Amazon's $220 Billion Bet on Chips It Built Itself

Amazon raised its 2026 capital expenditure forecast by $20 billion, to roughly $220 billion — and the stock climbed 15% in a single day. A few months earlier, almost the same plan had knocked 11% off the shares. The dollars barely changed. What changed was that Amazon finally explained what the money buys, and how fast it pays back. That disclosure, not the capex number, is where this story actually lives.

Here is the part that rewrites the usual AI narrative: Amazon is spending this fortune substantially on chips it designs itself, not on renting Nvidia's. Its custom-silicon business — Trainium AI accelerators plus Graviton processors — has passed a $25 billion annual revenue run rate and is growing in the triple digits. Trainium is purpose-built for the part of the market growing fastest: inference, the ongoing job of running a trained model for every user, rather than the one-time training runs that made Nvidia's GPUs famous. As AI shifts from training to inference, the contested share of the compute cycle keeps moving toward the kind of economics Amazon's own silicon targets.

Why self-built silicon changes the math

The mechanism, stated plainly: every AI workload that runs on Amazon-built chips stops paying Nvidia's margin. That captured cost is a large reason AWS operating margin jumped to about 39%, up roughly 650 basis points from a year earlier. Amazon is not just competing with Nvidia for customers; it is trying to stop being one of Nvidia's biggest customers — and to pocket the markup itself. It does not need to beat Nvidia on raw performance. It needs its own chips to be good enough on cost per workload, the dimension that decides which architecture wins as inference scales.

That is the theory. The numbers that make it more than a theory came in the July quarter. AWS revenue grew about 37% year over year — its fastest in 18 quarters — to roughly $42 billion, an annualized run rate near $169 billion. And the demand behind it is already contracted: Amazon's backlog of customer commitments hit $496 billion, with a weighted-average contract term of about 6.4 years, and management says capacity for 2027 is largely spoken for. This is the strongest evidence these are real orders, not a defensive buildout on a guess.

There is one number from management that matters most for a thinking investor. Amazon says its servers and networking equipment now pay for themselves in under three years, against a useful life of five to six years. If that is true, the $220 billion is not a sunk gamble but a series of investments that return before they wear out — which is why the market, once told, treated the raise as good news rather than fear.

The other side of the double signal

But a commitments surge is a dual signal, and this is the half that usually gets skipped. The buildout is being financed by borrowed money. Amazon's long-term debt roughly doubled in six months to about $129 billion, with $67 billion drawn in the first half of 2026, and trailing twelve-month free cash flow turned negative, an outflow near $7.6 billion. Amazon also recorded a large paper gain from revaluing its Anthropic stake — an accounting windfall that adds to reported earnings but not to cash. The reported profit and the cash reality are two different things right now.

None of this is a verdict against Amazon. It is a statement of what the bet depends on: the $496 billion of contracted demand has to convert into cash revenue faster than depreciation and debt service consume it. The entire question is payback — whether the backlog becomes cash before the interest and the aging servers change the economics. That is a timing bet on execution, not on demand.

This reframes the "AI stocks to buy" question productively. Amazon's AI story is not a moat you rent; it is a construction project you pay for up front. The opportunity is that Amazon is capturing the margin it used to hand to a chipmaker while growing cloud revenue at its fastest year-over-year pace in 18 quarters. The risk is that negative free cash flow and a doubled debt load remain uncomfortably live as long as the servers are still paying themselves off. Watch whether that contracted backlog keeps converting into revenue and whether operating income keeps growing faster than the interest bill. That conversion rate, not the capex headline, is the number that decides whether the $220 billion was a bargain or a burden.