Summary

  • J.P. Morgan states it originated $9.6bn across two construction loans for Project Stargate's Abilene, Texas campus as lead left, sole underwriter and sole structuring agent — a deal-specific, self-reported figure, not a portfolio total.
  • JPMorgan recorded an $863m net addition to its allowance for lending-related commitments in the third quarter of 2025, predominantly driven by new lending-related commitments made mainly in the second quarter of 2025: risk recognised on facilities before they are drawn.
  • The $38bn Oracle-linked Texas and Wisconsin package led with Mitsubishi UFJ took months to distribute, and accounts conflict on whether the tail was sold at par, discounted, or retained.
  • No reviewed JPMorgan disclosure disaggregates AI or data-center exposure, so every portfolio total in circulation is a third-party estimate and must be labelled as one.

The number that is not a portfolio

J.P. Morgan's own capital-markets commentary states that the bank originated $9.6bn across two construction loans for the Abilene, Texas campus of Project Stargate, acting as lead left, sole underwriter and sole structuring agent on both transactions (J.P. Morgan, Financing AI infrastructure and U.S. data centers). The same page sets out the taxonomy that matters for everything that follows: AI-infrastructure debt typically takes one of two forms, corporate balance-sheet debt or project-level debt.

That $9.6bn is an origination figure for named transactions. It is not an exposure figure, not a drawn balance, and not a portfolio ceiling. J.P. Morgan is the source of it, which makes it an institutional claim about the bank's own roles and volumes rather than an audited schedule. The page carries no data-center exposure total at all, and the absence is not an oversight — it reflects how the bank reports.

Project Stargate itself was announced by the U.S. government in January 2025 and contemplates up to $500bn of investment in U.S. data centers and energy infrastructure over four years, according to the same page. A programme ceiling of that size and a single bank's originated volume are different orders of magnitude, and the space between them is where the financing question actually lives.

What the bank reserved for, rather than what it lent

The more informative number in JPMorgan's disclosure is not the origination figure. It is the allowance.

The bank recorded an $863m net addition to its allowance for lending-related commitments in the third quarter of 2025, recognised in other liabilities on the consolidated balance sheet, and attributed it predominantly to the impact of new lending-related commitments made primarily in the second quarter of 2025 (JPMorgan Chase, 3Q25 results narrative). The same filing line appears in the Form 10-Q for the quarter ended 30 September 2025, a separate regulatory channel carrying the same figure and the same stated cause (SEC EDGAR, JPMorgan Chase 10-Q).

A lending-related commitment is money promised but not yet drawn. Reserving against it means the bank booked expected loss on facilities that were still undrawn when the reserve was taken. The three-quarter-2025 narrative also records a provision for credit losses of roughly $3.4bn and a $1.4bn net addition to the allowance for loan losses, split roughly $1.1bn consumer and $340m wholesale.

The supplement shows the level, and it shows why the numbers should not be swapped (JPMorgan Chase, 3Q25 financial supplement). The allowance for lending-related commitments stood at $2,798m at 30 September 2025 against $2,013m a year earlier — a 39% increase — with $2,757m at 30 June 2025. A separate roll-forward in the same document shows a $2,964m ending balance and an $862m nine-month provision. Total lending-related commitments were $1,714,006m at 30 September 2025, up 3% quarter on quarter and 9% year on year, with wholesale commitments of $596,028m.

Those figures do not reconcile to one another, because they are not the same measure. The supplement shows both a $31m quarterly provision for lending-related commitments and an $862m nine-month provision, while the narrative cites an $863m net addition to the allowance. Quarterly flow, nine-month flow and balance-sheet reserve level are three different quantities, and the two ending levels of $2,798m and $2,964m reflect different table scopes within the same document.

What none of this states is the share attributable to data-center lending. The build is firm-wide and attributed to new commitments generally. Treating the $863m as an AI figure would be an invention.

The syndication test

The clearest evidence about whether AI-infrastructure credit behaves like a distinct asset class is not in a reserve line. It is in how the largest deals distributed.

Bloomberg reported that JPMorgan and Mitsubishi UFJ Financial Group took on a $38bn loan package in August 2025 backing Oracle data-center projects in Texas and Wisconsin, and that months later — after more than two dozen banks and other investors joined to share the risk — the syndication was nearly done, with some lenders still looking to offload less than $1bn (Bloomberg, Oracle-tied $38 billion debt takes months to spread).

Business Insider, citing a person familiar, reported that JPMorgan encountered diminished interest as it sold off pieces of the package, while the same person said the two projects are fully financed and that the syndication was successful overall, and noted that banks and institutional investors have grown wary of concentrated Oracle exposure (Business Insider).

The two accounts are not contradictory so much as differently framed. Months of distribution, two dozen-plus participants and a residual tail are not the signature of a deep, liquid, standardised market. Concentration limits emerge as the specific obstacle: reporting on the Oracle-linked financings describes lenders applying single-counterparty and single-tenant caps during syndication in late 2025 and early 2026, alongside approximate deal sizes of $10bn for Crusoe's original Abilene site, $38bn of borrowed money across Vantage Data Centers' Texas and Wisconsin campuses, and about $18bn for a Stack Infrastructure project serving a fourth OpenAI facility in New Mexico (Hindustan Times). The same reporting describes lenders balking at an Abilene expansion on an Oracle lease, after which the developer leased the capacity to Microsoft instead — a tenant substitution that is itself a statement about how the risk was priced. Vantage told the same outlet that its Texas and Wisconsin loans were largely syndicated in the fourth quarter of 2025 with more than 50 lenders.

MarketScreener, citing the Financial Times, reported on 4 May 2026 that JPMorgan and MUFG were looking to distribute $38bn of Oracle-linked debt sometimes at a discount, and that banks were developing significant risk transfer structures to offload the riskiest tranches of concentrated loans while retaining part of the exposure (MarketScreener, citing the Financial Times).

On the question that a lender would care about most — whether the tail cleared at par, at a discount, or stayed on balance sheet — the public record does not settle. Bloomberg describes a nearly completed syndication with a sub-$1bn residual. Business Insider's source calls it successful overall. MarketScreener's account introduces discounts. Those are three different economic outcomes, and they are attributed to sources, not to a filing.

What management actually said, and what it did not

On the second-quarter 2026 earnings call, CFO Jeremy Barnum called the data-center underwriting space a bellwether for what lenders are doing, said the key questions are what happens with power supply and what happens with tenants, said the firm has a pretty precise framework to govern what it is and is not willing to do in that space, and said JPMorgan passed on some deals (JPMorgan Chase, 2Q26 earnings transcript). On the same call he noted that some relationship lending to start-up entities building data centers is occurring, and tempered the alarm: he was not warning that underwriting is about to collapse.

Jamie Dimon characterised competitor credit-underwriting deterioration as very mild, citing revenue-growth assumptions, add-backs of expenses, more payment-in-kind, weaker covenants and more rollover risk, and said the bank's own risk standards have not changed. The framing matters. A bellwether is an observation device, not a loss estimate, and a framework that permits passing on deals is not evidence of either strength or weakness — it is evidence of selectivity that cannot be measured from outside.

For scale, the second-quarter 2026 release reports net income of $21.2bn ($7.70 per share), reported revenue of $57.3bn, credit costs of $2.5bn with $2.4bn of net charge-offs and a $149m net reserve build, and average loans up 10% year over year, against approximately $1.9tn of credit and capital raised year to date — including $1.7tn for corporations and non-U.S. government entities (JPMorgan Chase, 2Q26 earnings press release). Data-center lending is not separable in any of those lines.

Aggregates are estimates, not disclosures

Against that vacuum, third-party aggregations fill the space. One analysis ranks JPMorgan fifth among data-center lenders with roughly $13.3bn of exposure — about $11bn in syndicated debt and about $2.3bn in property mortgages — and attributes to the bank a $7.1bn financing for Crusoe and Blue Owl's 1.2 GW Abilene campus, plus lead bookrunner status on $3.3bn of secured notes for Hut 8 (Bisnow).

That $13.3bn is an estimate with an undisclosed methodology. It is not a bank disclosure, its components do not map onto the bank's own reporting categories, and it should never be quoted as though JPMorgan published it. The same discipline applies to programme figures: a $500bn Stargate ceiling is an intention, a $9.6bn origination is a transaction, and an $863m commitment allowance is an accounting estimate. Each is real. None substitutes for another.

What would falsify the asset-class claim

The claim worth testing is that AI-infrastructure lending is a distinct asset class with distinct underwriting, rather than conventional corporate credit secured by unusually long-dated, single-tenant collateral.

Several observable conditions would settle it in one direction or the other. First, disaggregation: if an issuer begins reporting data-center or AI-linked exposure and drawn balances separately from general wholesale lending, and those drawn balances prove small relative to announced roles, the distinct-asset-class framing weakens considerably.

Second, loss experience: sustained low net charge-offs through a full construction-to-operating cycle would support the thesis that power supply and tenant credit were underwritten conservatively; the first concentrated impairment in a single-tenant campus would not disprove the thesis, but a cluster of them would. Third, distribution: if the next large packages syndicate quickly and at par, the concentration-limit friction of late 2025 and early 2026 was a temporary capacity problem.

If they keep requiring discounts, risk-transfer structures or multi-month tails — or if significant risk transfer becomes the standard exit rather than the exception — then the market is pricing this exposure differently from ordinary corporate credit, whatever the label.

What the current record supports is narrower and more defensible: JPMorgan is originating AI-infrastructure debt at scale, reserving against undrawn commitments on a firm-wide basis, and applying tenant and power-supply criteria with enough discipline to decline deals. What it does not support is any single number for the bank's AI exposure.