Good technology, bad credit

On September 18, 1873, Jay Cooke & Co, financier of the Union’s Civil War, collapsed, unable to sell the bonds of a railway built ahead of its traffic. On September 18 this year, NaBFID, India’s state-run infrastructure lender, said it had sanctioned over Rs 3,000 crore each for at least four data centres, with repayment holidays of up to five years. The same day, $18 billion of loans for an Oracle-leased AI campus in New Mexico, built for OpenAI, were quoted at 90c on the dollar, its banks struggling to sell them amid Oracle’s weakening credit and a blocked gas pipeline.

Read one way, India is catching the AI wave. Read another, its public balance sheet is joining a debt-fuelled boom of railway proportions.

Morgan Stanley expects global AI-related debt issuance of nearly $570 billion this year, more than double last year’s. Optimists note that Walmart ran negative free cash flow for its first 15 listed years yet enriched shareholders. But its new stores earned far more than they cost to finance.

Debt by another name

A joint venture borrowed $27.3 billion to build Meta’s Hyperion data centre. Meta took an initial four-year lease but gave a capped 16-year guarantee on the campus’s value. Bondholders lend against Meta’s promise, yet Meta books none of it as debt. Goldman Sachs puts hyperscalers’ total lease commitments at about $1.5 trillion, $1 trillion of it not yet begun and so off-balance sheet. Veterans of 2008 remember off-balance sheet is not off the hook.

The giants still look sober for their size: low leverage, top-notch ratings. But capex now outruns cash flow; Amazon’s free cash flow has turned negative, and Alphabet’s briefly so; Alphabet has sold $85 billion of shares. Oracle is the exception. Otherwise, the leverage sits around them: in vehicles, neoclouds, and labs.

The customer who pays with your own money

Walmart’s shoppers spent their own wages. Much of a hyperscaler’s backlog comes from still-unprofitable AI labs; OpenAI alone was 45% of Microsoft’s in January. The labs raise cash from Amazon, Nvidia, SoftBank, and, earlier, Microsoft, then spend much of it on their backers’ computing. It is a house whose valuer, mortgage bank, and builder share an office.

Hyman Minsky called borrowers who can pay only by raising fresh money “Ponzi” units. The labs mostly raise equity, but their compute contracts behave like debt. Can fast-growing revenues cover commitments many times larger? The alarm is loudest at Oracle, whose default insurance hit records near 200 basis points, and which S&P rates one notch above junk.

It’s the slowdown, not the crash

Herbert Stein, Richard Nixon’s adviser, said if something cannot go on forever, it will stop. In 2006, subprime defaults rose when house-price growth merely slowed. UBS expects hyperscaler capex growth to fall from 76% this year to 6% by 2028. Slower building means slower revenue for neoclouds and cooler valuations, both needing each funding round priced above the last. “As long as the music is playing, you’ve got to get up and dance,” Citigroup’s Chuck Prince said in 2007. Hyperscaler bosses, fearing rivals, are still dancing.

Not 2000, and maybe not 1873

The dot-com bust was mostly a share bubble. Shareholders can wait. Lenders cannot. Telecom is the closer precedent: its fibre was eventually used, but it made up over half the world’s corporate bond defaults in 2002. And railway track lasted generations; most data-centre spending goes on chips depreciated over six years that some reckon obsolete in three.

Sceptics say the risk sits with insurers and private credit, not deposit-taking banks. But JPMorgan and Morgan Stanley, nearing internal risk limits, have sought to offload data-centre loans. What banks cannot sell, they keep; what they sell may return, as private credit borrows heavily from them.

What it means for India

“There is nothing so disturbing to one’s well-being and judgement as to see a friend get rich,” wrote Charles Kindleberger. Foreign investors sold a record $29 billion of Indian shares in the first half, largely, says Jefferies, to chase Korean and Taiwanese AI hardware. It calls India the “inverse AI trade”: money leaving AI comes back here. True, in an orderly rotation; a 2008-style crunch strengthens the dollar and sells emerging markets first.

India is building capacity too, with over $300 billion pledged by 2030, Anarock reckons. A tax holiday till 2047 rewards foreign cloud firms serving global customers from Indian data centres, whose owners borrow `2-4 for each of their own. That makes them property businesses with a tech label. Bank credit grew 18.3% in the year to June, twice last year’s pace, as borrowers swapped bonds for loans. Infrastructure credit grew only 10.9%, but Minsky would recognise the mood.

The risk lies in the pipeline. Knight Frank counts 8.3 Gw planned against 1.6 Gw live; just 0.3 Gw is being built. NaBFID reckons the sector needs `1 trillion by 2031, likely lent mostly to a few groups at the top of a boom. India’s power-plant lending spree of 2006-11 began the same way. NaBFID plans to raise about 40% of this year’s `1-trillion borrowing abroad, much of it through an RBI swap window. But the swap lasts at most five years; its dollar bonds 10 and 15. NaBFID keeps the remaining currency risk, for savings that have all but vanished, exactly what a 2008-style crunch would test. With the Fed raising rates last month, the test need not wait for a crunch.

Look past the lease to whoever pays. A building let to Microsoft is safe while the lease runs; one let to a neocloud serving a loss-making lab is not. Indian lenders mostly finance the shell, power, and cooling, not chips. Their risk is not rotting collateral but tenants who never arrive, or halls built for air-cooled racks AI has outgrown. NaBFID’s five-year holidays, plus the extensions RBI rules permit, could let a failing project look healthy for years.

A data centre hosting the financial system cannot go dark when its owner defaults. That makes it too important to switch off, and its losses public. As Rudi Dornbusch warned, crises take longer to arrive than you expect, then happen faster than you thought possible.

Railways transformed America. They also bankrupted many who financed them, Jay Cooke among them. AI will change the world. That does not guarantee it will repay those lending to build it.

The writer is Former Executive Director, Nomura, alumnus of Harvard Business School, and a visiting faculty at B-schools.

Disclaimer: The views expressed are the author’s own and do not reflect the official policy or position of Financial Express.