The numbers behind the headline
Amazon Web Services reported 46.0 billion dollars in revenue for the third quarter of 2026, up from 33.0 billion dollars a year earlier, a 39.3 percent increase that the company called its fastest growth in nearly five years. Operating income rose to 17.8 billion dollars from 11.4 billion dollars, lifting AWS's operating margin to 38.7 percent from 34.6 percent. On an annualized basis, AWS is now running at a 184 billion dollar revenue rate.
Jassy put the number in context on the earnings call: "AWS is a 184 billion dollar annualized revenue run rate business, and yet the opportunity remains extraordinary." Management guided to growth rates above 40 percent in both 2027 and 2028, a rare instance of a business this size accelerating rather than slowing. Across all of Amazon, revenue reached 201.9 billion dollars, up 12 percent, with operating income up 61 percent to 28.1 billion dollars and a record 13.9 percent company wide operating margin.
Sold out through 2027, and into 2028
Six days before the earnings call, Jassy told Andreessen Horowitz cofounder Ben Horowitz and general partner Erik Torenberg on a podcast that AWS capacity is effectively gone for the near term. "2027 is largely sold out," he said, adding that "a good chunk of 2028 is sold out" as well. He described Amazon as struggling to keep pace with demand from AI labs and enterprise customers at the same time, and said the company "could invest a lot more and it wouldn't be a bad use of capital."
Jassy noted that Amazon always reserves some capacity for on demand use through EC2 and for its Bedrock inference service, so none of this is a complete sellout. But the backlog figure disclosed on the earnings call, 578 billion dollars and growing at a triple digit percentage year over year, backs up the picture. For a CIO planning a multi year AI workload, this is the clearest signal yet that reserved capacity pricing today may not be available at any price in twelve months.
The capital bill behind the growth
None of this growth is free. Amazon's purchases of property and equipment reached 77.3 billion dollars in the quarter, up from 35.1 billion dollars a year earlier, and the company still expects to spend roughly 220 billion dollars in cash capital expenditure for all of 2026. The company attributes the increase mainly to artificial intelligence infrastructure, and the pace of installation is now a talking point in its own right.
"We're installing an AI accelerator about every 14 seconds," Jassy said on the call, a line meant to convey the scale of the build out rather than a literal operational metric. Chief financial officer Brian Olsavsky added that "robotics and AI are delivering measurable gains through productivity and employee safety" inside Amazon's own operations, framing the spending as something that pays back inside the company as well as through AWS customer revenue.
Why the stock fell anyway
Despite the AWS acceleration, Amazon's earnings per share came in at 1.89 dollars against a 1.96 dollar estimate, and total revenue of 201.9 billion dollars landed just under the 202.1 billion dollar consensus. The miss was small in percentage terms, but it was enough: Amazon shares fell about 4.25 percent in regular trading to 243.79 dollars, then dropped roughly another 6 percent after hours to 228.90 dollars, a decline of close to 10 percent from the prior close.
The pattern is becoming familiar across the hyperscalers this earnings season. Growth numbers that would have been celebrated outright two years ago are now read alongside capital spending figures, and investors are pricing in the gap between the two rather than the growth rate alone. For a reader tracking vendor stability rather than stock price, the signal is less about sentiment and more about how much capital AWS is willing to keep committing to capacity it says is already sold out.
Betting on custom silicon
A meaningful share of that capital is going toward Amazon's own chips rather than Nvidia hardware. The company highlighted TorchNeuron, a new tool that lets developers run existing PyTorch code on its custom Trainium accelerators without rewriting it, removing one of the bigger practical barriers to adopting AWS silicon instead of GPUs. That barrier has mattered for years: teams that had already standardized their training pipelines around Nvidia's software stack had little incentive to rewrite working code just to try a cheaper chip, no matter how good the underlying economics looked on paper.
For enterprise technology leaders, this is worth testing rather than ignoring. Trainium workloads can be meaningfully cheaper per unit of compute than GPU equivalents when they work, but migrating production training or inference jobs carries its own risk if the tooling around a newer chip architecture is less mature than the Nvidia ecosystem. A sensible approach is to pilot a non critical workload on Trainium now, while AWS is actively subsidizing the migration path with tools like TorchNeuron, rather than waiting until GPU capacity becomes the harder constraint and the cost of switching is no longer the deciding factor. Treat the pilot as a hedge against scarcity, not just a cost optimization exercise, since the chips you can get matter as much as the chips you would prefer.
What this means for the roadmap
Treat Jassy's "sold out through 2027" comment as a procurement signal rather than a marketing line. If AWS capacity is genuinely constrained at this scale, reserved instance and committed use pricing locked in now is likely to look cheap in hindsight, and any team still assuming it can scale AI workloads on demand pricing next year should revisit that assumption before budgeting season closes and capacity gets allocated to whoever committed first.
At the same time, the gap between AWS's growth and Amazon's stock reaction is a reminder that the capital behind this build out is not unlimited, even for the largest cloud provider. Enterprises making multi year commitments should ask AWS directly how backlog growth maps to delivered capacity by region and by chip type, since a 578 billion dollar backlog is only useful to a customer if the specific capacity they need actually ships on schedule rather than sitting behind someone else's reservation.
