Three stories crossed my desk this week that look unrelated. One is about distressed bonds, one is about a compute deal, and one is about a fintech crossing a revenue milestone. Together they say something useful about where the money in AI actually comes from, and what that means for the companies buying it.
Signal one: CoreWeave's bonds went distressed
As The Verge reported, CoreWeave's bonds slipped into distressed territory last week, the point where investors demand 1,000 basis points more yield than government bonds. Its shares fell more than 20 percent in a few days, and Oracle, another heavy borrower in the buildout, lost almost 10 percent over the same period.
The numbers explain the nerves. CoreWeave owed more than $35 billion as of its second quarter and lost more than $1 billion in the first half of the year. A co-founder told the Financial Times the company wants investment-grade debt so it can borrow from insurers, whose capital he called "unlimited".
There is a counterweight. CoreWeave holds $103.7 billion in remaining performance obligations, contracts customers cannot easily cancel, with about 40 percent due in 2028. Microsoft, OpenAI and probably Nvidia generated almost three quarters of its second-quarter revenue.
My read: this is not a collapse story. It is a concentration story. A large part of the AI compute you rent sits on top of a few heavily indebted companies whose survival depends on a few very large customers continuing to pay. When you plan a three-year AI roadmap, that is part of the operational reality, whether it appears in your vendor's sales deck or not.
Signal two: the same $45 billion deal, twice
TechCrunch, citing Bloomberg, reported that Anthropic and Nscale are close to a second $45 billion compute deal, roughly a month after announcing the first one. Both companies are also pursuing IPOs, and Nscale is reportedly facing investor pushback on a proposed $30 billion valuation.
Two deals of that size, a month apart, tell you how compute is being bought now. Not as a service you scale up and down, but as multi-year commitments signed far ahead of demand. The labs are betting that usage will grow into the capacity. The infrastructure providers are borrowing against those bets.
For a business customer, the consequence is pricing. Models are priced to recover commitments that were made years in advance. That is why I keep telling clients to expect AI pricing to move through rate limits, tiering and quiet model routing, rather than through a simple price list. I covered the early version of this in big tech hiding its AI bill.
It also changes how I think about vendor choice. A provider that has to fill $90 billion of committed capacity will fight hard for your workload. That is a negotiating opportunity today. It is also a reason to make sure your architecture can move if their terms change tomorrow.
Signal three: a boring business crossed $1 billion
The contrast came from fintech. Mercury's CEO Immad Akhund told Newcomer the company has reached $1 billion in annualized revenue, up about 50 percent year over year, with more than 450,000 customers. It is also launching a payroll service.
Nothing about that is an AI headline, and that is exactly why it belongs here.
Mercury grew by selling a useful product to a very large number of small customers who pay every month. No customer is large enough to sink it. No compute contract has to be refinanced for it to survive.
Put it next to CoreWeave and you see two shapes of business.
One has enormous revenue commitments concentrated in three customers and funded by debt. The other has broad, granular revenue funded by customers. Both can work. Only one of them is relaxed when bond markets get nervous.
There is a second lesson in Mercury's numbers. A payroll launch is a classic move for a company with trust and distribution: sell one more useful thing to customers who already rely on you. That kind of expansion does not need a compute deal or a bond issue. It needs customers who stay.
What this means for buyers of AI
Most executives I talk to evaluate AI vendors on model quality and price per token. Very few ask how the vendor's own supply chain is financed. Given this week's news, that is a gap worth closing.
Start with dependency mapping. For each AI tool your company relies on, know which model provider sits underneath and, if possible, which infrastructure. A surprising amount of the market runs on the same few clouds.
Then design for movement. Keep prompts, evaluations and data pipelines portable so you can switch providers in weeks, not quarters. The companies that did this in 2025 had a much easier 2026 when pricing changed. I wrote about the exit side of this cycle in the AI IPO window cracking open.
Finally, learn from Mercury's shape in your own AI spending. Many small, useful deployments with clear returns are more resilient than one giant bet on a single platform. Diversification is not exciting. It is what keeps you calm on the days the bond market is not.
The AI boom is real. It is also financed with borrowed money, and debt always shows up in the price eventually.