Rechecking the five siphons
In June we said the AI buildout was funded by five transfers from the public. We also said the bet closed only on an expectation of federal rescue. Three months of filings, auctions and votes are now in. One siphon reversed. One grew past our estimate. A sixth appeared. And a private company agreed to cover the losses before the government was asked to.
The test we set in June
The June analysis made claims that could be checked. Hyperscaler capex would reach roughly $725 billion in 2026. Incremental bond issuance above the 2020 to 2024 baseline would run $80 to $120 billion a year and land on passive holders. Four of five index families would let an unprofitable SpaceX into the public's automatic holdings while S&P held the line. Community opposition, state abatements and ratepayer cost-shifting would keep growing. Office vacancy would keep climbing as AI thinned the middle layer that fills the floors. And the whole structure closed only on an expectation that the federal government would absorb a concentrated loss, as it had five times in eighteen years.
The record since then holds Q2 filings, two capacity auctions, a 417 to 3 roll call, a grid-operator freeze, a mega-IPO with a full price history, and a 10-Q naming a number nobody was modelling in June.
The capex held and the guidance went up
The June figure was roughly $725 billion of 2026 capex across Microsoft, Alphabet, Amazon and Meta. After the Q2 2026 calls, combined guidance sits at $720 to $745 billion. Every one of the four raised. Amazon moved to about $220 billion from $200 billion. Alphabet moved to $195 to $205 billion from $180 to $190 billion. Meta lifted the floor of its $125 to $145 billion range to $130 billion. Microsoft restated to about $175 billion on a calendar-2026 basis (TMT Finance; CreditSights).
The June estimate was accurate to within about 2%. The thing it did not anticipate is what the spending did to the cash statements.
Alphabet spent $44.9 billion on capital in Q2 2026 against $39.1 billion generated by operations, producing free cash flow of negative $5.9 billion. That is the first negative free cash flow quarter since the 2004 IPO. Trailing twelve-month free cash flow remains positive at $53.3 billion. The gap was covered by $49.6 billion of stock issued in June 2026 and $20.3 billion of senior unsecured notes in the quarter (Tom's Hardware). Meta cleared the same bar by $784 million.
The June piece described capex-to-revenue at 45 to 57% as historically unusual. Two of the four largest firms on earth have now funded a quarter of AI capital spending out of the capital markets. Operations did not cover it.
The credit signal we said to watch has fired
June listed three signals for the back door. One was a widening in hyperscaler bond spreads. That has happened. The shape of it sits in the mechanics of individual deals more than in headline yields.
Amazon, Alphabet, Meta and Oracle issued about $194 billion of bonds in 2026 through 7 July, against roughly $108 billion in all of 2025, a 79% increase (Quartz). Goldman expects the five hyperscalers including Microsoft to reach roughly $250 billion in 2026 and $400 billion in 2027. S&P Global counted $225 billion of hyperscaler and related-entity issuance through midyear, a 973.7% jump on the prior-year pace.
The absorption terms moved against the issuers on every measure that is published. The median spread on two to four year paper for the four issuers rose to 40 basis points from 30 in 2025. The median deal-level concession, the yield premium an issuer pays over its own existing curve to clear a new deal, rose to 12 basis points in 2026 from 2.25 in 2025. Order cover ratios ran 3.5 to 4.5 times through the first half and near five times in February. The first surprise hyperscaler deal of the second half cleared at 1.4 times (Yahoo Finance; Sage Advisory).
June put the incremental issuance channel at $80 to $120 billion a year. The 2026 print through early July was $194 billion for four issuers. The estimate was low by roughly a factor of two on an annualised basis.
Two independent counts put a further layer outside the bond market entirely. A Nikkei study measured $1.65 trillion of off-balance-sheet obligations across five US technology firms, eight times the level of four years earlier, covering long-term GPU and server purchase agreements plus data-centre lease arrangements. Moody's separately flagged $1.2 trillion of similar liabilities with more than $820 billion attached to data centres still under construction (Fortune). These are disclosed in the notes. The balance sheet does not carry them. That is ordinary practice. It also means an index-tracking bondholder's stated exposure understates the position.
The Bank for International Settlements put an AI capex bust and a circular-financing collapse among its top three threats to global financial stability in its Annual Economic Report on 28 June 2026, placing the episode in a line with the canal mania of the 1830s, the British railway bubble of the 1840s and the dot-com crash of 2000 (CNBC). The IMF's April 2026 Global Financial Stability Report reached the concentration question from the other direction. It named the small number of cloud, semiconductor and network providers underpinning most financial-system activity as the channel through which an AI failure becomes a financial-stability event. It separately noted data centres' effect on regional energy prices. The Federal Reserve's May 2026 Financial Stability Report recorded that 50% of market contacts named AI as a possible shock, against 9% a year earlier. That last figure comes from a US federal agency reporting during the period this site discounts as partially entangled. It is cited here on two grounds. It reports a survey of market participants. The agency is relaying what its contacts said. BIS and IMF also reach the same place independently of it.
The Oracle mirror
Oracle's Form 10-Q for the quarter ended 31 August 2026, filed 11 September, carries this sentence in Note 6.
As of August 31, 2026, we had $288 billion of additional lease commitments, substantially all related to data center arrangements, that are generally expected to commence between the second quarter of fiscal 2027 and fiscal 2029 and for terms of fifteen to nineteen years that were not reflected on our condensed consolidated balance sheets as of August 31, 2026 or in the maturities table above.
The recognised figures in the same note give the scale of what is missing. Total operating lease liabilities are $34.621 billion. Total finance lease liabilities are $9.185 billion. The commitments that do not appear are therefore about 6.6 times the lease obligations that do, on terms of fifteen to nineteen years, at a company Moody's and S&P rate BBB, two notches above high yield.
The rest of the note fills in the funding. Operating lease expense for the quarter was $1.101 billion against $575 million a year earlier. Finance lease expense was $275 million against $107 million. Unconditional purchase obligations, primarily long-term component supply for cloud infrastructure and power supply for data centres, total $34.150 billion. Oracle also exhausted a $20 billion at-the-market equity programme inside the quarter, issuing 141 million shares for $19.9 billion net.
Nvidia has guaranteed $105 billion of OpenAI's twenty-year leases at Pike County. Oracle holds a $300 billion five-year cloud agreement with the same customer and $288 billion of unstarted data-centre leases to serve it. Neither obligation sits on a balance sheet. Both are disclosed, correctly, in the notes.
Which number belongs to whom
Nvidia's $279 billion and Oracle's $288 billion sit three per cent apart. Both are called commitments and both concern data centres. They are unrelated obligations owed by different companies to different people. Every large figure in this piece is set out here once.
| Amount | Owed by | Owed to | For |
|---|---|---|---|
| $279bn | Nvidia | memory makers and foundries | memory and manufacturing capacity to build Nvidia products |
| $29bn | Nvidia | cloud providers | compute for Nvidia's own research |
| $25bn | Nvidia | landlords | engineering, design and test space, leases not yet begun |
| $25bn | Nvidia | AI model makers and infrastructure financiers | committed equity investments |
| $8bn | Nvidia | equipment vendors | its own plant and equipment |
| $366bn | Nvidia | total of the five rows above | |
| $108.5bn | Nvidia, only if called | SB Energy and AI-cloud landlords | guarantee of its customers' rent |
| $288bn | Oracle | landlords | data-centre leases not yet begun |
| $125.3bn | Oracle | lenders and bondholders | borrowings |
| $34.15bn | Oracle | suppliers | components and data-centre power |
| $638bn | owed to Oracle | its customers | contracted future revenue |
| ~$300bn | OpenAI | Oracle | five-year cloud contract |
| ~$1.4tn | OpenAI | seven vendors | multi-year compute commitments |
Nvidia's $279 billion buys silicon inputs. It rents nothing. Oracle's $288 billion rents buildings, with Oracle as the tenant on those leases. OpenAI rents compute from Oracle under a separate $300 billion contract. The building Nvidia guarantees is leased from SB Energy.
The vendor circularity has been mapped
This is a different circle from the one the June analysis drew. The two should not be run together. June's ouroboros was an asset-allocation contradiction, private-equity houses long AI infrastructure and long the office real estate that AI displacement would erode. What has now been counted is vendor circularity, the money going round between suppliers, customers and investors inside the AI supply chain. The ouroboros is dealt with separately below.
Sona Asset Management, a London credit specialist managing about $16.6 billion, examined 176 deals between 202 companies to establish how much of the AI economy is circular. It put roughly $3.6 trillion of financing through that map and flagged about 120 of the deals as ones where the same firms appear as investor, customer and supplier inside a single chain. Its own worked examples are short. Nvidia committed about $100 billion to OpenAI. OpenAI used it to buy cloud services from Oracle, Microsoft and Amazon, who are among Nvidia's largest chip customers. Amazon invested about $50 billion in OpenAI against OpenAI's commitment to buy about $100 billion of compute from Amazon.
Sona is a credit fund and profits if the structures it mapped unwind. The map is cited here as its analysis. The individual deals in it are separately disclosed by the counterparties.
How a circular deal works
A normal sale is funded by the buyer. In a circular one the seller funds some or all of it.
Three parties. Nvidia commits capital to OpenAI. OpenAI commits that capital to Oracle, Microsoft or Amazon for cloud capacity. Those firms buy Nvidia chips to supply it. Nvidia records the chip sale as revenue. The money completes a circuit back to the party that advanced it.
Two parties. Amazon invested about $50 billion in OpenAI against OpenAI's commitment to buy about $100 billion of compute from Amazon. The outlay returns as Amazon's own revenue at twice the nominal amount.
No cash. Nvidia guarantees OpenAI's rent. The guarantee makes the building financeable. The building houses Nvidia hardware that OpenAI pays for. Nvidia exchanges a contingent liability for demand, with no cash leaving at signing.
Each leg is arm's length, permitted under the accounting rules and disclosed in the counterparties' filings. The consequence is not misstatement. Revenue funded by the seller stops when the seller stops funding it. The demand is not independent of the supplier. One stream of expected spending is also recorded as revenue at more than one firm. That makes the addressable market appear larger than the underlying demand.
The 1999 telecom episode is the same structure at a twentieth of the size. It also shows the failure mode. Lucent extended about $8.1 billion of vendor financing to carriers with little revenue of their own. When capital markets closed in 2000 and 2001, 47 of those carriers went bankrupt. The equipment had no secondary market. The financing became write-offs. The revenue it had generated did not recur.
The leverage is concentrated where the ratings are weakest. Oracle carried about $125.3 billion of borrowings as of 31 August 2026 on top of the $288 billion of unstarted leases. Oracle, CoreWeave and Nebius together hold roughly $173 billion of debt at about 7.8 times quarterly revenue, against Microsoft at 0.4 times. CoreWeave's credit default swaps have traded around 855 basis points. About $18 billion of loans tied to an Oracle-leased New Mexico data centre were quoted at 89 to 91 cents on the dollar in September by syndicate banks.
The depreciation question
The June analysis cited Goldman's point that shortening modelled chip life from five years to three adds roughly $1 trillion to annual industry depreciation. The filings show the industry moving the other way for most of the period. One firm is now reversing.
Microsoft extended the depreciable life of server and network equipment from four to six years beginning fiscal 2023, with a stated reduction in depreciation expense near $3.7 billion. Alphabet went from four to six years on servers and five to six on network equipment. Meta moved most servers and network equipment to 5.5 years, lowering depreciation by $2.3 billion across the first nine months of 2025. Amazon and Oracle made comparable moves. Each extension adds roughly 4 to 5% to reported net income without changing a single cash flow.
Michael Burry's November 2025 estimate held those schedules against a two-to-three-year replacement cycle and put the difference at more than $176 billion of overstated earnings across 2026 to 2028, with reported operating income at Oracle and Meta more than 20% above his read of economic reality. That figure is a short seller's model and should be treated as one.
The disclosure that is not a model is Amazon's. Having been first to extend in 2020, Amazon shortened the useful life of a subset of its servers from six years back to five in 2025. It said why, citing "the increased pace of technology development". The firm with the longest operating history in the asset class revised its own estimate downward, in a filing, in the direction the skeptics argued for.
Nvidia agreed to pay OpenAI's rent if OpenAI cannot
The June argument was that the buildout's losses would land on the federal balance sheet, on the precedent of TARP, the repo facility, CARES and the SVB systemic risk exception. That argument assumed the first party to absorb a loss would be public.
Nvidia's Form 10-Q for the quarter ended 26 July 2026, filed 26 August, says otherwise. Every number below comes from the filing itself (SEC accession 0001045810-26-000075). The earnings release and the call are not the source for any of it.
Note 10 gives future commitments by fiscal year. Every figure below is billions of US dollars. The $279bn supply row is $279,000,000,000. The $366bn total is $366,000,000,000.
| Commitment | Rest of FY27 | FY28 | FY29 | FY30 | FY31 | FY32+ | Total |
|---|---|---|---|---|---|---|---|
| Supply and capacity | $92bn | $87bn | $88bn | $6bn | $5bn | $1bn | $279bn |
| Cloud service agreements | $3bn | $8bn | $7bn | $6bn | $4bn | $1bn | $29bn |
| Data centre leases not commenced | 0 | $1bn | $1bn | $2bn | $1bn | $20bn | $25bn |
| Equity investments | $18bn | $3bn | $2bn | $2bn | 0 | 0 | $25bn |
| Capital expenditures | $7bn | $1bn | 0 | 0 | 0 | 0 | $8bn |
| Total | $120bn | $100bn | $98bn | $16bn | $10bn | $22bn | $366bn |
For scale, the $366 billion total is larger than Nvidia's entire balance sheet. The same filing puts total assets at $320.3 billion. It is also roughly 1.1% of a year of US economic output.
In every row Nvidia is the party that pays. Supply and capacity, $279 billion. Money Nvidia has committed to its own component suppliers. The filing says the commitments are "for our data center infrastructure systems, primarily memory and manufacturing facilities, to produce our products for long-term demand across current and future product architectures". The obligation is not absolute. The filing adds that the agreements "may be cancelable, rescheduled, or adjustable for our business needs prior to placing firm orders". Changes to them "may result in additional costs".
What Nvidia has already paid out, as distinct from committed
The commitments table covers future obligations. Money already gone sits in Note 6, and the movement there is larger than the commitment line suggests.
Nvidia held $47.9 billion of non-marketable equity securities at 26 July 2026, described in the filing as "primarily in privately-held companies carried at cost less impairment, and adjusted for observable price changes". At 27 July 2025 the balance was $3.8 billion.
| Six months to 26 July 2026 | Amount |
|---|---|
| Opening balance, 25 January 2026 | $22,251m |
| Net additions | $31,005m |
| Unrealized gains | $7,504m |
| Reclassified out on public listing | ($12,712m) |
| Impairments and unrealized losses | ($150m) |
| Closing balance, 26 July 2026 | $47,898m |
Nvidia put $31 billion of cash into private companies in six months. It booked $7.5 billion of unrealized gains on those holdings in the same period, recognised in Other income, so a share of reported non-operating income is the markup on companies Nvidia funded. And $12.7 billion left the balance when holdings began trading publicly, which is the only line in the note that represents an exit rather than an entry.
Cumulative gross unrealized gains stood at $9.1 billion against $661 million a year earlier. Cumulative losses and impairments were $250 million.
Separately, Nvidia holds $3.3 billion of investments in infrastructure financiers accounted for by the equity method. Where those are variable interest entities the filing puts maximum loss exposure, including carrying values and future committed amounts, at $4.7 billion. Nvidia states it is not the primary beneficiary and does not consolidate them.
The total therefore reads differently from the commitments table alone. Equity paid in stands at $47.9 billion, with a further $25 billion committed and unpaid, $3.3 billion in infrastructure financiers, and $108.5 billion of contingent guarantee behind all of it. The filing does not identify any recipient, so none of this can be attributed to a named company.
Who Nvidia is committing that $279 billion to
The 10-Q does not name counterparties. It identifies the inputs as memory and manufacturing capacity. The suppliers of those inputs to Nvidia's data-centre products are a short and public list.
High-bandwidth memory. SK Hynix holds roughly 50 to 55% of the HBM market, Samsung 35 to 40% and Micron 5 to 10%. For Nvidia's Vera Rubin generation, SK Hynix is reported to hold about 70% of the HBM4 allocation.
Logic and advanced packaging. TSMC fabricates the compute dies and performs the CoWoS-L packaging that bonds them to the memory stacks. CoWoS capacity has been a standing constraint on how many accelerators can be built.
The cost weighting explains why the commitment is this large. A 12-layer HBM4 stack prices above $600 and a Rubin-class GPU carries eight of them. That puts more than $4,800 of memory into a single accelerator. HBM is therefore the highest-cost line in the bill of materials, ahead of Nvidia's own silicon.
That changes the position the June analysis assigned to Nvidia. June described it as the picks-and-shovels seller, the one participant certain to be paid whatever happened to AI revenue. On the 26 July filing Nvidia has committed $279 billion to buy the picks and shovels from someone else, before the end demand exists. The firms holding contracted revenue with no demand risk are now SK Hynix, Samsung, Micron and TSMC. Nvidia holds the obligation.
The commitment buys Nvidia's own manufacturing inputs. The filing is explicit that the purchases are "to produce our products". None of it is committed on a customer's behalf. Nvidia's exposure to OpenAI is separate and differently shaped, sitting in the $105 billion lease guarantee and in the $25 billion of committed equity investments in "AI model makers, infrastructure financiers, and other private companies".
The extended payment terms belong to a different set of customers again. The filing attributes the $24.6 billion increase in receivables to "extended payment terms on large multi-quarter agreements with certain investment-grade customers". OpenAI is not investment grade. The same filing says Nvidia's customers lack "investment-grade financing capacity". The receivables growth therefore reflects longer credit given to the rated hyperscalers.
What connects the $279 billion to the guarantees is the demand forecast. No contract links them. Nvidia committed that money against "long-term demand". It is separately underwriting part of the demand it is forecasting.
Nvidia does not name a single customer
Concentration is disclosed without identities. The categories matter more than the percentages.
A direct customer is whoever Nvidia bills. The filing lists the types as "AIBs, distributors, ODMs, OEMs, CSPs, AI model makers, and system integrators". A direct customer is therefore often the contract manufacturer assembling the servers, one step removed from the company that will run them. In the first half of fiscal 2027, three direct customers accounted for 16%, 15% and 13% of total revenue, 44% between them. Nvidia names none of them. It is not required to.
An indirect customer buys through those firms. The filing says indirect customers are "CSPs, AI clouds, AI model makers, enterprises, and public sector entities". It adds that Nvidia generates "a significant amount of our revenue from a limited number of indirect customers, some individually representing 10% or more of our revenue". None of those is named either. Geography is assigned by the direct customer's headquarters. The filing warns that "the end customer and shipping location may be different".
Nvidia writes that it estimates "one AI research and deployment company contributed a meaningful amount of our revenue by purchasing cloud services from our customers" in the second quarter and first half of fiscal 2027.
That describes revenue Nvidia attributes to a company that never buys from Nvidia. The company rents cloud capacity from a Nvidia customer. That customer buys Nvidia chips to supply it. Nvidia books the chip sale. "AI research and deployment company" is OpenAI's own standard description of itself. That makes the identification likely without making it certain. The filing does not name it.
It also explains the pattern in the Korean data. Semiconductors reached a record $46.65 billion of Korean exports in August, 47.5% of all goods exports, with memory roughly 90% of that. Nvidia's forward commitment lands substantially on Korean memory makers as cash. The Korean equity market fell 16% in two sessions in July while those shipments were rising.
Cloud service agreements, $29 billion. Money Nvidia has committed to cloud providers, for compute to run its own research. The stated purpose is research and development on its open models, named in the filing as Nemotron, Cosmos and GR00T, plus its autonomous vehicle software. Nvidia is buying cloud capacity from firms that buy its chips.
Data centre leases not commenced, $25 billion. Rent Nvidia has committed to landlords on buildings it has not yet occupied. The filing says they will be "primarily used for engineering, product design, and testing of our compute chips, networking products, and systems", beginning between the third quarter of fiscal 2027 and fiscal 2033, on terms of up to twenty years.
Equity investments, $25 billion. Money Nvidia has committed to put into other companies. The filing describes the recipients as "AI model makers, infrastructure financiers, and other private companies, subject to certain contingencies". That is a commitment to fund the buyers of its products and the entities financing them.
Capital expenditures, $8 billion. Nvidia's own equipment and facilities, for engineering and manufacturing operations.
Supply and capacity commitments went from $119 billion at the prior quarter to $279 billion in one quarter. The filing attributes the increase primarily to memory and manufacturing facilities.
| Guarantee | Most Nvidia could be required to pay |
|---|---|
| Land, power and shell guarantees for AI clouds | $3.5 billion |
| SB Energy Corp. guarantees | $105.0 billion |
| Total | $108.5 billion, or $108,500,000,000 |
At the prior fiscal year end the guarantee line was $3.5 billion on its own. Partners have placed $712 million in escrow against that original exposure.
Set the commitments and the guarantees against the balance sheet in the same filing. Total assets are $320.272 billion. Total shareholders' equity is $228.984 billion. Total liabilities are $91.288 billion. Forward commitments of $366 billion plus contingent guarantee exposure of $108.5 billion exceed the company's entire balance sheet.
The filing states the reason in its own risk factors. Customers lack "investment-grade financing capacity to secure the AI infrastructure necessary to grow". The company has "undertaken initiatives to address these challenges including securing and providing guarantees of land, power, shell, and capacity". On the guarantee specifically it says that on an OpenAI default or insolvency event that triggers a guarantee, "we may assume the applicable lease".
Working capital moved with it. Accounts receivable rose to $63.059 billion at 26 July 2026 from $38.466 billion at 25 January 2026, an increase of $24.593 billion. The company attributes that to "financing arrangements with certain investment-grade customers, including extended payment terms under large, multi-quarter agreements". Net income for the quarter was $59.688 billion. Nvidia repurchased $19.674 billion of stock and paid $6.047 billion in dividends in the same three months.
Google has taken the same position on the other frontier lab, guaranteeing billions of dollars of Anthropic's data-centre lease and power obligations against default (MediaPost). A Morgan Stanley-led consortium is weighing $15 billion for Nexus Data Centers on an Anthropic-linked campus in Hubbard, Texas with a 1.6 gigawatt on-site gas plant.
Both frontier labs now have their rent guaranteed by the company that sells them their inputs. The June piece expected the federal government to be the first party to absorb a loss. Two private companies took that position first.
Why anyone would write a guarantee like this
Nothing above explains the motive. A guarantee is a liability given away. Three of the best-capitalised companies in the world have given very large ones away inside two quarters. The reason is in Nvidia's own risk factors.
SB Energy builds, owns and operates the Pike County campus. OpenAI is the tenant on a twenty-year lease. The lenders are the banks and institutional investors financing SB Energy's construction. That construction runs to billions before the site produces any income.
Those lenders are repaid from the rent. The rent comes from OpenAI. Financing the building is therefore a judgement about whether OpenAI can pay for twenty years. OpenAI is a private company with a run rate above $40 billion, cash burn near $27 billion for 2026, a projection above $60 billion for 2027, and no credit rating. No lender underwrites a twenty-year obligation against that borrower. Nvidia states the consequence directly, saying its customers lack "investment-grade financing capacity to secure the AI infrastructure necessary to grow".
An unfinanced building is not built. An unbuilt building leaves the chips nowhere to go. The three firms standing behind these leases are the three in the chain with ratings. None of them is lending cash. Each is lending its credit rating. A guarantee costs nothing until it is called and stays off the balance sheet until then. That is the cheapest available way to move creditworthiness from where it sits to where the demand is.
Nvidia is protecting its own upstream commitments. It has pre-committed $279 billion to suppliers, up from $119 billion one quarter earlier, primarily memory and manufacturing capacity. If the end demand does not arrive it absorbs that itself. The guarantee manufactures the demand that consumes the commitment. It buys a second thing as well, twenty years of exclusivity across 8 gigawatts. That shuts AMD, Broadcom and Google's TPUs out of the site for the life of the lease. S&P affirmed Nvidia at AA specifically on the Ohio guarantee. Moody's holds it at Aa1. It carries roughly $106 billion of cash against $33.5 billion of funded debt. It can write this. Both agencies have said so in public.
Oracle bought backlog. Remaining performance obligations reached $638 billion at the end of Q4 FY2026, up 363% year over year, passing Microsoft. That is the metric the market rewards. Oracle was losing the cloud contest to firms with better balance sheets. The $288 billion of leases that will service the backlog sit in a note. The revenue commitment is announced. The cost commitment is disclosed where nobody trades on it. That asymmetry is the trade.
Google is defending three positions with one instrument. It holds 14% of Anthropic. It committed up to $40 billion in April 2026 at a $350 billion post-money valuation, and placed up to a million of its own TPUs. A five-year $200 billion cloud commitment comes back the other way. The lease guarantee protects an equity stake, a revenue contract and a silicon channel at once.
The same structure ran in 1999 at a twentieth of the size
Telecom equipment vendors financed their own customers through the last buildout. Lucent committed roughly $8.1 billion of vendor financing, Nortel about $3.1 billion, Cisco about $2.4 billion. McKinsey counted roughly $25.6 billion across nine suppliers by the end of 2000. Forty-seven competitive local exchange carriers then went bankrupt between 2000 and 2003. The equipment had no secondary market. The financing became write-offs.
Nvidia's disclosed guarantee exposure is $108.5 billion on its own. That is roughly four times the entire nine-supplier telecom episode, inside one company. The structural difference in Nvidia's favour is that it is vastly more profitable than Lucent ever was and the agencies have tested the exposure. The structural difference against it is that Lucent's customers were building networks other carriers could use, while the PORTS-Pike collateral is contractually restricted to hosting Nvidia hardware.
Where this leaves the loss
No loss has been absorbed. What exists is a chain of credit substitution that terminates at Nvidia, the last and largest private balance sheet in the chain, holding the largest position.
The exposures are correlated. That matters more than their size. If the guarantee is called it will be called at the moment the chips stop selling and the $279 billion of supply commitments fall due, against collateral that only a buyer of Nvidia compute can use. Those are not three exposures diversified against each other. They are one exposure booked three times.
The June argument assumed the first absorber of a concentrated AI loss would be public, on the precedent of five federal interventions in eighteen years. It is private. It is also now identifiable from a filing. That does not retire the argument. It adds a step to it. The public balance sheet is no longer the first stop. It is the second. Section 3 of EO 14318 is already written for when it is reached.
Pike County closes the federal-land loop
June flagged Executive Order 14318 and the Department of Energy's 24 July 2025 selection of four federal sites, three of them former nuclear-weapons-complex or uranium-enrichment land, as a channel that removed the interconnection queue for hyperscaler-scale buyers only. That was a structural read with no named project attached. It also understated the order.
The Federal Register text (document 2025-14212, published 28 July 2025) has EO 14318 doing five things beyond shortening a queue. Section 3 directs the Commerce Secretary to "launch an initiative to provide financial support for Qualifying Projects, which could include loans and loan guarantees, grants, tax incentives, and offtake agreements". Section 5(b) has the Council on Environmental Quality establish new categorical exclusions for qualifying projects. Section 7(a) directs the EPA to expedite permitting by "developing or modifying regulations" promulgated under the Clean Air Act, the Clean Water Act, CERCLA and the Toxic Substances Control Act "that impact the development of Qualifying Projects", with 7(b) directing it to identify Superfund and Brownfield sites for their use. Section 8(a) puts Endangered Species Act section 7 consultation for ten years of common construction activity on a programmatic basis. Section 8(b) sends the Army Corps to determine whether an activity-specific nationwide permit is needed under section 404 of the Clean Water Act. Section 9(b) authorises the Defense Secretary to "competitively lease available lands for Qualifying Projects" on military installations under 10 U.S.C. 2667.
That is a standing instruction to reshape four environmental statutes around one industry. The TSCA line in 7(a) bears directly on the PFAS reporting cycle the June analysis listed as a signal to watch, since the agency instructed to expedite the projects is the same agency that would set the disclosure rule they fall under.
The $105 billion guarantee has one. It covers leases for approximately 4.25 gigawatts of IT load at SB Energy's PORTS-Pike Technology Campus in Pike County, Ohio, the decommissioned Portsmouth Gaseous Diffusion Plant, on land the developer describes as private and federal. The Department of Energy and the Department of Commerce are named collaborators (Nvidia Newsroom; OpenAI).
SB Energy builds, owns and operates under a 20-year lease to OpenAI. The campus targets 8 gigawatts of IT capacity with at least 10 gigawatts of new generation. SB Energy says it will fund $4.2 billion of regional grid infrastructure. Nvidia is investing a further $1.5 billion in SB Energy directly and holds an option to extend credit support across roughly 3.8 additional gigawatts. It secured a term under which the campus "will exclusively host NVIDIA AI infrastructure, subject to limited exceptions". Construction runs to 2032 on a stated 35,000 construction jobs and 2,500 long-term operating jobs. Community funds total $120 million.
The chain runs in a single line. Federal land and federal permitting acceleration sit on one end. A chip vendor's investment-grade credit standing guarantees the tenant's rent in the middle. An exclusivity clause returns the whole site to that vendor's hardware on the other end. Every element of it is disclosed. None of it required an appropriation.
The 2,500 operating-jobs figure is materially above the per-site headcount the June piece described. Against 8 gigawatts it still works out near 300 jobs per gigawatt. The number is a developer projection for a campus whose first phase is expected in 2028.
The other three federal sites are now live
June named the four DOE sites solicited on 24 July 2025 and treated them as a channel. Three have since become projects.
On 29 July 2026 the DOE announced a privately funded AI campus at the former Paducah Gaseous Diffusion Plant in Kentucky, put at $100 billion, with NextEra Energy developing 2 gigawatts of new grid-connected gas generation, transmission upgrades and as much as 2.6 gigawatts of battery storage. No tenant was named. At the Savannah River Site, the National Nuclear Security Administration selected Amentum to negotiate a phased lease for a 1 gigawatt data centre supported by about 2 gigawatts of on-site generation, starting on natural gas and moving to nuclear. Idaho National Laboratory and Oak Ridge have solicited and not yet selected.
PORTS-Pike was not on that list. It is a fifth site, on the sister enrichment plant to Paducah. It is also the first to reach a signed twenty-year lease with a named tenant and a $105 billion private guarantee. Of the five, four are former nuclear-weapons-complex or uranium-enrichment land. Each carries existing grid capacity, an existing security perimeter and federal control of the surrounding acreage. That combination is what lets a hyperscaler-scale buyer step past the interconnection queue described below.
The sixth siphon runs through the interconnection queue
Grid operators price capacity years forward against forecast peak load. Forecast peak load is built from interconnection requests. Filing an interconnection request has been close to free. Developers hedge multi-year queue delays by filing functionally the same 500 to 1,000 megawatt project with three or four utilities at once, building wherever a transformer arrives first and abandoning the rest (Utility Dive).
Wood Mackenzie's Q3 2026 national pipeline analysis puts 72% of the 1,066 gigawatts of requested US data-centre capacity in the phantom category. National requests top 700 gigawatts against an existing US data-centre fleet drawing roughly 60 to 70 gigawatts.
Texas acted on it. ERCOT interconnection requests went from about 48 gigawatts in 2023 to more than 474 gigawatts, roughly 90% of them data centres, more than five times the state's record peak demand. Governor Abbott and the Public Utility Commission froze new data-centre grid connections and ordered an audit of each project's financing, expected electricity use, water requirements, cooling, ownership and on-site generation plans, due December 2026 (Utility Dive). Texas is the first major hub to stop connecting.
The cost of the filings is already collected. PJM's forward capacity cost went from $2.2 billion for the 2024 to 2025 delivery year to $16.4 billion for 2027 to 2028. The 2026 to 2027 auction cleared at $329.17 per megawatt-day against $28.92 for 2024 to 2025. PJM's own Internal Market Monitor found that removing all data centres would cut capacity payments by $9.33 billion, a 64% reduction. Counting only energised data centres cuts them by $7.74 billion, or 53%.
Households in PJM are paying today, through the capacity line of a regulated bill, for generation reserved against load that the grid operators' own audits say will not arrive. The developer who filed four applications for one project paid nothing for the option. Utilities have begun charging $50,000 per megawatt in non-refundable collateral. That prices the option for the first time.
The ouroboros did not close the way we said
June put Blackstone, Brookfield, KKR and Apollo on both sides of a single trade, long AI infrastructure through their fund vehicles and long the office real estate that AI-driven displacement would hollow out. The prediction had three parts. The two legs were negatively correlated by construction. An AI payoff would cause the CRE writedown that cancelled it. The LP would eat the difference while the GP earned fees on both.
The 2026 disclosures say something different. What they say is worse for the LP.
The houses rotated. They did not sit on both sides and absorb the contradiction. They moved. BREIT, Blackstone's flagship non-traded real estate vehicle, held 26.1% of its real estate asset value in QTS data centres as of 30 June 2026 against 12.5% in office. Data centre exposure is now more than double office exposure inside the fund the June piece used as the example of the trap. BREIT deployed a further $5.7 billion into pre-leased data-centre development across 2026. Brookfield closed the first tranche of a dedicated AI infrastructure fund and set out a $100 billion campus in Kentucky. Apollo deployed more than $40 billion into next-generation data centres and infrastructure in 2025 and continued into 2026.
The AI leg is propping the office leg up. As set out above, AI tenants supplied a large share of the office absorption that did occur. The two legs the June analysis expected to move against each other have so far moved together, in the same direction.
And the risk is being packaged for third parties. On 10 August 2026, Nvidia announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to "mobilize over $500 billion of third-party capital for the buildout of AI infrastructure over time", with Nvidia compute as the underlying asset. Nvidia committed no capital of its own in that announcement. Its stated role is to "create dedicated pools of capital at significant scale at attractive rates for NVIDIA customers" (NVIDIA Newsroom).
Third-party capital means limited partners. Pension funds, endowments, sovereign wealth funds and insurers, including the life insurers the same houses own. The four firms the June piece described as unknowingly trapped on both sides of one trade are, sixteen months later, formally partnered with the chip vendor to raise half a trillion dollars of other people's money for the leg they rotated into.
So the correction runs in two parts. The mechanism June described, GP fees on both legs and LP exposure to the net, was right. The identification of who sits on the losing leg was wrong. The houses were not caught. They sold down the office exposure and concentrated into data centres. They are now originating the next round of AI risk onto LP balance sheets.
Who is left holding the office tail is a separate and less well-documented question. It is not the large diversified houses. On the available evidence it is smaller owners, the regional banks with commercial-real-estate concentration the June analysis flagged, the municipal tax bases that depend on assessed office value, plus whichever vintage funds bought what the large houses sold. None of those counterparties publishes a quarterly deck.
Who has mapped which part of it
No one has mapped the whole ouroboros. Three parties have mapped pieces that do not meet.
MSCI measured the concentration inside the AI leg. It came the closest. Its work on data centres in private markets covers 584 data-centre companies and assets held by closed-end funds in its Private Capital Universe, valuing the aggregate at $122 billion as of Q3 2025, with real assets at 60% and equity strategies at 36%. The finding that matters is that 13.6% of those data-centre companies were held by funds spanning more than one asset class. An LP holding an infrastructure fund, a real-estate fund and a buyout fund can therefore own the same asset three times. MSCI's conclusion is that "asset-class diversification therefore does not eliminate thematic concentration" because the allocations stay "anchored to a common economic driver: computational demand". That is the coordination failure the June analysis described, measured. MSCI does not extend it to the office book held by the same managers. That is the step that would close the loop.
Newmark quantified the labour-to-office leg. Its through-2030 work models AI slowing office-using employment growth by about 2.1 percentage points, leaving growth essentially flat at +0.3% across 2025 to 2030. It puts US office vacancy at 21.5% in 2030 against 19.8% in a no-AI counterfactual, with a 19.5% upside case and a 23.5% severe case. It also has the human-only task share in professional and business services falling from about half today to about a third by 2030, roughly two thirds of that from outright automation. That is a real effect arriving slowly. It is consistent with the lease gating set out above and inconsistent with the near-term cascade the June piece asserted. CBRE Investment Management has published a comparable exercise.
Sona mapped vendor circularity. That is the supply-chain circle, a different object from the allocation one.
The missing work is the join. Nobody has taken a single manager, or a single LP, then set its AI infrastructure exposure against its office exposure, its private-credit exposure and the insurance balance sheet that holds the paper. Nobody has then asked what the combined position is short of. The data to do it exists in fund reporting, statutory filings and MSCI's own universe. It has not been assembled in public. The parties best placed to assemble it are the ones being paid on both legs.
The circular system, counted
The worked example is one building. PORTS-Pike generates a single stream of economic activity. Nvidia appears in it four separate times. It guarantees the tenant's lease for up to $105 billion. It holds $1.5 billion of equity in the landlord. It supplies the hardware the site is contractually restricted to hosting. It "may earn a share of revenue generated by sales of the supported capacity". Four roles, one campus, one underlying demand.
| Disclosed obligation | Amount | Where it appears |
|---|---|---|
| Nvidia future commitments | $366bn | Note 10, 10-Q |
| Nvidia maximum guarantee exposure | $108.5bn | Guarantees table, 10-Q |
| Nvidia equity investments committed | $25bn | Note 10, 10-Q |
| Oracle lease commitments not commenced | $288bn | Note 6, 10-Q |
| Oracle borrowings | $125.3bn | Balance sheet |
| Oracle unconditional purchase obligations | $34.15bn | Note 6, 10-Q |
| Oracle remaining performance obligations | $638bn | Earnings release |
| OpenAI multi-year compute commitments | ~$1.4tn nameplate | Counterparty announcements |
| Hyperscaler bond issuance, 2026 | ~$250bn expected | Goldman Sachs estimate |
| Off-balance-sheet obligations, five firms | $1.65tn Nikkei, $1.2tn Moody's | Filing notes |
| Nvidia and private-equity financing platforms | $500bn of third-party capital | 10 August 2026 announcement |
| Sona's mapped AI financing | $3.6tn across 176 deals | Sona |
These figures cannot be added together. Oracle's $638 billion of performance obligations is largely the same economic activity as OpenAI's $1.4 trillion of commitments. That is largely the same activity Nvidia's $279 billion of supply commitments exists to serve. That is partly what the $108.5 billion of guarantees underwrites. They cannot be summed because one stream of expected AI demand is booked as revenue by one party, as an obligation by another, as a contingent liability by a third, and as an asset by a fourth. Each disclosure is individually correct.
Two parties have measured the overlap directly. Sona put $3.6 trillion through 176 deals between 202 companies and flagged about 120 as circular. MSCI found 13.6% of data-centre companies in its private-capital universe held by funds spanning more than one asset class. The supply chain double-counts the demand. The allocation layer double-counts the asset.
The dot-com mirror
The structure is the one that ran in 1999. The scale is not. The differences all point the same way.
| 1999 to 2002 | 2026 | |
|---|---|---|
| Supplier credit to customers | $25.6bn across nine suppliers | $108.5bn, Nvidia guarantees alone |
| Same, in 2026 dollars | $48.8bn | $108.5bn |
| Buildout, share of GDP | ~0.12% a year (CLECs) | ~2.24% in 2026 alone |
| Market value at peak | $18.3tn NYSE and Nasdaq | ~$76tn total US |
| Market cap to GDP | 150% to 172% at peak | 234.6% on 20 September 2026 |
| Top-cohort index weight | dot-coms largely outside the S&P | Mag 7 at 34%, Nvidia alone at 7.5% |
| Drawdown | $9.3tn, 91% of GDP | a 51% repeat is ~$39tn, 120% of GDP |
| Federal budget entering | surplus, 1998 to 2001 | 5.8% of GDP deficit |
| Debt held by the public | ~34% of GDP | 101% of GDP |
| Policy rate available | 6.5%, cut to 1.0% | ~4.1% |
| USD share of allocated reserves | above 70% | 56.9%, a 31-year low |
| Office vacancy entering | 8.6% in 2000 | 18.3% in Q2 2026, near record |
| Office CMBS delinquency | not a stress point | 12.34% in January 2026, a record |
| Bank CRE concentration | below the 300% guidance | 54.8% of regional banks above it, median 312% |
| CRE debt maturing | no comparable wall | $875bn in 2026, $652bn in 2027 |
The inflation adjustment on its own understates the change. CPI ran 172.2 to 328.0 between 2000 and 2026, a multiplier of 1.905. That makes Nvidia's guarantee exposure about 2.2 times the entire nine-supplier telecom episode in real money. Set against the economy it is 0.34% of GDP against 0.25%, or 1.4 times. Both numbers are smaller than the raw comparison suggests. Neither is where the scale sits.
The scale sits in the buildout. The competitive local exchange carriers consumed roughly $60 billion across five years. The Big Four alone will spend about $725 billion in 2026. That is 2.24% of a $32.38 trillion economy in a single year, against roughly 0.12% a year for the CLECs. The ratio is close to nineteen times. It is the number that governs everything downstream.
Why the forecast is not the same
The collateral depreciates and the buyer set is narrower. Dark fibre had no secondary market in 2002 either. It was eventually lit, since glass laid in the ground does not become obsolete. A GPU is on a three-year architectural cycle. The PORTS-Pike campus is contractually restricted to hosting one vendor's hardware. The 2002 overhang was an asset waiting for demand. This one is an asset waiting for demand from a narrower set of buyers while losing capability against each new generation.
The concentration is inside the index. In 2000 the Nasdaq fell 78% while the S&P fell 50%. Most of the speculative names sat outside the large-cap benchmark. A passive S&P holder took half the damage of a Nasdaq holder. That separation no longer exists. The Mag 7 are 34% of the S&P and Nvidia alone is 7.5%, the highest single-cohort concentration in the index's modern history. The front door the June work described has since loaded the passive holder into the epicentre by construction. No broad-market position avoids it.
The absorber starts from a deficit. The United States met the dot-com bust from a budget surplus, with debt held by the public near 34% of GDP and 550 basis points of policy rate to deploy. It would meet this one with a 5.8% deficit in a year with no recession in it, debt at 101% of GDP and about 410 basis points. The CBO's own projection has the primary deficit smaller in 2036 than today while the total deficit grows. Interest is now the entire story.
Commercial real estate was an absorber last time and is a second shock now. Office vacancy stood at 8.6% in 2000 and rose to 16.9% by Q1 2003. The system took that deterioration from a healthy starting point. It stands at 18.3% today before anything has happened. Office CMBS delinquency hit a record 12.34% in January 2026. Among office loans that matured before 2026 with balances still outstanding, 83.7% are delinquent and 92.7% sit in special servicing. Banks and thrifts hold $1.844 trillion of commercial and multifamily mortgage debt. Of regional banks, 54.8% sit above the 300% concentration guidance, at a median of 312%. Commercial mortgages of $875 billion mature in 2026, with $652 billion behind them in 2027. The CRE maturity wall, the office lease-expiry wall running to 2028 and the AI capex verdict all land in the same window. In 2002 the banking system had a clean balance sheet to absorb an equity shock with. It does not have one now. What it has instead is two new ways to hold the exposure without pricing it.
And the currency cushion is thinner. This site has already priced it. The dollar's share of allocated reserves has gone from above 70% in 2000 to 56.9% at Q3 2025, a 31-year low, with central banks net sellers. The Mild Take's currency framework scores the US reserve role +5 near-term and -3 long-term, the largest near-and-long split in the dataset. It ranks US currency 171st of 193. The US composite sits at 173rd with a negative skew. The framework's stated reason for the split is that the privilege rests on trust in the institutions behind the unit. A privilege depending on institutional trust cannot outlast the trust.
The erosion has continued since we measured it
Those May figures are four months old. The same instrument has been re-run on live data through 19 September. The direction has not held steady. It has accelerated.
| Refresh date | Living | Assets | Currency | Reserve near | Reserve long |
|---|---|---|---|---|---|
| 29 May 2026 | -5.29 | -4.78 | -4.10 | +5 | -3 |
| 18 June 2026 | -5.32 | -4.75 | -3.94 | +5 | -3 |
| 30 July 2026 | -5.32 | -4.92 | -4.31 | +5 | -3 |
| 14 August 2026 | -5.40 | -5.02 | -4.56 | +5 | -3 |
| 25 August 2026 | -5.40 | -5.14 | -4.92 | +4 | -3 |
| 3 September 2026 | -5.40 | -5.19 | -5.06 | +4 | -3 |
| 19 September 2026 | -5.45 | -5.36 | -5.39 | +4 | -3 |
The currency read has fallen 1.29 points since late May. All of it arrived after mid-July. The assets read is down 0.58 over the same period. On 25 August 2026 the near-term reserve-currency score was cut from +5 to +4, the first time the cushion itself has been marked down since the framework began scoring it.
The reasons recorded against that cut are monetary. The sub-factor notes cite the dollar index, the gold price and US employment data, with gold at roughly $4,380 in September against a $4,640 three-month high set on 25 August. The AI-capex structures documented in this piece are not the stated cause. No claim is made here that they are. What matters for the comparison is the level and the direction, both of which are worse than the figures the May analysis published.
The fiscal position and the currency position are connected
The rescue instrument already written for this is Section 3 of EO 14318 read with Section 5(c), a federal guarantee of up to 49.9% of project cost, requiring no appropriation, no roll call and no environmental review. Under the Federal Credit Reform Act it also enters the budget at its estimated lifetime subsidy cost on a present-value basis. The amount guaranteed never appears. That estimate excludes market risk. The CBO puts the gap at an average 5.7 percentage points below a fair-value figure. A guarantee stays off the public balance sheet until it is called, exactly as Nvidia's $108.5 billion and Oracle's $288 billion stay off theirs.
What made the 2002 response payable was not the budget. It was the buyer. The United States funded its answer by issuing Treasuries into a world that held the dollar as more than 70% of its reserves. The paper cleared. That share is now 56.9% and central banks are net sellers. The thing that has changed is who is on the other side of the trade.
A guarantee avoids the vote by not being visible. The visibility is the part that carries a cost. The currency framework scores institutional trust. The privilege is priced off that score. Reading a deliberately unscored, unvoted and unreviewed commitment as an institutional-trust withdrawal is our own inference. The framework does not output it. On that inference the response draws down the same account it exists to protect.
The currency piece describes reserve status as eroding gradually and then suddenly. It notes that the framework cannot time the second part. The table above records the gradual phase, in our own numbers, across four months.
The position differs from 2002 in four respects. The earlier episode destroyed $9.3 trillion against an economy that had a surplus, a central bank with room and a currency whose privilege was intact. A proportionate repeat now removes something closer to $39 trillion, from higher valuations, held more concentrated, by holders who cannot step aside. The state meeting it enters with a 5.8% deficit, three times the debt and 240 fewer basis points of policy rate. The banking system meeting it is already above its own commercial-real-estate concentration guidance with $1.5 trillion of mortgages maturing across two years. And the currency that funded the last answer has gone from above 70% of world reserves to 56.9%, with our own reading of it down 1.29 points since May.
The seventh channel runs through life insurance
The five siphons transfer value before a failure. The bailout precedent describes what happens after one. Between them sits a mechanism the June analysis did not have. It converts an AI credit loss into a state taxpayer liability automatically, with no federal action and no vote at any level.
The chain has four links. Private credit, roughly a $3 trillion market and the financing arm of the large private-equity houses, is a major lender to data-centre construction. Those same houses own life insurers, Apollo having acquired Athene and KKR having acquired Global Atlantic, with private equity controlling at least $1.5 trillion of life-insurer assets. Those insurers invest policyholder float into their parents' private credit. In 2024, 49.5% of new investment by private-equity-owned insurers went into privately placed instruments, against 14% for independent insurers.
The fourth link is the one that matters. If default rates on that paper run high enough to push an insurer into insolvency, state guaranty associations pay policyholder claims. Surviving insurers are then assessed to reimburse the fund. Most states then let the assessed insurer recover the payment through a credit against its state premium tax. The common structure runs at 20% per year across five years, netting to full recovery in the states written that way. The registry of those statutes is maintained by NOLHGA, the guaranty associations' own national organisation. It shows a patchwork across the states. The American Prospect's August 2026 treatment puts the full-offset figure at 44 states (American Prospect). We have not verified that count against NOLHGA's state-by-state tables. It is reported here as their figure.
The direction is not in dispute even where the count is. To the extent the offset applies, the assessment lands on the state's premium-tax revenue. That is the state's general fund, reached through a mechanism that requires no legislature to act.
This closes the ouroboros the June analysis left open. That piece put Blackstone, Brookfield, KKR and Apollo on both sides of one trade, long AI infrastructure and long the office real estate that AI displacement erodes. The third leg is the insurance balance sheet. It holds the credit risk of the first leg. State guaranty associations stand behind it. The GP collects fees on all three. The loss, if it comes, lands on policyholders, on state general funds and on the taxpayers who fund them.
Two caveats belong with it. The 15% default threshold cited for insolvency is an analyst estimate, with no regulatory standing. State guaranty associations have also absorbed insurer failures before without visible fiscal disruption.
What made the banks comfortable letting CRE go
The chain above describes where the credit risk ends up. It does not explain why a bank would originate into it, or why the large holders appear relaxed about commercial real estate at the levels set out earlier. Two mechanisms changed in 2026. They solve different halves of the problem.
The liquidity half is federal. It moved in September. Banks pledge whole loans to the Federal Reserve's discount window through the Borrower-in-Custody programme. That lets the loans stay on the bank's own premises while serving as collateral. Commercial real estate is one of the standing BIC arrangement types. On 6 August 2026 the Fed announced changes to that programme, effective 8 September 2026, under the heading of discount window modernisation. Pre-enrollment inspections were largely eliminated. Collateral schedules moved online through Discount Window Direct. Institutions no longer report when the balance of pledged loan collateral declines by more than 10 percent. Periodic inspections narrowed to three items. Separately, the loan collateral margins were reset effective 1 July 2026, with the tables rebuilt to publish the minimum, weighted-average and maximum margin observed on historical pledges in each category.
The distinction that matters is what the discount window does. It lends against the asset. It does not buy it. The bank keeps the loan and keeps the loss. What it stops having is the liquidity pressure that would force a sale. A sale is what produces a mark. So the federal change does not remove the exposure. It removes the event that would require the exposure to be priced. That is the same structure as everything else in this piece.
The credit half is private. It is where the risk actually goes. Significant risk transfer lets a bank keep a loan book while selling the credit risk on it synthetically. That releases regulatory capital against the same assets. The global pool of bank synthetic securitisations has passed $670 billion, with sales forecast to grow about 11% a year. Commercial real estate is now an explicit reference pool. BNP Paribas has discussed an SRT against roughly €1 billion of CRE loans. Aareal Bank completed a synthetic transfer referencing a €2 billion CRE portfolio, about 6% of its book. The natural buyers of the mezzanine tranches are private credit funds.
Those are the funds in the chain above. So the two halves connect. A bank can convert a commercial real-estate loan into central bank liquidity without selling it. It can then sell the credit risk on the same loan to a private credit fund owned by a house that also owns a life insurer. If that insurer fails, a state guaranty association pays its policyholder claims and recovers the cost through state premium-tax credits.
Neither step requires anyone to sell a building. Neither step requires anyone to book a loss. The loan stays on the balance sheet at its carried value. The capital against it is released. The credit risk has moved to an entity that a state guaranty association stands behind.
That reading of why the large holders appear comfortable is ours. The Federal Reserve's stated purpose for the BIC changes is programme efficiency. No part of the announcement references commercial real estate conditions. The SRT market is a response to Basel capital rules. What is documented is that the two mechanisms exist, that both changed materially in 2026, and that they meet in the same set of balance sheets.
What a margin is, in plain terms
A bank holds a commercial real-estate loan carried on its books at $100. Selling that loan would establish what it is actually worth. If a buyer pays $70, the bank books a $30 loss. The loss appears in its results. The price becomes a reference point every other holder of similar paper is measured against.
Pledging it to the discount window produces cash without any of that. 1. The Federal Reserve makes its own estimate of the loan's fair market value, recalculated monthly, from the loan's cash flows and credit characteristics plus market data. That estimate is not the bank's carrying value. Call it $80. 2. The Fed then applies a margin to its own estimate. The margin is a percentage reflecting how much that category of loan has historically moved in price over the period the Reserve Banks assume it would take to liquidate it. 3. Collateral value equals the fair-value estimate multiplied by the margin. That is what the bank may borrow against.
The published margins for commercial real-estate loans at in-scope institutions, effective 1 July 2026, run from 95% at one-year maturity down to 51% at twenty-five years for floating-rate, minimal-risk, fully amortising paper. The fixed-rate equivalent runs to 59% at thirty years. Construction loans at normal risk run from 85% at one year down to 17 to 21% at thirty years.
So on a mid-range CRE loan the bank might pledge a $100 loan, against which the Fed estimates $80 and lends perhaps 60% of that. The bank receives roughly $48 in cash. The loan remains on its balance sheet at $100. No loss is recognised.
The part that matters is what happened to the $80. A price was discovered. The Federal Reserve marked that loan monthly and the bank knows the number. It is simply not published. It does not enter the bank's income statement. No other holder of similar paper is repriced by it. Price discovery moved from the market, where it is public and contagious, into a bilateral relationship with the central bank, where it is neither.
"Margins reset effective 1 July 2026" therefore means the Fed republished those percentages across every loan category, maturity band, rate type and risk grade. The tables were also rebuilt to show the minimum, weighted-average and maximum margin actually observed on historical pledges in each category. An institution can now forecast how much cash a given pool will raise before it pledges anything. That is the change. It is not a rescue. It is a published, predictable price for turning an unsellable loan into cash without selling it.
Where the risk reaches a public ledger
The Fed does not take the credit risk. It lends against collateral it has already discounted twice. It is senior. The margin exists precisely so the Reserve Banks do not lose money if the borrower fails. The public exposure through the discount window is the residual risk that the haircut was set too low. It is real and it is small. Nothing in the BIC change moves a commercial real-estate loss onto a federal balance sheet.
What the mechanisms in this piece mostly do is keep risk off public ledgers. That is the design. Nvidia's $108.5 billion sits in a note. Oracle's $288 billion sits in a note. Significant risk transfer moves credit risk to private funds. A federal guarantee under EO 14318 enters the budget at FCRA subsidy cost. Section 5(c) keeps it under the threshold for environmental review.
Already on public ledgers, as realised cost. State tax expenditure, running above $1 billion a year each in Georgia, Virginia and Texas with fourteen states not disclosing. Foregone federal receipts, at $70 billion of breaks across five firms with $27 billion of it accelerated depreciation. Ratepayer capacity charges, with PJM's forward capacity cost going from $2.2 billion to $16.4 billion on forecasts built from interconnection requests that are 72% phantom. Federal land, committed at five sites.
Contingent, recorded nowhere yet. EO 14318 Section 3 guarantees enter at subsidy cost when used. State guaranty-association assessments reach state general revenue through premium-tax offsets only if an insurer becomes insolvent. Then there is the deposit insurance fund.
That last one is the direct commercial-real-estate path and it belongs here. Banks and thrifts hold $1.844 trillion of commercial and multifamily mortgage debt. The FDIC's Deposit Insurance Fund stood at $157.4 billion at 31 March 2026, with a reserve ratio of 1.48% against a designated target of 2%. The fund covers insured deposits at a failed bank and is replenished by assessments on the surviving banks. That is the same recovery structure as the insurance guaranty associations. The March 2023 systemic risk exception established that the coverage can be extended past the $250,000 cap when regulators decide the failure is systemic.
So the answer to where the risk lands is not one place. The realised transfers are already on state and federal books and on household utility bills. The risk itself has been deliberately parked in notes, private funds and contingent facilities. It reaches a public ledger only on a trigger. The triggers are a called guarantee, an insurer insolvency, or a bank failure. None of them has fired.
The transfer runs past the Fed
The intuition is that the Federal Reserve lends more than the collateral is worth, leaving the Fed to carry the difference on a failure. The arithmetic runs the other way. The Fed discounts twice, first to its own fair-value estimate and then by the margin. It hands over roughly $48 against a loan the bank carries at $100. For the Fed to lose a dollar, that loan has to be worth less than half its carried value at the moment of seizure. The Fed also does not fund these loans out of anything it has to hold back. It creates the reserves. The binding constraint is the quality of the collateral, never the availability of the money.
Advances from the Reserve Banks and the Federal Home Loan Banks are secured. The Federal Deposit Insurance Act prevents the FDIC from disallowing either claim. The FHLBs hold a super-lien and are repaid ahead of both the FDIC and the Fed, frequently under a blanket lien across all of a member's assets. That lien must be subrogated before anything can be pledged at the discount window at all. The consequence is stated plainly in the policy literature. FHLB lending limits the assets the FDIC has access to when resolving a failed bank. It exposes the FDIC to more loss while the FHLBs stay fully protected.
So the effect of pledging is not that the central bank absorbs the risk. It is that the good collateral leaves the estate before the FDIC gets to it. The receiver is left with whatever was not worth pledging. The Fed is made whole. The FHLBs are made whole ahead of the Fed. The shortfall lands on the deposit insurance fund.
2023 priced that sequence. The Fed lost nothing. The FHLBs lost nothing. The Deposit Insurance Fund absorbed roughly $31.5 billion across three banks, at $16.1 billion for Silicon Valley Bank, $2.4 billion for Signature and $15.6 billion for First Republic. Around $16.7 billion of it is attributable to covering uninsured deposits under the systemic risk determination of 12 March 2023. Statute requires the FDIC to recover that through special assessments on the surviving banks, who price it back into their own customers. The remainder hit the fund directly.
The public ledger holding the risk is the FDIC. It ranks behind the Federal Reserve and the Federal Home Loan Banks. The fund stood at $157.4 billion at 31 March 2026, a reserve ratio of 1.48% against a 2% target, against banks and thrifts holding $1.844 trillion of commercial and multifamily mortgage debt. It carries a $100 billion line of credit with the Treasury behind it.
The distinction changes what to watch. A rising discount window balance is not the public taking on commercial real-estate losses. It is collateral moving to the front of the queue, ahead of the insurer of last resort, before anything has failed.
Is the repo market buying bad CRE paper
The Standing Repo Facility does not take commercial real-estate collateral. Its eligible securities are US Treasuries, agency debt and agency mortgage-backed securities. Agency CMBS does not appear on that list. The facility also lends against collateral. It buys nothing in any case. One caveat on confidence. The New York Fed's FAQ states the three categories and refers to a separate Repo Securities Schedule for the detail. We were not able to read that schedule. The treatment of agency CMBS inside it is asserted from the standard reading and not verified.
The Fed did buy commercial mortgage-backed securities outright. It still holds them. Between March 2020 and November 2021 the Desk purchased agency CMBS, defined as securities secured primarily by multifamily mortgages and guaranteed as to principal and interest by Fannie Mae, Freddie Mac or Ginnie Mae. Holdings peaked at $9.943 billion on 14 April 2021. The Desk was directed to stop purchasing on 4 November 2021. The System Open Market Account still held $7.531 billion on 22 July 2026. That is commercial mortgage risk on the central bank's own balance sheet. It is small, closed to new purchases and running off. Treat it as a precedent.
The large public ledger for commercial real estate is not the Fed at all. It is the guarantee. Fannie Mae and Freddie Mac carry a combined guaranty book of roughly $7.8 trillion, with Fannie's alone at $4.136 trillion in May 2026. The multifamily portion of it is commercial real estate by any definition. Both have been in federal conservatorship since 6 September 2008, backed by Treasury senior preferred stock purchase agreements. Treasury has already injected $187 billion into them once. A guarantee written by an entity the federal government controls and stands behind is a federal contingent liability by construction, whatever the accounting says.
The number that matters for how close that is to firing is the delinquency rate. It is currently low. Fannie's multifamily serious delinquency rate sat at 0.58% in May 2026, down six basis points on the month. The distress documented earlier in this piece is concentrated in office. The multifamily modifications that CRED iQ tracks sit in private-label CMBS and CRE CLOs, a different book from the agency one, with different underwriting.
So the shape is this. Office credit risk sits with banks. It reaches the public through the deposit insurance fund if a bank fails. Multifamily credit risk sits substantially with two federally controlled guarantors. It is already on a public ledger before anything goes wrong. Neither route runs through the repo facility.
What the carrying value is holding up
The loan stays on the books at $100 and that figure does work beyond the loan itself. It sits in the asset base and it is risk-weighted. The capital ratio is capital divided by those risk-weighted assets. Marking the loan down takes the loss out of capital and out of the denominator at the same time. Capital is the far smaller number. The ratio falls.
A bank with $100 of capital against $1,000 of risk-weighted assets runs a 10% ratio. Writing $18 off that book leaves $82 against $982, or 8.35%. A markdown of under two percent of assets costs roughly 165 basis points of capital ratio. Run it the other way and the leverage is clearer. At a 10% target, each dollar of capital carries about ten dollars of risk-weighted assets. An $18 loss left unrecognised is therefore holding up something near $180 of balance-sheet capacity.
The reserving against that is thin. Credit-loss allowances across all bank loans ran about 1.84% in early 2025. The figure for commercial real estate specifically was 1.61%, below its 1.65% peak in the third quarter of 2024. Office CMBS delinquency sits at a record 12.34%. Gupta and co-authors estimate a 39% decline in the long-run value of New York office buildings. Loans held for investment are carried at amortised cost. None of this is irregular. It is what the accounting standard permits.
The Federal Reserve has measured the behaviour this produces, using its own supervisory data. Crosignani and Prazad's New York Fed staff report, number 1130, first published in October 2024 and revised in June 2026, works from loan-level filings on CRE mortgages at stress-tested banks. Their finding is that weakly capitalised banks extend distressed CRE loans and grant payment relief specifically to preserve capital in place of recognising losses. Conditional on comparable borrowers, properties and loan terms, distressed mortgages at less-capitalised banks are more likely to receive maturity extensions, more likely to receive a lower rate or a switch to interest-only, less likely to default. They are assigned a lower probability of default by the lender itself.
Reported credit quality is an artefact of the extending. The authors note that aggregate CRE non-performing loans remained low by historical standards as of the fourth quarter of 2025, despite the valuation decline across the same period. The number is low because the loss has not been booked.
The extensions build the wall they were meant to avoid. As of the fourth quarter of 2024, CRE loans maturing within three years accounted for 37% of marked-to-market capital at less-capitalised banks against 27% at better-capitalised ones. In 2020 there was no significant gap between the two groups. The banks least able to absorb a loss have concentrated the most rollover risk into the nearest years.
The capacity is not being used for new lending. Extend-and-pretend ties capital up in legacy exposures. The authors estimate it produced a 1.1% contraction in aggregate CRE mortgage origination, with no equivalent effect in commercial and industrial lending. So the balance-sheet capacity the carrying value preserves is not financing anything. It is holding the old position open.
The exposure is concentrated where the capital is thinnest. CRE mortgages are 9% of assets at the stress-tested banks in the supervisory sample and 31% of total assets at regional banks, with the median bank holding company carrying CRE at 47.8% of total loans as of the fourth quarter of 2025.
What actually forces the loss
The gap between $100 and what the loan is worth does not matter until something makes the bank recognise it. Three things can.
Maturity. The borrower cannot repay the balloon and cannot refinance at current rates and values. The bank extends. That is the documented response. The loan returns in two or three years at a larger size, in a more concentrated wall.
The examiner. A supervisory classification forces reserving. That is discretionary and slow. It is also the channel most exposed to political weather.
Deposit outflow. Money leaves and the bank needs cash. With nothing left to pledge it has to sell. The sale sets a price. The price forces the mark. The mark hits capital. The capital hole triggers the run that finishes it. That is the sequence Silicon Valley Bank followed.
Only the third sits outside the bank's control. That is precisely the one the discount window neutralises. Pledging converts a loan into cash without a sale. The forcing event never arrives.
That connects back to the collateral priority set out above. Pledging buys time against the only trigger a bank cannot manage. It also moves the best collateral behind a senior lien, out of reach of the receiver. A bank that survives longer by pledging more leaves less for the receiver. The mechanism that prevents a failure in one quarter raises the cost of a failure later. That cost falls on the deposit insurance fund. What is new is the concentration and the correlation between the assets. The collateral is also unusual. A data centre holds its value only if the AI revenue arrives.
The demand side still cannot be measured
The June analysis set the enterprise return question as the thing the capex ultimately rests on. It reported the 2024 MIT NANDA finding that 95% of GenAI pilots produced no measurable return. Fifteen months later, adoption has risen sharply and measurement has not.
McKinsey's State of AI 2026 puts 78% of organisations using AI in at least one business function, against 55% two years earlier. Roughly three quarters of the largest enterprises run at least one AI system in production. Against that, between 5% and 8% of companies report a measurable, at-scale financial return. Fewer than one in ten can point to a generative-AI deployment delivering measurable sustained value at scale. More than two thirds rely on estimates in place of measured financial results. PwC's January 2026 survey of chief executives found 56% reporting neither revenue growth nor cost reduction from AI over the preceding twelve months, with 12% reporting both.
Gartner continues to forecast that more than 40% of agentic AI projects will be cancelled by the end of 2027, while also projecting that 40% of enterprise applications will embed task-specific agents by the end of 2026. Those two forecasts are consistent with each other and with everything above. Deployment is broad. Attribution is rare.
None of that settles whether the returns exist. It does establish that after three years and roughly $1.5 trillion of cumulative capital commitment, the buyers of the output still cannot say in aggregate what it earned them. The capex rests on the same expectation it rested on in June.
Siphon by siphon
Siphon 1, utility ratepayers. Partly validated. The causation is narrower than we wrote. The national residential average is 18.34 cents per kilowatt-hour as of September 2026, up 5.0% year over year, against a rise of about 27% since 2019. One-year state moves run 16% in Illinois, 13% in Virginia and 12% in Ohio. June presented a 6% data-centre-attributable delta as a national figure. The defensible attribution is jurisdictional. Inside PJM, the market monitor puts 53 to 64% of the capacity price spike on data centres. Inside ERCOT, forward power prices rose 11 to 17% on comparable data-centre growth. SemiAnalysis reads that contrast as evidence that PJM's forecasting and auction design carry a large share of the move (SemiAnalysis). Both readings can be true. The load is real. The auction converts a forecast of that load into a bill. The design of the conversion decides how much lands on a household.
Siphon 2, adjacent communities. Validated, at several times the June figure. June cited $64 billion of projects blocked or delayed between May 2024 and March 2025. Data Center Watch counted 75 projects worth about $130 billion blocked or delayed in the first quarter of 2026 alone, the most in any quarter since tracking began in 2023, then 45 projects worth $68 billion in the second (NBC News; Bloomberg). Active opposition groups more than doubled to 833 across 49 states. More than 300 state bills were filed in the first six weeks of 2026. Moratorium bills appeared in 14 states in the first quarter. Senator Sanders and Representative Ocasio-Cortez filed a federal version. At least five local ballot measures go to voters in California, Michigan, Nevada and Wisconsin. An Ohio citizen initiative would bar data centres drawing 25 megawatts or more. SemiAnalysis argues the blocked-and-delayed denominator conflates announcements with executable projects. It reports that its own year-end 2026 hyperscaler self-build forecast moved about 1% in six months. Those two findings sit together without contradiction. Opposition is stopping speculative announcements at a high rate and has not yet moved the hyperscalers' own construction schedule.
Siphon 3, taxpayers. Validated, with a larger federal number. ITEP's 2026 work covers five firms, one more than the June count. Amazon, Alphabet, Meta, Microsoft and Oracle reported $422 billion of profit and paid an effective federal income tax rate of 4.5%, taking $70 billion in breaks against the 21% statutory rate, of which $27 billion came from accelerated depreciation (ITEP). At the state level, Georgia's projected annual sales-tax loss went from $186 million in 2023 to nearly $1.9 billion by 2026. Indiana disclosed $655 million lost through its data-centre exemption as of 2025, with more than 83% of the 2025 subsidy going to Amazon. Fourteen states still publish no aggregate revenue-loss figure. S.5054 would strip AI data centres of bonus depreciation and require electricity and water reporting. That attacks the largest single line of the federal channel directly.
Siphon 4, investors. Validated, with SpaceX as the demonstration. Covered below.
Siphon 5, other businesses. The input side validated. The demand side is not yet measurable by the series we used. Grid crowding and hardware allocation held. On the office side, Q2 2026 US vacancy fell 30 basis points to 18.3%, the largest quarterly decline since 2015, with prime vacancy down 40 basis points to 12.3% and Midtown Manhattan prime at 2.2% (CBRE). Our first reading of that was that the June thesis had failed. Four properties of the series say otherwise.
Lease gating. Occupancy is a contractual fact. A firm that cuts a third of its staff pays rent to the end of its term. More than 32% of all US office leases expire between Q2 2026 and year-end 2028. Newmark counts roughly 1.4 billion square feet of pre-pandemic leases rolling between 2025 and 2027. No vacancy print before that window can carry information about 2026 headcount decisions.
The absorption is the subject of the experiment. CBRE counts 30.4 million square feet leased by AI companies since 2019, 21 million of it in the San Francisco Bay Area and 9.4 million across Manhattan, Boston and Seattle (CBRE). Technology's share of US office leasing went from 16.8% across 2025 to 22.7% in Q1 2026. Of San Francisco's 1.6 million square feet of Q1 net absorption, AI companies were close to 40% of deal volume. Colliers puts AI at more than a third of tech-sector demand, against 12% in 2025. Using office absorption to test the AI-displacement thesis while AI tenants supply a large share of that absorption measures the buildout against itself.
Denominator withdrawal. CBRE tracks 81 million square feet in the US conversion and demolition pipeline, about 90,300 residential units. Completed conversions ran roughly 13.1 million square feet in 2024 and 12.8 million in 2025. That made 2025 the first year since 1988 in which removals outpaced completions. Only 2.2 million square feet completed in Q2, the lowest first half since CBRE began tracking in 1990. Converting a vacant building to housing lowers the vacancy rate arithmetically, with no tenant involved.
Price. The rate says nothing about the rent. On a 2026 lease carrying meaningful concessions, net effective rent runs 25 to 35% below face rent, with six to eighteen months free and tenant-improvement allowances from $10 to $100 per square foot. CBRE's series has effective rents up 5.2% for top-tier buildings and down 1.2% for lower-tier. The tightening is in trophy space, where the AI tenants are. Class B and C, the subject of the June argument, is still repricing downward. Sublease availability falling to its lowest since Q2 2020 is consistent with recovery and equally consistent with sublease space rolling into direct vacancy or off-market at term end.
The June piece wrote that the cascade was "already visible in the office-vacancy numbers". It was not visible then and it is not visible now. The correct statement is that the instrument cannot resolve the signal until the leases roll. We should not have cited it in either direction.
The labour series we cited was the wrong one. Aggregate JOLTS layoffs and discharges sit near 1.1 to 1.2%, close to a historic low. That is the figure we reported. That aggregate spans roughly 160 million jobs dominated by high-churn sectors whose churn fell. Inside it, the information sector's layoff rate went from 1.3% to 2.4% over the year, the largest increase of any sector and close to five times the national average. Hiring in professional and business services fell by 188,000, the largest single-sector decline in the report. The quits rate fell to 1.9%, last seen in 2014.
The announced-cut series points the same way. Challenger, Gray & Christmas counted 139,156 technology job cuts through June 2026, up 83% on the 76,214 through June 2025. AI was named in 101,743 announcements across all sectors through June, about 23% of all cuts, against 54,836 for the whole of 2025. It led every stated reason for five consecutive months through July. Layoffs.fyi counted 128,536 tech job losses across 299 companies globally by 10 September 2026, more than the whole of 2025 with a quarter still to run.
Oracle is the single cleanest case. Its annual report filed 22 June 2026 puts global headcount at 141,000 as of 31 May 2026 against about 162,000 a year earlier, a reduction near 21,000 people and 13% of the company, with $1.84 billion of severance and exit costs. It opened another round in mid-September. The most leveraged AI-infrastructure builder in the cluster, rated BBB and two notches above junk, shed an eighth of its staff in the year it was borrowing hardest to build.
The age data is unchanged. Unemployment for ages 20 to 24 was 7.1% in August 2026 against 4.1% overall. Stanford finds employment for 22 to 25 year olds in the most AI-exposed occupations down 13% since 2022, with software developers in that band down nearly 20% from 2024 peaks. Against that, entry-level headcount grew 12% at the firms making the largest AI investments over the two years after adoption. Economists reading the same series attribute much of the graduate weakness to a frozen hiring market (SIEPR; NPR). The displacement is running through both doors, a frozen entry level and a sectoral layoff rate that doubled. Neither empties a leased floor before the term ends.
The index gate, with a full price history attached
S&P Dow Jones reaffirmed twelve-month seasoning, GAAP profitability and the float requirement on 4 June 2026 and declined the six-month waiver. It has held. Nasdaq and FTSE Russell had already widened. CRSP admitted SpaceX five trading days after listing. Nasdaq-100 added it on 7 July 2026, fifteen trading days after listing, under the fast-entry rule adopted in May 2026 that also eliminated the 10% minimum float requirement.
| Date | Event | Price |
|---|---|---|
| 11 June 2026 | IPO priced | $135 |
| 12 June 2026 | First day open, then close, about $2.1T | $150, then $161 |
| Within three days | Peak close, about $226 intraday | about $211 |
| 7 July 2026 | Nasdaq-100 inclusion effective | about $160 |
| Mid-July 2026 | Below the IPO price | toward $125 |
| 20 September 2026 | Last, about 34% below the high | $152.71 |
JPMorgan estimated about $4.3 billion of forced passive buying from Nasdaq-100 trackers, with estimates up to $27 billion across Nasdaq-100 and Russell trackers, against a public float of 281 million of 7.57 billion shares, roughly 3.7% (ETF.com; CNBC). The eighteen banks that underwrote the deal published a median $225 target as the stock began its decline.
Index funds bought near $160 on 7 July because a rule said to. Within about ten days the stock traded toward $125. Index funds bought at roughly $160 on the inclusion date. The price fell about 22% over the following ten days.
What we got wrong
The office-vacancy cascade, twice over. June called the cascade already visible in the vacancy numbers. Our first pass at this recheck called it reversed. Both readings were wrong in the same way. Vacancy is gated by lease expiry. The absorption that did occur is substantially AI tenants. The inventory denominator is shrinking through conversion. The headline rate carries no information about effective rent. The series cannot answer the question in either direction before the 2026 to 2028 expiry window clears.
Citing an aggregate layoff rate against a sectoral claim. We reported private-sector layoffs near their lowest level since 2013. The thesis concerns white-collar technology employment, where the information-sector layoff rate doubled over the same period and technology announcements ran 83% above prior year. The aggregate was the wrong instrument and we should have disaggregated before citing it.
National ratepayer attribution. We presented a data-centre-attributable price delta as a national figure. It holds inside PJM on the market monitor's own numbers and does not transfer cleanly to ERCOT.
The RTX Spark claim. The June hardware-moat section leaned on Nvidia's stated capability of 120 billion parameters at one million tokens of context on a consumer Windows laptop shipping in the autumn. Nvidia now says the first RTX Spark Windows PCs arrive in October 2026. No third party has published a benchmark. That claim was vendor marketing and should not have carried weight in the argument. The thesis it supported survives on hardware that shipped and on models that are independently measurable. Both are covered below.
The single revenue verdict mark. June set one test, roughly 30% quarter-on-quarter sustained across cloud AI lines and model labs together. The two halves have separated. The labs clear it. OpenAI went from about $24 billion of run-rate revenue in June to more than $40 billion in August. Anthropic went from $47 billion in late May to $65 billion in late July. The hyperscaler cloud lines do not. Google Cloud grew 82% year over year in Q2 2026 to $24.8 billion, accelerating from 63%. AWS grew 37% to $42.2 billion, its fastest in eighteen quarters. Azure grew 43% in constant currency past $100 billion annualised. Those are roughly 8 to 16% quarter on quarter. A single blended number hid a real divergence. The divergence matters. The capex sits on the cloud side of it.
The issuance estimate. Low by about half.
South Korea separated the two questions
South Korea ran the experiment that no amount of argument settles. The question is whether an AI equity repricing and an AI demand slowdown are the same event.
The KOSPI went from about 5,000 in March 2026 to above 9,000 by June. In late July it fell more than 16% across two sessions, removing roughly $2 trillion of market capitalisation. Samsung was down 45% and SK Hynix down 55% from their June peaks at the trough, two firms that together hold close to half the index's market capitalisation. On 31 July the index posted a 17.9% single-day gain, the largest in the exchange's history. It closed at 6,820.02 on 31 August (CNBC; Al Jazeera).
The transmission was leverage. Korean retail investors had taken large positions in newly available leveraged single-stock ETFs on the two memory makers. When prices faltered in late June, the funds were forced to liquidate into a falling market. That triggered margin calls on the underlying holders, producing more forced selling (Noahpinion).
The physical business did not participate. Semiconductors were 43.8% of total Korean exports in June 2026 and a record $46.65 billion in August, 47.5% of goods exports that month. Memory is roughly 90% of semiconductor exports, against about 70% in prior cycles. Nominal GDP grew 21.9% in the first half of 2026 with almost 70% of the expansion coming from semiconductors. HBM3E in 36 gigabyte form has traded on the spot market at about $2,100 against long-term contract prices four to five times lower. Both makers raised HBM3E supply prices by close to 20% for 2026 (Seoul Economic Daily).
A market can remove $2 trillion from the price of two companies in two sessions while those companies sell every unit they can make at rising prices. The equity claim and the cash flow are separable. Korea separated them in public.
The read that matters for the US is the direction of the exposure. South Korea is now the most AI-concentrated national economy in the world by export share, by index concentration and by contribution to nominal growth. The Bank of Korea has flagged the AI chip trade as a financial stability concern on those grounds. Korea sells the memory that goes into the racks a year before the racks earn anything. A pause in US hyperscaler capex therefore arrives in Korean GDP before it arrives in a US earnings call.
The open-weight floor is measurable, the consumer hardware claim is not
The June argument was that the cloud's pricing power depends on frontier capability staying out of reach locally. It said that condition was being dismantled. Splitting that into what can be measured and what cannot gives two different answers.
On models, the gap is measurable and it is closing. The reported distance between the leading open-weight model and the leading proprietary model on Artificial Analysis's intelligence index has been quoted at 4 to 9 points across the middle of 2026, depending on index version and release date, against 13 points twelve months earlier. Four of the five leading open-weight entries come from Chinese labs, DeepSeek, Moonshot, Zhipu and Alibaba. Kimi K3 shipped with open weights on 27 July 2026. DeepSeek-V4.1-Flash shipped on 10 September 2026 under an MIT licence. SWE-bench Verified is close to saturated among frontier open models. The discriminating evaluations have moved to SWE-bench Pro, Terminal-Bench and DeepSWE. The index-version skew across those quotes is real and the direction across all of them is the same.
On consumer hardware, the claim we used cannot currently be checked. We have replaced it with work we can check ourselves.
We run one gemma4:26b runner on a single consumer-tier GPU in the lab. It serves Open WebUI, a coding agent, a Home Assistant voice agent, a doorbell agent and voice search concurrently, at 65,536 tokens of context across three parallel slots, from one resident copy of the weights. Separately, Legion runs Qwen3-235B-A22B in a browser by streaming experts per tensor across a peer mesh, validated live on 23 July 2026 and written up in a 235B model in a browser. Neither of those is a benchmark score and neither is a vendor claim. They are two production deployments in a home lab. That is the relevant altitude for an argument about whether cloud rental is the only way to reach useful inference.
The bailout question has a date on it
The June framing was that the buildout is rational because the federal government has absorbed concentrated losses five times in eighteen years. The question now is whether that absorption arrives before the 3 November midterms.
On the record through 20 September, the political movement runs the other way, hard and on a roll call.
The Ratepayer Protection Act passed the House 417 to 3 on 16 September 2026, sponsored by Representatives Castor and Evans, after clearing the Energy and Commerce Committee 52 to 0 in July. It requires large-load data centres to pay the cost of the new energy infrastructure built to serve them. States without a data-centre-specific standard must open hearings (Utility Dive). The three votes against came from the left, on the grounds that it does too little.
The Senate did not pass it by unanimous consent the next day. Senator Heinrich blocked it as a half measure. He is promoting the GRID Savings Act of 2026, giving FERC rulemaking jurisdiction over facilities drawing 150 megawatts or more (CNBC). The disagreement in the Senate is over how much further to go.
Electricity prices are a named midterm issue in polling and in campaign material, with data-centre siting attached to it. Candidate sentiment toward these facilities has moved negative in both parties (Brookings). No member of Congress facing voters in six weeks has an incentive to appropriate money for a data-centre operator.
The counterparty concentration belongs in one place. OpenAI has signed multi-year compute and data-centre commitments with a nameplate value above $1.4 trillion, across Broadcom, Oracle, Microsoft, Nvidia, AMD, Amazon and CoreWeave, against a revenue run rate that passed $40 billion in August 2026. It filed a confidential draft S-1 on 8 June 2026 and was valued at $852 billion by its $122 billion round on 31 March 2026. Reporting through the summer put a targeted listing near $730 billion, below that private mark, with internal planning moving toward 2027.
Every structure in this piece terminates at that balance sheet. Nvidia's $105 billion guarantee is on OpenAI's leases. Oracle's $288 billion of unstarted data-centre leases are substantially to serve it. Sona's circularity map runs through it. A single private company with roughly $40 billion of revenue is the counterparty to obligations that three public companies have disclosed in their notes.
Two further facts bear on the shape of what happens after. Nothing currently requires rescuing. Nvidia earned $59.688 billion in the June quarter. The labs are growing revenue faster than the clouds. The distress that exists is concentrated in the leveraged neoclouds. CoreWeave's credit default swaps have traded around 855 basis points. Oracle, CoreWeave and Nebius together carry about $173 billion of debt at roughly 7.8 times quarterly revenue, against Microsoft's 0.4 times.
And the federal posture has already changed from guarantor to shareholder, though not in the shape most coverage gives it.
The government holds 9.9% of Intel, 433.3 million shares bought at $20.47 in August 2025, funded by $5.7 billion of remaining CHIPS Act grants and $3.2 billion of Secure Enclave money, with a five-year warrant over a further 5% if Intel ceases to own a majority of its foundry business. That is confirmed in Intel's own filings.
The chip-revenue arrangement is the one to be careful with. Widely repeated reporting says Nvidia and AMD pay the Treasury 15% of China AI chip revenue. That was the August 2025 proposal covering the H20. It was superseded in January 2026 by a 25% arrangement covering the H200, under which Commerce cleared roughly ten Chinese firms to buy up to 75,000 units each. Nvidia's own 10-Q describes the mechanism as a 25% tariff on importation into the United States. It says it has been unable to pass that tariff to customers. It reports shipments under the programme at "less than 1% of Data Center revenue in our most recent quarter", against a $0.4 billion charge for H200 excess inventory and purchase obligations. A Commerce official told Congress on 14 July 2026 that shipments were "very few". The channel exists on paper. It is currently immaterial to the revenue it was described as taxing. Repeating the 15% figure in 2026 restates a rate that was never the operative one. OpenAI's own April 2026 policy paper proposed a Public Wealth Fund seeded by equity donations of 1 to 5% from AI companies, against a valuation of $852 billion set by its $122 billion round on 31 March 2026.
The instrument that would deliver a rescue without a vote is Section 3 of EO 14318, quoted above, read against Section 5(c) of the same order. Section 5(c) states that "Federal financial assistance representing less than 50 percent of total project costs shall be presumed not to constitute substantial Federal control and responsibility", placing such assistance outside the definition of a major federal action under NEPA. A qualifying project clears $500 million of capital expenditure or 100 megawatts of incremental load.
Taken together, the Commerce Secretary may extend a loan guarantee covering up to 49.9% of a data-centre project's cost under an initiative that already exists, with no appropriation, no roll call, no environmental impact statement. The 50% threshold is not a ceiling on federal exposure. It is the line below which federal exposure is presumed not to require examination.
A rescue of the recognisable kind is not on the table before the midterms. There is nothing yet to rescue and no member wants to cast the vote. The version requiring neither a vote nor a review is already authorised. It is calibrated to sit below the threshold at which anyone would have to look at it. The entity that would need it is now identifiable from a 10-Q. The state has taken a position in the trade. A promise to absorb losses is no longer its only instrument. Whether that position is a good deal for the public depends entirely on the price at which it was entered. Nobody has published that.
What to watch next
Whether Nvidia's guarantee line grows. It went from $3.5 billion to $108.5 billion in roughly two quarters. The company holds an option on 3.8 further gigawatts at Pike County alone. The next 10-Q prints in late November.
The ERCOT audit, due December 2026. It is the first official count of how much of an interconnection queue is real. Whatever share it finds is the share of every PJM and ERCOT ratepayer's capacity charge that was collected against nothing.
Which ratepayer bill the Senate passes, if any. The House version shifts the cost of new infrastructure. The Heinrich version gives FERC rulemaking authority over large loads. Those produce different worlds.
Cloud AI segment growth against lab revenue growth. The capex sits on the cloud side. The revenue acceleration is on the lab side. Q3 prints in late October.
Whether office vacancy resumes climbing in Q3 and Q4. The reversal in Q2 was partly supply withdrawal. Two more quarters of decline against continued entry-level hiring freezes would retire the demand-contraction leg of Siphon 5.
Bonus depreciation. S.5054 would remove the single largest line of the federal tax channel. Its progress is a direct read on whether the bipartisan energy consensus extends to the tax code.
Whether any other hyperscaler follows Amazon on useful life. Amazon shortened a subset of servers from six years to five and said why. A second firm doing the same converts the depreciation argument from a short seller's model into an industry disclosure.
Private-credit marks on data-centre paper. The $18 billion Oracle-leased New Mexico loans quoted at 89 to 91 cents are the first visible mark. Private credit is held at model prices until something forces a trade. Life-insurer statutory filings and any state insurance department inquiry into private-placement concentration are where the seventh channel would first show.
Whether enterprise ROI measurement improves at all. Three years in, 5 to 8% of firms report measurable at-scale return. If the 2027 surveys print the same number against another trillion dollars of capex, the demand-side question is answered by default.
Source discipline for this piece
The site methodology rates every input on factuality, independence-of-subject and slant. It sets three trust tiers. Two of its rules bind unusually hard here. US official data from 2025 onward is treated as partially entangled. The hard exclusion rule covers any entity reporting on itself, applied to allies and adversaries alike. The subject of this piece is an industry with a direct relationship to the sitting administration, six weeks before a midterm. The tiering is therefore set out in full.
| Source class | Tier | Why, and how it is handled here |
|---|---|---|
| SEC filings (Nvidia 10-Q, Oracle 10-Q, Intel 8-K, Oracle 10-K) | Trustworthy for this subject | A company disclosing its own $108.5 billion of guarantees or $288 billion of off-balance-sheet leases is stating something against its own interest, under liability, to a regulator. That is the strongest class of corporate statement available. Every figure in this piece that carries an argument comes from one. Press characterisation of a filing is never used in its place |
| Federal Register statutory text (EO 14318) | Trustworthy as to text | What an order says is a legal fact, distinct from what an agency claims it achieves. Quoted verbatim from document 2025-14212, not from the White House summary or from coverage |
| Multilateral (BIS Annual Economic Report 2026, IMF GFSR April 2026) | Trustworthy | Independent of the US government by construction. The site's preferred angle on systemic risk, used to corroborate every conclusion that would otherwise rest on US sell-side research |
| US federal agencies from 2025 (Fed FSR, EIA, BLS JOLTS, DOE, EPA, Commerce) | Middle, partially entangled | Usable with mandatory cross-check. DOE site announcements are the administration describing its own programme and are reported as announcements, with the absence of a named tenant at Paducah stated. Labour and price series are cross-read against private trackers that disagree with them |
| Grid operators and state regulators (PJM Internal Market Monitor, ERCOT, PUCT) | Middle, usefully adverse | A market monitor is structurally adversarial to the operator it monitors, which is the property that makes its data-centre attribution usable. The ERCOT freeze is a Texas executive action in an election year and is reported as an order with an audit outstanding, not as a finding |
| Commercial real-estate brokerages (CBRE, JLL, Colliers, Newmark, Cushman) | Middle, conflicted on the subject | These firms are paid on transactions in the asset class they report. A brokerage publishing that vacancy is falling has a commercial interest in that conclusion. This matters because our office-vacancy correction rests on their data. Their figures are used only where the direction cuts against that interest, on conversions removing inventory, on concessions, and on lower-tier effective rents falling |
| Advocacy and specialist trackers (ITEP, Good Jobs First, Data Center Watch, Challenger Gray, Wood Mackenzie, Layoffs.fyi) | Middle | Each has a stated position or a commercial customer. ITEP and Good Jobs First publish their method against filed 10-Ks and state reports and are cited as their own analysis. Data Center Watch is industry-adjacent and counts announcements, so it is paired with SemiAnalysis as the adversarial read on the same denominator. Layoffs.fyi is crowdsourced and carries the least weight of the three layoff series used |
| Private research with a position (SemiAnalysis, Sona Asset Management, Michael Burry) | Middle, position disclosed | SemiAnalysis is technically strong and oriented toward the industry it covers; it is used precisely where it contradicts this piece. Sona is a credit fund that profits if the circularity it mapped unwinds. Burry's depreciation figure is a short seller's model and is labelled as one, with Amazon's own filed useful-life reversal carrying the argument instead |
| US financial press (Bloomberg, CNBC, Fortune, Forbes, Reuters wire) | Middle, ownership-tilt flagged | Used for event confirmation and never as the sole basis for a number that a filing could supply. Reuters and AP wire sit higher than outlet commentary under the methodology |
| Aggregator and content-farm coverage | Excluded from any claim that carries an argument | Several open-weight capability figures circulating in 2026 trace to sites of this kind. The open-weight gap is therefore quoted as a range across index versions rather than as a point, and our own deployments carry the feasibility claim instead |
What this discipline changed. Two claims did not survive it. The widely repeated line that Nvidia and AMD remit 15% of China chip revenue to the Treasury describes a superseded 2025 proposal. Nvidia's own filing describes a 25% import tariff on shipments amounting to under 1% of Data Center revenue. And our own first reading of the office-vacancy series was built on brokerage data without accounting for the brokerages' position or for lease gating. Both are corrected above.
Where the discipline still runs thin. Interconnection-queue integrity has no multilateral aggregator. The phantom-load figures therefore rest on one consultancy, the ERCOT filings and utility commission dockets. Off-balance-sheet AI obligations are counted by Nikkei and by Moody's on different perimeters that are not reconciled to each other. Private-lab revenue is investor-reported and unaudited throughout. The Korean section leans on Korean-language business press and Korea Customs Service export releases. That follows the site's language rule. The leveraged-ETF transmission mechanism is a reconstruction from commentary. No regulator has published a finding on it. The Financial Supervisory Service has not published one.
Source-canon corroboration
| Claim type | Canon resource | Methodology angle |
|---|---|---|
| Nvidia and Oracle commitments, guarantees, leases, receivables | SEC EDGAR primary filings | The filing supersedes the earnings release, the call, and all coverage of both |
| AI capex as a financial-stability question | BIS Annual Economic Report 2026, IMF GFSR April 2026 | The multilateral references the site prefers over US sell-side research and over US agency assessment during the partially-entangled period |
| Executive action and its legal reach | Federal Register; Integrity Index, Transparency International | Statutory text is primary; the canon supplies the lobbying and state-capture layer on why the thresholds sit where they do |
| Capacity auction attribution | PJM Internal Market Monitor; ERCOT and PUCT dockets | A monitor adverse to its own operator is the useful property |
| Federal corporate tax avoidance | ITEP 10-K analysis, cross-read to the filings | Specialty US source with a stated position and a published method |
| Korean memory concentration | Nikkei Asia, Bank of Korea, Korea Customs Service | Asia-language press is the site's preferred lens on memory supply chains |
| Community opposition counts | Data Center Watch, Ballotpedia, Ground News spread-check | Industry-adjacent counter paired with its adversarial read |
| Insurance guaranty mechanics | NOLHGA statutory tables; state insurance codes | The industry's own registry is the authority on the statutes, and it shows a patchwork the secondary account compresses |
| Open-weight capability | Artificial Analysis; model cards and licences; our own deployments | Leaderboards are version-skewed and outside canon, so a range is quoted and first-party deployments carry the feasibility claim |
Sources
NVIDIA Newsroom, "NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR..." (10 August 2026). Blackstone Real Estate Income Trust portfolio and Q2 2026 disclosures (sector allocation as of 30 June 2026). Federal Reserve Discount Window, Collateral Valuation and margin tables (effective 1 July 2026). Federal Deposit Insurance Corporation, special assessment pursuant to systemic risk determination. Federal Deposit Insurance Corporation, 2023 failure cost estimates for Silicon Valley Bank, Signature Bank and First Republic Bank. Federal Reserve Bank of New York, Standing Repo Facility FAQs and Agency Commercial Mortgage-Backed Securities programme terms, operation results and SOMA holdings data. Fannie Mae Monthly Summary (May 2026) for guaranty book size and multifamily serious delinquency. Congressional Research Service R44525 on Fannie Mae and Freddie Mac in conservatorship. Matteo Crosignani and Saketh Prazad, "Extend-and-Pretend in the U.S. CRE Market", Federal Reserve Bank of New York Staff Reports no. 1130 (October 2024, revised June 2026), using FR Y-14Q Schedule H.2 supervisory data. S&P Global Market Intelligence on US bank loss-allowance coverage ratios. Congressional Research Service on the Federal Home Loan Bank System and on discount window policy issues (collateral priority and FDIC recoveries). Federal Deposit Insurance Corporation, First Quarter 2026 CFO Report and Quarterly Banking Profile (Deposit Insurance Fund balance and reserve ratio). Federal Reserve Discount Window, Borrower-in-Custody programme standards announcement (6 August 2026, effective 8 September 2026) and updated collateral margins (effective 1 July 2026). Federal Reserve Bank of New York, BIC Program Requirements (revised June 2026). BIS Quarterly Review and market reporting on significant risk transfer volumes and CRE-referenced transactions (2026). US Bureau of Labor Statistics Consumer Price Index (CPI-U annual averages, 2000 and 2026). International Monetary Fund World Economic Outlook (US nominal GDP) and COFER (reserve currency composition, Q3 2025). Congressional Budget Office, The Budget and Economic Outlook: 2026 to 2036 (February 2026). Federal Reserve FOMC Summary of Economic Projections (September 2026). MSCI, "The Rise and Concentration Risk of Data Centers in Private Markets" (Q3 2025 data, 584 companies). Newmark, "AI and the Future of Office: Quantifying Workforce Change and Space Demand Through 2030". CBRE Investment Management, "Gen AI's Impact on U.S. Employment and Office Space". Sona Asset Management AI circular-economy mapping (176 deals, 202 companies). S&P Global Ratings and Moody's Ratings actions on NVIDIA Corporation (2026). Oracle Corporation remaining-performance-obligation disclosures, FY2026 quarterly releases. McKinsey and contemporaneous reporting on telecom vendor financing exposure, 1999 to 2001. Oracle Corporation, Form 10-Q for the quarterly period ended 31 August 2026 (filed 11 September 2026), Note 6, Leases and Other Commitments. Sona Asset Management AI-financing mapping as reported September 2026. American Prospect, "The AI Bailout Could Be Baked Into the Bubble" (3 August 2026) on private credit, private-equity-owned life insurers and state guaranty associations. McKinsey, The State of AI 2026. PwC Global CEO Survey (January 2026). Gartner agentic-AI forecasts (2026). Microsoft, Alphabet, Meta, Amazon and Oracle useful-life disclosures in their respective 10-K filings. US Department of Energy releases on the Paducah and Savannah River site selections (July 2026). Nvidia Corporation, Form 10-Q for the quarterly period ended 26 July 2026 (SEC accession 0001045810-26-000075), Notes 8, 9, 10, 14. Nvidia Corporation, Form 8-K earnings release of 26 August 2026. Bank for International Settlements, Annual Economic Report 2026 (28 June 2026). Federal Reserve Board, Financial Stability Report (May 2026). PJM Interconnection Base Residual Auction results and Independent Market Monitor reports for the 2026/2027 and 2027/2028 delivery years. Electric Reliability Council of Texas and Public Utility Commission of Texas, large-load interconnection freeze and audit orders (September 2026). Wood Mackenzie, Q3 2026 national data centre pipeline analysis. Data Center Watch, Q1 2026 and Q2 2026 reports. Good Jobs First, data centre subsidy reporting (April 2026). Institute on Taxation and Economic Policy, "Big Tech Is Raking in Enormous Tax Breaks for AI Development" (2026). CBRE, Q2 2026 US Office Market Report; CBRE, "AI Investment Fuels Office Leasing Recovery Across Major Markets"; CBRE office conversion and demolition pipeline tracking (Q1 2026). Newmark lease-expiration analysis. Colliers San Francisco office reports (Q1 and Q2 2026). Challenger, Gray & Christmas monthly Job Cut Announcement Reports (March to July 2026). Layoffs.fyi tracker (to 10 September 2026). Oracle Corporation Form 10-K for fiscal year ended 31 May 2026 (filed 22 June 2026), headcount and restructuring disclosures. US Bureau of Labor Statistics, Job Openings and Labor Turnover Survey (June and July 2026). Executive Order 14318, "Accelerating Federal Permitting of Data Center Infrastructure", Federal Register document 2025-14212, published 28 July 2025, sections 2, 3, 5, 7, 8 and 9. US Energy Information Administration retail electricity price series. Alphabet Inc., Amazon.com Inc., Meta Platforms Inc. and Microsoft Corporation Q2 2026 results and capital expenditure guidance. Nvidia Newsroom and SB Energy releases on the PORTS-Pike Technology Campus (August 2026). OpenAI, "OpenAI joins PORTS-Pike project". US House of Representatives roll call on H.R. Ratepayer Protection Act (16 September 2026). Congress.gov records for S.4214, S.5054 and H.R.8488. Korea Customs Service monthly export releases (June and August 2026). Bank of Korea financial stability commentary (2026). Artificial Analysis intelligence index (v4.1 and v4.3.2). Stanford Institute for Economic Policy Research, "What is really happening to jobs". SemiAnalysis, "Are AI Datacenters Increasing Electric Bills for American Households?" and "Stop Saying Half of 2026 US Datacenter Capacity Is Canceled".