CoreWeave’s third-quarter earnings should have been a clean win. Revenue hit $3.66 billion, up 168% from a year ago. The backlog of contracted work stands at $132.5 billion. The AI compute it rents out is, by every measure, in ridiculous demand.
Then you read the rest of the filing.
The company posted a $709 million net loss. It paid $846 million in interest in a single quarter. Its debt-to-equity ratio has climbed to 702%, up from 291% a year ago. And the bond market has already voted: some of CoreWeave’s bonds are yielding more than 15%, a level that signals distress, not strength.
So CoreWeave is changing how it pays for the boom. It is asking customers to pay up front. It is borrowing against specific contracts. And it is flipping toward short-term deals that command premium prices — the financial equivalent of selling concert tickets at scalper rates.
That is the story of AI infrastructure in late 2026. The money is real, the demand is real, and the cash is not.
The Numbers Behind the Headlines
CoreWeave is a neocloud — a company that rents out GPU computing power to the firms building and running AI. It pivoted from crypto mining in 2023 and became one of the most important pieces of plumbing in the AI economy. Microsoft and OpenAI are among its customers.
The Q3 numbers are genuinely impressive on the surface. $3.66 billion in revenue, up 168% year over year. A $132.5 billion backlog. About $9 billion in customer prepayments and deposits expected over the next three months, including more than $3 billion tied to new contracts for capacity the company has already built — meaning that money requires no additional hardware spending.
Then the other side of the ledger. A $709 million net loss. $846 million in quarterly interest expense — that is more than twice the net loss, just to service the debt. About $35 billion in debt as of June 30. The debt-to-equity ratio at 702% is not a typo.
CEO Michael Intrator put it plainly on the earnings call: “We will not use the unsecured debt market to fund our business at current yields, nor do we need to.”
Read that again. The CEO of one of the most important AI infrastructure companies on Earth is saying the debt market is too expensive to use. That is the tell.
The New Playbook: Your Money, Up Front
So what is CoreWeave doing instead? It is turning to its customers for the cash.
The company said it expects about $9 billion in customer prepayments and deposits over the next three months. It is borrowing against specific customer contracts — loans secured by the revenue those contracts will generate. In August it closed a $2.6 billion facility called DDTL 5.5, its first loan designed specifically for shorter-term customer contracts, which lets it draw money in installments as needed.
And it is reallocating capacity toward short-term deals, where the pricing “justifies it.” Intrator said the company is selling short-term capacity at more than $40 million per megawatt annually, while pricing for new infrastructure on four- and five-year contracts is moving to $20 million to $25 million per megawatt.
Let that sink in. Short-term compute costs roughly 60% to 100% more per megawatt than long-term compute. CoreWeave is deliberately choosing the expensive, flexible, risky option — because it needs the cash now.
The company has also raised prices twice: about 25% in July and another 10% later in the quarter. Prices are still climbing, in a way that “more than offsets” cost increases, Intrator said.
There is a logic to this. If demand is this hot, short-term deals let you capture it. But there is a cost: lenders value predictable cash flow above all, and short-term contracts are less predictable. So CoreWeave is trading balance-sheet strength for flexibility. That is a choice you make when you are stretched.
The Market Already Voted
The bond market has been making the same diagnosis for weeks.
In mid-October, yields on some CoreWeave bonds surged above 15%. The yield on its 9.75% bond due in 2031 was 15.5% on November 10, after spiking to 16.7% the previous Thursday. Its five-year credit default swaps — the cost of insuring against default — were at distressed levels, around 1,152 basis points.
Bond yields move inversely to prices. When yields blow out like that, it means investors are dumping the debt and demanding a huge premium to hold it. They are not convinced CoreWeave will earn enough to pay it all back.
The broader context is sobering. According to Barclays, $37 billion of the $80 billion in data center bonds issued this year have been rated junk — roughly twice the share of 2025. The AI buildout is being financed with the riskiest debt in the market, and the market is starting to price that risk.
CoreWeave’s rivals are on the same treadmill. Nebius and IREN have also relied on hefty Microsoft prepayments to pay for computing capacity. This is not one company’s problem. It is the industry’s business model.
This Is Bigger Than One Neocloud
It is tempting to treat CoreWeave as a single bad apple — a company that borrowed too much, too fast. But the pattern is everywhere in AI right now.
OpenAI’s quarterly revenue reportedly crossed $10 billion for the first time, up almost 50% from $6.7 billion in the second quarter — a revenue gap I argued earlier this year says more about the industry’s economics than any single quarter’s headline. That sounds like a company crushing it. Except its adjusted operating margin narrowed from -150% in Q2 to -109% in Q3. It is still losing more than a dollar for every dollar it earns. Anthropic, meanwhile, said its annual revenue run rate had grown by $10 billion in a single month to $81 billion, and the company is preparing a $1.4 trillion IPO.
The revenues are enormous. The profits are not there yet. And the gap between the two is being filled with other people’s money.
That money keeps moving in circles. Customers prepay neoclouds. Neoclouds buy chips from Nvidia and AMD. Chip startups keep raising billions — Britain’s Fractile just closed a $720 million Series C at a valuation reported to be around $7 billion, backed by the UK’s Sovereign AI Fund, Thrive Capital, Founders Fund, Lightspeed, and Redpoint. Governments are writing checks to make sure they have “a stake in the hardware that will underpin their future economy.”
Nobody is wasting money, exactly. But everyone is betting that the demand keeps climbing faster than the debt comes due. That is a bet, not a business model.
I wrote recently about the $3.2 trillion in off-balance-sheet AI obligations that big tech companies have accumulated — commitments that sit outside the traditional debt numbers. CoreWeave’s situation is the on-balance-sheet version of the same problem: the bill is real, it is just structured so you cannot see it easily.
What It Means if You Are Building on AI
Here is where this stops being a Wall Street story and becomes a practical one.
If you are a company running workloads on rented AI compute — or a government office, a school, or a development team evaluating cloud costs — CoreWeave’s playbook should change how you think about contracts.
First, prepayments shift risk to you. If you pay up front for capacity, you are lending your supplier money at zero interest. If that supplier re-prices, delays, or restructures, your capital is already committed. That is not a normal vendor relationship. It is a financing arrangement with a vendor logo on it.
Second, short-term contracts at $40 million per megawatt versus $20 million to $25 million for long-term deals mean flexibility has a price — and it is a steep one. If you are budgeting for AI compute over three to five years, assume the price will not stay where it is today. In either direction.
Third, concentration is the real risk. When one company controls a large share of the GPU cloud market and its balance sheet is leveraged at 702%, the failure mode is not just “higher prices.” It is disruption. If CoreWeave or a peer stumbles, the compute you depend on does not politely wait for you to migrate.
I think about this in chess terms. Right now, the AI industry is playing for tempo — rapid expansion, rapid funding, rapid price increases. It is sacrificing structure for speed. In chess, that works until it doesn’t. The player who is down material but has the better position often wins the endgame. The AI infrastructure buildout is not there yet, but the structure is getting worse while the tempo gets faster.
The Chess Move
So what should you actually do?
Diversify your compute. If you can, split workloads across at least two providers — a hyperscaler and a neocloud, or two neoclouds. The cost of portability is lower than the cost of being stuck.
Read the fine print on prepayments. If a vendor asks for money up front, ask what happens to it if they cannot deliver. Is it refundable? Is it secured? Or is it just a loan you made to a leveraged company?
Keep a local fallback. I have written before about running local AI for daily tasks — and the honest truth is that local models have gotten good enough that a meaningful slice of everyday work no longer needs a rented GPU. For a government office or a school with a fixed budget, a local model is not a downgrade. It is insurance.
And watch the bond market, not the press releases. The bond market is the part of the financial system that has to price risk for real. When CoreWeave’s bonds yield 15%, the market is telling you something that a $3.66 billion revenue number will not.
The AI boom is not a bubble in the sense that the demand is fake. The demand is real. The workloads are real. The revenue is real.
But the cash flow is borrowed, the balance sheets are stretched, and the people building the infrastructure are now asking their customers to finance it. That is not a sign of an industry in perfect health. It is a sign of an industry that has outrun its own means and is improvising the rest.
CoreWeave’s $9 billion in expected prepayments is not a victory lap. It is a bridge loan from the customer base. And bridges only work if there is solid ground on the other side.
The question for the next few quarters is whether the demand is real enough to hold the weight. The market has already started to place its bet.