News

Banks seek to curb AI debt exposure as backlash mounts

Banks that financed the AI build-out are trimming exposure. ING and Société Générale have put as much as $800mn of data-centre debt up for sale in recent weeks, including loans tied to Oracle's Stargate initiative with OpenAI. ING is auctioning $500mn of loans linked to CyrusOne, EdgeConneX, Vantage Data Centres and Oracle-backed Project Jupiter in New Mexico, which is struggling for electricity; SocGen is selling $300mn of project-finance debt for Oracle's Project Lighthouse (Wisconsin) and Project Frontier (Texas). The three Oracle projects raised more than $56bn of construction debt from dozens of banks last year. The Wisconsin and Texas loans are quoted above 90 cents on the dollar (described by one person as 'a portfolio management exercise'), while New Mexico is expected to price in the mid-80s after Oracle issued a force majeure notice over power-permit delays. Banks are also exploring credit insurance and risk transfer and structuring more deals in 144A format; a $63bn debt package for Anthropic and $40bn of SpaceX chip financing are set to trade in the 144A bond market in coming months, and some major lenders dropped out of the Anthropic deal over concentration. The Bank of England is asking UK banks for data-centre loan details and the ECB is working on a supervisory review. Nuveen's Don Dimitrievich calls the capital involved 'historic by any definition'; Linklaters' David Lucking says it is 'capital-inefficient for banks to just hold on to the debt'. The context is the Democrats' midterm sweep and data-centre opposition. Caveats: largely anonymous sourcing; ING, SocGen, the BoE and the ECB declined to comment; parts were previously reported by Bloomberg and 9fin; pricing above 90 cents suggests no acute credit stress; confidentiality clauses limit hedging via CDS or risk transfers.

Why it matters

Bank appetite for data-centre loans is a key input for AI infrastructure companies and neoclouds raising debt. If lenders start distributing risk to bond and private-credit investors and supervisors start probing exposure, financing for AI build-outs gets more expensive and more selective, especially for weaker projects such as power-constrained sites.

Read the full article: Financial Times

More top stories