Microsoft's $25B Bond Sale: The AI Build-Out Now Runs on Borrowed Money

Microsoft hadn't sold US high-grade bonds since 2017. This week it went to the market for $25 billion and lenders offered $115 billion. The same week, Bloomberg reported that AI borrowing pushed global bond sales to a record, and Alphabet's disappointing quarter gave the AI trade its first real test.

Taken alone, a company with Microsoft's balance sheet raising debt isn't news. In context, it shows how the AI build-out is now being paid for. For most of the past three years the hyperscalers funded data centres, GPUs and power contracts out of operating cash flow. That is changing. The largest and safest borrowers in corporate America are now borrowing as well, and that pulls the bond market, interest rates and credit appetite into the economics of AI capacity. Anyone planning on the assumption that AI keeps getting cheaper should pay attention.

What actually happened

The facts, as reported this week:

• Microsoft launched a $25B US investment-grade bond sale, its first in that market since 2017, according to Bloomberg. Peak orders reached $115B, roughly 4.6 times the amount offered.
• Global bond sales hit a record, with Bloomberg attributing the surge to an "AI borrowing binge".
• Alphabet reported disappointing revenue and income and its shares fell (Bloomberg). Reuters then reported Nasdaq futures sliding as mixed Big Tech results "test the AI trade."

Our briefing marks the Microsoft and record-issuance items Unverified: both come from single-outlet reporting that we have not independently confirmed. Several details that matter for analysis were not in the reporting we reviewed. These include the tranche structure and maturities, final pricing and spreads over Treasuries, and Microsoft's stated use of proceeds. Until those are public, treat any specific claim about what this money pays for as inference. That includes the claims in this article.

Why a company with Microsoft's cash flow borrows at all

The obvious question is why Microsoft needs to borrow. It generates enormous free cash flow and has long held one of the strongest credit ratings of any US company. For most of the past decade it has been a net reducer of debt, letting old bonds mature without replacing them.

Three explanations fit the evidence, and they aren't mutually exclusive.

1. Capex has grown faster than cash flow can comfortably cover. Inference AI infrastructure spending is front-loaded. Data centres, power interconnects and accelerator orders are committed years before the revenue they support arrives. Even a very profitable company can reach a point where funding all of that, plus buybacks and dividends, out of cash flow starts to squeeze something. Debt closes the gap without cutting shareholder returns or slowing the build.

2. Cheap capital is a competitive weapon. Inference A AAA-grade issuer borrows close to government rates. When competitors are funding similar build-outs with more expensive capital, through lower-rated bonds, private credit or equity, each of Microsoft's dollars goes further. Borrowing when the market is willing to lend at investment-grade spreads locks in a cost advantage for the life of the bonds.

3. Market timing. Inference A $115B order book shows lenders currently want high-quality, long-duration corporate paper. Issuers borrow when demand is strong, not when they're desperate. A first deal in nine years suggests Microsoft's treasury team judged the window to be open, and that it might not stay open.

Whichever explanation weighs most, the effect is the same. The company that sets the pace for enterprise AI infrastructure has added leverage to its funding mix.

What $115B of demand tells you

Order books exaggerate real demand, because investors inflate their orders when they expect to be scaled back. Even so, 4.6 times coverage on a $25B deal says something clear: credit investors are currently happy to underwrite the AI build-out.

That is worth noticing, because the equity market said something different the same week. Alphabet's results disappointed, its shares dropped, and Reuters framed the Nasdaq slide as a test of the AI trade. Bondholders and shareholders are taking different bets on the same companies:

• A bondholder in a AAA-grade issuer is betting the company stays solvent and pays its coupons. That is a very safe bet on Microsoft whether or not AI investment returns what's promised.
• A shareholder is betting the capex produces returns above the cost of capital. That bet is getting harder to win as spending rises and, in Alphabet's case at least this quarter, results disappoint.

Inference So strong bond demand is not a vote of confidence in AI economics. It is a vote of confidence in Microsoft's ability to repay. Both can be true at once: lenders are happy to finance the build-out even while shareholders start asking whether it will pay off. Executives shouldn't read the order book as proof the AI investment case is settled.

How credit markets become part of your AI cost curve

This is the part that matters for anyone buying AI capacity rather than building it.

For two years the standard planning assumption has been that the cost of AI drops sharply every year. Inference gets cheaper, models get more efficient, and capacity becomes abundant. That assumption still holds on the technology side. But capacity reaches you through a supply chain that someone has to finance, and the financing is changing.

Inference When hyperscalers fund capacity from cash flow, the main constraint on supply is their profitability and appetite for capex. When they fund it partly with debt, a second constraint appears: the price and availability of credit. That brings in variables most AI roadmaps don't model:

• Interest rates. Higher long-term yields raise the cost of each new data centre. Some of that cost eventually shows up in cloud pricing, commitment terms or discount levels.
• Credit appetite. Today's market is eager. If something cools it, such as a weak earnings cycle, a high-profile AI project that fails to deliver, or a broader risk-off move, borrowing gets more expensive for everyone, starting with weaker issuers further down the chain.
• Concentration risk in credit portfolios. Our Sector Watch noted that debt desks and credit investors are now directly underwriting the AI build-out. If the AI trade sours, that exposure becomes a financial-stability topic, not only a tech-sector one.

Microsoft is the safest borrower in this story. The bigger risk sits with less creditworthy participants: data-centre developers, neoclouds, and any lab or infrastructure provider financing capacity on thinner margins. Inference Your AI vendors' financial health now depends partly on how they fund their capacity, and it's fair to ask them about it.

The week's other signal: capacity is being built ahead of demand

The bond sale makes more sense alongside the rest of this week's briefing. Anthropic is reportedly moving into consulting. Nvidia's reported $12.93B deal involving Hugging Face is drawing scrutiny. Both are moves to own the route by which customers adopt AI.

Inference Put together, it looks like this. The industry is borrowing to build capacity, then going downstream into services, distribution and implementation to make sure that capacity gets used. That is what you'd expect from an industry where supply is being built ahead of proven demand. It isn't necessarily a bubble signal. Railways, telecoms and cloud all went through a period of debt-funded overbuild followed by a long stretch of adoption. But it does mean the companies selling you AI have strong reasons to lock you in, because their financing depends on utilisation.

What's unresolved

Several questions will decide whether this week's events are a footnote or a turning point.

• Is this a one-off or the new normal? One Microsoft deal after nine years could be opportunistic. A run of hyperscaler issuance every quarter would mean debt has become a structural part of AI funding. Watch the next two quarters.
• What are the terms? Long maturities would mean Microsoft is matching debt to long-lived infrastructure, which is sound. Heavy short-dated issuance would suggest a gap is being bridged. We don't have the tranche detail yet.
• Does the equity wobble spread? One disappointing Alphabet quarter is not a trend. If capex keeps rising while revenue disappoints across several hyperscalers, the gap between bond and equity sentiment will close, and it usually closes in the bond market's direction.
• Who's further down the chain? Microsoft can borrow almost without limit. The more important question for the overall build-out is what happens to borrowers with weaker credit, who will be the first to feel it if conditions tighten.
• How much of this reaches buyers? Higher financing costs may be absorbed in hyperscaler margins, passed on through pricing, or show up as stricter commitment terms. No one has said yet, and the answer will differ by vendor.

If you're a CEO

The AI build-out you're relying on is now partly financed with debt, and debt comes with conditions. Your board has heard the "AI keeps getting cheaper" story for two years. This week added a qualifier: it gets cheaper as long as credit markets stay open.

That doesn't mean you should slow down. Lenders put $115B on the table, and capacity isn't going away. What it changes is how you time your commitments. Long-term AI contracts signed while vendors are competing hard for usage, backed by eager lenders, will probably look good in hindsight. Commitments you put off until conditions tighten may not.

Expect two questions from investors. One is whether your AI spending is earning its return, because Alphabet's quarter has made that question unavoidable. The other is how exposed you are to a few vendors whose economics now depend on utilisation. Have answers to both.

The question to be ready for in your next board meeting: "If our main AI vendors' cost of capital rose two points next year, what happens to our AI unit economics, and have we locked in terms that protect us?"

If you're a CIO/CTO

Add a line to your vendor risk model: how is this provider financing its capacity? For Microsoft (Azure, Azure OpenAI, Copilot), Google Cloud and AWS, the answer is a fortress balance sheet. For the neoclouds, GPU-as-a-service resellers and smaller inference providers you may have added for price or availability, it could be private credit or GPU-collateralised loans. Those are the first to break if credit tightens.

Practical moves:

• Review multi-year commitments (Azure MACC, Google Cloud commits, AWS EDP) for price-protection and capacity-guarantee clauses. Vendors chasing utilisation have room to negotiate right now.
• Keep model-routing portability. A routing layer that can move inference between providers, together with "rent the model, own the context" discipline, is your hedge if one provider reprices.
• Model a scenario where inference costs stay flat. Most FY27 business cases assume falling per-token costs. Run one where they don't, and see which use cases still work.

The read: Stay with your primary hyperscaler. Its funding is the safest in the market. Before renewing any capacity contract with a smaller GPU provider, ask how it is financed, and keep the option to switch.

If you lead AI transformation

This week's news doesn't change what you should pilot. It changes the business-case assumptions behind every pilot. Most adoption plans were built on a cost curve that only goes down. The financing shift adds uncertainty to that curve, and programmes that only pay off at 2028 prices are exposed.

Three implications:

• Re-baseline the business case. Pilots should show value at today's prices, not at projected ones. If a use case only works with a 50% drop in inference cost, label it a bet, not a plan.
• Make finance a partner. Your CFO's treasury team understands credit cycles better than your AI centre of excellence does. Bring them into vendor-commitment decisions. This is also a new skill for transformation leads: reading vendor economics, not only vendor roadmaps.
• Plan governance for vendor lock-in. Vendors under pressure to fill capacity will push bundled services, consulting (as Anthropic reportedly is) and proprietary tooling. Your governance framework should say when to accept the bundle and when to keep layers separable.

The pattern across the week: vendors are borrowing to build ahead of demand and then working hard to capture that demand. That gives you negotiating leverage now. It also carries concentration risk later.

The experiment to run this month: Pick your three largest AI use cases in production or late pilot. Re-run their ROI at flat inference prices and again at a 20% price increase. Any use case that stays positive at both goes on the scale-up list, and you should negotiate multi-year terms for it while vendors are eager to sign.

---

This post is also published on our Substack newsletter at edge-ai.forum. Subscribe for the weekly roundup direct to your inbox — fresh AI news, executive context, and devices + robotics every Friday morning.