Amazon Is Borrowing to Build AI. That Changes the Whole Trade.

On July 7, 2026, Amazon walked into the US bond market and asked for at least $25 billion. Eight tranches (separate slices of the same deal, each with its own maturity and coupon), stretching from three years out to forty. Per the filing, the pieces ran $750 million of floating-rate notes due 2029, then fixed-rate chunks at 2029, 2031, 2033, 2036, 2046, 2056, and a $2.25 billion bond that does not mature until 2066.^[4] The stated purpose was general corporate needs. The actual purpose is a $200 billion capital expenditure budget, most of it AWS data centers full of AI accelerators.^[5]

Amazon also told the market it does not plan to issue any more debt this year.^[1] Which is a strange thing to say out loud, unless you think investors need reassuring. And they do. Because this is the third raise in twelve months. In March, Amazon pulled $37 billion out of the US market and then did a record euro debut of 14.5 billion euros, roughly $16.8 billion, the largest corporate deal ever done in that currency.^[10,11] Bank of America counts Amazon at $92 billion of debt issued in 2026 alone, more than Alphabet, Meta, or Oracle individually.^[3]

Here is the part that should reorganize how you think about the AI trade. Two years ago, the whole bull case rested on one structural fact: hyperscalers were funding the buildout from operating cash flow. They didn't need anyone's permission. That is no longer true.

The cash flow ran out first. The bonds came after.

Free cash flow (operating cash flow minus capital expenditures, the money left over after you build the thing) is the number that broke. Amazon generated $25.9 billion of it on a trailing-twelve-month basis a year ago. As of the most recent reporting it is $1.2 billion, and the reason is a $59.3 billion year-over-year increase in purchases of property and equipment.^[3] That is not a rounding error. That is the entire cash engine of the company being redirected into concrete, power, and silicon.

Amazon is not alone. Across the big four, 2026 capex is running toward roughly $700 billion, up from about $410 billion in 2025.^[14] Epoch AI's work on this puts the crossover point plainly: aggregate hyperscaler free cash flow goes through zero around the third quarter of 2026.^[13]Not "compresses." Crosses zero. The group that spent a decade as the most cash-generative cohort in the history of public markets is, in aggregate, about to spend more than it makes.

The debt market has noticed. Bank of America raised its 2026 forecast for hyperscaler investment-grade issuance by 25%, to $175 billion, and pegs total AI-related bond supply at $270 billion this year against $136 billion in 2025.^[9] AI paper has gone from around 1% of the investment-grade market in 2024 to roughly 18% in the first half of 2026.^[9] Tech hyperscalers are now displacing banks as the top issuers in the US investment-grade index.^[17] If you own a broad bond fund, you already own the AI trade. You just may not have known you were long it.

The demand signal in this deal is the actual news

Everyone will write the headline as "Amazon raises $25 billion." The interesting number is what it cost. Amazon marketed the three-year at 65 basis points over Treasuries and the 40-year at 145.^[4]To get the book filled it had to hand back 18 to 21 basis points of extra yield on the longest bonds, and per Bank of America the deal's new-issue performance was the weakest of any hyperscaler since Meta's $30 billion sale in October 2025.^[3] Orders peaked around $62 billion and shrank to roughly $41 billion as the spread was tightened.^[4]

Compare that to Meta's book of about $125 billion nine months earlier.^[12] Same asset class. Same story. Radically different appetite. Bond investors were not saying no. They were saying yes, at a price, and the price went up.

That is what a functioning market does, and honestly it is healthier than the alternative. The version of this story that should scare you is the one where $25 billion of 40-year AI paper gets six times oversubscribed without a single question. Repricing is discipline. It just means the cost of the buildout is now a variable, not a given.

The depreciation question is where the bodies are buried

Here is the mechanic that most people skip. When Amazon buys a GPU (graphics processing unit, the chip class that trains and serves AI models), the cash goes out the door immediately, but the cost hits the income statement slowly, spread over the asset's assumed useful life. Stretch that life from five years to six and you move billions of expense out of this year's earnings and into a future one. Nothing about the physical chip changes. Reported profit does.

Amazon has already been honest about this once. Effective January 1, 2025, it cut the estimated useful life on a subset of servers and networking gear from six years back down to five, and told investors to expect roughly $0.7 billion less operating income in 2025 as a result. The CFO's explanation on the call was that a useful-life study had observed "an increased pace of technology development, particularly in the area of artificial intelligence and machine learning."^[16] Read that again. Amazon looked at the AI hardware cycle and concluded its own equipment wears out faster than it used to.

Michael Burry has built the loudest version of the bear case on exactly this. His estimate is that hyperscalers will understate depreciation by about $176 billion between 2026 and 2028, and that by 2028 Oracle will be overstating earnings by roughly 27% and Meta by about 21%.^[15] I think the headline number is theatrical and the underlying observation is correct. Those are different things.

So is this prudent leverage or late-cycle behavior?

My honest read: it is prudent leverage that is one bad quarter away from looking like late-cycle behavior, and the thing separating the two is AWS's demand curve.

The bull evidence is real and I am not going to hand-wave it. AWS grew 28% in the first quarter of 2026 to $37.6 billion, its fastest rate in fifteen quarters, with a backlog (contracted revenue not yet recognized) of $364 billion.^[7]That is not a company building into a void. Andy Jassy's defense of the $200 billion plan was that this "isn't some sort of quixotic top-line grab," and on the numbers he is right.^[8] Amazon is capacity-constrained, not demand-constrained. When you can sell everything you can plug in, borrowing at 100 basis points over Treasuries to plug in more is arithmetic, not hubris.

The bear evidence is structural. Operating cash flow is variable. Coupon payments are not. Amazon has now committed to fixed obligations running to 2066 against a revenue stream whose durability past about 2029 nobody can actually underwrite. If AI inference gets radically cheaper per token, which is the one thing that has happened reliably every year since 2023, the revenue side of that equation compresses while the debt side does not move at all. Alphabet has issued century bonds into this.^[18]A hundred-year instrument funding a five-year asset is a sentence that should make you sit up regardless of which side you're on.

What I'd actually watch

Forget the capex headline. It is a lagging, pre-announced number. These four move first.

1. The spread, not the coupon.Watch where Amazon's long bonds trade in the secondary market against Treasuries. Credit spreads reprice weekly. Equity narratives reprice at earnings. Bond investors get the information first and they act on it in public.
2. Any change to useful-life assumptions. It shows up in the 10-K and 10-Q as a boring paragraph about estimates. If a hyperscaler extends GPU lives in 2027, treat it as an earnings-management signal, not an engineering one. If one shortens them, believe it, because nobody voluntarily reduces their own reported profit.
3. Backlog conversion, not backlog size. $364 billion of backlog is a great number only if it converts on schedule and at margin.^[7] Watch AWS operating margin against the depreciation load. That is where the capex bill actually lands.
4. Whether the "no more debt in 2026" promise holds. Amazon said it explicitly.^[1] If it comes back to the market in the fourth quarter anyway, the buildout is outrunning the plan, and the plan was already the most aggressive one in corporate history.

The character of the trade has changed

The AI trade of 2023 to 2025 was an equity story about companies so rich they could self-fund a technological revolution out of petty cash. That story is over. The AI trade of 2026 is a credit story. It runs through spreads, coverage ratios, depreciation schedules, and the willingness of pension funds to own 40-year paper backed by chips with a three-year product cycle.

I don't think Amazon is in trouble. It is one of the highest-rated corporate credits on earth, it is selling every unit of compute it can build, and borrowing against a $364 billion backlog is defensible in a way that borrowing against a slide deck is not. But the option value is gone. When you fund from cash flow, a bad year means you slow down. When you fund from debt, a bad year means you slow down and still pay the coupon. Amazon traded flexibility for speed. That is a rational trade if you believe you are in a land grab with a closing window, and every hyperscaler clearly believes exactly that.

The thing to hold in your head is that they can all be right about the demand and still be wrong about the financing. Those are separate bets. The bond market just charged Amazon 18 extra basis points to remind everyone.