Summary
- AWS second-quarter sales increased 37% year on year to $42.2 billion, the fastest growth in 18 quarters.
- AWS operating income rose to $16.6 billion from $10.2 billion a year earlier, showing that the acceleration was not merely a revenue event.
- Amazon’s trailing free cash flow deteriorated from an $18.2 billion inflow to a $7.6 billion outflow after a $66.1 billion increase in purchases of property and equipment, net of incentives and proceeds, mainly for AI.
- Management raised expected 2026 company-wide cash capital expenditure to about $220 billion; the total also includes robots, semiconductors and satellites and is not an AWS-only budget.
- Amazon says AWS’s AI business and its chips business each exceed a $25 billion annualised revenue run rate, but run rate is not recognised annual revenue or a disclosed profit line.
- Capacity utilisation, backlog conversion, power requirements, asset lives and return on invested capital are the evidence needed to judge whether the current spending creates durable economics.
Revenue is arriving, but cash leaves before the capacity is fully used
AWS’s 37% sales growth is the quarter’s clearest operating signal. At $42.2 billion, the segment is expanding from a very large base and at its fastest reported rate in 18 quarters. Operating income of $16.6 billion, up from $10.2 billion a year earlier, shows that the additional revenue also passed through a profitable segment rather than being bought solely through lower margins.
Cash follows a different timetable. A data-centre building, electrical system, network, server or accelerator is paid for before it serves years of customer workloads. Accounting profit recognises depreciation over an asset’s useful life, while free cash flow records a much earlier investment outlay. A rapidly growing infrastructure business can therefore report stronger segment income at the same time that group cash conversion weakens.
Amazon’s trailing free cash flow moved from an $18.2 billion inflow a year earlier to a $7.6 billion outflow. The company attributes the change primarily to a $66.1 billion year-on-year increase in purchases of property and equipment, net of proceeds and equipment incentives, mainly reflecting AI investment. This does not show that AI demand is weak. It shows that Amazon is building ahead of the revenue that those assets may generate.
The economic question is the duration of that gap. If capacity fills quickly at attractive prices, today’s outflow can support future cash generation. If equipment is underused, becomes obsolete quickly or requires more supporting power and network investment than expected, the cash cost can outrun the revenue curve. The quarterly result identifies the wager but does not disclose its return schedule.
The $220 billion figure is company-wide, not a data-centre invoice
According to Associated Press reporting of the earnings call, chief executive Andy Jassy now expects approximately $220 billion of cash capital expenditure in 2026, up from the $200 billion plan set out in February. He cited higher memory costs as the principal reason for the increase.
The scope boundary is essential. Amazon’s total covers infrastructure beyond AWS and AI data centres, including robots, semiconductors and satellites. Calling all $220 billion “AI spending”, “AWS capex” or “data-centre investment” would assign assets to a segment without evidence. It would also confuse cash capital expenditure with a single-quarter charge.
The number does establish scale. A company preparing to spend roughly the annual output of a medium-sized economy on long-lived and technical assets must manage procurement, construction, power, deployment and utilisation across many programmes. Small errors in timing or useful-life assumptions can move billions of dollars of cash and depreciation.
Investors need an asset bridge: buildings and electrical systems, servers and accelerators, networking, custom silicon, logistics automation, satellite infrastructure and other property. They also need the portion already placed in service, the portion under construction and the portion committed but not yet paid. Without that breakdown, the total indicates investment intensity but cannot explain AWS capacity or returns.
AI demand pulls ordinary cloud services behind it
Amazon places AWS at a $169 billion annualised revenue run rate and says the AI business within AWS and Amazon’s chips business have each passed a $25 billion annual run-rate level. Those figures are evidence of current scale and momentum, not recognised annual revenue. They also do not provide standalone margins or eliminate overlap between how products are consumed.
Management’s more useful operating argument is that AI and core cloud services reinforce each other. Training and inference require accelerators, but post-training workflows, reinforcement learning and agent tool use also consume CPUs, storage, databases and networking. A customer deploying an AI application may therefore increase spending across the existing AWS estate rather than purchase an isolated accelerator service.
This matters for infrastructure economics. A data centre built for mixed workloads can earn revenue from several layers of a customer’s system. Custom chips may reduce hardware costs or improve supply control. Storage and database usage can persist after a burst of model training. The resulting demand can be broader and less episodic than a simple count of GPUs.
The company does not provide an audited split between AI, chips and core-service profitability. It does not show whether new customers are taking full-stack services or whether a small group of model developers accounts for much of the growth. The complementarity thesis is plausible and supported by product logic, but customer concentration and service-level margin remain unreported.
Commitments are not the same as utilisation
Amazon cites multi-year, multi-gigawatt Trainium commitments from Anthropic and OpenAI. The language indicates substantial prospective demand for its own accelerator platform. It does not disclose the timetable, power delivery, deployment sequence, take-or-pay terms or realised usage.
Infrastructure return depends on more than signing demand. A commitment may ramp over years. Power can arrive later than servers, or a building can be ready before a customer’s software. Memory and accelerator supply can constrain complete systems. A customer may reserve capacity but consume it unevenly.
Utilisation affects both revenue and efficiency. An accelerator earns nothing while idle but still depreciates, occupies powered space and may require cooling and network capacity. High utilisation can improve economics, although congestion and overcommitment can damage service quality. The operator must balance spare capacity for reliability against the cost of unproductive assets.
The most informative disclosure would connect contracted demand, installed capacity and billed consumption. It could use ranges rather than reveal customer-sensitive details: accelerator fleet utilisation, power capacity placed in service, backlog scheduled within 12 or 24 months and the share protected by minimum commitments. Amazon’s current materials do not provide that bridge.
Segment profit should be separated from the Anthropic valuation gain
AWS operating income of $16.6 billion is a reported segment measure. Amazon’s group net income contains a different item: $53.4 billion of non-operating pre-tax other income primarily related to the valuation of its investment in Anthropic.
The valuation gain is not cash earned by operating data centres, nor is it AWS revenue or evidence that new infrastructure is already producing a return. It reflects a remeasurement of an investment. Treating it as cloud operating performance would overstate both the quarter’s recurring earnings and the immediate productivity of capital expenditure.
This separation also clarifies the cash-flow picture. A non-cash valuation gain can lift reported net income while heavy purchases of equipment push free cash flow negative. The two measures answer different questions: one records the changed carrying value of an investment; the other shows cash remaining after operating activity and capital purchases under Amazon’s definition.
A reader judging infrastructure performance should follow AWS revenue and operating income, depreciation, cash capital spending and future cash generation. The Anthropic gain is material to Amazon’s financial statements, but it belongs outside that operating bridge.
Higher memory prices expose a supply-chain channel
Jassy’s attribution of the capex increase mainly to memory costs shows how demand growth reaches beyond data-centre construction. AI systems require high volumes of specialised memory as well as accelerators, CPUs, networking and storage. A higher component price can raise cash spending even when the planned quantity of capacity changes less dramatically.
That can affect returns in several ways. Amazon may pass costs into service pricing, absorb them to preserve customer economics, redesign systems or use purchasing scale to negotiate supply. Each response has a different effect on margin, deployment speed and competitive position.
Custom silicon gives AWS more control over some parts of the stack, but it does not remove dependence on memory, fabrication, packaging or power infrastructure. Nor does a large purchase programme guarantee that every component arrives in the right sequence. An incomplete system cannot generate the intended service revenue.
The company does not disclose memory quantities, contract prices or the proportion of the $20 billion plan increase attributable to price rather than volume. The safe conclusion is narrower: component inflation contributed materially to the revised company-wide expectation, increasing the cash required for the present investment cycle.
The return dashboard starts with capacity, power and asset life
AWS’s quarter demonstrates operating demand: sales accelerated and segment profit increased. The free-cash-flow outflow demonstrates investment intensity. Neither measure alone establishes the return on assets being bought now.
The first missing variable is capacity placed in service. Buildings under construction and chips in inventory should not be treated like active systems. The second is utilisation after commissioning, including the speed at which contracted workloads ramp. The third is power: capacity that lacks an energised electrical path cannot produce cloud revenue.
Asset life is equally important. Buildings and electrical infrastructure can serve multiple generations of computing equipment, while accelerators and memory may face a shorter economic life. A blended capex total obscures this difference. Depreciation assumptions and refresh cycles determine how much future revenue is required to earn an adequate return.
Finally, Amazon needs to connect the cloud funnel to cash: remaining performance obligations or commitments, conversion to billed usage, operating margin after depreciation, maintenance spending and free cash flow as the build rate changes. The current numbers say that AWS demand is strong enough to justify an exceptional expansion. They do not yet show when that expansion funds itself.
The contradiction is therefore productive, not cosmetic. Faster AWS growth gives Amazon evidence to spend. Negative trailing free cash flow shows the size and timing of the payment. The next several quarters will reveal whether utilisation catches the assets before technology and cost move again.