Goldman Sachs’ credit strategists counted $296 billion of AI hyperscaler debt supply this year as of November 5. Their forecast for all of 2026 was $250 billion. They now expect $440 billion in 2027, and say almost all of the increase will come from private credit, which is mostly unrated. Those numbers are why AI financing became the market’s favorite worry this autumn.
The worry is fair as far as it goes. Long-lived buildings are being financed against GPUs that lose value fast, and a campus leased to one tenant on a short contract carries real renewal risk. But “AI debt” is not one asset. It is at least two very different credits that happen to share a headline, and the market has started to price them differently. That is the healthy part.

Start with the top of the stack. US companies sold $1.88 trillion of investment-grade bonds through November 4, a record that beat the $1.86 trillion of 2020. Hyperscalers accounted for more than $370 billion of it, about a fifth, with Amazon and Alphabet the largest borrowers. These are issuers with enormous operating cash flow. Morgan Stanley’s framework puts global data center capex at roughly $2.9 trillion through 2028 against about $1.4 trillion of hyperscaler operating cash flow. So yes, there is a financing gap, about $1.5 trillion. But more than half the bill is covered from internal cash before a single bond is sold, and the bonds that are sold come from companies whose default probability is close to the sovereign’s.
The bottom of the stack looks different, and it is being treated differently. Junk-rated data center bond issuance reached roughly $37 billion this year against about $8.6 billion in all of 2025. By mid-October average yields on those bonds were close to 10%, and the jump happened in about a week. Yields on the better data center paper rose toward 7% from under 5% earlier in the year. That is not a market that has stopped asking questions. It is a market charging a leveraged developer with one tenant a lot more than it charges Amazon.
That repricing is the mechanism that keeps a buildout from turning into a bubble. In 1999 the telecom credit market lent to everyone at similar spreads until it lent to no one. In 2026 the spread between a hyperscaler’s own bond and an off-balance-sheet campus SPV is already wide, and it widened in weeks. Weak projects will find capital expensive or unavailable. The projects that clear a 10% cost of debt are the ones with signed take-or-pay contracts and power already secured, and those are exactly the ones you want built.
The structures themselves have improved too. Meta’s $27 billion joint venture with Blue Owl for the Hyperion campus in Louisiana put a private-credit balance sheet behind a facility whose sole user is one of the most profitable companies in the world. Lease backstops, residual-value guarantees and step-in rights are now standard asks. A lender to a data center in 2026 usually has a hyperscaler’s credit somewhere in the chain. That was not true of the fiber-to-nowhere loans of the dot-com era.
Demand is the other half of the coverage ratio. Nvidia guided its October quarter to $108 billion of revenue, and its own 10-Q showed supply commitments of $279 billion as of late July, mostly for memory and manufacturing capacity. The big five hyperscalers are projecting about $725 billion of capex in 2026, up 77%. Capacity in the tightest markets is full on arrival: Johor’s colocation vacancy rate is 0.7%. Debt raised against capacity that customers are queuing for is a very different instrument from debt raised against hope.
For investors the useful questions are specific. Who is the tenant, and what is the contract tenor against the debt maturity? Is the power connected or merely applied for? Does the borrower have a parent with cash flow, or only a GPU fleet as collateral? Debt service coverage, customer concentration and utilization will separate the winners from the casualties. Some casualties are coming, especially among single-tenant developers that financed at the top. That would be the credit cycle doing its job.
The constructive case does not require every borrower to survive. It requires the market to keep doing what it did in October: lend cheaply to cash-generating platforms, lend dearly to speculative builders, and refuse the worst deals outright. On the evidence of this autumn’s spreads, it is doing exactly that. AI compute is turning into a utility-like layer of the economy, and capital markets are learning to finance it the way they finance utilities, with a premium for leverage and a discount for contracts. Related reading on this site: how 5% Treasury yields change the AI math.