Fitch's credit analysis of $90B in AI data center JV bonds reveals a dangerous pricing error — investors are betting on brand names, not contract terms.
Fitch's credit analysis of $90B in AI data center JV bonds reveals a dangerous pricing error — investors are betting on brand names, not contract terms.

The $90 billion AI data center joint-venture bond market is pricing a dangerous assumption: that a Google logo on the lease guarantee is worth the same as ironclad contract language. Fitch Ratings' analysis shows otherwise.
"Investors look at whether the parent company providing the lease guarantee is Google or Meta, without considering that rent payments only start after construction is complete," Anubhav Arora, senior director at Fitch Ratings, said.
Oracle, the weakest balance sheet among the five hyperscalers, earned Fitch's highest credit score in this asset class — because its Michigan data center project committed to unconditional rent payments regardless of construction status. By contrast, three Google-backed FluidStack joint ventures totaling nearly $11 billion in bonds received ratings ranging from investment grade to near-junk, Citi analysts found, depending on whether the contracts included rent discounts, tenant exit rights or construction cost caps.
The divergence matters because the market has been treating these bonds as a homogeneous asset class, with tightly clustered yields. Goldman Sachs credit strategists expect spreads to widen as investors wake up to the contractual differences — a repricing that could raise financing costs for future AI infrastructure projects at a time when six hyperscalers have already issued about $244 billion in debt this year, more than double last year's pace.
The FluidStack case study illustrates the problem. Citi analysts reviewed three data center JV projects backed by Alphabet's Google, each with a different lease guarantee structure. The first project carried Google's full lease guarantee — if the cloud operator defaults, Google steps in directly. The second included a 25% rent discount on Google's guarantee. The third tied the rent formula to operating costs and allowed FluidStack and Google to exit the lease if construction was delayed beyond six months — a clause that, for bondholders, means no rental income and no recourse if the project stalls.
Only one of the three — the $3.25 billion HUT 8 DC LLC project in St. Francisville, Louisiana — received an investment-grade rating. It combined Google's full lease guarantee with a maximum-price construction contract and no tenant exit rights.
The five largest hyperscalers have issued roughly $200 billion in corporate bonds this year, and six AI hyperscalers combined have raised about $244 billion — double the 2025 pace. Goldman Sachs chief credit strategist Amanda Lynam estimates the five giants have roughly $510 billion of remaining capacity in the US investment-grade bond market, a fraction of the $5.8 trillion in total capital expenditure Goldman projects.
That math makes JV bonds an indispensable financing tool. Tech companies have already diversified into euro, Swiss franc and sterling bond markets, and are tapping private credit channels through firms such as Apollo Global Management. But the terms of those deals remain highly bespoke, Arora said, and growing labor constraints in data center construction could give landlords more bargaining power, further fragmenting credit quality.
For fixed-income investors, the question is no longer whether AI infrastructure is worth funding. It is whose contract language stands on which side of the next repricing.
This article is for informational purposes only and does not constitute investment advice.