The Bond Market’s AI Indigestion
The AI buildout has reached the part of the evening when someone looks at the bill. Data centers, chips, cooling, electricity: extraordinary capital, and investors are starting to ask who supplies it. When financing gets expensive, projects get harder to justify and stocks reprice fast. But there's another question worth asking while everyone studies the financing: how much business is already signed, and can the companies using that infrastructure actually earn money on it?
The financing concern has substance. J.P. Morgan Asset Management reported hyperscaler bond issuance of seventeen billion in twenty twenty-four, one hundred nine billion in twenty twenty-five, and one hundred ninety-four billion in the first half of 2026, three hundred twenty billion in two and a half years, with more ahead. These are bonds from the large cloud companies, whose businesses extend well beyond AI, so treat them as a measure of financing pressure, not a precise count of AI-only debt.
For scale, SIFMA reported twelve point one trillion in U.S. corporate bonds outstanding in Q2 twenty twenty-six, and fifty one point four trillion on its broader fixed-income measure excluding MBS and ABS. Those are outstanding balances; the three twenty is issuance over time. The comparison gives you proportion, not remaining capacity. Individual managers also face concentration limits, competing opportunities, and clients who may not want any corner of the portfolio leaning on the same technology story.
That helps explain the indigestion. J.P. Morgan described softening demand at the margin, with technology credit spreads moving above the broader investment-grade benchmark. Buyers respond to heavy supply with higher yields, smaller allocations, or more time between meals. The bond market has consumed a Thanksgiving-sized serving of AI financing, and whoever's carrying another tray of data centers may need to wait.
Meanwhile, there's already substantial business on the books. CoreWeave reported roughly one hundred four billion in revenue backlog at June thirtieth, twenty twenty-six, excluding more than twenty-five billion in additional commitments secured early the following quarter. The definition includes contracted revenue subject to delivery and service-availability requirements, those conditions matter, but the scale gives us something concrete to examine beyond the next financing headline.
Anthropic's October prospectus offers another window: minimum spending commitments of forty-four point nine billion with Microsoft, one hundred fourteen point five billion with Google, and one hundred ten billion with Amazon, running into twenty thirty-three and twenty thirty-six. Together, two hundred sixty-nine point four billion. These are multiyear computing-service commitments, and their value should never be casually stacked onto suppliers' backlogs, where the same business may already appear.
I wanted annual numbers for the next three years, though public disclosures don't give us a complete, non-overlapping industry schedule for twenty twenty-seven through twenty twenty-nine. For a transparent illustration, spreading those three Anthropic commitments evenly across their stated contract periods produces about thirty-three point four billion annually in each of those years. An eighty percent realization assumption brings that to roughly twenty-six point seven billion per year, about eighty point two billion combined. These are my modeled allocations for a selected slice of signed business; actual payments will follow different schedules.
That haircut lets us explore a less-than-perfect outcome while still retaining substantial activity. Because these contracts purchase computing services, the numbers represent potential spending realized, not the dollar value of buildings physically completed that year. Contract protections, delivery conditions, financing, and customers' ability to pay still determine what ultimately happens. A signature gives the analysis substance. The payment requires someone to keep showing up with money.
Now consider what Anthropic has shown us about earning that money. Its IPO disclosures reported eleven point six billion in Q2 revenue and nine hundred seventy-four million in operating profit, roughly an eight percent operating margin, alongside a substantial net loss. The November second update then showed revenue rising to about seventeen billion in Q3 with a five hundred sixty-seven million operating loss. One profitable quarter demonstrates an encouraging possibility; the subsequent loss shows how uneven the economics remain as capacity and expenses expand.
OpenAI could add another piece when its financials go public. The company announced a confidential draft IPO submission in June and left timing open. A future filing would let investors examine revenue, costs, cash requirements, and management's assumptions directly. It could strengthen the profitability argument, expose weaknesses, or do some of both. The document has to earn the conclusion.
Here's the hypothesis I find worth considering. Existing contracts sustain activity through a stretch when financing is harder to arrange, and operating results gradually give lenders better evidence about customers' ability to pay. Bond portfolios absorb the supply, new savings enter the market, and yields get attractive enough to bring buyers back. If those pieces align, another financing wave could support another phase of infrastructure investment and renewed strength in selected AI stocks.
Several things would need to go right. Customers must turn computing capacity into revenue. Suppliers must deliver economically. Power availability must keep pace. Higher rates, disappointing utilization, cancellations, or equipment costs could overwhelm improvements elsewhere. Even a successful industry can produce disappointing investments when the purchase price assumes too much success.
In managing portfolios, I want exposure to durable themes backed by financial discipline: the balance sheet, the customer, the contract, the valuation, and the concentration already sitting in the book. Infrastructure, power, networking, and software each offer different ways to participate, each with its own risks. On the client side, liquidity needs and the financial plan determine how much uncertainty we can sensibly carry while the industry develops.
When pessimism is the prevailing conversation, asking what could go right is useful work. When enthusiasm is universal, asking what could go wrong deserves equal effort. This hypothesis gives us a framework to test as contracts convert into revenue and companies report results. The next opportunity may start taking shape while the market is still asking for a little more time to digest lunch.
If you'd like to discuss how the AI buildout fits your portfolio and financial plan, call or email me. We can review where the opportunities appear credible and where the assumptions deserve another look. There's room for a thoughtful investment conversation while the bond market loosens its belt.
Your capital, our expertise, a bespoke creation.