back
Whoever Cosigns AI's Rent Will Own Its Slowdown
EdgeFiles

Whoever Cosigns AI's Rent Will Own Its Slowdown

About the Author

Sources

Credit backstops look like insurance until AI revenue runs late, and then they start to look like ownership.

On the site of a former aluminum smelter in Hawesville, Kentucky, the data center developer TeraWulf is preparing three data center buildings for the AI lab Anthropic. The leases run 20 years, cover about 401 megawatts of computing capacity, and are expected to bring TeraWulf roughly $19 billion in contracted revenue. Anthropic's rent starts in stages as each data hall is delivered, beginning in late 2027 and continuing through the second quarter of 2028.

In November, Google agreed to stand behind that rent.

Google is hardly a neutral party here. It owns a stake in Anthropic, and in April Anthropic signed an agreement with Google and the chipmaker Broadcom for multiple gigawatts of capacity on tensor processing units (TPUs), the AI chips Google designs, starting in 2027. Now Google has agreed to backstop Anthropic's lease obligations so that TeraWulf can borrow against them. The form of the agreement TeraWulf filed with the Securities and Exchange Commission contains a detail I keep coming back to. For the first six years of coverage, Google cannot walk away from a lease as long as the Google TPU pods installed in the building can draw at least half of the power reserved for Anthropic's equipment.

One company now sits on three sides of the same building. It owns part of the tenant, it supplies chips to the tenant, and it guarantees the tenant's rent.

None of this is hidden. It is spelled out in securities filings, and it is a clear picture of how the AI buildout is now being financed.

What worries me is the timing. An earnings miss can change the price of a stock before lunch. A lease draws on a company's cash for fifteen or twenty years, and a guarantee can demand payment just as the guarantor has its own reasons to hold on to cash. The industry can be right about how useful AI will become and still be wrong about when it will earn enough to pay for what it has already ordered.

If that happens, the companies that lent their credit to the buildout will not simply absorb losses. Their contracts give them the right to take over leases and buy into landlords, and their cash will decide which projects get a second chance. Credit is how the boom is being built. It may also decide who owns the slowdown.

From the stock market to the contracts

The people paid to watch the financial system have noticed. In the spring, half of the market contacts surveyed by the Federal Reserve Bank of New York named artificial intelligence as a risk to U.S. financial stability over the next 12 to 18 months. By the fall, 87 percent did.

That made AI the most frequently cited risk in the Federal Reserve Board's Financial Stability Report, published on November 13.

The fall survey drew 23 responses in September and October, the spring survey 20 in March and April, all from professionals at broker-dealers, banks, investment funds, and advisory firms.

These are the contacts' views, not the Fed's. The same contacts saw cyberattacks and the sustainability of federal debt as the risks with the most severe potential consequences. Their AI concerns came in three parts: elevated stock valuations that weaker earnings could bring down, cyber damage from state actors or from frontier AI models themselves, and the increasingly complex ways the buildout is financed. Markets talk about the first. The third is the one that lasts.

The Fed's own analysis, in a box titled "Leverage in the Data Center Buildout," puts numbers on it. Debt issued for data centers and other advanced computing infrastructure passed $680 billion by October 2026, roughly five times the $130 billion issued in all of 2023. Its share of investment-grade nonfinancial corporate debt issuance rose from less than 6 percent in 2023 to about 31 percent this year. The estimate covers public deals and private securities sold under Rule 144A, which lets companies place securities with large institutional investors without a public offering. The Fed says its data do not capture all private deals, so the full figure is likely higher.

The borrowing has also changed shape, from mostly corporate bonds to a mix of public and private debt, structured credit, and project finance.

The more revealing numbers sit off the balance sheet. The Fed examined seven companies: Amazon, Alphabet, Broadcom, Meta Platforms, Microsoft, Nvidia, and Oracle. By mid-2026, leases they had signed but not yet started carried more than $1 trillion in future payments. Their other obligations came to more than $2 trillion, including take-or-pay contracts, which require payment whether or not the buyer takes delivery, and residual value guarantees, which promise that an asset will be worth a set amount later. Those obligations have grown faster than the companies' debt. The Fed cautions that the figures are undiscounted and not directly comparable with balance-sheet debt. It also says the companies' economic exposure to a downturn may be substantially larger than their on-balance-sheet liabilities suggest.

Oracle shows what this looks like from the tenant's side. In its quarterly filing for the period ended August 31, 2026, Oracle disclosed $288 billion in additional lease commitments, substantially all for data centers. The leases are expected to start between late 2026 and mid-2029 and to run fifteen to nineteen years, and none of it was on Oracle's balance sheet. On the same date, Oracle reported about $125 billion in notes payable and other borrowings. Undiscounted, the leases it has signed but not yet started are worth more than twice the debt it reports.

Credit without a loan

None of this requires a strong company to lend anyone money. It only has to sign something.

The most common instrument is the lease. A cloud company commits to years of capacity in a building that does not exist yet, and the developer uses that lease to secure financing for construction. The Fed describes exactly this sequence. The lender gets a claim on a project with a creditworthy tenant, and the cloud company locks in capacity long before it pays rent. A lender to the developer and a shareholder of the tenant now depend on the same contract, from opposite ends.

For the lender, the rent is what services the loan. For the shareholder, it is a claim on the tenant's future cash.

The second instrument is a guarantee on hardware. Makers of graphics processing units (GPUs), the chips that train and run most AI models, often guarantee a residual value for GPUs used as collateral, the Fed notes, to support purchases of their own chips.

The buyer gets financing, the lender gets a floor under its collateral, and the chipmaker sells more chips.

The third runs in a circle. Federal Reserve Governor Michael Barr described it in a speech in Milan on November 16. Cloud providers and GPU makers invest in smaller AI firms, and those firms use the money to rent or buy compute from the same investors. Barr put the cumulative equity investments at more than $100 billion. Money like that can get a young firm started. For the buildout to become economically viable, Barr said, those firms will eventually have to replace their suppliers' funding with revenue from paying customers.

Kentucky folds the same logic into a single relationship. The lease makes the building financeable, the guarantor also supplies chips to the tenant, and the guarantor is an investor standing behind a company it already owns part of.

What Google signed in Kentucky

The details matter, because a guarantee is a bundle of options, and options decide who holds power when things go wrong.

TeraWulf's subsidiary Raylan Data is the landlord on three leases with Anthropic. On November 4, TeraWulf, Raylan, Anthropic, and Google signed three credit support agreements, one per lease, under which Google agreed to backstop certain of Anthropic's obligations. TeraWulf said the support was arranged for the anticipated debt financing of the campus. Google's parent, Alphabet, is rated AA+ by S&P Global Ratings and Aa2 by Moody's on its senior unsecured debt, near the top of the corporate scale.

Google's support for each lease takes effect only when the last data hall under that lease commences. From then on, if Anthropic misses a payment or becomes insolvent, Raylan notifies Google, and Google chooses. It can pay the rent that is due and take over the lease as tenant. Under certain circumstances, it can instead let the lease terminate and pay a termination fee that shrinks over a defined term. Paying the fee gives Google the right to acquire an equity interest in Raylan, sized as the fee divided by the project's total development and construction costs. If Anthropic misses a payment without becoming insolvent, Google can put off its choice for a set period, as long as it covers the unpaid rent and the rent that keeps coming due.

The form of agreement narrows the exit. During the first six years of coverage, Google cannot choose termination unless the TPU pods installed in the building are capable of drawing less than 50 percent of the power reserved for Anthropic's equipment. In those years, unless that threshold has been crossed, silence counts too. If Google lets its 30-day review window pass without answering, it is treated as having taken over the lease. Google's commitment, in other words, is firmest where its own chips do most of the work. A guarantor that inherits a building running on its own hardware inherits something it can use.

The arrangement ends if Anthropic, or a qualifying affiliate that guarantees its obligations, achieves and keeps investment-grade credit ratings for at least three months. Until then, Google supplies the investment-grade credit the lease needs.

This is not the first time Google has lent its name to a TeraWulf project, and the earlier deal shows how such credit gets paid for. When Google backstopped TeraWulf's leases with the cloud provider Fluidstack, TeraWulf issued Google warrants to buy 73.58 million TeraWulf shares at one cent per share, according to its latest quarterly report. Google agreed to pledge those warrants to the lenders financing construction until the backstop takes effect. The guarantee was paid for in equity.

A lender can count Google's support when it decides how much to lend. The contract also tells the lender what it still has to price. The money goes out during construction, and Google's backstop does not cover that period. If the final hall under a lease arrives late, the date Google's support begins moves with it, while the developer's spending does not.

Then there is the moment the backstop would actually be used. Google is most likely to be asked to pay when demand for AI capacity has disappointed, which is also when its own cloud and chip businesses would be under pressure. Bankers call this wrong-way risk, exposure that grows as the party protecting you weakens. A guarantee from one of the strongest balance sheets in the world is still most likely to be tested when that balance sheet has other calls on it.

How a late payoff travels

The optimistic path needs little explanation. If demand keeps pace with construction, tenants grow into their commitments and the backstops are never called.

Suppose instead that businesses adopt AI more slowly than the buyers of infrastructure expected. A lab that has signed for capacity still owes the rent, even if its own revenue falls short of plan.

The obvious way to conserve cash is to defer the next order. That decision lands on developers who were counting on new orders to fill unused space, and who now approach refinancing with less revenue in sight. Barr said the firms with lower credit quality that depend most on outside funding had already seen their cost of capital rise significantly in recent months, and that they were turning to more complex asset-backed financing. A slowdown would find them already paying more.

That puts pressure on two assumptions every lender made at the start: how much cash the project will generate, and what the collateral will fetch if it has to be sold. Chips are the uncomfortable part. Barr noted that lease rates for older generations of chips appear to be holding up, but he called the trend worth monitoring, because specialized chips can lose value quickly once newer ones overtake them. If several borrowers need to sell similar equipment at the same time, recoveries will depend on whether anyone else is still buying.

The power infrastructure tells a different story. Barr expects substations, high-voltage connections, and grid interconnections to be useful in some places regardless of what happens to AI. In others, he said, it may take years for the excess capacity to become useful. His comparison was dark fiber.

Telecom built the network and lost it

That comparison deserves more than a passing mention, and the Fed makes it too. Its box sets the cumulative capital spending and debt issuance since 2020 of five cloud companies, Amazon, Alphabet, Meta Platforms, Microsoft, and Oracle, against the broadband telecom cycle of the late 1990s and early 2000s. The lines have tracked each other closely. With AI investment accelerating, the Fed says the current cycle may surpass the telecom cycle's pace relatively soon.

Telecom left the country with an enormous amount of fiber in the ground and a long list of carriers that could not pay for it. Much of that fiber stayed dark for years. It outlived the companies that laid it, and buyers with patience and cash picked it up for a fraction of its cost and lit it as demand caught up. The infrastructure was right, and the timing was wrong. The losses landed on the builders and their creditors. The gains went to whoever held cash when the assets came up for sale.

The AI buildout is wired differently, and the difference is the point of this piece.

In telecom, the buyers of distressed assets showed up after the bankruptcies. In AI, some of them are already named in the paperwork. Google's option to take over Anthropic's lease, its right to buy into Raylan if it pays a termination fee, and the warrants it holds in TeraWulf are pre-negotiated paths from guarantor to owner.

If AI revenue arrives late, the first choice over buildings financed on terms like these will not be made at an auction. It has already been granted, by contract, to the company that cosigned them.

Where the losses would land

A guarantor's strength protects a project's creditors. It does not keep the stress inside the AI industry.

The Fed says the on-balance-sheet debt of the large cloud and chip companies is held mainly by mutual funds, pension funds, and insurance companies. Financing tied to leases and guarantees often runs through special purpose entities, legal vehicles set up to hold a single project or pool of assets, which issue both senior and subordinated debt. The riskier subordinated debt and equity tranches sit with high-yield bond funds, life insurers, business development companies (investment funds that lend to midsize private companies), and other private debt vehicles. A cloud company's bond and a subordinated claim on a data center carry very different contractual rights. Both come under pressure if the same cloud company has to support the project while defending its own cash flow.

Banks are exposed as well. The Fed puts the largest U.S. bank holding companies' exposure to large technology firms, data center developers, advanced computing special purpose entities and private credit vehicles, and utilities at close to $1 trillion. It evaluates those exposures in its annual supervisory stress test, and the report describes the banking system as sound and resilient, with high regulatory capital ratios.

Barr said he is less concerned about direct losses, given the still-limited amount of debt behind the boom. His concern is leverage elsewhere. Hedge fund leverage appears quite high, and it is provided in part through prime brokerage, the financing and trading services large dealers sell to hedge funds. A sharp reversal, he said, could lead to rapid deleveraging and potentially self-reinforcing fire sales. That is the route by which trouble in data center finance becomes trouble for banks.

A strong sponsor can absorb losses that would overwhelm a developer, which limits the damage to other creditors. The harder case is a broad slowdown in which the same few sponsors face several calls on their cash while their own growth disappoints. If they cut new investment, the pullback reaches equipment suppliers and construction firms. Tighter credit then makes the next project more expensive to finance, even if its tenant is still paying.

Who can afford to wait

A data center can stay useful long after the loan that built it comes due. Power shifts in the gap between the life of the asset and the life of its financing.

A company with cash and strong credit can wait for demand to catch up. A developer carrying construction debt cannot, and neither can a lab whose rent is fixed while its revenue is not. In a downturn, the patient balance sheets decide which projects get another round of funding, which leases get assumed, and who gets to use the capacity.

In Kentucky, the contract even ties Google's commitment for six years to whose chips are in the building.

Consider what the equity alone buys. When Google's stake in Anthropic surfaced in court filings in 2025, it came with no voting rights, no board seats, and no board observer rights. The guarantee in Kentucky comes with the right to step in as tenant. Ownership gave Google a share of the upside. Credit gives it a say over the buildings.

That changes how I read the structure of this industry. The Fed counts exposures, and control does not show up in its tables. The guarantees being signed today are financial instruments, but they are also options on the future shape of AI: which labs stay independent, which chip architectures fill the buildings, and which cloud companies end up owning capacity that others planned. Those options are being written while capital still flows freely, which is when they are cheapest to grant.

I would ask any board weighing an AI infrastructure commitment to put a date on the payoff it is counting on. Then ask the harder question. If the revenue arrives two years late, whose cash keeps the project alive in the meantime, and what does that party get for it?