ING Groep's $500 Million Data-Center Loan Auction Shows Who Actually Holds AI Credit Risk

ING's total loan book was €903.8 billion at Sep 30, 2026, and €312.7 billion of credit outstandings sat in Wholesale Banking, where project finance lives. A $500 million portfolio is worth less than a tenth of one percent of either figure, so leaves ING's earnings and capital ratios essentially untouched. That is the point. The auction's output is not cash and not capital; it is a price, and prices are what everybody else in this chain is short.

A February 2026 study from the Federal Reserve Bank of Chicago found that banks' exposure to AI-adjacent industries averaged 0.8% of total assets. It reads like reassurance and it is closer to a receipt. Banks hold so little AI credit because they have sold it — into risk transfers, into insurer policies, into securitizations, and now into live auctions. The study's own caveat is the sharper point: banks also lend to private credit funds, so hides much of the AI credit exposure outside the measured 0.8%.

ING is running every channel at once. It completed two significant risk transfers in 3Q2026, freeing €3.0 billion of RWA, and booked a €111 million management overlay for an IFRS model update in its project finance portfolio on top of €208 million of Wholesale Banking risk costs, or 39 basis points of average customer lending. The bank is repricing the same book it is distributing, which is the clearest signal that project finance risk has moved from a pricing question to a placement problem.

JPMorgan Chase and Mitsubishi UFJ Financial Group arranged roughly $38 billion of construction financing for two Oracle-linked campuses — $23.25 billion for a Texas facility and $14.75 billion for a Wisconsin one — priced about 250 basis points over the benchmark. Spreading it took months across more than two dozen banks and investors, and as of Oct. 15 lenders were still trying to place less than $1 billion of it into 2026. The market absorbs AI construction credit far more slowly than banks originate it, and in the same week Bloomberg reported that hedge funds had been heavy sellers of AI debt. A sale process is how a bank finds out which of those two forces is winning.

Read down the right-hand column and the pattern is clear: every buyer that can absorb this paper prices it less often than a bank does. A BDC marks its portfolio each quarter; a private credit fund marks it when an investor asks or when paper is bought; an insurer re-underwrites annually; a securitization tranche marks daily but only against a rating set at issuance. The result pushes the AI credit cycle into slower, lumpier price discovery than the equity market prices the same story. That is why the Nov. 19 print matters more than a quarter of earnings from any single buyer.

Spread the same 100 basis points across both books and the arithmetic is manageable at one end and brutal at the other. On Oracle's $155.9 billion of debt, 100 basis points costs about $1.56 billion a year, or 4.1% of trailing EBITDA of $37.7 billion. On CoreWeave's $51.6 billion, the same 100 basis points costs about $516 million, or 13.1% of trailing EBITDA of $3.94 billion. CoreWeave's debt went from $29.8 billion at Dec. 31, 2025 to $51.6 billion at Jun. 30, 2026 while book equity stayed near $5.0 billion. That makes the marginal cost of construction financing a solvency question for the neoclouds and a manageable cost of business for the hyperscalers.

The one part of the chain nobody is arguing about is the equipment. Vertiv reported third-quarter net sales of $4.31 billion, up 61%, with organic orders growing more than 100% and full-year guidance raised to $14.65 billion of net sales. GE Vernova booked $29 billion of orders, up 96% organically, and a $197 billion backlog it expects to push above $200 billion in the fourth quarter, with gas equipment backlog and slot reservations of 122GW heading to at least 130GW by year-end. The physical build is racing ahead of the price of the money funding it.

The verified facts point one way: a bank with a 13.3% capital ratio put a dozen real AI construction loans into a public auction because it would rather own the fee than the risk; the equipment order books behind those loans are still growing at a double-digit clip; and the credit market for the same assets repriced sharply over the past two months. The inference is that AI data-center debt has quietly become a traded, marked asset class with a discoverable bid price, and that the marginal owner is no longer a bank. The speculation — and it is only speculation until Nov. 19 — is how wide the gap between last quarter's origination terms and today's clearing terms turns out to be.