The fastest AWS quarter in five years
AWS booked $42.2 billion in the second quarter, up 37 percent year over year, its strongest expansion since the quarter that closed in December 2021. Analysts had modeled roughly 31 percent, so the beat was wide rather than marginal. Operating income climbed to $16.6 billion, a 64 percent increase, and operating margin widened to 39.4 percent from 32.9 percent a year earlier. Andy Jassy put the scale plainly on the earnings call: "AWS is now a $169 billion annualized revenue run rate business, which, for perspective, would place it 24th on the Fortune 500 list if it was a stand-alone company." That is the sort of line a CEO delivers when the numbers have stopped needing interpretation.
Five consecutive quarters of accelerating growth tells us something more durable than a single beat. For most of 2024 and into 2025 the story around AWS was a slow grind against Azure and Google Cloud, with the market questioning whether Amazon had missed the generative AI turn. That framing has now collapsed. Growth is accelerating off a base large enough that each incremental point represents several billion dollars of annual revenue. For technology leaders who spent the past two years hearing that AWS was the laggard in AI infrastructure, the practical implication is simple: the capacity crunch you are negotiating against is not confined to the smaller providers.
Jassy says the shortage runs through 2027
The most consequential sentence of the quarter was not about revenue. Amazon raised its 2026 capital expenditure plan to $220 billion, up from a prior $200 billion, and Jassy immediately qualified what that money buys. "Even at that amount, we will still not have enough capacity to meet all the demand we have in 2026," he said, adding, "And I believe this dynamic will also be true in 2027, too." Read that as a two year guidance on scarcity from the largest infrastructure operator in the industry. When the supplier tells you in advance that supply will not clear, the planning assumption for any workload that depends on accelerated compute has to change.
We have watched too many enterprise AI programs plan as though capacity is a procurement detail to be settled after the architecture is agreed. That sequence no longer works. If AWS expects to be short into 2027, the practical constraint on your 2027 roadmap is the reservation you make in the next two quarters, not the model you select or the framework your team prefers. Teams running production inference at any real volume should be treating capacity commitments the way they treat data center leases: long dated, contractual, and owned by someone senior enough to sign. The alternative is discovering in the middle of a launch that the instances you assumed were available belong to someone who booked earlier.
Memory prices are now visible in the capital line
Amazon attributed a meaningful share of the $20 billion capex increase to higher memory costs. That detail deserves more attention than it received. Equipment purchases rose $66.1 billion year over year, and the composition of that spending has shifted toward high bandwidth memory and the DRAM that surrounds accelerated compute. Memory has become a genuine bottleneck in the AI supply chain, and the pricing power has moved to a small group of suppliers who cannot expand fabrication capacity on the timeline that hyperscaler demand requires. When the largest buyer in the world revises its budget upward because a component got more expensive, that component is repricing for everyone underneath it.
For enterprise buyers, the transmission mechanism is straightforward. Input cost inflation at the hyperscaler eventually appears in instance pricing, in the generosity of discount programs, and in how aggressively account teams defend committed use terms. We would expect the era of automatic annual price reductions on compute to stay paused, and we would plan budgets accordingly. Anyone modeling a three year AI infrastructure spend on the assumption that unit costs fall the way they did between 2015 and 2022 is building a forecast on a trend that has already broken. Model flat to rising unit costs on memory heavy workloads and treat any reduction you win as upside.
A $496 billion backlog changes the negotiating table
AWS reported a backlog of $496 billion, a figure that reframes what the business actually is. That number represents contracted future revenue, much of it from customers who committed years of spending to secure access to capacity. It is close to three times the current annualized run rate. A backlog of that size tells you that the sophisticated buyers in the market have already concluded that capacity is the constraint and have paid to lock it in. The customers still buying on demand, quarter to quarter, are competing for whatever is left after those commitments are served.
This has a direct consequence for how CIOs should approach renewal conversations. The leverage in a cloud negotiation has historically come from credible threat of migration, and vendors responded with discount schedules. In a supply constrained market, the vendor's scarce resource is capacity and the customer's scarce resource is certainty. Those trade against each other differently. We would go into the next renewal willing to extend term length and commitment depth in exchange for guaranteed capacity allocations and priority access to new instance families, and we would put those guarantees in the contract rather than the relationship. A discount you cannot use because there are no instances available is worth nothing.
Negative free cash flow is the cost of the buildout
Amazon's free cash flow swung to negative $7.6 billion from positive $18.2 billion a year earlier. For a company of Amazon's cash generation, that is a deliberate choice rather than a distress signal, and the market rewarded it with a seven percent after hours gain. Still, it marks a genuine change in the financial character of the cloud business. The hyperscalers have moved from harvesting an installed base to funding an industrial buildout, and the capital intensity now resembles utilities or semiconductor manufacturing more than it resembles software. Anyone underwriting vendor risk should update their mental model to match.
Management also signaled where they think this ends. On the call they said AWS "could become a few hundred billion dollars revenue business and now believe it will be at least double that and very possibly be $1 trillion annual revenue business for us in time." We treat forward statements at that scale with appropriate skepticism, but the direction of the bet is informative. Amazon is spending as though cloud consumption will absorb an order of magnitude more compute than exists today. If they are right, the enterprises that secured position early will have bought at prices that look cheap in hindsight. If they are wrong, the overcapacity lands on Amazon's balance sheet rather than yours.
What we would change in the next two quarters
Three actions follow from this quarter. First, move capacity planning out of the engineering backlog and into the finance calendar, with named ownership and a committed spend envelope that runs through 2027. Second, audit which of your AI workloads genuinely require the newest accelerators and which are running on them out of habit, because the cheapest capacity is the capacity you do not need to reserve. Jassy also noted that "you see increasingly more and more companies being interested in the open models as well," and open weight models running on older or cheaper silicon remain the most reliable way to take pressure off a constrained budget.
Third, revisit the multicloud question honestly. The instinct in a shortage is to spread bets across providers, and there are good reasons to keep a second option live. The countervailing reality is that commitment depth is what earns priority in an allocation queue, and splitting spend across three vendors can leave you a mid tier customer at each one. We would concentrate committed capacity where the workloads actually run and maintain genuine portability at the data and orchestration layer rather than the compute layer. That gives you negotiating credibility without paying the tax of running the same platform three times.
