Research publicly reported data-center construction and term loans led by JPMorgan, Morgan Stanley, MUFG, SMBC, Mizuho, Goldman Sachs, and others.

Banks are the origination machine for the AI data-center buildout, but single-name tickets of $15–38 billion have already pushed them against internal concentration limits. The response is not to stop lending; it is to underwrite, then push risk into syndication, project bonds, CMBS/ABS, and experimental single-asset significant risk transfers (SRTs). Public data do not support a precise global bank-exposure number. Tracked retained books at the named lenders are on the order of $10–17 billion each, while system-wide estimates of bank credit tied to AI physical infrastructure run from the low hundreds of billions (narrow regulatory definitions) to roughly $800 billion (broader academic tallies).

Mega construction loans, and who actually leads them

The binding constraint is no longer finding a bank willing to look at a hyperscale campus. It is finding a syndicate that can hold a four-year construction loan whose size exceeds what any one balance sheet wants to keep. Japanese project-finance banks and U.S. bulge brackets now share the lead on the largest tickets, with pricing that started tight and has widened as Oracle-linked supply piled up.

The landmark package is the Vantage Data Centers financing for Oracle-leased campuses in Texas and Wisconsin. JPMorgan and MUFG led a roughly $38 billion senior secured package—about $23.25 billion for Texas and $14.75 billion for Wisconsin—with Wells Fargo, BNP Paribas, Goldman Sachs, SMBC, and Société Générale in the underwriting group. Both facilities were structured as four-year loans with two one-year extension options, priced around 250 basis points over the benchmark (SOFR), interest-only during construction and amortizing once operations start. [1] [2] [3] An earlier report framed a related Vantage Texas campus loan near $22–23 billion under the same lead banks. [4] [5]

A second Oracle-tied campus, in New Mexico, drew about $18 billion from roughly 20 banks, with SMBC, BNP Paribas, Goldman Sachs, and MUFG as administrative agents. Pricing was discussed at the same SOFR+250 level, again with a four-year tenor plus extensions, and the leads planned a retail syndication. [6] [7]

Other publicly reported construction and term deals show the same cast:

• JPMorgan led a $7.1 billion financing for Crusoe and Blue Owl’s Abilene, Texas campus (Stargate), where construction began in June 2024. [8]
• Bank of America was structuring agent and a major lender on about $14 billion of debt for a Related Digital / Oracle campus tied to Stargate; the bank later announced a broader $250 billion, 18-month (January 2026–July 2027) digital-infrastructure financing initiative covering on-balance-sheet loans, project bonds, and project capital. [8] [9]
• Natixis, MUFG, and Société Générale, with RBC and SMBC as arrangers, launched a $3 billion Meta-backed Ohio package ($2.1 billion plus $900 million) at SOFR+250, four years plus a one-year extension. [10]
• Morgan Stanley and MUFG arranged a $3.1 billion broadly syndicated term loan for CoreWeave to buy and install Nvidia GPUs—the first GPU financing done in that format. [11]
• SMBC says it structured and led about $25 billion of North American hyperscaler data-center construction financing from 2022 through 2024, excluding deals where it was only a participant. [12]

What this means for entrants: Lead roles now require both project-finance underwriting (power, construction, lease assignment) and a distribution franchise large enough to move multi-billion hold positions. Pure balance-sheet lenders without a takeout story are being asked to step back.

Pricing and syndication: from FOMO to a distribution problem

Initial clearing levels for investment-grade-tenant construction loans clustered near SOFR+250. That level has not held for incremental Oracle-linked paper, and the time required to place the $38 billion package is the clearest evidence that syndication capacity, not origination appetite, is the scarce resource.

More than a dozen banks lent against Oracle’s long-term leases on the $38 billion Texas/Wisconsin facilities and the $18 billion New Mexico campus at SOFR+250. By late January 2026, TD Cowen reported that borrowing costs on newer, not-yet-sold Oracle-linked data-center debt had widened to 300–450 basis points over SOFR—closer to high-yield than to classic infrastructure. [13] The $38 billion package, underwritten from around August 2025, took until around April 2026 to approach full distribution across more than two dozen banks and other investors, with less than $1 billion still seeking a home. Retail lenders needed to commit at least $300 million to earn the top fee tier. [14] [15] Business Insider reported diminished incremental interest as JPMorgan worked through the tail of that syndication, even as bankers described the projects as fully financed. [16]

Secondary marks have already slipped on at least one large ticket. Syndicate banks, including Santander and Jefferies, were quoted around 89–91 cents on the dollar on roughly $18 billion of loans tied to the Oracle-leased New Mexico campus, according to reporting cited by Disruption Banking. [17]

Weaker or construction-stage credits clear in the bond market at much higher coupons. Morgan Stanley arranged a $3.2 billion TeraWulf construction bond, Google-backed, at a 7.75% yield, explicitly to avoid staged bank drawdown reviews. [11] Applied Digital sold 6.75% and 7.00% senior secured notes due 2031, with Goldman as representative. [18]

Demand is no longer uniformly strong. JPMorgan CFO Jeremy Barnum said in July 2026 that the bank had passed on some data-center deals over power-supply and tenant questions. [19] The Information reported on September 30, 2026 that Société Générale and MUFG had pulled back from certain data-center loan deals. [20] Apollo’s Torsten Slok has noted order books on some AI-linked offerings thinning from about five times oversubscribed in February to less than two times by July. [21] By contrast, SoftBank’s separate $15 billion one-year AI bridge—led by Mizuho, SMBC, and JPMorgan, with Goldman and MUFG among participants—was more than twice oversubscribed, showing that relationship bridge capacity still exists when the credit is the sponsor rather than a single campus. [22]

What this means for entrants: Underwriting at SOFR+250 and distributing six to nine months later into a SOFR+350 market is a negative-carry problem. New lenders should assume flex language, delay-draw mechanics, and a pre-committed bond or private-credit takeout, not a hope that the bank market will absorb the hold.

How concentrated the bank books actually are

No major U.S. bank has disclosed a standalone data-center loan number. Third-party reconstructions and regulatory snapshots therefore diverge by definition—retained mortgages versus commitments versus “AI-adjacent” commercial and industrial (C&I) loans that include software borrowers.

AtriumData.ai, as reported by Bisnow in August 2026 (rankings as of around March), estimated that U.S. data centers and related infrastructure carry at least $1.3 trillion of debt. That total is broader than bank loans: utility and infrastructure loans about $428 billion, syndicated facilities about $225 billion, and hyperscaler corporate credit about $225 billion. The top 15 lenders accounted for about $196 billion, or 15% of the total. Named exposures in that cut:

• Goldman Sachs about $16.5 billion (likely understated given its arranger role; at least 48 syndicated deals)
• JPMorgan about $13.3 billion ($11 billion syndicated, $2.3 billion mortgages)
• MUFG about $11.8 billion
• SMBC about $11.5 billion across roughly 60 syndicated deals
• Morgan Stanley about $9.3 billion
• Mizuho about $8 billion
• Bank of America about $12.8 billion; Citi about $10.5 billion; Wells Fargo about $16.4 billion

The three Japanese banks together were about $31.6 billion. [8] [8] These figures are retained or tracked exposures, not peak underwriting commitments, and they likely lag the late-2025 mega-deals still being distributed in 2026.

Infralogic’s 2025 project-finance rankings show the same shift in flow. The five U.S. bulge-bracket banks lent about $13.28 billion on data-center project finance in 2025 versus about $600 million in 2024. JPMorgan ranked first in that slice at $6.72 billion. MUFG participated in 146 project-finance deals overall in 2025, far more numerous if smaller than the U.S. houses’ tickets. [23]

Federal Reserve Bank of Chicago staff, using supervisory data through late 2025, put large-bank C&I commitments to “AI-adjacent” borrowers (software, infrastructure, data-center construction, and loans secured by data centers) at about $450 billion, of which about $150 billion was outstanding. That commitment share rose from about 9% of C&I commitments in 2015 (about $250 billion) to 13% in late 2025. Direct outstanding exposure to AI-adjacent industries was about 0.8% of total assets. Across large banks, C&I outstanding to that group averaged about 9% of Tier 1 capital, with committed exposure closer to 25%. MSCI Real Capital Analytics counted only $14.9 billion of bank lending to data centers in the year through the third quarter of 2025—a property-transaction measure, not project-finance commitments. [24] [17]

A Columbia Business School estimate cited by American Banker puts bank exposure to AI physical-infrastructure debt—mortgages, syndicated loans, and project debt—at about $800 billion, with many banks described as near internal concentration limits and leverage on some projects as high as 90% debt. [25] [26] That figure is an academic aggregation, not a regulatory total, and it is wider than the Chicago Fed’s C&I snapshot.

Working estimate (inference, triangulating the sources above, not a single published total): retained bank credit specifically against data-center real estate and construction is most plausibly in the low-to-mid hundreds of billions globally. Adding AI-adjacent C&I commitments and loans to private-credit vehicles that themselves finance campuses pushes the banking system’s economic exposure into the mid-hundreds of billions, and possibly toward the $800 billion academic estimate if utility lending and broadly defined infrastructure are included. The $1.3 trillion Atrium figure is system debt, not bank debt. Banks remain the origination bottleneck even when they are not the final holder.

What this means for entrants: The visible $8–17 billion books understate risk because the same hyperscaler tenants (Oracle, Meta, Microsoft, Google, Amazon, and neoclouds such as CoreWeave) recur across campuses. Single-name and single-tenant limits bind before sector limits do.

Risk transfer: SRTs, CLOs, and ABS/CMBS takeouts

Banks are not waiting for construction loans to season. They are trying to sell the riskiest slice of a single campus, refinance delay-draw loans into bonds as soon as they are drawn, and use CMBS and ABS as the permanent takeout once a lease is in place. The transfer is real, but it is incomplete, and some of the buyers are already long the same assets in equity or private credit.

The Financial Times reported on May 3, 2026 that JPMorgan, Morgan Stanley, and SMBC were seeking ways to distribute data-center exposure, including a variant of significant risk transfer aimed at large, concentrated loans rather than a diversified reference portfolio. JPMorgan and MUFG had already spent more than six months distributing the $38 billion Oracle-linked construction debt; some banks tried to sell pieces at a discount to non-bank lenders. Linklaters had seen deals in the $500 million range backed by a single borrower, with the bank typically retaining a slice. [27] [27] Octus reported in July 2026 that data-center-specific SRTs were pricing in the high single digits (about 7–9%), tighter than a broader infrastructure SRT range of 8–12%, even though sources said development and construction risk often made up 50–70% of the referenced portfolios. Critics described buyers as investors who already hold equity in the same assets—an “asset-gathering” loop rather than clean risk transfer. One cited example was a TD Bank repeat SRT on a $2 billion pool (0–8% first-loss bought by Blackstone) that included data centers plus other assets and ramp features. [28]

Bond takeouts are the preferred exit when a hyperscaler backstop exists but does not cover construction. A Morgan Stanley-led consortium planned to refinance about $15 billion of debt on a Google-backed, Anthropic-leased Texas campus into the bond market as soon as delay-draw loans were funded. The bonds were expected to be speculative-grade because Google’s support activates only after completion, leaving investors with overrun and delay risk. Banks had also spent months seeking buyers for more than $50 billion of construction debt on several Oracle-leased projects. [29] [21]

Securitized takeouts are scaling but remain a fraction of the construction pipeline:

• About $17 billion of data-center CMBS has been issued since the start of 2025, more than triple the prior two years, and data centers are roughly 8% of new commercial-property bond deals. Most are single-asset, single-borrower. Citigroup expects about $18–20 billion of data-center CMBS next year. AAA data-center bonds have averaged about 165 basis points over their floating-rate benchmark, wider than office, retail, or industrial. Blackstone closed a $2.05 billion CMBS loan on a QTS portfolio in February 2026. [30] [31]
• Barclays counted $11.9 billion of data-center ABS issued so far in 2026. Spreads on hyperscaler ABS were about 150 basis points over the five-year Treasury, and the paper has started trading more like hyperscaler corporates than like generic ABS. [32]
• Apollo tallied $81.4 billion of U.S. data-center securitization (ABS and CMBS) since 2018, of which $18 billion was in the first half of 2026. Europe and the UK remain negligible by comparison ($1.7 billion and $2.3 billion since 2018). [33]
• Morningstar DBRS recorded $9.51 billion of U.S. data-center securitization issuance in the first quarter of 2026 alone, about three-quarters ABS historically, with CMBS regaining share via large SASB deals. [34]
• On the CLO side, Bank of America strategists estimated in September 2026 that CLO exposure to chip and data-center term loans could reach $100 billion, versus roughly $4.5 billion today. PGIM anchored a $500 million Allstate CLO, arranged by BNP Paribas, that caps AI collateral at 15%—the first explicit cap of its kind. [35]

Morgan Stanley has been the most visible architect of the non-bank structures: about $65 billion of corporate bond deals for data centers or other AI investment led or co-led since October 2025, including the Meta–Blue Owl Hyperion project bond tied to a roughly $27 billion Louisiana joint venture. [36] [11]

What this means for entrants: Construction-phase capital is a bridge, not a home. The economic spread is earned by whoever can commit the construction loan and simultaneously pre-place the SRT tranche or the post-completion bond. Holding the residual first-loss slice of a single Oracle or neocloud campus is a concentration decision, not a yield decision.

What regulators and analysts are actually saying

Official commentary has moved from “the stock of AI debt is still modest” to “the flow is now large enough to matter,” with the sharpest warnings aimed at opacity, circular financing, and the re-entry of transferred risk through non-bank channels—not at a classic bank capital shortfall today.

Bank for International Settlements. Bulletin No. 120 (January 7, 2026) argued that AI investment is large enough to require a shift from operating cash flow to debt, with private credit playing a rapidly increasing role. Outstanding private-credit loans to AI-related companies had risen from near zero to over $200 billion; risks looked moderate but hinged on earnings expectations that equity markets had priced more aggressively than debt markets. [37] [38] The March 2026 Quarterly Review (Eren, Krohn, and Todorov) documented hyperscaler gross bond issuance topping $100 billion in 2025, mostly longer than five years, and described joint-venture and special-purpose-entity financings with private credit as “shadow borrowing”—economically like debt, largely off corporate balance sheets. [39] The BIS annual report, as summarized in late June 2026, put AI-related capital expenditure by the five largest hyperscalers at more than $1 trillion in 2025–2026 combined, flagged poorly disclosed lease structures that could allow the same asset to be pledged more than once, and noted that private-credit funds had quadrupled lending to AI and IT companies over five years, with AI around 15% of those loan books, while banks are increasingly exposed by lending to those funds. [40]

Bank of England. The July 2026 Financial Stability Report said AI-related use of public markets, private credit, leveraged finance, and structured finance had accelerated and was set to rise further, from a stock that was still modest at the start of 2026. It highlighted off-balance-sheet vehicles, securitized data-center structures, and the risk that long-dated debt is financing buildings that may obsolesce if they cannot support newer hardware. On SRTs, the Bank noted that transferred credit risk can re-enter the banking system through banks’ exposures to non-banks that hold the SRT positions, making the distribution of risk more complex and less transparent. [41] [17] On September 30, 2026 the Financial Policy Committee said the likelihood that interconnected vulnerabilities crystallize had risen, citing a rapid increase in AI-related debt. It referenced Morgan Stanley’s estimate of about $450 billion of global AI-related debt issuance by early September 2026, roughly double 2025, and JPMorgan’s projection of about $4.1 trillion of AI-related debt issuance from 2026 to 2030. It also cited Morgan Stanley’s view that about $700 billion of data-center capital expenditure in 2026–2028 could be financed by private credit, inside a roughly $2.9 trillion global data-center spending forecast through 2028. The countercyclical capital buffer was left at 2%. [42] [43] [44]

IMF. The April 2026 Global Financial Stability Report treated AI-related equity concentration and interconnected financing as an amplification channel, but judged the near-term financial-stability impact modest. It noted hyperscalers had raised more than $100 billion in bonds since January 2025, supplemented by leveraged loans and sometimes circular intercorporate arrangements, against a projected AI-related capital-expenditure path that some estimates put near $3.4 trillion by the end of the decade. [45] [46] A secondary citation of Box 1.3 in that report, in a 2026 academic draft, says the IMF recorded $46 billion of data-center securitizations since 2018 (about 70% ABS, 30% CMBS) and an expectation that data-center securitization could reach $150 billion by 2028. That box figure was not independently re-extracted from the IMF PDF in this research pass, so it should be treated as reported rather than re-verified. [47]

Federal Reserve system. There is no single Board of Governors concentration limit or public aggregate for data-center loans. The Chicago Fed’s April 2026 note is the clearest supervisory snapshot: AI-adjacent C&I is a rising but still moderate share of large-bank balance sheets, and a small fraction of all AI-related debt (JPMorgan analysts’ roughly $1.2 trillion of AI-company-issued debt was cited as context). The note frames the issue as tail risk and as hard-to-measure indirect exposure through loans to private-credit lenders. [24] The Kansas City Fed’s October 2, 2026 Economic Bulletin showed AI firms’ share of investment-grade bond issuance rising from 2% in 2023 to 10% year-to-date 2026, with about $330 billion of AI-related investment-grade issuance already in 2026—ten times full-year 2023—and noted that data-center construction is a growing but still modest share of bank commercial real estate. [48]

Analyst and rating-agency overlay. Morningstar DBRS (August 27, 2026) said concentration is one of the most significant credit issues for banks in the sector, and that syndication, securitization, and risk-transfer structures shift exposure to institutional investors while reducing transparency and complicating system-wide assessment. [49] Goldman Sachs analysts tracked nearly $500 billion of AI-related debt sold through early August 2026. [21] The Reserve Bank of Australia’s October 2026 stability review, relevant as an outside central-bank read, warned that off-balance-sheet data-center funding and circular financing could create hidden exposures, citing external estimates of $1–1.5 trillion of related financial obligations. [50]

What this means for entrants: Regulators are not, on the public record, telling banks to stop originating data-center loans. They are watching whether distribution creates a less transparent version of the same concentration—in insurers, private-credit funds, CLO equity, and SRT buyers who are financed by the same banks. A lender that cannot show both a hold limit and a credible, non-circular takeout will be the one supervisors ask about first.

---

Recent Findings Supplement (October 2026)

Recent data-center financing shows continued large-scale lending by JPMorgan, Goldman Sachs, MUFG, and others, alongside emerging pullbacks, tougher terms, and growing use of risk transfers amid rising scrutiny.[1][2]

Major recent deals (announced or advanced in summer-fall 2026): JPMorgan and MUFG are in advanced talks to underwrite/lead a ~$22 billion loan for Vantage Data Centers’ planned $25 billion Texas campus (1,200 acres, phased completion mid-2026 to 2028), with Goldman Sachs in the broader syndicate and $3 billion equity from Silver Lake/DigitalBridge.[1][1] Goldman Sachs is leading/premarketing a $3.5 billion high-yield bond for Applied Digital’s Louisiana campus (15-year lease, potential Meta/Amazon tenant) and a ~$1.15 billion secured note for Digital Drive’s Virginia campus (CoreWeave tenant).[3] Other pipeline items include ~$15 billion Nexus bridge financing (Texas, Anthropic lease) and Morgan Stanley-led bond takeouts for Crusoe’s $5.5 billion bridge plus a $509 million Goldman-led GPU SPV loan.[3] An $18 billion syndicated facility (BNP Paribas/Goldman-led) supports a major New Mexico Oracle-related project.[2]

Bank of America’s $250 billion 18-month initiative (Jan 2026–July 2027) includes a closed April 2026 $16 billion Michigan Oracle/OpenAI data center financing. Morgan Stanley announced a parallel $1.5 trillion 10-year infrastructure push, while Goldman participates in a $500 billion AI infrastructure collaboration.[4]

Lender pullbacks and syndication dynamics have shifted since mid-2026. Société Générale and MUFG stepped away from certain data center loan deals (reported late September 2026).[5][2] Order books for some offerings thinned sharply (5x oversubscribed in February 2026 to <2x in July). Lenders are demanding tougher terms amid power delays, local opposition, and project risks (e.g., Oracle force majeure notice on a New Mexico site, flagged by Morgan Stanley analysts in late September as weighing on debt markets).[2][6] Pricing examples include JPMorgan-led deals discussed around SOFR+500 bps with 98–98.5 OID and rising bond yields (e.g., certain Meta-backed from ~6.58% earlier to 7.53% in July; CoreWeave-linked slices near 10%).[3][7]

Risk transfer tools (SRTs, securitizations) are seeing increased activity to manage concentration. Deutsche Bank is structuring its first project-finance SRT on a ~€2 billion ($2.3 billion) book including data center loans (September 2026).[8] JPMorgan, Morgan Stanley, and SMBC have explored synthetic risk transfers (SRTs) and private sales of data center exposures to free capacity.[9][10] Securitization/ABS markets are absorbing volumes, with ~$11.9 billion data center ABS YTD (Barclays data) and expectations of $30–40 billion annual absorption; CMBS and other structures are also used for takeouts.[11][2]

Aggregate exposure estimates and concentration remain elevated. Goldman Sachs analysts tracked nearly $500 billion in AI-related debt issuance through early August 2026.[2][12] Morgan Stanley forecasts $2.9 trillion global data center capex through 2028 (with a $1.5 trillion financing gap after hyperscaler cash flows).[2] AtriumData.ai estimates at least $1.3 trillion in U.S. data center debt (utility/infrastructure loans the largest slice).[13] Japanese megabanks (MUFG, SMBC, Mizuho) hold significant project finance exposure, including gigawatt-scale AI data center deals; Japan’s FSA highlighted concentration risk management in its July 2026–June 2027 priorities (published mid-September).[14] Chicago Fed estimates place direct AI-adjacent exposure at ~0.8% of large-bank assets (C&I ~9% of Tier 1 capital; committed closer to 25%).[10]

Regulatory and analyst commentary has intensified on concentration and systemic risks. BIS head Pablo Hernández de Cos (September 10, 2026) flagged AI infrastructure debt/private credit financing as opaque and interconnected, with top-five tech firms investing >$1 trillion in 2025–2026 and potential vulnerabilities from valuations and concentration.[15] Bank of England’s Financial Policy Committee (late September 2026 record) noted rising likelihood of interconnected vulnerabilities crystallizing, citing rapid AI debt growth (~$450 billion by early September per Morgan Stanley) and broader cyber/operational risks.[16] Reserve Bank of Australia’s October 2026 Financial Stability Review warned of hidden risks from off-balance-sheet structures, opaque interlinkages, and circular financing in data centers (roughly one-third debt-funded globally), though leverage remains contained for now.[17][18]

These developments indicate banks are scaling lending while actively managing balance-sheet pressure through syndication, pullbacks from marginal deals, and risk transfers—yet regulatory bodies are increasingly highlighting concentration and opacity as potential fault lines. Newer data (post-August 2026) centers on pullbacks, specific large Texas/Vantage talks, SRT execution examples, and coordinated warnings from BIS/BoE/RBA/FSA.