For three years the AI story was a silicon story. The question that moved markets was who could secure Nvidia allocations and stand up the next gigawatt campus. That framing is now out of date. The binding question in mid-2026 is no longer who has the chips or the land. It is who will lend against them, on what terms, and what happens to the lender if the compute does not earn its keep.
The AI buildout has quietly become a credit story. The capital no longer comes mainly from cash flow and equity. It comes from the bond market.
The pivot from equity to debt
Through 2024 the hyperscalers funded their data-center surge largely out of operating cash. That cushion is gone. Capex outran it. So the largest balance sheets in technology went to the debt markets at a scale the corporate bond market has rarely seen.
Meta set the tone. On 30 October 2025 it sold $30 billion of bonds in a six-part deal, the biggest investment-grade offering of the year, drawing an order book near $125 billion. Days earlier Alphabet had raised $25 billion. In February 2026 Oracle priced what Bloomberg called the largest corporate bond sale in history — $25 billion across eight tranches. In June 2026 Nvidia, of all issuers, raised $25 billion in its largest-ever debt deal.
These are not isolated trades. Morgan Stanley estimates AI-related debt issuance will nearly double to roughly $570 billion in 2026, with $250–300 billion coming from hyperscalers and their joint ventures alone. AI and tech paper has grown from a rounding error to nearly 12% of the investment-grade market in barely two years. The asset class did not exist at this scale before; now it is the single largest new source of supply in global credit.
The structure matters more than the size
The headline bond deals are the visible part. The more consequential shift is structural, and it is designed to keep the debt off the issuer’s balance sheet.
Meta’s Hyperion campus in Louisiana is the template. In October 2025 Meta and Blue Owl Capital closed a $27 billion financing — the largest private-credit transaction ever executed. The mechanics are the point. Meta owns 20% of the joint venture; Blue Owl-managed funds own 80%. A special-purpose vehicle arranged by Morgan Stanley issued $27 billion of A+ rated debt plus equity, anchored by PIMCO and BlackRock. Meta keeps operational control of the data center. It does not carry the debt.
This is project finance wearing a tech-company name. And it is spreading into corners the SEC never sees. A growing share of data-center development is funded through 144A placements — bonds sold privately to institutions without SEC registration. Applied Digital, CoreWeave, Hut 8 and Related’s data-center arm have together added more than $40 billion in 144A debt since November. JPMorgan projects data-center securitization alone could run $30–40 billion a year in 2026 and 2027.
The buildout is being financed by instruments that are progressively harder to see, harder to price, and further from the issuer’s reported leverage.
Where the credit market is already nervous
Pricing is beginning to register the risk. Oracle is the clearest tell. Its five-year credit-default swap widened from about 40 basis points at the start of 2025 to roughly 200 by March 2026 — a near-record level that implies a default probability no investment-grade name should carry. Oracle raised FY2026 capex guidance 43% to $50 billion and ran free cash flow deeply negative. Its 10-year tranche priced around 145 basis points over Treasuries, wide for the rating. With roughly $120 billion of Oracle paper sitting in the Bloomberg US high-grade index, that repricing is not Oracle’s problem alone. It is every index fund’s problem.
That is the concentration trap. When one sector goes from negligible to a tenth of the investment-grade market in two years, passive funds and institutional mandates inherit a technology bet they never chose to make — the credit-market echo of an over-concentrated equity index. The diversification a bond allocation is supposed to provide quietly erodes.
The deeper question is the collateral. These deals are underwritten against contracted compute revenue and the residual value of GPUs and buildings. Contracts assume tenants stay solvent and keep paying — and a meaningful share of demand traces back to a handful of AI labs whose own revenue is young. GPUs depreciate fast and their resale value is untested at scale. The buildings are durable; the economics inside them are not yet proven across a full cycle.
Implications
The financing shift changes what an AI slowdown would actually break. In an equity-funded boom, a disappointment shows up as falling share prices and trimmed capex plans — painful but self-correcting. In a debt-funded boom, a disappointment shows up as stressed coupons, downgrade cascades, and forced selling by funds that must hold investment-grade paper. The transmission runs straight into pensions, insurers and money-market-adjacent vehicles. That is a wider blast radius.
Three things are worth watching from here.
First, spreads on AI paper relative to the broader index. As long as the gap stays modest, the market is treating this as ordinary corporate credit. A sustained widening — the Oracle pattern generalizing — would signal that lenders have started pricing these as the project-finance bets they are.
Second, the off-balance-sheet share. The more capex migrates into SPVs, private credit and 144A structures, the less reported leverage tells you about real exposure, and the harder a stress event is to map in advance.
Third, the identity of the ultimate lenders. Private credit, insurers and securitization vehicles are absorbing risk that used to sit with banks under tighter supervision. Morgan Stanley reckons $800 billion of private credit will be needed for AI infrastructure, power and fiber through 2028. Who holds it when the music slows is the question regulators have not yet answered.
The chips were always the easy part. They could be bought. The bill for buying them at this scale is now a fixed obligation on someone’s balance sheet — and increasingly on a balance sheet built specifically so you cannot see it. The AI trade has become a credit trade. Credit trades end differently than equity ones.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.