Nvidia is guaranteeing OpenAI's rent
In June we argued the AI buildout was priced on an expectation of federal rescue. Somebody did agree to cover the losses in the three months since. It was not the government. Nvidia has agreed to pay OpenAI's rent if OpenAI cannot, up to $105 billion, disclosed in a 10-Q.
On page 18 of Nvidia's quarterly report for the three months ended 26 July 2026 there is a two-line table headed "maximum gross exposure related to our guarantees". The first line is $3.5 billion. The second is $105 billion. The total is $108.5 billion.
At the previous fiscal year end that table had one line on it. The number was $3.5 billion.
The $105 billion is credit support for lease payments on roughly 4.25 gigawatts of computing capacity in Pike County, Ohio, extended on behalf of an affiliate of OpenAI. Nvidia's own filing gives the reason in its risk factors. Its customers lack "investment-grade financing capacity to secure the AI infrastructure necessary to grow". So Nvidia is lending them its own. The filing also states what happens if that does not work. On an OpenAI default or insolvency event, a triggered guarantee means Nvidia "may assume the applicable lease".
Nvidia has guaranteed twenty years of rent on buildings whose tenant exists to buy Nvidia chips.
What the guarantee replaced
Three months ago we argued from the federal record of absorbing concentrated losses five times in eighteen years. That precedent makes a trillion-dollar capital programme rational even where it fails to close on its own arithmetic. The upside sits with equity holders. The downside ends up on public balance sheets.
That structure still holds. What has changed is who pays first. The same Nvidia 10-Q, set out in full here, discloses $366 billion of future commitments across five lines.
| Commitment | Amount | Owed to | For |
|---|---|---|---|
| Supply and capacity | $279bn | memory makers and foundries | memory and manufacturing capacity to build Nvidia products |
| Cloud service agreements | $29bn | cloud providers | compute for Nvidia's own research |
| Leases not commenced | $25bn | landlords | engineering, design and test space |
| Equity investments | $25bn | AI model makers and infrastructure financiers | committed investments |
| Capital expenditures | $8bn | equipment vendors | Nvidia's own plant |
| Total | $366bn |
Two of those lines point back at Nvidia's own customers. The $29 billion buys cloud capacity from the firms that buy its chips. The $25 billion of equity is committed to "AI model makers, infrastructure financiers, and other private companies". That is money going to the buyers of its products and to the firms financing them.
The balance sheet in the same filing reports total assets of $320.3 billion and total liabilities of $91.3 billion. The $366 billion of commitments, with $108.5 billion of guarantees behind it, exceeds everything Nvidia owns.
That table covers future obligations only. Money Nvidia has already sent downstream sits in a different note.
What Nvidia has already put into its customers
Nvidia held $47.9 billion of non-marketable equity securities at 26 July 2026. The filing describes them as "primarily in privately-held companies". Twelve months earlier the balance was $3.8 billion.
The movement is in the note. Across the six months Nvidia made $31.0 billion of net additions, recorded $7.5 billion of unrealized gains, and reclassified $12.7 billion out of the balance when holdings began trading publicly. Cumulative gross unrealized gains reached $9.1 billion, against $661 million a year earlier. Those gains run through Other income. Part of Nvidia's reported non-operating income is therefore the markup on private companies it funded.
Two smaller lines sit alongside. 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 future committed amounts at $4.7 billion, and states that Nvidia is not the primary beneficiary and does not consolidate them.
So the downstream figure is not the $25 billion in the commitments table. It is roughly $73 billion of equity paid in and committed, plus $108.5 billion of contingent guarantee, against a stake balance that went from $3.8 billion to $47.9 billion in a year. Accounts receivable moved the same direction, from $38.5 billion in January to $63.1 billion in July.
Where the $279 billion goes
The supply line is the largest and it stood at $119 billion one quarter earlier. The filing says the purchases are "primarily memory and manufacturing facilities", bought "to produce our products". None of it is committed on a customer's behalf. Nvidia's exposure to OpenAI sits in the $105 billion lease guarantee and in the $25 billion of committed equity.
The filing does not name the counterparties. The suppliers of those inputs are public. SK Hynix holds roughly 50 to 55% of the high-bandwidth memory market, Samsung 35 to 40% and Micron 5 to 10%, with SK Hynix reported at about 70% of the HBM4 allocation for Nvidia's Vera Rubin generation. TSMC fabricates the compute dies and does the packaging that bonds them to the memory.
A 12-layer HBM4 stack prices above $600 and a Rubin-class GPU carries eight of them. More than $4,800 of memory goes into one accelerator. Memory is the highest-cost line in the bill of materials, ahead of Nvidia's own silicon.
June called Nvidia the picks-and-shovels seller, the one participant certain to be paid. It has now committed $279 billion to buy the picks and shovels from someone else, before the end demand exists. The firms with contracted revenue and no demand risk are SK Hynix, Samsung, Micron and TSMC.
What links that commitment to the guarantee is the demand forecast. Nvidia committed the money against "long-term demand". It is separately underwriting part of the demand it forecasts.
Nvidia names no customer
Three direct customers were 16%, 15% and 13% of first-half revenue, 44% between them, all unnamed. A direct customer is whoever Nvidia bills. The filing says the category includes "ODMs, OEMs, CSPs, AI model makers, and system integrators". It is frequently the contract manufacturer assembling the servers, one step removed from the company that will run them.
The receivables belong to a different group again. The filing attributes the $24.6 billion increase to "extended payment terms on large multi-quarter agreements with certain investment-grade customers". OpenAI is not investment grade, by the same filing's account of its customers. That credit went to the rated hyperscalers.
Nvidia estimates that "one AI research and deployment company contributed a meaningful amount of our revenue by purchasing cloud services from our customers". That is revenue Nvidia attributes to a company it does not sell to. The company rents cloud from a Nvidia customer, the customer buys Nvidia chips to supply it, Nvidia books the sale. "AI research and deployment company" is OpenAI's own description of itself. That makes the identification likely without making it certain. The filing names nobody.
The same position at Google and Oracle
Google has taken the same position on the other frontier lab, guaranteeing Anthropic's data-centre lease and power obligations against default. Both labs now have a landlord who is paid, in the last resort, by the firm that sells them their inputs.
Oracle's own filing is the mirror image. Its 10-Q for the quarter ended 31 August 2026 discloses $288 billion of additional lease commitments, substantially all data centres, on fifteen to nineteen year terms, "not reflected on our condensed consolidated balance sheets". The lease liabilities that are on the balance sheet total $43.8 billion.
Oracle's $288 billion and Nvidia's $279 billion are unrelated. Nvidia's buys silicon inputs and rents nothing. Oracle's rents buildings, with Oracle as the tenant. OpenAI rents compute from Oracle under a separate $300 billion contract. The campus Nvidia guarantees is leased from SB Energy. Every large figure here is listed once with its owner.
A London credit fund has since mapped the vendor circularity, putting $3.6 trillion of AI financing through 176 deals between 202 companies and flagging about 120 as ones where the same firms appear as investor, customer and supplier.
A normal sale is funded by the buyer. In these 120 the seller funds some or all of it.
The three-party version. 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 has completed a circuit back to the party that advanced it.
The two-party version. Amazon invested about $50 billion in OpenAI. OpenAI committed to buy about $100 billion of compute from Amazon. Amazon's outlay returns as its own revenue at twice the nominal amount.
The guarantee version, moving no cash at all. Nvidia guarantees OpenAI's rent. The guarantee makes the building financeable. The building houses Nvidia hardware that OpenAI pays for. Nvidia has exchanged a contingent liability for demand.
Each leg is arm's length, permitted under the accounting rules and disclosed. The consequence is not misstatement. It is that the demand is not independent of the supplier. Revenue funded by the seller stops when the seller stops funding it. One stream of expected spending is also recorded as revenue at more than one firm. That makes the market appear larger than the underlying demand.
None of this required an appropriation. None of it required a vote. It is all disclosed. It was all done in about two quarters.
Why anyone would write a guarantee like this
None of that explains the motive. The motive is in the 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 earns anything.
Those lenders are repaid out of the rent. The rent comes from OpenAI. So the decision to finance the building is a judgement about whether OpenAI can pay for twenty years. OpenAI is a private company with roughly $40 billion of revenue, about $27 billion of cash burn this year, and no credit rating. Nvidia states the consequence itself. Its customers lack "investment-grade financing capacity to secure the AI infrastructure necessary to grow".
The guarantee substitutes one credit for another. Lenders assessing the Pike County rent are now assessing Nvidia, rated Aa1 by Moody's and AA by S&P, in place of an unrated borrower. Without that substitution the building is not financed. The chips then have no site to occupy.
The three firms in the structure that hold credit ratings are lending those ratings instead. None is lending cash. A guarantee costs nothing until it is called and stays off the balance sheet until then. That makes it the cheapest way to move creditworthiness from where it sits to where the demand is.
Each has its own reason. Nvidia has pre-committed $279 billion to its own suppliers and needs the demand that consumes it. The guarantee also buys twenty years of exclusivity across 8 gigawatts, shutting out AMD, Broadcom and Google's TPUs. Oracle bought backlog, taking remaining performance obligations to $638 billion, up 363% and past Microsoft, while the leases that service it sit in a footnote. Google is defending a 14% stake in Anthropic, a $200 billion cloud commitment and a TPU channel with one instrument.
The structure ran before. Through 1999 to 2001 Lucent committed about $8.1 billion of vendor financing, Nortel $3.1 billion and Cisco $2.4 billion, roughly $25.6 billion across nine suppliers. Then 47 carriers went bankrupt. The equipment had no secondary market. The financing became write-offs.
Adjusted for inflation that episode is $48.8 billion in today's money. Nvidia's $108.5 billion is about 2.2 times it, or 1.4 times measured against GDP. The raw multiple flatters the point. The scale is in the buildout instead. The competitive local exchange carriers took roughly 0.12% of GDP a year. The Big Four will spend 2.24% of GDP on AI capex in 2026 alone, close to nineteen times the rate.
Nothing has absorbed a loss yet. What exists is a sequence of credit substitutions ending at Nvidia, the last and largest private balance sheet in it. 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 only a buyer of Nvidia compute can use. Those three exposures are not diversified against each other. Each depends on the same demand.
The June argument assumed the government would be the first to pay when a concentrated AI loss arrived. The filing puts Nvidia ahead of it. Public money is committed only after Nvidia's $108.5 billion is exhausted.
The site is a decommissioned uranium plant
June flagged Executive Order 14318 and the Department of Energy's release of federal sites as a channel open only to buyers clearing a $500 million or 100 megawatt threshold. That argument had no project attached to it. It now has four, with Paducah and Savannah River both awarded since.
The campus Nvidia is guaranteeing is the PORTS-Pike Technology Campus, built on the decommissioned Portsmouth Gaseous Diffusion Plant, on land the developer describes as private and federal, with the Departments of Energy and Commerce named as collaborators. SB Energy builds and operates it under a twenty-year lease to OpenAI. The campus targets 8 gigawatts of computing capacity against at least 10 gigawatts of new generation. Nvidia has also invested $1.5 billion in the landlord and holds an option to guarantee a further 3.8 gigawatts. It secured a term under which the site "will exclusively host NVIDIA AI infrastructure, subject to limited exceptions".
The order that made the site available does more than shorten a queue. Section 7(a) directs the EPA to expedite permitting by "developing or modifying regulations" under the Clean Air Act, the Clean Water Act, CERCLA and the Toxic Substances Control Act where they "impact the development of Qualifying Projects". Section 8 puts Endangered Species Act consultation on a ten-year programmatic basis. Section 9(b) opens military installations to competitive lease. June read EO 14318 as a permitting accelerator. That was too narrow. It is a standing instruction to reshape four environmental statutes around one industry. The TSCA line touches the same PFAS reporting cycle the June analysis was waiting on.
The site uses federal land. Nvidia's credit rating secures the tenant's lease. An exclusivity clause reserves the site for Nvidia hardware. June described each of those three as a separate mechanism. They now appear in one signed transaction.
The queue nobody paid to file
The sharpest new transfer was not in the June piece at all. Grid operators price generating capacity years ahead against forecast peak demand. That forecast is built from interconnection requests. Filing one has been close to free. Developers hedge queue delays by filing the same 500 to 1,000 megawatt project with three or four utilities at once, then building wherever a transformer arrives first.
Wood Mackenzie's third-quarter 2026 analysis puts 72% of the 1,066 gigawatts of requested US data-centre capacity in the phantom category, against an existing national fleet drawing roughly 60 to 70 gigawatts.
Texas stopped. ERCOT requests went from about 48 gigawatts in 2023 to more than 474, five times the state's record peak demand, roughly 90% of them data centres. The governor and the Public Utility Commission froze new connections in September and ordered an audit due December 2026.
The bill has already gone out. 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. PJM's own market monitor calculates that removing all data centres from the forecast would cut capacity payments by $9.33 billion, a 64% reduction.
Households are paying, on 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.
A seventh channel runs the other way, from the credit to the public. Private credit lends to data-centre construction. The private-equity houses behind it own life insurers, Apollo holding Athene and KKR holding Global Atlantic. Those insurers put policyholder float into their parents' paper, at 49.5% of new investment in privately placed instruments against 14% for independent insurers. If that paper defaults hard enough to impair an insurer, state guaranty associations pay the claims. The assessment that funds them is largely recovered through credits against state premium tax. The loss lands on state general revenue without any legislature voting on it.
What held and what did not
Capex held almost exactly. June projected roughly $725 billion across Microsoft, Alphabet, Amazon and Meta for 2026, against guidance now at $720 to $745 billion. All four raised. Alphabet spent $44.9 billion in the June quarter against $39.1 billion from operations, producing the first negative free cash flow quarter since its 2004 IPO.
The bond estimate was too low. June put the incremental issuance landing on passive holders at $80 to $120 billion a year. Amazon, Alphabet, Meta and Oracle printed about $194 billion through 7 July alone, on terms that moved against them. Orders covered the February deals nearly five times over. The first unscheduled deal of the second half was covered 1.4 times.
The tax and community channels grew. ITEP now counts five firms reporting $422 billion in profit at an effective federal rate of 4.5%, with $27 billion of the $70 billion in breaks coming from accelerated depreciation. Data Center Watch counted about $130 billion of projects blocked or delayed in the first quarter of 2026 and $68 billion in the second, against the $64 billion over eleven months that June cited.
The index gate behaved as predicted. The Nasdaq-100 admitted SpaceX on 7 July at around $160 under a fast-entry rule adopted in May. The stock was below its $135 IPO price within about ten days. Index funds bought at roughly $160 and the price fell about 22% within ten days.
The enterprise return that all of it rests on still cannot be measured. McKinsey has 78% of organisations using AI in at least one function, against 55% two years ago. Between 5 and 8% report a measurable at-scale financial return. PwC found 56% of chief executives reporting neither revenue growth nor cost reduction from AI in the previous twelve months.
The floor we said would empty
June argued that AI would thin the middle communication-relay layer that fills Class A offices. It called the cascade already visible in the vacancy numbers. US office vacancy then fell 30 basis points in the second quarter to 18.3%, the largest quarterly decline since 2015, with prime vacancy down 40 basis points to 12.3%.
We read that as the thesis failing. It is not evidence either way, for four reasons we should have checked before citing it.
A tenant that cuts headcount keeps paying rent until the lease ends. More than 32% of all US office leases expire between the second quarter of 2026 and the end of 2028. Newmark counts roughly 1.4 billion square feet of pre-pandemic leases rolling between 2025 and 2027. Displacement cannot appear in a vacancy series before the lease rolls. The instrument does not read the signal yet.
The absorption that did happen is substantially the buildout itself. CBRE counts 30.4 million square feet leased by AI companies since 2019, 21 million of it in the San Francisco Bay Area. Technology's share of US office leasing went from 16.8% across 2025 to 22.7% in the first quarter of 2026. In San Francisco's 1.6 million square feet of first-quarter 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. The floors are being filled by the same capital programme whose effects on the other floors we were trying to measure.
The denominator is also shrinking. CBRE tracks 81 million square feet in the US office conversion and demolition pipeline, roughly 90,300 residential units. Conversions completed about 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. Space taken out of inventory lowers the vacancy rate without a tenant signing anything.
Price is the fourth thing the headline rate does not carry. On a 2026 office lease with meaningful concessions, net effective rent runs 25 to 35% below face rent, with six to eighteen months of free rent and tenant-improvement allowances from $10 to $100 per square foot. CBRE's own series has effective rents up 5.2% for top-tier buildings and down 1.2% for lower-tier. The recovery is concentrated in trophy space. That is where the AI tenants are. Class B and C, the subject of the June argument, is still repricing downward.
The labour series we cited was the wrong one. Aggregate JOLTS layoffs sit near 1.1%, close to a historic low. That is the series we reported. Inside that aggregate the information sector's layoff rate went from 1.3% to 2.4% in a year, the largest increase of any sector and close to five times the national average, while hiring in professional and business services fell by 188,000 and the quits rate dropped to 1.9%, a level last seen in 2014. Challenger, Gray & Christmas counted 139,156 technology job cuts through June 2026, up 83% year over year, with AI named in 101,743 announcements across all sectors and leading every stated reason for five consecutive months. Oracle's own annual report, filed 22 June 2026, shows global headcount down from about 162,000 to 141,000 across fiscal 2026, roughly 21,000 people and 13% of the company, against $1.84 billion of severance and exit costs. It began another round in mid-September.
Oracle is the most leveraged AI-infrastructure builder in the group and is rated two notches above high yield. It reduced headcount by an eighth in the year it borrowed most heavily. Office vacancy cannot register that reduction until Oracle's leases expire.
The banks holding that office debt have two new ways to carry it without pricing it. Both changed this year. A commercial real-estate loan can be pledged to the Federal Reserve's discount window, where the Fed makes its own monthly fair-value estimate and lends a published percentage of it. Nothing is sold. No public price is set. Participation in that programme was eased on 8 September and the margins were reset on 1 July. The credit risk on the same loan can then be sold synthetically to a private credit fund, in a market whose pool has passed $670 billion and now references CRE directly. The arithmetic is worked through in the analysis.
That does not put the loss on the Federal Reserve. It discounts twice and lends about $48 against a loan carried at $100. That puts it near-last in line. What pledging does is move the good collateral out of the estate before the FDIC reaches it. Reserve Bank and Federal Home Loan Bank advances are secured. The Federal Home Loan Banks are paid before both the FDIC and the Fed. The Deposit Insurance Act bars the FDIC from disallowing either claim. In 2023 the Fed lost nothing and the FHLBs lost nothing. The Deposit Insurance Fund absorbed about $31.5 billion across Silicon Valley Bank, Signature and First Republic. That is the ledger holding the risk. It stands at $157.4 billion against banks holding $1.844 trillion of commercial mortgage debt.
The carrying value is doing work beyond the loan. It sits risk-weighted in the asset base. At a 10% capital ratio each dollar of capital carries about ten of risk-weighted assets. An $18 loss left unrecognised therefore holds up near $180 of balance sheet. Banks reserve 1.61% against commercial real estate, against office CMBS delinquency at 12.34%. The Federal Reserve has measured what that produces. A New York Fed staff report using loan-level supervisory data finds weakly capitalised banks extending distressed CRE loans specifically to preserve capital. Those banks assign the loans a lower default probability than better-capitalised banks assign to comparable ones. They have pushed commercial-property loans maturing within three years to 37% of their capital against 27% at stronger banks, measuring capital at what their assets would fetch today. The extensions also cut new CRE origination by 1.1%. The capacity is not financing anything. It is holding the old position open.
June also predicted that the private-equity houses funding the buildout would be caught holding the office it emptied. They were not. They rotated. BREIT held 26.1% of its real estate value in QTS data centres as of 30 June 2026 against 12.5% in office. It put a further $5.7 billion into pre-leased data-centre development this year. On 10 August, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR announced platforms with Nvidia to "mobilize over $500 billion of third-party capital" for AI infrastructure. Third-party capital means pensions, endowments and insurers. The houses sold the leg we said would sink them. They are now raising other people's money for the one they kept.
South Korea separated the two questions
South Korea ran the experiment that argument cannot settle. The question is whether an AI repricing and an AI slowdown are the same event.
The KOSPI went from about 5,000 in March to above 9,000 in June. In late July it fell more than 16% across two sessions, removing roughly $2 trillion of market value, with Samsung down 45% and SK Hynix 55% from their June peaks at the trough. On 31 July the index rose 17.9% in a single day, the largest gain in its history. It closed August at 6,820.
The transmission was leverage. Retail investors held newly available leveraged single-stock ETFs on both memory makers. When prices faltered the funds liquidated into a falling market. That triggered margin calls on the underlying holders, forcing more selling.
The physical business never participated. Semiconductors were a record $46.65 billion of Korean exports in August, 47.5% of all goods exports, with memory at roughly 90% of that. Nominal GDP grew 21.9% in the first half, almost 70% of the expansion coming from semiconductors. HBM3E has traded on the spot market at four to five times its contract price.
Two trillion dollars left the price of two companies in two sessions while those companies sold every unit they could make at rising prices. Korea separated the claim on future cash flow from the cash flow itself, in public.
South Korea is now the most AI-concentrated national economy in the world by export share, by index concentration and by contribution to growth. It sells the memory a year before the racks it goes into earn anything. A genuine pause in US capex therefore reaches Korean GDP before it reaches a US earnings call.
A rescue before November
The midterms are on 3 November. Through 20 September the political movement runs hard against the industry.
The Ratepayer Protection Act passed the House 417 to 3 on 16 September, after clearing committee 52 to 0 in July. It makes large-load data centres pay for the new energy infrastructure built to serve them. The three votes against came from the left, on the grounds that it does too little. The Senate declined to pass it by unanimous consent the next day, blocked by a senator who wants FERC given rulemaking authority over any facility drawing 150 megawatts or more.
Both parties are competing to be harder on data centres, six weeks out, on a roll call. No appropriation for a data-centre operator survives that.
Nothing currently needs rescuing either. Nvidia earned $59.7 billion in the June quarter. The distress that exists is concentrated in the leveraged neoclouds. CoreWeave's credit default swaps have traded near 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.
We have also measured the cushion thinning, on our own instrument, while writing this. The US currency read in the ratings framework has fallen from -4.10 in late May to -5.39 on 19 September, with all of the move arriving after mid-July. On 25 August the near-term reserve-currency score was cut from +5 to +4, the first markdown of the cushion itself. The recorded reasons are monetary. None of them is AI. The series is in the analysis.
The federal posture has meanwhile changed from guarantor to shareholder. The government holds 9.9% of Intel, bought in August 2025 with converted CHIPS Act money, plus a warrant over a further 5%. The chip-revenue arrangement that gets repeated alongside it needs care. The 15% figure was a 2025 proposal covering the H20. It was superseded in January 2026 by a 25% arrangement on the H200. Nvidia's own filing describes that as an import tariff it cannot pass on, against shipments amounting to "less than 1% of Data Center revenue".
The instrument that matters is older and quieter. Section 3 of Executive Order 14318 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". A qualifying project is one clearing $500 million of capital expenditure or 100 megawatts of new load. Section 5(c) then sets the disclosure floor. "Federal financial assistance representing less than 50 percent of total project costs shall be presumed not to constitute substantial Federal control and responsibility". That places it outside the definition of a major federal action under the National Environmental Policy Act.
Those two sections together allow the federal government to guarantee up to 49.9% of a data-centre project's cost through an initiative that already exists, with no appropriation, no roll call and no environmental impact statement. The 50% figure sets the threshold for review. It does not limit the amount committed.
A bailout requiring an appropriation does not arrive before November, for two reasons that both expire. No participant has failed. No member of Congress wants to cast that vote six weeks before an election. The version requiring neither a vote nor a review is already authorised, at a size set below the threshold for either.
The floor under the price
The June argument was that cloud pricing power depends on frontier capability staying out of local reach. The measurable part of it moved faster than we expected. The reported distance between the leading open-weight model and the leading proprietary model on Artificial Analysis's intelligence index has been quoted between 4 and 9 points through the middle of 2026, against 13 points a year earlier. Four of the five leading open-weight entries come from Chinese labs.
The consumer hardware claim we leaned on has not held up as evidence. June cited Nvidia's stated capability of 120 billion parameters at a million tokens of context on a Windows laptop shipping in the autumn. Nvidia now says the first of those machines arrive in October. No third party has published a benchmark. It was a marketing claim, carrying weight it had not earned.
What we can put behind the claim is our own hardware. One 26-billion-parameter model on a single consumer-tier GPU in this lab serves a chat front end, a coding agent, a voice assistant, a doorbell agent and voice search at once, at 65,536 tokens of context across three parallel slots, from one resident copy of the weights. Separately, Legion runs a 235-billion-parameter mixture-of-experts model inside a browser by streaming experts across a peer mesh, documented in a 235B model in a browser. Neither is a leaderboard score. Both are in production.
What to watch
The fuller watchlist sits with the evidence. Nvidia holds an option on 3.8 further gigawatts at Pike County alone. That would roughly double the campus to eight. The $105 billion is also the filed figure and not the ceiling reported. Press accounts since describe a broader Nvidia and OpenAI arrangement in negotiation, up to $100 billion of credit support and as much as $350 billion of GPU financing, none of which appears in any filing yet. The next quarterly filing lands in late November. The nine lease guarantees step up as each data centre is placed in service, the first in fiscal 2029. The ERCOT audit reports in December, the first official count of how much of an interconnection queue is real. Whatever share it finds is the share of every capacity charge collected against nothing.
Cloud revenue and lab revenue have diverged. 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, with a first operating profit of $559 million. Google Cloud grew 82% year over year, AWS 37%, Azure 43%. The labs are compounding faster than the clouds. The capex sits on the cloud side of that gap.
In June we identified the public as the party that absorbs a concentrated AI loss, on the precedent of five federal interventions in eighteen years. Nvidia's filing identifies a private party ahead of the public. Nvidia pays up to $108.5 billion and may then assume the lease. Public exposure begins after that. It depends on whether Nvidia can carry the obligation while its own supply commitments fall due.