Data Center Financing Goes Junk: Where Sub-Investment-Grade AI Money Sits

In April 2026 a joint venture backed by Fluidstack sold $5.7 billion of five-year high-yield bonds at 6.25% to build roughly 430 MW of AI capacity in rural Indiana. It was the largest AI-driven financing the high-yield market had completed. Data center financing now has a junk-rated tier, and its price is set by whoever stands behind the lease, not by the borrower.

What the junk tier of data center financing is

High-yield debt is debt rated below investment grade, meaning below BBB minus at S&P or Baa3 at Moody's. The market calls it junk. It pays more because the borrower is more likely to default, and it is bought by funds whose mandates allow that risk.

For most of this cycle, AI data center financing ran through investment-grade channels. Microsoft, Meta, Alphabet and Amazon sold unsecured corporate bonds on the strength of their own balance sheets. That tier is covered in the breakdown of AI debt and private credit exposure, and it is not the subject here.

Where does junk debt sit in AI data center financing? It sits below the hyperscalers, at the level of individual project companies, former bitcoin miners and GPU cloud specialists. These borrowers have no investment-grade rating of their own. They borrow against a single campus and a lease signed with a much stronger tenant, usually for five years.

That last detail is the whole story. The borrower is junk-rated. The rent check often is not. The high-yield market is being asked to price the gap between the two, and it is pricing it at volumes nobody planned for at the start of 2026.

My view is that "junk" is the wrong word for most of these deals and the right word for some of them. The work is telling which is which, and the coupon tells you most of what you need.

What Meridian Arc's $5.7 billion actually bought

The issuer is Meridian Arc HoldCo LLC. It is an indirect wholly owned subsidiary of Meridian Compute LLC, according to Latham & Watkins, counsel to Next Frontier. Meridian Compute is a joint venture between subsidiaries of Frontier Development Holdings LP, known as Next Frontier, and Fluidstack Ltd.

The notes carry a 6.25% coupon and mature in 2031. They were sold privately to institutional buyers rather than registered with the SEC, which is standard for high-yield deals of this kind. Cahill, counsel to the initial purchaser, described it as the largest high-yield bond offering for a data center project to date.

The structure in three layers

The proceeds buy land, then build, then reserve. Latham lists them in that order. First, an approximately 140-acre property across multiple parcels in New Lebanon, Sullivan County, Indiana. Second, a 245 MW turnkey data center, a 185 MW turnkey data center and an on-site electrical substation.

Third, the proceeds fund debt reserves and pay fees. The two buildings sum to about 430 MW of critical IT load, which matches Cahill's figure. Critical IT load is the power available to the servers themselves, excluding cooling and other site overhead.

So the bondholders are funding a construction project, not an operating business. On the day the notes priced, there was no revenue at the site. Interest is covered by reserves until the buildings deliver, which is why the reserve line matters more than it looks.

What "supported by Google" does and does not tell you

Cahill's release says the campus obligations are "supported by Google." Neither public release says how. It could be a lease guarantee, a backstop on a tenant's rent, a warrant arrangement or something else. The offering memorandum will say, and it is not public.

The closest public precedent involves the same operator. In October 2025, TeraWulf's subsidiary WULF Compute priced $3.2 billion of 7.75% senior secured notes due 2030. According to TeraWulf's pricing release, the collateral included a designated lockbox account of Fluidstack USA I. Until the expansion was finished, it also included a pledge by Google of warrants to buy TeraWulf stock.

That is a specific, limited form of support. It is not a Google guarantee of the bonds. If you read "supported by Google" as "Google owes the bondholders," you are reading something the public documents do not say. I would treat it as the single most important undisclosed term in the deal.

High-yield bonds for AI are priced off the tenant

Lay the final coupons side by side and a pattern appears. The deals carrying hyperscaler support price near the bottom of the range. The deals where a GPU cloud specialist pays the rent, or is itself the borrower, price near the top.

The benchmark matters here. The average coupon on newly issued US dollar high-yield debt was close to 7.2% in late 2025, per Janus Henderson citing BofA Global Research, as reported by Cointelegraph. Meridian Arc priced about a full point inside that. CoreWeave, borrowing on its own name in 2025, paid about 2 points over it.

The CoreWeave figures show what a GPU cloud borrower pays without a stronger party behind it. Its wider debt structure is examined in the map of CoreWeave's debt stack and its stress points.

The same borrower, two prices

The cleanest test available is one company that appears twice. Applied Digital paid 9.2% on notes issued in November 2025, per the Cointelegraph compilation. In September 2026 it marketed a $3.5 billion high-yield bond for its Delta Forge 1 campus in Louisiana, backed by a 15-year lease to Meta or Amazon.

Early price whispers on that deal were in the mid-7% to 8% range, according to Octus. Same borrower, roughly a point or more cheaper, and the main thing that changed was the name on the lease. Whispers are not final terms, and rates moved in between, so treat the gap as directional.

The contrast runs the other way too. Digital Drive premarketed about $1.15 billion of secured notes for a Virginia site with CoreWeave as tenant. Whispers there sat in the high-8% to 9% range. A hyperscaler on the lease is worth something close to 1 to 1.5 points of coupon in this market, and I would not underwrite a deal that assumed otherwise.

Final terms come from the issuers' counsel or the issuer. Every row marked whispers or talk is pre-pricing guidance reported by Octus on 9 September 2026 and may differ from where the deal closed.

Term loans for data centers: the buyout market's new borrower

Bonds are not the only junk-rated channel. In June 2026, The Information reported that Morgan Stanley was pitching data center developers on institutional term loans, the floating-rate market that has historically funded private equity buyouts rather than infrastructure.

The coverage, summarised by AI Weekly, says Morgan Stanley expects about $20 billion of AI-related deals to reach these markets in 2026. The original headline put the figure at $15 billion. The two numbers come from the same report, and I could not open the original to settle which is current, so read it as $15 billion to $20 billion.

The longer view comes from JPMorgan. Strategists led by Tarek Hamid estimated in November 2025 that the AI build would need about $1.5 trillion of investment-grade bonds over five years. The high-yield bond and loan markets would provide around $150 billion over the same half decade, per Bloomberg Law.

Why a developer wants a loan, and why a buyer should hesitate

A term loan floats over SOFR, the benchmark overnight rate, so the borrower carries rate risk rather than paying for fixed money. It can usually be repaid early without the call penalties bonds carry. That suits a developer who expects to refinance into cheaper debt once a building is leased and running.

The buyers are the attraction for the banks. The AI Weekly summary notes that this market has established buyers in collateralised loan obligations and credit funds. Collateralised loan obligations, or CLOs, are pools that buy hundreds of loans and sell slices of the pool to investors.

The Volta deal shows what the collateral can look like. Octus reported that part of the $5 billion JPMorgan-led package funds GPU purchases, with loan talk around SOFR plus 500 basis points at an issue price of 98 to 98.5. A lender secured on GPUs is secured on an asset whose useful life is disputed, as covered in the analysis of GPU depreciation and AI earnings.

I would want a much wider spread for GPU collateral than for a powered building. A building with a grid connection holds value across chip generations. A rack of accelerators does not.

Where sub-investment-grade money sits in the AI capital stack

The phrase "AI debt" hides at least five different instruments with different buyers and different collateral. The junk tier is two of them, sometimes three. Here they are from safest to riskiest, as I read them.

The comparison between the second and third rows surprised me. Meta's Hyperion vehicle sold A+ rated bonds at 6.581% in October 2025, as set out in the analysis of Meta's Hyperion bond and who carries its risk. Six months later Meridian Arc sold high-yield bonds at 6.25%.

That does not make junk cheaper than investment grade. Hyperion's notes run to 2049 and Meridian's to 2031, and longer money costs more. It does show how thin the extra pay for the junk tier has become when a hyperscaler sits somewhere in the structure.

The bottom row of the table is where the junk tier is heading. Octus reported a roughly $15 billion bridge financing for Nexus Data Centers' Anthropic-leased site in Texas, and a bond takeout of $5.5 billion of bridge financing for Crusoe. Bridges are designed to be refinanced. The question is into which market, and the high-yield market is the obvious one. Crusoe's own economics are covered in the breakdown of Crusoe's valuation and contracted capacity.

The five-year bond against the fifteen-year lease

MacKay Shields notes that the bulk of data center high-yield issuance this year carries five-year maturities. Meridian Arc matures in 2031. WULF Compute matures in 2030. The leases these bonds lean on often run much longer, such as the 15-year term on Applied Digital's Louisiana campus.

None of this is hidden. It is a deliberate trade. Five-year money is cheaper and easier to place than fifteen-year money in the high-yield market, so the issuer takes refinancing risk to save coupon today.

That trade only works if the refinancing market is open in 2030 and 2031. It also means a cluster of AI campuses will need new money in the same two years, from the same pool of buyers who bought them the first time.

What a refinancing in 2031 depends on

Three things, in my order of importance. First, whether the tenant is still paying and still wants the space, which depends on the AI demand that justified the build. Second, whether the hyperscaler support still applies, since the TeraWulf warrant pledge falls away once the expansion completes.

Third, where spreads sit that year. A campus that refinances in a calm market rolls over quietly. One that comes due in a stressed market pays whatever the market demands, and the equity holders absorb it. I would rather own the lease than the bond, because the lease survives the refinancing and the bond is the refinancing.

What the high-yield index is absorbing

The concentration is the part most readers will miss. MacKay Shields reported that AI data center deals made up 45% of US high-yield net issuance in the first half of 2026. Net issuance counts new money, not refinancing, so this is the part of the market that is actually growing.

Anyone holding a broad high-yield index fund owns a growing slice of AI campus debt, whether they chose to or not. Concentration in one theme is how a sector-specific problem becomes an index problem.

The same outlook flagged that valuations look stretched. The spread to worst on the ICE BofA US High Yield Index finished the second quarter at 294 basis points, below the 325 to 525 range MacKay Shields calls the post-crisis "non panic" range. The firm warned that data center issuance could outpace demand and push spreads wider.

Defaults are not the problem yet. MacKay Shields put the high-yield default rate at 1.5%, against a 25-year average of 3.4%. The risk is price, not loss: tight spreads plus heavy supply means new buyers are being paid less for a theme they already own a lot of.

Octus quoted a banker saying that more information on regulatory, political and geographic risks was making it harder for issuers to set price talk. That is the market starting to discriminate between deals, which I read as healthy rather than alarming.

Where this argument is weakest

The central claim here is that tenant credit sets the coupon. The evidence for it is thinner than the conclusion sounds, and you should know exactly how thin.

The sample is small and badly timed

The coupon comparison rests on about seven final coupons and a handful of price whispers. They were issued across roughly 18 months, during which base rates and spreads both moved. Some of the gap between CoreWeave's 9.25% in May 2025 and Meridian's 6.25% in April 2026 is market timing, not tenant quality.

The Applied Digital comparison uses whispers, not final terms. Whispers often tighten before pricing, which would strengthen the point, and sometimes widen, which would weaken it. A fair reader should hold the 1 to 1.5 point estimate loosely.

Junk is a label, not a verdict

The strongest counter-case is that much of this debt is not really junk risk at all. A secured bond on a finished campus with a 15-year hyperscaler lease behaves more like infrastructure debt than like a buyout loan. The rating reflects a new borrower with no history, not a weak asset.

On that reading, the high-yield market is offering a pricing mistake, and the buyers are being paid a premium for novelty rather than for risk. Low defaults support that view today. If the leases hold, the people calling this a junk-bond bubble will have been wrong, and the buyers will have been paid well for taking a label.

I also could not see the Meridian Arc offering memorandum. If "supported by Google" turns out to mean a full lease guarantee, the deal is safer than this post implies. If it means a narrow warrant pledge like TeraWulf's, it is riskier than the coupon suggests.

Frequently asked questions

Where to start if your vendors sit on this debt

If you run a business that buys GPU capacity, your provider or its landlord may be one of these borrowers. Two checks are worth an afternoon.

First, find out who owns the building your compute runs in, and whether that entity has issued secured notes. Pricing releases like TeraWulf's are public and short. Read the collateral paragraph, because it tells you who the bondholders can reach if the rent stops.

Second, note the maturity year of that debt in your vendor risk register. A capacity contract that runs past 2030 depends on a refinancing your vendor has not done yet. Ask in your next renewal what happens to your service terms if the facility changes hands.

Related reading