The artificial-intelligence boom has entered the bond drawer. A Bloomberg story republished by EnergyNow puts the expected data-center buildout above $3 trillion and follows the money through bonds, leases, special-purpose vehicles, private credit, power projects, and refinancing risk [1]. That is a different newspaper story from the one told by the equity chart. It asks not who bought Nvidia or OpenAI exposure, but who already owns AI credit without knowing the name on the risk.

The paper's June 2 account of Anthropic's IPO optionality beside its compute bill said AI finance was no longer only a laboratory or venture-capital story. Its May 27 account of record stocks beside fragile oil made the same point in another market: public prices can look serene while operating risk gathers underneath. Data-center debt is the fixed-income version of that split.

The public argument on X is easy to summarize: AI is either the next industrial revolution or a leveraged fever dream. The mainstream business frame is narrower and more polite. Bloomberg describes a scramble to finance the concrete, transformers, fiber, land, and power behind AI demand [1]. Both frames miss the household-sized consequence. A mutual fund, pension plan, insurance account, or target-date portfolio can carry the boom through credit instruments whose marketing label says infrastructure, real estate, utilities, or high grade.

The distinction matters because hyperscaler demand does not automatically make every financed building safe. A data center can have a tenant, a power assumption, a lease, a refinancing date, and a set of lenders who are not all taking the same risk. If a project depends on one customer, one grid interconnection, or one assumption about future GPU utilization, the bondholder owns a different asset than the headline promises. It is not just a building. It is a claim on future AI revenue translated into credit.

That translation is where opacity enters. Equity investors at least know when they have bought a chipmaker, a cloud company, or an AI lab proxy. Bond investors can inherit the exposure through municipal-style structures, private placements, securitized leases, or corporate issuers that sit one step away from the marquee tenant. The risk migrates from the story everyone debates to the document few readers open.

There is also a power problem. A data center is an electrical asset before it is an intelligence asset. The EnergyNow/Bloomberg account treats power plants and grid availability as financing variables, not scenery [1]. That turns utility queues, gas turbines, transmission lines, and cooling systems into credit facts. A project that cannot get power on schedule cannot service debt on schedule merely because AI demand remains fashionable.

The best way to read this market is to separate three claims. First, AI demand is real enough to pull capital into physical infrastructure. Second, financing that infrastructure requires instruments that can move risk away from the companies most associated with the boom. Third, those instruments can reach portfolios whose owners never voted to buy an AI thesis. The first claim is an industrial story. The second is a Wall Street story. The third is a retirement story.

The debt market has always been good at making one thing look like another. Mortgages became structured products. Aircraft became leases. Toll roads became yield. Now data centers are becoming a way to own AI indirectly, with the reassuring language of infrastructure wrapped around the volatile economics of compute. That does not make the bonds bad. It makes the label insufficient.

The useful question is not whether AI is a bubble. It is where the bubble, if one forms, would surface. In stocks, it announces itself on a screen. In data-center bonds, it may appear later, as refinancing, tenant concentration, utility delay, or a covenant no one quoted in the launch story. The portfolio statement will not say it plainly. The documents will.

That is why ordinary investors need a better vocabulary than optimism or doom. A data-center bond can be secured, unsecured, project-linked, lease-backed, or indirectly exposed through a borrower whose credit story depends on tenants and electricity. Each label changes the loss path. In a benign version, AI demand keeps buildings full and lenders collect steady income. In the harder version, capacity arrives late, power costs rise, tenants renegotiate, or refinancing happens after rates and sentiment have moved. The market will not fail all at once. It will sort the careful documents from the euphoric ones.

Debt has a habit of making booms look prudent until the payment calendar arrives. The AI buildout may prove durable. It may also prove uneven. Either way, the reader who owns bond funds should learn to ask a new question: not whether the portfolio owns AI stocks, but whether it owns AI promises with interest due.

-- THEO KAPLAN, San Francisco