Amazon's Cloud Business Just Broke Its Own Cash Machine to Keep Up With AI Demand

Amazon spent $77.3 billion on property and equipment in the third quarter of 2026, more than double the $35.1 billion it spent in the same quarter a year earlier, according to the company's October 29 earnings release. Net of proceeds from equipment sales and incentives, that works out to roughly $75.9 billion in a single three-month stretch — a pace that, if sustained, would put Amazon on track for over $300 billion a year. The company reaffirmed its full-year 2026 guidance of approximately $220 billion instead, which still means the fourth quarter alone could see tens of billions more.

The spending bought growth: AWS revenue rose 39% year over year to $45.98 billion, putting the cloud unit on a roughly $184 billion annualized run rate. Operating margin expanded to 38.7% from 34.6%. By almost every operating measure, AWS had one of its best quarters on record. But the capital outlay behind it was large enough to flip Amazon's cash position: trailing-twelve-month free cash flow swung to an outflow of $42.4 billion, down from a $14.8 billion inflow a year earlier — a $57.2 billion reversal that the company's own filing attributes primarily to AI-related infrastructure investment.

A capacity crunch that predates the earnings call

The spending numbers landed four days after Amazon CEO Andy Jassy had already previewed the underlying problem. In an October 23 interview with venture firm Andreessen Horowitz, Jassy said AWS's large-scale capacity blocks for 2027 are largely committed, with a meaningful share of 2028 capacity already under contract as well — a claim independently reported by the French outlet Siècle Digital and by the analysis site TS2 Tech, both naming the a16z interview directly, in the days before the earnings release; the aggregator LatestLY reported the same 'sold out through 2027' framing without identifying its source as the a16z conversation. Jassy carved out an exception for on-demand services like EC2 and Bedrock, and declined to disclose megawatts, processor counts, or what fraction of 2027 capacity remains open, so customers have no way to know whether a specific chip generation or region is still available to them.

That ambiguity matters because of how slowly AWS can respond to it. Facility construction runs roughly two years from groundbreaking to revenue-generating capacity, while server and networking hardware can be ordered and deployed in a matter of months. That mismatch means the capacity Amazon will have available in 2027 was substantially locked in by decisions made in 2025, and any shortfall this earnings call reveals can't be fixed by spending more today — only by having broken ground earlier. Buildings are expected to run for around 30 years and hardware for five to six, so a single data center site typically cycles through five or six generations of servers before it's retired, which is part of why Amazon is willing to keep committing capital even as near-term returns compress the cash statement.

What the earnings release added — and left out

The October 29 release filled in some of what Jassy's interview left vague, though not all of it. Amazon pointed to new or expanded commitments with Meta, Nvidia, and OpenAI without putting dollar figures on any of them, and separately disclosed that Fortune 500 spend on its Bedrock AI service is up sevenfold and token volume processed through it up seventeenfold, though it gave no comparable dollar figures for AWS's broader AI or custom-silicon businesses specifically. It also listed 'resource and supply volatility, including for memory chips' as a named risk factor — language that lines up with the memory-price spiral that has been pushing up DRAM and HBM contract prices across the industry this year, though Amazon did not quantify how much of its own capex increase, if any, is attributable to that specific input. Separately, the company said it would put $1 billion into a new 'Built Together' program for communities hosting its data centers, a modest figure next to its overall spending but a signal that Amazon expects continued friction over water, power, and land use near new sites.

What the release did not do is reconcile the capacity story with the cash-flow story in a way that tells outside observers how close Amazon is to a ceiling. Jassy has now said in back-to-back venues — the a16z interview and, by extension, the earnings call four days later — that even at Amazon's roughly $220 billion 2026 capex pace, AWS will not have enough capacity to satisfy 2026 demand, and he expects the same gap to persist into 2027. Read against the capex and free-cash-flow numbers, that amounts to Amazon telling investors it is choosing to run its core profit engine at a cash deficit rather than risk turning away AI workloads it believes will be larger and stickier than anything it has hosted before. Whether that bet pays off depends on whether the enterprise and AI-lab demand Jassy is describing actually survives a stretch where the industry's broader capital intensity — and its willingness to keep financing it with debt and equity rather than operating cash — is now facing its own scrutiny.