The AI Buildout Is Becoming a Bond Market Story

For three years the AI story was a stock story. The companies building the future sold shares, raised mega-rounds at vertical valuations, and funded their compute out of cash flow and equity. That era is ending — not because the spending is slowing, but because it has grown too large for equity to carry alone. The next phase of the buildout is being financed with borrowed money, and the scale of the borrowing is now its own headline.

The number

According to a Morgan Stanley forecast reported on June 10, 2026, global debt issuance tied to artificial intelligence is on track to more than double to nearly $570 billion in 2026. The pace is already there in the data: AI-related issuers had sold close to $236 billion of debt globally by May 31 — roughly four times the amount raised over the same stretch of 2025. With hyperscaler capital spending projected to cross $1 trillion in 2027, the bank expects bond sales to accelerate through the second half of this year.

The definition is broad on purpose. It captures debt raised to fund AI infrastructure across the hyperscalers — Amazon, Alphabet, Meta, Microsoft, Oracle — plus semiconductor firms and data-center developers, in both dollar and non-dollar markets. The throughline is physical: this is money borrowed to pour concrete, buy GPUs, secure power, and stand up the data centers that frontier models run on. Intelligence may be the product, but the balance sheet is buying buildings and silicon.

The deal that shows the mechanics

The clearest example of how this works is the financing assembled for Anthropic. Apollo and Blackstone are arranging a private-credit transaction of about $36 billion — potentially one of the largest such deals ever and quite possibly the biggest chip-financing transaction to date. The borrowed funds buy Google's custom TPUs, the tensor processing units that Google develops with Broadcom. Those chips are then leased back to Anthropic for deployment across data centers in New York, Texas, Louisiana, and Indiana.

The structure is worth unpacking because it is becoming the template. Anthropic, however explosive its growth, is still a young company whose own credit could not easily support $36 billion of debt. So Broadcom — which has the manufacturing relationship and the cash flows — backstops payments on the largest portions of the deal, lending the transaction a credit foundation the AI lab could not provide on its own. The financiers get a securitized claim on hard assets (the chips) plus a backstop from a profitable supplier; Anthropic gets compute without raising another $36 billion of dilutive equity.

This is happening alongside, not instead of, Anthropic's equity story. The company recently raised $6.5 billion at a $965 billion post-money valuation, briefly overtaking OpenAI on paper. The lesson is that even a company commanding a near-trillion-dollar valuation is now reaching for leverage to fund compute — because the appetite for GPUs and TPUs has outrun what equity rounds, however large, can practically supply.

Why it changes the risk picture

When AI was funded by equity, the downside of overbuilding was contained: shareholders who chose to ride a volatile sector absorbed the losses if demand disappointed. Debt distributes the exposure differently and far more widely.

Morgan Stanley's own framing is the part that should make ordinary savers pay attention. For bond investors — including the index funds and target-date funds sitting inside everyday 401(k) accounts — the AI buildout is no longer just a stock story; it is becoming the largest single position in the investment-grade credit market. That is a structural shift. Exposure to whether AI capex pays off is migrating from people who deliberately bought a high-growth narrative into the default retirement allocations of people who never made that choice.

The risk is not that any single deal goes bad. It is the correlation. Much of this debt is collateralized by the same class of assets — data centers and the GPUs and TPUs inside them — whose value depends on the same assumption: that demand for AI compute keeps climbing fast enough to justify a trillion dollars a year of spending. If that assumption holds, the debt is serviced and the buildout looks prescient in hindsight. If demand stalls or the hardware depreciates faster than the loans amortize, the losses do not stay contained inside a few tech stocks. They surface in the credit market, where the exposure is now broad and partly involuntary.

The circularity problem

There is a more subtle worry threaded through these structures. A chip vendor backstopping the debt used to buy its own chips; a cloud provider financing the GPUs it then rents back to a lab; suppliers, lenders, and customers increasingly the same small set of names. Each individual arrangement is rational. In aggregate they create a web in which the health of the lender, the supplier, and the borrower are bound to the same underlying bet, and a stress in one corner can transmit quickly to the others. Concentration like this is invisible while volumes are rising and ruinous if they reverse.

What to watch

The forecast is the easy part; the tells are in the terms. Watch whether issuance actually accelerates in the second half as Morgan Stanley expects, or whether spreads on AI-linked debt start widening — the first sign that credit investors are demanding more compensation for the correlation they are taking on. Watch how many more deals adopt the Anthropic template of a profitable supplier backstopping a young borrower. And watch the depreciation schedules: the entire edifice assumes today's GPUs and TPUs hold their value long enough to pay back the loans they financed.

The AI race used to be measured in benchmark scores and valuations. Increasingly it will be measured in coupons, maturities, and credit spreads — and that is a market with a much longer memory for the difference between a build and a bubble.