100 million GPU hours [TBV; the source garbles the currency]. The EU’s InvestAI, launched on 11 February 2025, aims to mobilise €200 billion including a €20 billion fund for AI gigafactories; the Commission calls its €10 billion AI Factories programme “already the largest public investment in AI in the world” [IC for the announcement; MS for the claim]. A EuroHPC call for up to seven gigafactories opened on 30 July 2026 [TBV]. The US National AI Research Resource pilot launched in January 2024 [TBV].
Individual and small-business use. Stripe’s research of about May 2026 reported that 63 per cent of new C corporations formed through Stripe Atlas in the second quarter of 2026 were solo-founded, an all-time high; that top-decile solo founders earned about 61 times the median founder’s revenue in the first six months; and that median six-month revenue fell 23 per cent in 2025 [TBV, secondary reports of Stripe’s study]. Tom Morgan’s observation, held over from After the Crash: A BioConomy Reading, that high-agency people use AI to bring their gifts to market is consistent with the solo-founder count and with the widening gap between top and median outcomes. The record found does not tie either figure to AI use directly.
Bounded systems. Gary Marcus, reporting on his meeting with the Pope, distinguished bounded systems such as GPS navigation and AlphaFold, which “work with a particular problem, a particular database, in a particular way,” from open-ended agents [IC that he said it; Marcus is a Global Ambassador for the encyclical Magnifica Humanitas and an interested party on the policy question]. The distinction finds partial support. AlphaFold’s database is a static public good once computed. AIFS runs on an existing national computing facility in minutes. InkubaLM at 0.4 billion parameters and Te Hiku’s speech models run on modest hardware. Two cases cut the other way: Farmer.Chat depends on a frontier API, and Apertus required a national supercomputer to train. Marcus’s figure of about 10,000 incidents from open-ended agents, including the Hugging Face, French refrigeration and Japanese cyberattack cases, remains [TBV]; no incident database was found to corroborate the count. The refrigeration attacks themselves are recorded on the early November snapshot [IC]; their attribution to an AI agent is not.
AI use register
The same table is available as CSV.
| organisation | kind | track | application | place | money | compute | governance | grade | sources |
|---|---|---|---|---|---|---|---|---|---|
| Palantir Technologies | defence contractor | Defence | Maven Smart System targeting and intelligence analysis; AIP | United States; governance by affected people not recorded | US gov revenue US$953m Q3 2026; Maven ceiling US$795m+ (TBV) | Fine-tuned open-weight and frontier models | Program of record (TBV); no published evaluation found | [IC] | Palantir 8-K 2 Nov 2026; fed-spend |
| Palantir Technologies | defence contractor | Border and migration | ImmigrationOS for US immigration enforcement | United States | Not verified | Unknown | Not found | [TBV] | Not reached |
| Anthropic | Frontier lab | Defence | Claude Gov; agentic prototypes on classified networks | United States | Up to US$200m ceiling, July 2025 | Frontier model | Contract limits on mass domestic surveillance and autonomous weapons; designated supply chain risk 2026 | [IC] | Breaking Defense 2025; CNBC 25 Sep 2026 |
| OpenAI | Frontier lab | Defence | OpenAI for Government; classified deployment replacing Anthropic | United States | Up to US$200m ceiling, 2025 | Frontier model | Pentagon “lawful use” terms | [IC] | Breaking Defense 2025; NPR Mar 2026 |
| hyperscaler | Defence | Agentic AI workflows; cloud and TPUs | United States | Up to US$200m ceiling, July 2025 | Frontier model and cloud | Not found | [IC] | Breaking Defense 2025 | |
| xAI | Frontier lab | Defence | Grok for Government | United States | Up to US$200m ceiling, July 2025 | Frontier model | Not found | [IC] | Breaking Defense 2025 |
| Nvidia, Microsoft, AWS, Reflection AI | hyperscaler | Defence | Deployment on IL6 and IL7 classified networks | United States | Not stated | Frontier and cloud | ”Lawful operational use” | [TBV] | ai2027-tracker |
| Planet Labs PBC | startup | dual use | Satellite imagery and AI analytics for defence and intelligence | Global; customers govern use | D&I about US$81m, about 70% of US$116.1m, Q2 FY2027 | Task-specific models and Planet AI app | Public benefit corporation; app queries kept out of training (MS) | [IC] | Planet 8-K Sept 2026; wiki Planet brief |
| National Geospatial-Intelligence Agency | government agency | Defence | Global Monitoring Service; maritime surveillance | United States | US$8m OTA Aug 2026; US$21.9m extension 2026 | Commercial EO analytics | Not found | [IC] | Planet 8-K filings 2026 |
| Swedish Armed Forces | government agency | Defence | Sovereign Pelican satellite | Sweden | About US$14m revenue recognised Q2 FY2027 | Satellite plus analytics | National ownership | [TBV] | Quilty Space |
| Israel Defense Forces | government agency | Defence | AI-assisted target generation reported by +972 Magazine | Gaza | Not verified | Unknown | IDF responses not retrieved | [TBV] | Not reached |
| Microsoft | hyperscaler | Defence | Cloud services to Israeli military intelligence, reported and later restricted | Israel and Gaza | Not verified | Cloud service | Company review reported | [TBV] | Not reached |
| Anduril Industries | defence contractor | Defence | Autonomous systems and Lattice software | United States and allies | Not verified | Task-specific | Not found | [TBV] | Not reached |
| Helsing | defence contractor | Defence | AI for European defence systems | Europe | Not verified | Unknown | Not found | [TBV] | Not reached |
| Shield AI | defence contractor | Defence | Autonomy software for aircraft | United States | Not verified | Task-specific | Not found | [TBV] | Not reached |
| Scale AI | startup | Defence | Data labeling and evaluation for defence | United States | Not verified | Unknown | Not found | [TBV] | Not reached |
| Clearview AI | startup | commercial surveillance | Facial recognition from scraped images | Global; subject to EU fines | Not verified | Task-specific | Regulatory actions not retrieved | [TBV] | Not reached |
| Flock Safety | startup | domestic security and policing | Automated licence plate reader network | United States | Not verified | Task-specific | Not found | [TBV] | Not reached |
| EMBL-EBI and Google DeepMind (AlphaFold DB) | research institute | prosocial | Open database of predicted protein structures | Global | Not found | Task-specific model, precomputed | Open access; 3m+ users in 190+ countries | [IC] | EMBL 2024; NAR 2024 |
| ECMWF (AIFS) | government agency | prosocial | Operational machine learning weather forecasts | Europe and global users | Not found | Task-specific model on existing HPC | Open data policy | [IC] | ECMWF Newsletter 186 |
| Google DeepMind (GraphCast, WeatherNext) | Frontier lab | prosocial | Machine learning weather models | Global | Not found | Task-specific | Open weights for GraphCast (TBV) | [TBV] | Not reached |
| Carbon Mapper | NGO | prosocial | Methane plume detection from Tanager-1 | Global | Over US$130m philanthropic | Satellite spectrometer and analytics | Public data portal | [IC] | Heatmap 2024; NASA JPL 2024 |
| Climate TRACE | NGO | prosocial | Emissions inventory from satellite and AI | Global | Not verified | Task-specific | Open data (TBV) | [TBV] | Not reached |
| Global Fishing Watch | NGO | prosocial | Vessel tracking and dark fleet detection | Global oceans | Not verified | Task-specific | Open data (TBV) | [TBV] | Not reached |
| Global Forest Watch (WRI) | NGO | prosocial | Tree cover loss alerts | Global; used by local monitors | Not verified | Task-specific | Open data (TBV) | [TBV] | Not reached |
| Allen Institute for AI (Skylight, EarthRanger) | research institute | place-based | Maritime monitoring; protected area operations | Coastal states; conservation areas, partly governed locally | Free to partners (TBV) | Task-specific | Not verified | [TBV] | Not reached |
| Rainforest Connection | NGO | place-based | Acoustic forest monitoring | Tropical forests | Not verified | Small task-specific | Not verified | [TBV] | Not reached |
| Digital Green (Farmer.Chat) | NGO | place-based | Farmer advisory chatbot | India, Kenya and others | Gates grant INV-047346; GitLab Foundation US$350,000 (2023) | Frontier model API (GPT-4) with retrieval | Not found | [IC] | arXiv 2409.08916; GitLab Foundation |
| Te Hiku Media (Papa Reo) | community organisation | place-based | Te reo Māori speech recognition | Aotearoa New Zealand; governed by iwi media | NZ$13m over 7 years from 2020 | Small task-specific model on A100s | Kaitiakitanga License | [IC] | Te Hiku Media 2019; papareo.nz |
| Lelapa AI (InkubaLM) | startup | place-based | Small language model for five African languages | Africa | US$2.5m seed, Feb 2023 (TBV) | 0.4bn parameters | Open weights and datasets | [IC] | UrbanGeekz 2024; Tracxn |
| Masakhane | community organisation | place-based | Community African-language NLP research | Africa | Not verified | Small task-specific | Community governance (TBV) | [TBV] | Not reached |
| Swiss AI Initiative (Apertus) | research institute | prosocial | Open multilingual LLM | Switzerland; public institutions | About CHF 20m public funding (TBV) | 8bn and 70bn parameters on Alps | Open weights, data and architecture | [IC] | ETH Zurich 2025; Sifted |
| EuroHPC (AI Factories, InvestAI) | government agency | prosocial | Public AI compute for research and startups | European Union | €10bn AI Factories; €200bn InvestAI target | Public HPC | EU rules | [IC] | OECD.AI; European Commission 2025 |
| NSF (NAIRR pilot) | government agency | prosocial | Shared research compute | United States | Not verified | Public and donated compute | Not verified | [TBV] | Not reached |
| Gates Foundation and Google | NGO | place-based | AI tools for smallholder farmers | Global South | US$100m reported, Sept 2026 | Unknown | Not found | [TBV] | Secondary report |
| Stripe Atlas solo founders | startup | prosocial | Individual and small-business company formation | Global, mostly US | 63% of Q2 2026 Atlas C corps solo-founded | Mixed | Not applicable | [TBV] | Secondary reports of Stripe 2026 |
| GIDA (CARE Principles) | research institute | place-based | Indigenous data governance principles | Global Indigenous nations | Not applicable | Not applicable | Collective benefit and authority to control | [TBV] | Carroll et al. 2020 |
Measuring the split. The documented figures do not support a ratio. What the record allows is an order-of-magnitude statement with wide error. Named defence and security AI money in this register runs to billions of US dollars a year: Palantir’s US government revenue alone was US$953 million in one quarter, and the four CDAO frontier agreements carry US$800 million of combined ceilings. Named place-based and prosocial project money runs to tens or low hundreds of millions: US$130 million philanthropic for Carbon Mapper, NZ$13 million for Papa Reo, US$2.5 million for Lelapa AI, about CHF 20 million for Apertus. Public compute programmes sit between them at billions (the EU’s €10 billion AI Factories), and serve research generally. Both tracks are two or more orders of magnitude below hyperscaler capex of about US$650 to 800 billion in 2026, most of which serves commercial cloud and model training whose end use is not disclosed. A comparison would need defence budget lines for AI by country (the US budget justification books and NATO’s DIANA figures), philanthropic AI grant totals by purpose (Candid or OECD DAC data), and hyperscaler disclosure of capacity sold to government customers.
Hypotheses
H1. Scale. AI capital expenditure, as a share of US GDP, has exceeded the peak of telecom capital expenditure in 1999 to 2001. For: Apollo puts hyperscaler capex at 1.4 per cent of GDP in 2025 against a 1.2 per cent telecom peak, with about 2 per cent guided for 2026. Against: the numerator is global spending; Goldman’s narrower AI capex measure was 0.8 per cent in late 2025. Grade [IC]. Recommended status: academic-inference. A practitioner can state that AI capex relative to US GDP has passed the telecom peak on the common measure, naming the measure.
H2. Financing rhymes. The AI build-out uses financing structures that resemble telecom vendor financing and capacity swaps. For: Nvidia’s investment in OpenAI tied to chip deployment, AMD’s warrants, Oracle’s contract and Goldman’s one-third debt share all show suppliers funding customers and parties booking revenue from one another. Against: Nvidia’s letter of intent closed at US$30 billion; the large buyers generate cash; no fraud finding exists, while the fibre-era swaps ended in enforcement. Grade [IC]. Recommended status: documented-fact for the deals, structural-inference for the resemblance. A practitioner can say the financing structures resemble telecom vendor financing in form.
H3. Forecasts rhyme. The demand forecasts used to justify AI investment share the features of the “doubling every 100 days” claim. For: forecasts of trillions in capex come from parties that sell the capacity or finance it, and revenue lags spending by about 46 per cent on Allianz’s measure. Against: forecasters are many; usage and revenue are measured and reported quarterly; no single claim has been shown false the way Coffman and Odlyzko showed the 100-day claim false. Grade [TBV]. Recommended status: working-hypothesis. A practitioner can say AI demand forecasts come mostly from interested parties, and should not say they are false.
H4. The residue differs. The durable residue of the AI build-out lies in power, sites and water more than in accelerators. For: accelerators are booked over five to six years and argued to last two to three; buildings are booked over decades; interconnection is scarce. Against: power contracts and permits can lapse; sites far from demand may strand; the value of a shell depends on a tenant. Grade [TBV] (life figures [IC], resale evidence thin). Recommended status: structural-inference. A practitioner can say the long-lived part of the AI build-out is power, land and buildings, and that its resale value is unproven.
H5. Cheap capacity went to the platforms and the state. For: commentary records Level 3’s consolidation and later platform purchases; backbone surveillance programmes were disclosed in 2006 and 2013. Against: the shares were not established, and research and community networks did acquire some capacity. Grade [TBV]. Recommended status: working-hypothesis. A practitioner should not state the shares until the primary record is checked.
H6. Defence is the buyer of last resort. For: Planet’s defence share rose from 57 to 70 per cent in a year; frontier labs signed defence agreements and launched government units. Against: Palantir’s commercial revenue overtook its government revenue in the third quarter of 2026; US$200 million ceilings are small against frontier-lab revenue; the Pentagon excluded one frontier lab. Grade [IC] for the facts; the hypothesis is contradicted as a general claim. Recommended status: working-hypothesis. A practitioner can say defence demand is growing at firms whose product is sensing and analysis, and cannot say it is growing across AI companies generally.
H7. Place-based AI is bounded and cheap. For: InkubaLM at 0.4 billion parameters, Te Hiku’s speech models, AIFS running in minutes, and the precomputed AlphaFold database. Against: Farmer.Chat depends on GPT-4; Apertus needed a national supercomputer; Carbon Mapper’s satellite comes from a firm 70 per cent dependent on defence revenue. Grade [IC] for the cases, inference for the generalisation. Recommended status: structural-inference. A practitioner can say many place-based systems run on small models, and should check each one’s dependence on a frontier API or a dual-use supplier.
H8. Institutions decided the fibre outcome and will decide the AI one. For: the 1996 Act set rights of way and open access during the build-out; no comparable rule for compute was found; the Anthropic case shows procurement law setting the terms of AI use. Against: the fibre shares are [TBV], so the causal link is unshown; EU AI Factories and Apertus show public compute exists without an acquisition rule. Grade [TBV]. Recommended status: working-hypothesis. A practitioner can say no open-access rule for AI compute is on record, and that proposals to take over stranded capacity exist.
Evidence the framing must face
- The parallel may be shallow. Gavin Baker is quoted contrasting 97 per cent dark fibre at the 2000 peak with “no dark GPUs” today [TBV]. 9fin found “demand for AI is real and growing” [IC]. No utilisation survey of data centres or accelerators was found. Commentary names one-year H100 rental prices as the glut signal to watch [TBV].
- The buyers differ. 7GC notes that Amazon, Microsoft and Alphabet, which account for most capex, carry free cash flow and legacy businesses that “insulate them from telecom-era vulnerabilities” [MS]. Lucent’s customers were cash-poor new carriers. A repricing may fall on shareholders without forced sales. The counterweight is that debt now funds about a third of capex and that neoclouds such as CoreWeave carry high-yield debt.
- Accelerators do not wait. Amazon’s own 2025 shortening of server lives is the strongest evidence for this point [IC]. If accelerators are written off within three years, the inheritance from them is close to nothing.
- Defence and prosocial uses may share the same base. Planet’s defence revenue pays for the constellation that also built Tanager-1, which carries Carbon Mapper’s public methane data [IC]. Published arguments on whether defence spending funds civilian capability, and arguments against, were not retrieved.
- The deployment period after fibre was broadly beneficial. Odlyzko records traffic continuing to double each year after the crash, and commentary describes the dark fibre carrying broadband, video and cloud computing [IC for traffic]. Mobile money and open-source software also ran on that capacity. Surveillance was one use among many.
Perez’s own published view weighs against treating AI as a new installation period. She wrote that AI “must be understood as belonging to a larger, more mature technological revolution that began a half-century ago,” the information and communications revolution [IC, quoted by the MIT Initiative on the Digital Economy]. In her model the installation period is when “finance plays a crucial role in unleashing the economic potential of the new technologies,” installation “ends in a ‘Frenzy’ phase,” and “the switch between these periods happens after a major financial crash” [IC, Soundings interview]. Read with her AI essay, the AI build-out is a late surge inside a revolution whose turning point she placed after 2000 and 2008. Ben Thompson of Stratechery observed that GPU spending differs from the railroads and fibre Perez studied [MS, commentator]. Howard Marks of Oaktree wrote that it is “unlikely that AI will be the first transformative technology that isn’t overbuilt” [MS, investor].
The forks ahead
The record shows these institutional choices, proposed, adopted or absent, bearing on who inherits AI capacity:
- Public acquisition of stranded data centres (proposed). A Vanderbilt paper of March 2026, After the AI Crash, says “Congress should turn data centers that become stranded assets into a public cloud and sustain AI research and development (R&D) for public purposes that companies may stop funding” [IC that it is proposed]. The Canadian Centre for Policy Alternatives writes that “governments could acquire these assets and their computing power at a discount” to construct “a fully public AI model” [IC that it is proposed].
- Public compute programmes (adopted). EU AI Factories (€10 billion), InvestAI (€200 billion target), the EuroHPC gigafactory call, Switzerland’s Apertus on a public supercomputer, and the US NAIRR pilot [IC for EU and Swiss; TBV for NAIRR].
- Open access for compute (absent). No rule requiring owners of AI compute to sell capacity to others on equal terms was found.
- State terms of use through procurement (adopted, contested). The Pentagon’s “lawful use” terms and the Anthropic designation, with a district court setting aside the designation under one statute (27 August 2026) and the D.C. Circuit upholding it under another (25 September 2026).
- Lethal autonomous weapons (pending). The CCW Review Conference, 16 to 20 November 2026, on the GGE’s elements of an instrument.
- Local permitting (adopted case by case). Local votes against data centres and new governors in Ohio, Texas and Georgia, per the snapshot and CNBC.
- Community data licensing (adopted locally). The Kaitiakitanga License and the CARE Principles set terms for data, and do not reach compute.
The Pope’s fifth point calls for “institutions capable of putting these principles into practice, particularly effective international institutions,” with “shared standards, independent oversight and clear mechanisms of accountability” [IC, address of 30 October 2026]. Against it, the record shows national and regional public compute adopted, international weapons rules pending, and no institution yet adopted for the transfer of private AI capacity on repricing. For a bioregion, the nearest of these forks is local permitting: the data centre measures of November 2026 were decisions about how much power and water a facility may draw from a place [structural inference].
Summary of graded claims
| claim | grade | recommended status | safe to state publicly |
|---|---|---|---|
| AI capex relative to US GDP exceeds the 2000 telecom peak | [IC] | academic-inference | with caveat |
| AI financing includes supplier investments in customers | [IC] | documented-fact | yes |
| AI demand forecasts rhyme with “doubling every 100 days” | [TBV] | working-hypothesis | no |
| Accelerators are depreciated over 5 to 6 years and Amazon shortened a subset | [IC] | documented-fact | yes |
| The durable residue is power, sites and water | [TBV] | structural-inference | with caveat |
| Fibre shares went mostly to platforms, carriers and the state | [TBV] | working-hypothesis | no |
| Defence is becoming AI’s buyer of last resort | [IC] facts, hypothesis contradicted | working-hypothesis | no |
| Planet’s defence share rose to about 70 per cent | [IC] | documented-fact | yes |
| Place-based AI mostly runs on small models | [IC] cases | structural-inference | with caveat |
| No open-access rule for compute exists | [TBV] | working-hypothesis | with caveat |
| Proposals exist to turn stranded data centres into public compute | [IC] | documented-fact | yes |
| Marcus’s 10,000 agent incidents | [TBV] | speculative-projection | no |
| The CCW conference has produced an outcome | none | not applicable | no; it meets 16 to 20 November |
Related
- Research
- Disruption Snapshot: Early November 2026
- After the Crash: A BioConomy Reading
- The S-Curve Thesis and the Retention Counter-Tradition
- Planet Labs: Ownership, Mission and the Terms of Ask Planet
- Planet AI App
- Deceleration Indicators
- Surveillance Capitalism
- The Emancipation Architecture
- AI Policy
- Glossary: Dual Use, Data Sovereignty, S-Curve, Deceleration Phase, Alpha Window
Sources
Pages in this wiki:
- Carroll, S. R. et al. (2020). The CARE Principles for Indigenous Data Governance.
- Zuboff, S. (2019). The Age of Surveillance Capitalism.
Works without a page in sources/ yet:
Primary records:
- Palantir Technologies (2026, 2 November). Form 8-K Exhibit 99.1, Q3 2026 results. https://www.sec.gov/Archives/edgar/data/0001321655/000132165526000045/a2026q3ex991pressrelease.htm
- Palantir Technologies (2026, 3 August). Form 8-K Exhibit 99.1, Q2 2026 results. https://www.sec.gov/Archives/edgar/data/0001321655/000132165526000039/a2026q2ex991pressrelease.htm
- Planet Labs PBC (2026). Form 8-K Exhibit 99.1, Q2 FY2027 results. https://www.sec.gov/Archives/edgar/data/0001836833/000119312526381874/pl-ex99_1.htm
- Planet Labs PBC (2026). Form 8-K Exhibit 99.1, Q1 FY2027 results. https://www.sec.gov/Archives/edgar/data/0001836833/000119312526257401/pl-ex99_1.htm
- Pope Leo XIV (2026, 30 October). Address to participants in “The Future of AI: A Dialogue between the Vatican and the Scientific World.” https://www.vatican.va/content/leo-xiv/en/speeches/2026/october/documents/20261030-incontro-pav.html
- Pope Leo XIV (2026, 15 May). Magnifica humanitas. https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html
- UN Geneva (2026). Seventh Review Conference of the High Contracting Parties to the CCW. https://www.ungeneva.org/en/meetings-events/events/observances-special-events/2026/11/seventh-review-conference-high
- UNODA (2026). CCW GGE LAWS final report, CCW/GGE.1/2026/3 (advance copy). https://docs-library.unoda.org/Convention_on_Certain_Conventional_Weapons_-Group_of_Governmental_Experts_on_Lethal_Autonomous_Weapons_Systems_(2026)/CCW-GGE.1-2026-3-Advance_copy.pdf
- European Commission (2025, February). Commission launches new InvestAI initiative. https://digital-skills-jobs.europa.eu/en/latest/news/commission-launches-new-investai-initiative-mobilise-eu200-billion-investment-ai
- ECMWF Newsletter 186. Machine learning services at ECMWF. https://www.ecmwf.int/en/newsletter/186/news/machine-learning-services-ecmwf-new-capabilities-users
Economic history and scholarship:
- Couper, E. A., J. P. Hejkal and A. L. Wolman (2003). Boom and Bust in Telecommunications. Federal Reserve Bank of Richmond Economic Quarterly 89(4). https://www.richmondfed.org/~/media/richmondfedorg/publications/research/economic_quarterly/2003/fall/pdf/wolman.pdf
- Coffman, K. G. and A. M. Odlyzko (2000). Internet growth: Is there a “Moore’s Law” for data traffic? https://www-users.cse.umn.edu/~odlyzko/doc/internet.moore.pdf
- Odlyzko, A. M. (2003). Internet traffic growth: Sources and implications. https://www-users.cse.umn.edu/~odlyzko/doc/itcom.internet.growth.pdf
- Odlyzko, A. M. Bubbles, gullibility, and other challenges for economics. https://www-users.cse.umn.edu/~odlyzko/doc/mania03.pdf
- Odlyzko, A. M. (2009). Report to the CRTC. https://www-users.cse.umn.edu/~odlyzko/misc/crtc-2009-report.pdf
- Lazonick, W. and E. March. The Rise and Demise of Lucent Technologies. MPRA Paper 22012. https://mpra.ub.uni-muenchen.de/22012/1/MPRA_paper_22012.pdf
- Perez, C. Interview by J. Rutherford, Soundings. https://carlotaperez.org/wp-content/downloads/media/interviews-and-podcasts/Soundings%20interview%20by%20J.Rutherford.pdf
- MIT Initiative on the Digital Economy. Two Reasons Why AI Isn’t a Tech Revolution Yet. https://medium.com/mit-initiative-on-the-digital-economy/ai-is-revolutionary-but-it-isnt-a-tech-revolution-yet-29cc5e6e6fe0
- Varadi, M. et al. (2024). AlphaFold Protein Structure Database in 2024. Nucleic Acids Research. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10767828/
- Singh, N. et al. (2024). Farmer.Chat. arXiv 2409.08916. https://arxiv.org/html/2409.08916v2
- Vanderbilt University (2026, March). After the AI Crash. https://cdn.vanderbilt.edu/vu-URL/wp-content/uploads/sites/412/2026/03/23144242/After-the-AI-Crash.pdf
- Canadian Centre for Policy Alternatives. What happens to data centres when the AI bubble pops? https://www.policyalternatives.ca/news-research/what-happens-to-data-centres-when-the-ai-bubble-pops/
Analysts and investors (interested parties):
- Apollo Academy (2026, February). How Much Is $646 Billion? https://www.apolloacademy.com/how-much-is-646-billion/
- Apollo Academy (2026, February). Hyperscaler capex chart book. https://www.apolloacademy.com/wp-content/uploads/2026/02/Hyperscaler-capex-022226_v2.pdf
- Apollo (2026). The AI Capex Boom Is Building Twice as Fast as the Housing Boom. https://www.apollo.com/wealth/insights-news/insights/daily-spark/the-ai-capex-boom-is-building-twice-as-fast-as-the-housing-boom
- Goldman Sachs (2025). Why AI Companies May Invest More than $500 Billion in 2026. https://www.goldmansachs.com/insights/articles/why-ai-companies-may-invest-more-than-500-billion-in-2026
- Goldman Sachs Exchanges (2026). How AI Debt Is Reshaping Credit Markets. https://www.goldmansachs.com/insights/goldman-sachs-exchanges/how-ai-debt-is-reshaping-the-credit-market
- Allianz Research (2026, 25 March). AI capex cycle: war-proof for now. https://www.allianz.com/content/dam/onemarketing/azcom/Allianz_com/economic-research/publications/specials/en/2026/march/2026_03_25_AI.pdf
- 7GC (2025). AI Capex and the Telecom Bubble: A Comparative Analysis. https://www.7gc.co/insights/ai-capex-and-the-telecom-bubble-a-comparative-analysis
- Meridian Report (2026). The $733 billion year. https://meridianreport.org/article/ai-infrastructure-capex-report-2026
- Marks, H. Is It a Bubble? Oaktree Capital. https://www.oaktreecapital.com/insights/memo/is-it-a-bubble
- Thompson, B. (2025). The Benefits of Bubbles. Stratechery. https://stratechery.com/2025/the-benefits-of-bubbles/
- VLSI.kr (2026). The USD 1 Trillion Circular Loop. https://www.vlsi.kr/en/ai-circular-investment-loop-en/
Journalism:
- CNBC (2026, 6 November). CoreWeave bond yield tops 16% as AI credit stress rises. https://www.cnbc.com/2026/11/06/coreweave-bond-yield-tops-16percent-as-ai-credit-stress-rises.html
- CNBC (2026, 6 November). AI debt concerns grow as ING reportedly seeks buyers for $500 million of data center loans. https://www.cnbc.com/2026/11/06/ing-ai-data-center-loans-debt.html
- CNBC (2025, 14 November). How long before a GPU depreciates? https://www.cnbc.com/2025/11/14/ai-gpu-depreciation-coreweave-nvidia-michael-burry.html
- CNBC (2026, 25 September). U.S. appeals court upholds Pentagon designation of Anthropic as supply chain risk. https://www.cnbc.com/2026/09/25/pentagon-anthropic-ai-risk-appeals-court.html
- 9fin (2026, 3 November). Credit holds on as neoclouds, tech equities fall. https://www.9fin.com/insights/credit-neoclouds-tech-equities
- Breaking Defense (2025, July). Anthropic, Google and xAI win $200M each from Pentagon AI chief for “agentic AI.” https://breakingdefense.com/2025/07/anthropic-google-and-xai-win-200m-each-from-pentagon-ai-chief-for-agentic-ai/
- NPR (2026, 6 March). Pentagon labels AI company Anthropic a supply chain risk. https://www.npr.org/2026/03/06/g-s1-112713/pentagon-labels-ai-company-anthropic-a-supply-chain-risk
- CNN (2026, 26 March). Judge blocks Pentagon’s effort to “punish” Anthropic. https://www.cnn.com/2026/03/26/business/anthropic-pentagon-injunction-supply-chain-risk
- TechCrunch (2026, 28 August). Anthropic gets its first court win. https://techcrunch.com/2026/08/28/anthropic-gets-its-first-court-win-over-the-pentagons-supply-chain-risk-label/
- ABC News (2026, 25 September). Federal appeals court upholds Pentagon designation of Anthropic. https://abcnews.com/Business/anthropic-appeals-court-declines-block-pentagon-blacklisting/story?id=136755690
- Human Rights Watch (2026, 7 September). UN Talks on Killer Robots End with Calls for Negotiations Growing. https://www.hrw.org/news/2026/09/07/un-talks-on-killer-robots-end-with-calls-for-negotiations-growing
- The Globe and Mail (2002). Nortel clients give warning of finances. https://www.theglobeandmail.com/report-on-business/nortel-clients-give-warning-of-finances/article20454748/
- Newsweek. The Stupid Loan Bubble. https://www.newsweek.com/stupid-loan-bubble-146353
- Fortune (2001, 5 February). Lessons from the Lucent Debacle. https://money.cnn.com/magazines/fortune/fortune_archive/2001/02/05/296152/index.htm
- Heatmap News (2024, August). Carbon Mapper methane satellite. https://heatmap.news/sparks/carbon-mapper-methane-satellite
- NASA JPL (2024, October). First greenhouse gas plumes detected with NASA-designed instrument. https://www.jpl.nasa.gov/news/first-greenhouse-gas-plumes-detected-with-nasa-designed-instrument/
- EMBL (2024, October). AlphaFold wins Nobel Prize in Chemistry. https://www.embl.org/news/science-technology/alphafold-wins-nobel-prize-chemistry-2024/
- ETH Zurich (2025, 2 September). Apertus: a fully open, transparent, multilingual language model. https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html
- Sifted. Switzerland national LLM. https://sifted.eu/articles/switzerland-national-llm
- Te Hiku Media (2019). Te Hiku Media awarded $13m. https://tehiku.nz/te-hiku-tech/papa-reo/11261/te-hiku-media-awarded-13m
- UrbanGeekz (2024, August). Lelapa AI InkubaLM. https://urbangeekz.com/2024/08/lelapa-ai-inkubalm/
- GitLab Foundation. Empowering Kenyan farmers. https://www.gitlabfoundation.org/our-journey/empowering-kenyan-farmers-how-ai-is-revolutionizing-agriculture-with-digital-green
Commentary, trackers and aggregators (used only for [TBV] claims):
- Marcus, G. (2026, 31 October). Report from the Vatican. Marcus on AI. https://garymarcus.substack.com/p/report-from-the-vatican
- Fabricated Knowledge. Lessons from History: The Rise and Fall of the Telecom Bubble. https://www.fabricatedknowledge.com/p/lessons-from-history-the-rise-and
- Dealroom summary of The Information (2026). Cracks Emerge in AI’s Debt-Fueled Data Center Boom. https://dealroom.co/news/info-wtb2ya-cracks-emerge-in-ais-debt-fueled-data-center-boom/
- Curionic (2026, August). Inside AI’s Circular Financing Web. https://www.curionic.net/2026/08/ai-circular-financing-nvidia-openai-oracle-coreweave-2026.html
- RRSRCH. Five hyperscaler useful-life changes since 2022. https://rrsrch.com/posts/50
- GPU Insights (2026). Server Useful Life: What Five 10-Ks Disclose. https://gpuinsights.net/server-useful-life-gpu-depreciation-2026/
- AI 2027 Tracker. Department of Defense scales up AI lab contracting. https://ai2027-tracker.com/predictions/dod-ai-contracting/
- Fed-Spend (2026). Federal AI and Cybersecurity Contract Awards 2026. https://fed-spend.com/blog/federal-ai-cybersecurity-contract-awards-2026
- Cloud Security Alliance (2026). UN Autonomous Weapons Talks: Human Oversight Erodes. https://labs.cloudsecurityalliance.org/research/csa-research-note-un-autonomous-weapons-treaty-oversight-202/
- Quilty Space. Planet Labs (PL): 2027 Q2 Earnings Review. https://www.quiltyspace.com/product-page/planet-labs-pl-2027-q2-earnings-review-financial-analysis
- Solo Founders. Solo founders are 63% of new startups in 2026: Stripe. https://solofounders.com/blog/solo-founders-are-63-of-new-startups-in-2026-stripe/
- Tracxn. Lelapa AI profile. https://tracxn.com/d/companies/lelapa-ai/__9srW_PDa-AS9Sts08WDTsnbZyzLI3Zqcj68wOcbuptE
Added on 13 November 2026:
- Ballotpedia News (2026, 6 November). Voters Favored Limiting Data Center Development in 40 of 41 Local Ballot Measures on Nov. 3. https://news.ballotpedia.org/2026/11/06/voters-favored-limiting-data-center-development-in-40-of-41-local-ballot-measures-on-nov-3/
- Carolan, M. (2002, 28 January). Global Crossing files for Chapter 11, reorganizes. eWeek. https://www.eweek.com/news/global-crossing-files-for-chapter-11-reorganizes/
- CNN Money (2002, 21 July). WorldCom files for Chapter 11 bankruptcy. https://money.cnn.com/2002/07/19/news/worldcom_bankruptcy/index.htm
- Light Reading (2002, 21 July). WorldCom files for bankruptcy. https://www.lightreading.com/business-management/worldcom-files-for-bankruptcy
- Taft Law (2026, August). Federal court rules government’s Anthropic supply chain designation was unlawful retaliation. https://www.taftlaw.com/news-events/law-bulletins/federal-court-rules-governments-anthropic-supply-chain-designation-was-unlawful-retaliation-what-government-contractors-need-to-know/
- Taft Law (2026, September). D.C. Circuit rules Pentagon can exclude Anthropic from supply chain. https://www.taftlaw.com/news-events/law-bulletins/d-c-circuit-rules-pentagon-can-exclude-anthropic-from-supply-chain/
Provenance
Conflict of interest. This brief was drafted by Claude, made by Anthropic. Anthropic is a frontier AI company that builds or leases data centre capacity and whose investors include hyperscalers whose capital spending this brief examines. The public record used here shows that it holds a CDAO prototype agreement with a US$200 million ceiling (July 2025), that it offers “Claude Gov” models for national security customers, and that the Pentagon designated it a supply chain risk in 2026 after it kept contract limits on mass domestic surveillance and autonomous weapons. The wiki’s early November snapshot records that Anthropic filed for an IPO in October 2026, so the repricing this brief describes bears directly on its valuation. Planet’s privacy policy names Anthropic as a model provider for the Planet AI app. Every claim about Anthropic here comes from the public sources listed, and Anthropic is tested on the same terms as other firms; no claim rests on Claude’s own knowledge of the company. The Anthropic dispute is reported because it is the clearest court record of procurement setting the terms of AI use. Readers should weigh that choice with the conflict in mind.
Written on 7 November 2026 from a Claude deep research run under version 0.1 of dark-fiber-and-the-ai-overbuild-deep-research.md in the BioConomy project. The run made 18 web searches, fetched Gary Marcus’s report and SEC filings, and used one research subagent for the prosocial register. It could not fetch the wiki’s own pages; while this page was drafted, the snapshot, the Planet brief and the essay were read from the site’s files, and the snapshot’s figures were reconciled with them (the snapshot records the data centre measures, the Nasdaq fall and the Bank of England productivity estimate as [IC]). The Pope’s address was checked against two independent transcriptions and Vatican News; the vatican.va page itself was not opened.
Editorial decisions. The working title “Dark Fiber and the AI Overbuild: Who Inherits the Capacity” was changed because the AI half of the record raises a prior question, whether anything cheap is left to inherit. epistemic_status is structural-inference because the page’s organising claims about AI inheritance are inferred from the structure of the record; its financing, accounting and contract figures for 2025 to 2026 rest on SEC filings, court reporting and named analysts. Several well-established fibre-era facts (Room 641A, Upstream, Google’s dark fibre purchases, the Grubman settlement, SAT-3 and SEACOM) carry [TBV] because their primary records were not reached; this is stricter than their standing in the historical literature. Analyst figures from firms that sell or finance AI capacity (Apollo, Goldman Sachs, Bank of America) were used for scale and graded [IC] when two of them agreed. Conflicts left unresolved: the AWS commitment (US$38 billion against US$138 billion), the effect of Amazon’s 2025 depreciation change, and the CCW state counts (76 against 130). The register’s place column records governance by affected people only where a source records it. The split between tracks is given in orders of magnitude, and no ratio is offered. Place-specific material (South African data centre build-out and grid access, the SAT-3/WASC and SEACOM price history for South Africa) was left out for the BioHubs directory. Figures from the run’s research report were cross-checked for internal consistency only; quotations should be checked against their sources before the page is relied on.
Edited on 13 November 2026, after the verification pass for Dark AI and the Place-Based Inheritance. The WorldCom and Global Crossing bankruptcies and WorldCom’s US$3.8 billion disclosure were upgraded to [IC] on contemporaneous press records (eWeek, CNN Money, Light Reading); court filings still weren’t reached. The Anthropic paragraph now names the statute behind each ruling, gives the district court ruling its date of 27 August (it was reported on 28 August), and records the D.C. Circuit’s stay. Ballotpedia’s vote count was added beside the snapshot’s. The page was converted to British spelling under the site rule of 11 November 2026; the title, slug, quotations, proper names and source titles keep their American spelling, the US term “program of record” is kept as it is, and “dark fibre” was added as an alias.