Other People’s Money

THE NUMBER: 9.25%. That’s the coupon a Blue Owl-backed data center outside Richmond, Virginia paid on Wednesday to borrow $1.1 billion for five years. The building is 76 megawatts, and every one of them is leased to CoreWeave for fifteen years. Lenders charged about 2.7 percentage points more than they charge other junk-rated borrowers with the same rating. The same week, Amazon, Alphabet and Meta could borrow for five to seven years at roughly 60 basis points over Treasuries. Keep both numbers in your head. The gap between them is the most important price in AI right now, and it isn’t on any leaderboard.

The price of a stranger’s credit

Start with the Virginia deal, because it’s the cleanest look we’ve had at what lenders really think.

The project is called Digital Drive. Blue Owl affiliates, Cedarwood and PowerHouse Data Centers put it together. Goldman, Citi and Deutsche Bank sold the bonds. The collateral is a building plus a fifteen-year, $2.94 billion lease with one tenant, and the tenant is CoreWeave, whose own credit S&P rates below investment grade. S&P’s one-line summary of the risk was blunt: “Exposure to a single speculative-grade tenant remains a key risk.”

That’s the whole deal. The bondholders aren’t underwriting the building. A data center with no tenant is a very expensive warehouse full of air handlers. They’re underwriting CoreWeave’s promise to pay rent for fifteen years, and they want 9.25% for it. And CoreWeave’s rent, in turn, depends on customers like OpenAI and Microsoft keeping the GPUs busy.

It’s a chain of IOUs, and every link charges the one behind it for the privilege.

This isn’t the first time the market has flinched at this chain. In February, a $4 billion financing for a CoreWeave site in Lancaster, Pennsylvania couldn’t find specialty lenders. One senior executive at a large private-credit shop summed it up in five words: “We saw it. We passed.” Blue Owl said the project was fully funded, on time and on budget. BMO’s analyst called the difficulty a red flag. The deal this week got done, which is the good news. The price it got done at is the news.

An act of God in Doña Ana County

On Thursday we found out what happens further up the chain.

Oracle sent Blue Owl, its landlord and developer, a force majeure notice on Project Jupiter. That’s the campus Oracle is building on 1,400 acres in Doña Ana County, New Mexico, to supply computing power to OpenAI under the Stargate deal. It’s roughly 2.5 gigawatts and up to $165 billion of planned investment. The plan was to open in 2028.

Force majeure is the contract clause for events nobody could have planned for: wars, earthquakes, floods. Lawyers call them acts of God. Oracle’s act of God was a pipeline. The campus is supposed to run on gas-fed fuel cells, and New Mexico rejected the gas pipeline in March and again in July. The developer rerouted it and pushed its completion to February 2027. The air-quality permit for the fuel cells is due in November.

First, a correction, because a lot of the first-day coverage got it slightly wrong. This isn’t Oracle walking away and leaving the lenders holding the bag. Under the lease, Bloomberg reports, getting power to the site is Oracle’s job, Oracle pays the debt costs, and Oracle “cannot terminate the lease under any circumstances.” During construction, Blue Owl earns about 9% on its equity, and that steps up to about 11% once the campus is finished. The notice lets Oracle stretch out the cheaper phase and push back the step-up. Blue Owl says its financial commitments don’t change.

So why does it matter? Because Oracle agreed to carry the power risk to win OpenAI’s business, and now that the power risk has arrived, it’s reaching for the clause built for hurricanes. Oracle’s own statement said such notices are “commonplace in developments of this scale.” ZeroHedge’s reply was the best line of the day: could Oracle publish the full notice, since it’s so commonplace?

And the lenders are already acting as if it isn’t commonplace. The FT reported last week that $18 billion of Jupiter construction loans are quoted at 89 to 91 cents on the dollar, and that Santander and Jefferies are stuck holding more than they planned because the syndication stalled. Oracle’s credit rating sits one notch above junk after S&P cut it in July. Roughly half of Oracle’s contracted backlog, by most counts, is one customer: OpenAI.

Borrowed credit, lent to a borrower, to build for a customer that’s also borrowing.

Look at who’s late

The easy read is that AI demand is cracking. That’s the wrong floor.

In June, SemiAnalysis went after the popular claim that half of 2026’s US data center capacity would be delayed or canceled. Its answer: the forecast for data centers the big hyperscalers build for themselves moved about 1% over six months. The delays it could name clustered somewhere else: the Stack/Oracle site in New Mexico (now 2029), Nebius in New Jersey, Core Scientific’s campus in Denton. Newer builders. Renters. Borrowers.

Sightline’s count, via FXStreet earlier this month, shows the rest of the picture: 148 gigawatts of AI data center capacity announced since 2024, about 12 gigawatts running, about 21 under construction. Everything else is a press release looking for a lender. Transformers have lead times of 128 to 144 weeks. Utilities now want serious money down before they’ll connect you.

So demand isn’t the question. The question is who gets to meet it, and the credit market has started answering.

Three piles

Sort the players by where their money comes from, and the AI industry falls into three piles.

The builders. Google, Microsoft, Amazon and Meta. Top-tier credit, gigantic operating cash flow, and, most importantly, their own demand. When Google builds a data center, Google fills it. The bond market still lends to these companies at tight spreads, even as they borrow at a pace nobody has seen. Goldman’s numbers, reported by Reuters this week, say hyperscaler borrowing is headed to about $420 billion next year, up roughly 60%. There’s some strain even here: 78 of the 91 hyperscaler bonds issued this year were trading at higher yields in late July than when they were sold, and order books shrank from nearly five times the size of the deal in February to under two in July. And even the builders like to keep some of it off the books. Meta put its $27 billion Hyperion campus in Louisiana into a joint venture where Blue Owl funds own 80%, and a Meta-linked Texas data center bond this month was expected to price around 7.5%. But the builders choose other people’s money. They don’t need it.

The tenants. OpenAI, Anthropic, CoreWeave, and Oracle as landlord to much of it. These companies have the best models in the world, or the contracts to host them, and not enough balance sheet to build what they’ve promised. FXStreet’s tally puts OpenAI at roughly $600 billion of contracted compute against something like $130 billion of available liquidity. That gap gets filled with somebody else’s credit, and somebody else’s credit now costs 9.25%.

The equity kings. Musk’s companies, with SpaceX stock that went public this summer and Starlink cash, plus Nvidia, which increasingly finances its own customers. They build with equity and cash flow, which is why they can do things that look insane on a bond spreadsheet. More on that below.

Same boom, very different costs of money, and the cost of money compounds.

Other people’s money

In 1914, three years before he joined the Supreme Court, Louis Brandeis published a book called Other People’s Money and How the Bankers Use It. His target was the Morgan-era bankers who took depositors’ money and used it to build and control the railroads, the steel mills and the utilities. His most quoted line comes from that book: sunlight is said to be the best of disinfectants.

Seventy-seven years later Hollywood borrowed the title. In Other People’s Money (1991), Danny DeVito plays Larry Garfield, “Larry the Liquidator,” a raider who never builds anything. He circles New England Wire & Cable, the proud old family company run by Gregory Peck, and waits. Peck gives the speech about loyalty and the town and the workers. DeVito gives the speech about buggy whips. DeVito wins the shareholder vote, because a company that can’t earn its cost of capital belongs to whoever can.

Put the two together and you have this week. Brookings published a paper on Wednesday by Columbia’s Stijn Van Nieuwerburgh estimating $10.3 trillion of AI infrastructure investment from 2025 through 2032, about 3.6% of US GDP a year. His conclusion wasn’t that AI is a bubble. It was that the financing is moving off the balance sheet, into joint ventures, private credit, special-purpose vehicles and leases, where “correlated exposures” become “hard to observe before a downturn.” His prescription is pure Brandeis: better measurement and more transparency, now, while the capital structure is still being poured.

The FT’s Alphaville mapped what’s in the shade. A dataset from Sona Asset Management covers 176 AI financing deals worth $3.6 trillion among 202 companies. About 120 of them are flagged as highly circular: money going around a loop of suppliers, customers and investors. The reddest names on the map were Oracle and CoreWeave. You got a fresh example on Thursday. Anthropic committed $11.6 billion over seven years to Akamai’s cloud, and Akamai handed Anthropic a warrant for up to 5% of Akamai. The vendor paid the customer in stock to sign. Akamai’s shares jumped 16% after hours.

We wrote on August 5, in Not For Credit, that credit hates binary outcomes and demands a fortress before it lends into AI. This week the bond market named its price for lending without one.

We’ve seen this movie

Van Nieuwerburgh’s paper has one comparison that should make every allocator sit up. Relative to the size of the economy, he says, the AI buildout would be bigger than the canal boom, the railroad boom, electrification, the interstate highways and the telecom buildout.

Railroads are the rhyme.

In the 1870s the most celebrated financier in America was Jay Cooke. He had sold Union war bonds to ordinary citizens through a network of small bankers, insurance agents and real estate men, and then he turned the same machine on the Northern Pacific Railway. Cooke ended up owning most of the company and borrowing against bonds he hadn’t sold yet. The line got bogged down on the northern plains, the bonds stopped moving, and on September 18, 1873, Jay Cooke & Company closed its doors. The New York Stock Exchange shut for ten days. The Panic of 1873 kicked off years of depression.

Here’s the part that matters, though. The railroad got finished. The Northern Pacific reached the coast a decade later, the country got its railroads, and the ones who got hurt were the ones who financed it on promises. That’s what we meant on September 1 in Field of Dreams: be bullish on the build and bearish on the paper. The build is real. The paper is the part you can lose money on.

When the ground says no, look up

Here’s a question I kept turning over last night: will we get data centers in Greenland or in space first?

The logic behind both is the same as everything above. The binding constraint on AI isn’t chips anymore. It’s a yes from somebody who controls land, water, power or a permit. New Mexico said no to a pipeline, twice. In the UK, the grid operator told Nscale that its planned “largest supercomputer” in Essex can’t get enough power until the early-to-mid 2030s. Pew says 54% of Americans now think data centers are bad for the environment. Bernie Sanders filed a bill on Wednesday to ban superintelligence outright, with prison terms up to twenty years.

So people are looking for places where nobody says no.

Greenland got a new US agreement on Tuesday, but it’s about military bases and a veto over investment from outside NATO in mining and infrastructure, not data centers. What it does do is make sure any Greenland data center gets built with Western money. The pitch already exists. A former Pence aide’s company, GreenMet, wants 300 megawatts near Kangerlussuaq by mid-2027 and 1.5 gigawatts by 2028. It plans to run on liquefied gas brought in by barge until a new hydro plant is built, and as of January it had no land and no approvals. A few things to know before you buy the brochure. Greenland isn’t Iceland: Iceland sits on the Mid-Atlantic Ridge, which is where its geothermal comes from, and Greenland is old, stable rock. Its real asset is glacial hydropower, and hydro takes years to build. There’s no national grid, because every town has its own power. And the 57,000 people who live there have their own parliament, which banned uranium mining in 2021 and isn’t in a mood to be told what to do by Washington.

Space has the better headline. Google’s Project Suncatcher launches its first test satellite of TPUs on October 1. Musk says compute in space will “round up to 100%” of all compute. The economics say otherwise, for now. Launch costs about $3,600 per kilogram, and Google’s own plan needs about $200, which depends on Starship getting there in the 2030s. Power in orbit costs about $14,700 per kilowatt-year versus $570 to $3,000 on the ground. A one-gigawatt orbital facility would run about three times the price of a ground one, and the chips burn out in about five years.

My answer: Greenland, easily, for anything measured in megawatts, and probably around the end of the decade. Space wins the photo op next week. But the next real gigawatt goes neither place. It goes wherever someone says yes and there’s power ready to plug into: Texas, Wyoming, Alberta, Quebec hydro, Norway, the Gulf. Notice who can afford to wait for orbit, though. A company with SpaceX stock and its own rockets. That’s the equity-king pile again.

Tenants of their competitors

Here’s the second derivative, and it’s the part the “AI bubble” crowd misses.

If the bond market keeps pricing borrowed credit at 9.25% while the builders borrow at 60 basis points over Treasuries, capacity goes to the builders. Not because they have better models. Because their money is cheaper, their demand is their own, and their data centers already exist.

Which means the labs with the best models, the ones on everybody’s leaderboard this week, end up renting from the companies they’re trying to beat. OpenAI rents from Microsoft, Oracle and CoreWeave. Anthropic rents from Google, Amazon and now Akamai. Every token they sell carries a landlord’s margin, and the landlord runs a competing model. We’ve written for months that the model is getting cheaper. The part nobody says out loud is that the concrete is getting more expensive, and not everyone pays the same price for it.

DeVito’s line in the movie is that he doesn’t care what the company makes, only what it’s worth. The bond market is saying something colder about AI: it doesn’t care how smart your model is, only whose balance sheet is behind it.

What to do about it

Know whose concrete you’re standing on. If your product runs on a model or a cloud that rents its capacity, find out who the landlord is and how concentrated their customers are. A capacity crunch at a tenant shows up for you as rate limits, price increases and delayed features.

Read your own force majeure clause. Oracle just used one on a permit denial, not a hurricane. Check what your biggest contracts count as an act of God, who carries the risk of power, permits and delays, and whether payments can be suspended. Most contracts were written for the weather. Draft for the permitting office.

Don’t lock in long compute commitments with borrowers. Prices on frontier models fell again this week (GPT-6 Sol by half). Capacity keeps getting cheaper on the balance sheets that are hitting their dates. Keep terms short, keep two providers running, and pay for the premium model only where your own tests show it wins.

And one for the allocators, with the usual caveat that this is analysis, not advice: if you own private credit, a business development company or an infrastructure fund, ask what share of the book is single-tenant data centers, who the tenants are, and what the leases say about power and permit delays. The spread between 60 basis points and 9.25% is the market telling you where the risk sits. Brandeis would tell you to go read it in the sunlight.

Follow the balance sheet

Jay Cooke’s railroad got built. It just got built by somebody else, a decade later, with cheaper money.

The AI buildout will get built too. The question isn’t whether. It’s who owns it when it’s done, and the answer is going to come from the credit desk, not the leaderboard.

Follow the balance sheet, not the benchmark.

CoreWeave-Tied Data Center Raises $1.1 Billion in Junk Bonds (Bloomberg, Sep 23, 2026)

AI computing power financing is being repriced (All Weather Finance, Sep 2026, bond terms and S&P quote)

The Implicator, The One Number and The Outside Read (Sep 24, 2026)

CoreWeave’s B+ rating leads to Blue Owl’s failed $4bn data center financing (Investing.com, Feb 2026)

Oracle Cites ‘Force Majeure’ to Shield Itself on Controversial Data Center (Bloomberg, Sep 24, 2026)

Oracle triggers force majeure on data centre project over power delays, source says (BNN Bloomberg, Sep 24, 2026)

Oracle sends force majeure notice on its New Mexico Stargate data center (TechCrunch, Sep 24, 2026)

Oracle issues force majeure notice to Blue Owl (DatacenterDynamics, Sep 24, 2026)

Oracle’s $18 billion data center debt under pressure, FT reports (Reuters, Sep 18, 2026)

Oracle’s statement and zerohedge’s reply (X, Sep 24, 2026)

Stop Saying Half of 2026 US Datacenter Capacity Is Canceled (SemiAnalysis, Jun 18, 2026)

148 gigawatts announced, 12 running: AI financing outran the grid (FXStreet, citing Sightline Climate, Sep 3, 2026)

Hyperscaler debt binge pushes yields up as investor demand cools (Yahoo Finance, Jul 29, 2026)

Big Tech’s AI-Debt Binge Set To Hit New Eye-Popping Milestone (Daily Caller, citing Goldman Sachs via Reuters, Sep 22, 2026)

Meta’s Hyperion joint venture with Blue Owl (widely reported, Oct–Nov 2025)

Van Nieuwerburgh, Financing the AI Buildout (Brookings Papers on Economic Activity, Sep 23, 2026)

FT Alphaville on the Sona Asset Management AI deal-circularity dataset (via The Implicator, Sep 24, 2026)

Anthropic signs $11.6 billion cloud deal with Akamai, gets warrant for up to 5% stake (Reuters, Sep 24, 2026)

Louis D. Brandeis, Other People’s Money and How the Bankers Use It (1914)

Other People’s Money (dir. Norman Jewison, Warner Bros., 1991)

Jay Cooke & Company and the Panic of 1873 (PBS American Experience; Wikipedia)

Three things to know about Trump’s Greenland deal (Atlantic Council, Sep 2026)

U.S. strikes Greenland deal at UN (CNBC, Sep 22, 2026)

Pence aide pitches gigawatt-scale data center campus for Greenland (DatacenterDynamics, Jan 2026)

Construction of UK’s ‘largest supercomputer’ delayed over power supply problems (The Guardian, Sep 24, 2026)

Americans’ views of data centers have turned more negative (Pew Research Center, Sep 22, 2026)

Bernie Sanders proposes banning ‘superintelligence’ (The Verge, Sep 23, 2026)

Project Suncatcher first orbital test (Google on X, Sep 24, 2026)

Why the economics of orbital AI are so brutal (TechCrunch, Feb 11, 2026)

Anthropic releases Claude Opus 5.5 and OpenAI counters with two cheaper GPT-6 models (SiliconANGLE, Sep 22, 2026).