In 2016, a small group of commodities traders bought a single graphics card, plugged it in, and set it on a pool table in a lower Manhattan office overlooking the East River. The card mined Ethereum (a cryptocurrency that, at the time, could still be “dug up” with off-the-shelf gaming hardware), and the steady whir of its fan marked the start of a hobby that was about to become much more.
And it did, quickly. The pool table gave way to a garage (one co-founder’s grandfather’s garage in New Jersey, to be specific), which became the group’s first makeshift “data center.” By 2017, the hobby had a name, Atlantic Crypto, and an actual business plan. Then crypto did what crypto does, and prices collapsed in 2018. Margins evaporated, and miners everywhere dumped their hardware for whatever they could get. Atlantic Crypto did the opposite; in a very Warren Buffett sort of way, it was greedy when others were fearful. It bought distressed GPUs during the “crypto winter,” growing from a few hundred cards to tens of thousands spread across seven facilities.
Now, that is the kind of move that looks either brilliant or deranged depending on how things go, and you usually don’t find out which for a few years. In 2019, the founders realized the same chips that solved crypto puzzles were also very good at rendering visual effects, running scientific simulations, and training a then-niche technology called machine learning. They renamed the company CoreWeave and started renting their chips to anyone who needed the computing power.
Then, in the summer of 2022, they made their biggest bet, buying roughly $100 million of Nvidia’s brand-new H100 chips, before most of the world knew what they would be used for. A few months later, ChatGPT launched. Suddenly, everyone needed GPUs, nobody had enough, and a former crypto miner from New Jersey sat on one of the planet’s most coveted stockpiles.
Renting Out the Gold Rush
Today, CoreWeave (CRWV) is the poster child for what the industry calls “neoclouds,” companies that do one thing: rent specialized AI computing power. Amazon, Microsoft, and Google run general-purpose clouds that host everything from your bank’s website to your Netflix queue. CoreWeave runs 51 data centers packed with Nvidia chips, all built for a single purpose, to train and run AI models.
The customer list reads like a who’s who of the AI race: Microsoft and OpenAI (the bulk of it), Meta, Anthropic, and IBM, among others. And the growth has been staggering. Revenue went from $16 million in 2022 to $1.9 billion in 2024 and $5.1 billion in 2025. Management is guiding to $12.4 to $13.2 billion this year, with a contracted backlog of more than $104 billion. Nvidia, its supplier, is also a customer and owns roughly 10% of the company.
The simplest way to understand the business is to think of CoreWeave as a landlord. It builds the building (a data center), fills it with extraordinarily expensive fixtures (GPUs), and signs long-term leases with tenants. Most of those leases are “take-or-pay,” meaning the tenant pays whether or not they actually use the space, just like your apartment lease doesn’t care if you spent the summer at the lake. Management says it only breaks ground once signed contracts cover the build cost over the life of the deal.
And like most real estate developers, CoreWeave doesn’t pay cash. It borrows…heavily. In 2023, it took out a $2.3 billion loan collateralized by the H100 chips themselves, reportedly the first time high-end processors had been used as loan collateral. Since then, much of its borrowing has run through special-purpose entities tied to specific customer contracts. As the CEO explained at a conference in September, lenders essentially look through CoreWeave to the credit quality of the customer on the other side of the contract.
Hold onto that sentence. It matters later.
The Bond Market Has Something to Say
Building 51 data centers in four years is not cheap. As of June 30, CoreWeave carried about $35 billion in debt on its balance sheet, or roughly $52 billion including lease obligations. Interest expense in the second quarter alone was $640 million, more than double the $267 million from a year earlier, contributing to a quarterly net loss of $626 million. And the company plans to spend somewhere between $31 and $35 billion on capital expenditures this year. S&P rates its credit B+, which is a polite way of saying junk.
However, none of this is new information. What changed this week is how the market chose to look at it.
On Tuesday evening, a chart showing yields on CoreWeave’s 2032 bonds spiking toward 15% circulated on social media, along with questions about whether AI borrowing was inflating a private credit bubble. It caught the eye of Elon Musk, who replied with a single word: “Yikes.” By Wednesday’s close, the stock had fallen 15.5%, dragging down Nebius, Oracle, and Dell among other AI infrastructure providers.
For context, the bonds referenced are $1.25 billion of unsecured notes CoreWeave sold just four months ago, paying a fixed 9.625% coupon. Bond coupons don’t change (we discussed this last week), but prices do. And when the price falls, we know the effective yield rises. Those bonds are now trading below 80 cents on the dollar, pushing the yield to roughly 14.8%.
The credit default swap market (essentially insurance against a company not paying its debts) told an even louder story. The cost of five-year protection on CoreWeave debt was 368 basis points (bps) on October 6. By Friday, October 9, it was 643 bps. By Wednesday, it had crossed 1,000 bps, a level Reuters noted implies roughly a 63% chance of default within five years. In plain English, insuring $10 million of CoreWeave debt now costs more than $1 million a year.
Two caveats are worth noting. CDS markets are thinly traded, so a handful of transactions can move the quotes. And CoreWeave has not missed a single payment. But the bond market usually speaks quietly, and without much fanfare. This week, it was much louder.
Who’s Paying the Rent?
Now, back to that sentence about lenders looking through to the customer. Here’s why it matters.
If CoreWeave’s debt is really underwritten by its tenants’ ability to pay, the obvious next question is “who are the tenants?” In 2024, Microsoft alone accounted for 62% of CoreWeave’s revenue. But a meaningful share of the capacity Microsoft rented reportedly went to OpenAI as overflow compute. One independent analysis I’ve been following traced those relationships through and estimated that roughly 28% of CoreWeave’s AI book is tied directly to OpenAI, with another 37% routed to OpenAI through Microsoft. Put together, about two-thirds of CoreWeave’s business ultimately rests on a single company.
And OpenAI is a very particular kind of tenant. Its revenue is real and growing quickly, but it is not profitable, and it has committed to hundreds of billions of dollars in computing contracts across CoreWeave, Oracle, Microsoft, Amazon, and others. For now, it pays for those commitments largely with money raised from investors.
Think about it from the landlord’s perspective. Your biggest tenant is a brilliant, fast-growing startup that pays its rent from its most recent fundraise. The lease is ironclad, right up until the tenant can’t pay. A $104 billion backlog is impressive, but a backlog is a promise, not cash, and a promise is only worth as much as the party making it.
There’s also a timing wrinkle. Many of the largest AI compute contracts signed in 2025 and 2026 don’t actually start billing until the data centers are finished, which for much of this cycle means 2027 and 2028. Until then, the obligations sit quietly on paper. Many are arguing that is precisely when the real test arrives.
Why Do We Care?
So why spend a whole column on an AI landlord? Because CoreWeave is one of the purest, most leveraged expressions of the AI infrastructure buildout. It sits at the exact intersection of chipmakers, AI labs, and the lenders financing it all. If cracks are going to show up anywhere first, they are likely to show up here. As history shows, cracks often appear earlier in the bond market than the stock market. Lenders get paid a fixed coupon if things go right and lose principal if things go wrong. That asymmetry tends to make them react more quickly.
For investors, the lesson isn’t to run from anything with “AI” in the press release. The technology is very real, and many of the companies building it will be enormously valuable. But leverage has a way of turning a great business into a bet on timing, and concentration has a way of turning a diversified-looking book into a wager on a single name. Both are worth understanding in detail before you take a leap of faith.
CoreWeave may well grow into its obligations, and the demand may arrive right on schedule. OpenAI may raise its next round without a hitch, and this week may end up as a footnote. But it’s worth remembering how this story started, with one graphics card on a pool table, owned outright. Today, it’s 51 buildings and roughly $52 billion in obligations. The chips are doing similar work, but the massive mortgage is new.
Markets / Economy
- The S&P posted its first daily 1% decline in 53 days on Wednesday as AI infrastructure spooked the market. The S&P finished the week down -1.9%, the Nasdaq down -4.5%, and the small-cap Russell 2000 down -3.0%.
- Core CPI, which excludes food and energy, rose 0.3% from the previous month in September 2026, the same pace as August and ahead of median market expectations of 0.2%.
- The U.S. inflation rate accelerated to 3.7% in September, up from 3.4% in each of the previous two months and above market expectations of 3.6%, driven by higher energy prices.
- Producer prices soared by 1.3% from the previous month in September, the sharpest increase since the start of the Russian invasion of Ukraine in March 2022, and the second-highest increase on record.
Stocks
- U.S. equities were in negative territory. Technology and Industrials led the decline, while Consumer Staples and Energy outperformed. Value stocks led growth stocks, and small caps beat large caps.
- International equities closed higher for the week. Developed markets fared better than emerging markets.
Bonds
- The 10-year Treasury bond yield decreased 15 basis points to 5.09% during the week.
- Global bond markets were in positive territory this week.
- Government bonds led for the week, followed by corporate bonds and high-yield bonds.
