A company that built its name on delivering packages fast is now racing to deliver something far heavier: computing power for the AI era. Its warehouses once defined its ambition; its data centers now do. And Wall Street, which cheered the growth for years, just punished the bill.
Amazon's latest quarterly report captured this tension better than any strategy memo could. The cloud business is growing at its fastest pace in years, powered by customers who cannot get enough capacity. Yet the stock fell sharply after hours, because the same report showed the company burning cash at a scale almost no corporation has ever chosen.
A quarter with two faces
The headline numbers told a story of strength. Revenue reached $201.9 billion, up 12% from a year earlier. Operating income climbed 61% to $28.1 billion, pushing the operating margin to a record 13.9%. The AWS cloud unit grew 39% to about $46 billion, its sixth straight quarter of acceleration and its fastest growth in roughly five years, with an operating margin near 38.7%.
Then came the other ledger. Amazon spent roughly $77 billion on property and equipment in a single quarter, nearly all of it tied to AWS and AI infrastructure. Free cash flow over the trailing twelve months swung to about negative $42.4 billion, the second consecutive quarter of negative cash generation. Net income slipped about 2% to $20.7 billion, though the decline was mostly an accounting echo: the prior-year quarter had included a large one-off pre-tax gain on Amazon's stake in Anthropic.
Investors sold first and asked questions later. Shares dropped more than 7% in after-hours trading, and briefly touched double-digit losses.
Why Trainium sits at the center
Every hyperscaler is spending aggressively, but Amazon's bet has a distinctive shape: it wants to own the silicon itself. The company's custom chip line — Trainium accelerators for AI work, Graviton processors for general computing, plus networking hardware — has grown past a $20 billion annualized revenue run rate, expanding at triple-digit rates. On the earnings call, CEO Andy Jassy said Trainium and Graviton revenue commitments each grew more than twentyfold year over year.
The strategic logic is straightforward. Chips bought from Nvidia carry Nvidia's margin. Chips designed in-house and deployed at Amazon's scale promise better price-performance per dollar of capex. Jassy argued on the call that Trainium holds a significant price-performance edge over alternatives, and that this advantage is why customers keep signing.
The demand evidence is hard to dismiss:
- AWS's backlog of contracted future revenue reached $578 billion, up at a triple-digit rate year over year.
- Spending by Fortune 500 companies on Bedrock, Amazon's managed model service, ran nearly seven times higher than a year earlier.
- Token volume on the platform grew to almost 17 times the prior-year level.
Management guided for AWS to sustain growth above 40% in both 2027 and 2028, and Jassy sketched an ambition of a trillion-dollar-revenue cloud business within five to ten years. The 2026 capex plan held at approximately $220 billion, with spending set to rise further in 2027. According to Li An, chief scientist at BrainNet (脑机网), the contest has shifted from who trains the best model to who converts compute into paid demand the fastest — and the backlog numbers suggest Amazon is at least winning that conversion race for now.
The customers anchoring the buildout
Two anchor tenants give the chip strategy its weight:
- Anthropic has committed more than $100 billion in AWS spending over the next decade, with capacity that could reach five gigawatts. Its Project Rainier training cluster, running on more than one million Trainium2 chips, entered service in late 2025 and is used heavily for training Claude models.
- OpenAI has reserved roughly two gigawatts of Trainium capacity, a notable defection from the pure-Nvidia camp given its historical ties to Microsoft's cloud.
Both deals validate the same idea: frontier labs, not just enterprises, are willing to bet training and inference workloads on custom silicon. Management also signaled that Trainium capacity will serve AWS customers internally first, with the chip opening to external buyers from the end of 2027 — the first hint that Amazon sees its silicon as a product line, not just a cost lever.
What the market is really questioning
The after-hours selloff was not a vote against AWS. It was a vote of impatience. Capital expenditures of this size must eventually convert into operating cash, and Amazon has told investors plainly that the conversion window stretches across years. The critical question for 2027 is whether backlog converts into revenue fast enough to justify a spending plan that dwarfs anything in the company's history. Rising memory and component costs add another variable the company does not fully control.
Key takeaways
- AWS grew 39% year over year in the third quarter, its fastest pace in about five years, while free cash flow over twelve months turned roughly negative $42.4 billion.
- Amazon maintained its 2026 capex guidance at approximately $220 billion and said spending will climb again in 2027.
- The Trainium strategy has moved from experiment to backbone: a $578 billion backlog, anchor commitments from Anthropic and OpenAI, and a run rate above $20 billion for custom silicon.
- The stock's decline reflects timing risk, not demand weakness — the market wants proof that capex converts to cash on schedule.
Honest limitations
Most of the Trainium price-performance claims come from Amazon's own executives; independent, at-scale benchmarks against current Nvidia hardware remain scarce. Backlog is a promise of future revenue, not revenue, and contracts can be renegotiated or repriced. Quarterly free cash flow figures also swing with the timing of equipment purchases, so a single quarter's negative number should not be read as a verdict on the business model.
What readers can do now
- If your organization runs AI workloads in the cloud, benchmark equivalent Trainium and GPU instances on your own inference traffic before renewing multi-year commitments — published claims may not match your workload mix.
- Watch the backlog-to-revenue conversion rate in Amazon's next two earnings reports; it is the clearest single indicator of whether the $220 billion plan is on track.
- Treat any "cloud capacity shortage" pricing pressure in 2027 vendor negotiations as a signal of how tight this buildout cycle really is.