Last week, two figures emerged about a single AI data center campus in Saline Township, Michigan, figures that could become a much larger story.
The project is being built by developer Related Digital to power applications for OpenAI, with Oracle as its primary tenant. The price tag for all this has jumped to $16 billion, up from the $10 billion estimated just last fall. Also in the mix: PIMCO, one of the world’s largest bond fund managers, which is in talks to provide $14 billion of that $16 billion as debt.
But the rising cost is not the most notable aspect of the deal. Neither is the “twisted” story of the financing, which Bloomberg reports has involved months of “break” negotiations, amid increased scrutiny of Oracle’s creditworthiness.
What’s really noteworthy is the light the deal sheds on AI data center debt and how it can find its way into retail investors’ portfolios (specifically, their retirement savings) without them realizing it. This can also happen through bond funds, which most ordinary people consider to be the “safe” end of the market, a world away from risky technology bets or straight AI plays like, say, Nvidia.
Regardless of how they are perceived, bond funds represent a huge market and a significant portion of the retirement savings of millions of Americans. According to the Investment Company Institute, Americans had about $8 trillion in 401,000 accounts in 2023, of which about $420 billion was in bond funds, which represent a little more than 5% of 401,000 assets overall, while the percentage increases in the portfolios of older workers. More importantly, however, the $420 billion figure does not take into account target-date funds, which account for more than 40% of the total holdings of $401 billion and have substantial bond holdings that increase as the target date approaches.
What the Michigan data center deal highlights is a little-discussed aspect of the broader development of AI: that some of this data center debt can find its way into incredibly large bond pools. In other words, ordinary Americans are likely already funding it, in part, and through the most conservative corners of their retirement savings, without even knowing it.
PIMCO’s presence in the agreement is a good example. The opposite of a niche player in the bond world, PIMCO manages about $2 trillion in assets for central banks, sovereign wealth funds, pension funds, insurance companies and individual investors, placing it among the world’s largest bond fund managers. Its funds appear on major 401k platforms nationwide, while the PIMCO Income Fund alone manages approximately $225 billion in assets and is among the most widely held bond funds in U.S. retirement accounts.
The Michigan deal would be PIMCO’s second major data center financing in less than a year. Last year, the firm helped anchor an $18 billion debt package for Meta’s Hyperion data center in Louisiana, and posted $2 billion in paper profits as prices for that debt rose after the shutdown. The Michigan deal could follow a similar playbook: PIMCO would purchase the bonds, which would be structured as 144A securities (more on such bonds in a moment) and possibly syndicate them to other institutional buyers.
The firm has also had no qualms about addressing this corner of the debt market in public statements. PIMCO’s own investment outlook for 2026 makes clear that the company is actively seeking exactly this type of exposure. While expressing caution about private credit in general, the company created a specific category that it finds attractive: “project financing or loans secured by data centers being built with existing leases to investment-grade tenants.” The Michigan campus, with Oracle as a tenant, the lease in place and the deal structured as project financing, fits those requirements.
The channel through which such debt reaches retail investors is complicated and structural, but not accidental. The Bloomberg Global Aggregate Index, the PIMCO Income Fund’s benchmark, explicitly includes US dollar 144A investment grade securities as a component. 144A securities are privately placed debt instruments that retail investors cannot purchase directly, but can be held in mutual funds that retail investors hold in their 401ks. By design, any fund compared to the Bloomberg Global Aggregate is very likely to hold them. PIMCO’s own brochure confirms this.
In this sense, it is reasonable to assume that some bonds used to finance AI data centers are also held in funds that serve retail investors. Whether the specific securities generated by the Michigan deal ever appear in a given retail fund, the deal illustrates the channel through which AI infrastructure debt could reach ordinary people: through the part of their portfolio they probably chose precisely because they thought it was safe and low risk.
Another catch: Oracle is the anchor tenant of the Michigan data center, but Oracle is not borrowing this money directly. The debt is held by Related Digital, the developer, through a special purpose vehicle secured against the data center asset itself. Oracle leases the facilities once they are built, and those lease payments are what pay off the debt.
The chain works like this: OpenAI pays Oracle for computing. Oracle makes lease payments to Related Digital. Related Digital uses those payments to repay the bonds. Bondholders receive their payment; At that time, if 144As were held in a bond fund, fund holders would receive a small proportional interest payment.
But this means that there are two links that must be maintained, not one. OpenAI has to keep paying Oracle, and Oracle has to keep paying Associated Digital. If OpenAI reduces its compute consumption, renegotiates, or has financial problems, Oracle’s revenue is reduced, but Oracle still owes the lease. If Oracle stops paying, Related Digital defaults.
The SPV structure is specifically designed to keep this debt off Oracle’s balance sheet. Which raises an obvious question: why would it be necessary?
Oracle’s current debt-to-equity ratio appears to be about 400%, perhaps higher, and that’s only the debt that’s actually on its books. The company, founded by Larry Ellison, spent decades as one of the most profitable and trusted enterprise technology companies, selling database software and related services to corporations that had few practical alternatives. It was never a sexy consumer technology brand, never a household name, and yet it was an extraordinarily complicated business; Once a company built its operations on Oracle’s database, the change was costly enough that most would never switch.
Still, in more recent years, as that legacy business has come under pressure, Oracle has moved aggressively to reinvent itself as an AI infrastructure and has poured into building the data centers that AI needs to function. That turn requires debt on a scale the old Oracle never had. As liabilities have grown and financial structures have become more complex (across the industry, not just at Oracle), some of that debt has migrated off the balance sheet entirely, making Oracle’s true total liabilities really difficult to measure from the outside.
By all appearances, financing only works if several aggressive assumptions are met simultaneously.
OpenAI needs so much long-term computing, and has or will have the revenue to pay for it, from a company that has never made a profit. Oracle can pay off tens of billions of debt while its cash flows, at least arguably, are already strained. And the underlying hardware does not become obsolete before the debt is repaid, which is a relevant risk given that AI chips are renewed approximately every two to three years, even when the bonds carrying this debt have maturities measured over much longer periods.
None of these assumptions are necessarily unreasonable. But they all have to be correct, at the same time, for the math to work, just like many other 144A bonds. The kind that could be finding its way into the retirement savings of everyday Americans.
And whether this deal turns out to be a success story or a cautionary tale, it’s very difficult to think that individual investors who allocate their savings to bond funds would actually want to hold this type of debt. Which makes the details revealing and surprising, but not for the reasons the players involved might expect.
Bottom line: If the most conservative part of your portfolio happens to have potentially risky AI debt, then it’s arguably No the most conservative part of your portfolio.
