Amazon AWS Q2 2026 earnings and AI infrastructure capex

Amazon's raised capex guidance came bundled with an unusually explicit defense of the math behind it. Jassy laid out a payback framework in which servers and networking equipment break even in a little under three years, then keep generating value for another two to three years on top of that before needing replacement, while the data centers that house them monetize for 30-plus years ^[1]. Run that timeline forward and a single data center can cycle through roughly five to six full generations of server hardware over its life ^[1]- a framing built specifically to counter the bear argument that AI infrastructure spending has outrun any realistic return.

What makes the pitch land differently than the usual hyperscaler capex justification is the second half of it: Jassy isn't describing a company racing to catch up with demand, he's describing one that structurally can't. Even after lifting 2026 capex guidance to roughly $220 billion, Jassy told investors AWS still won't have enough capacity to meet all of its 2026 demand, and expects the same shortfall to persist into 2027 ^[2]. That reframes the record spending less as a growth bet than as a rationing problem - Amazon is capacity-constrained rather than demand-constrained, which is the more bullish read that sent the stock sharply higher.

The headline number - a roughly $20 billion jump in 2026 capex guidance, from about $200 billion to about $220 billion - wasn't attributed to GPU scarcity or data center construction delays. Amazon pointed instead to rising memory costs, specifically DRAM and HBM contract pricing, as the key driver of the increase ^[2]. That's a meaningful shift in the story: six months earlier, in February 2026, Amazon had guided to the original $200 billion figure with Jassy expressing confidence in that plan ^[3], and the jump since then reflects a supply-side cost shock in memory markets more than any change in Amazon's own ambitions.

Community discussion around the earnings framed this as a broader supply-chain story rather than an AWS-specific one. Memory fab capacity takes years to bring online, and the current DRAM/HBM pricing spike is being felt across every hyperscaler building AI infrastructure at once, not just Amazon. That context matters for reading the $220 billion figure: it is as much a statement about a tightening upstream chip market as it is about AWS's own growth trajectory, and it suggests further capex surprises could come from the memory supply chain rather than from AWS demand itself.

Buried beneath the AWS revenue headline is a smaller but arguably more strategically important number: Amazon's custom silicon business, spanning the Trainium AI chip and Graviton CPU lines, crossed a $25 billion annual revenue run rate and is growing at a triple-digit year-over-year clip ^[4]. That run rate is being underwritten by genuinely large customer commitments - Anthropic has committed to up to five gigawatts of Trainium capacity over multiple years, and OpenAI has committed to two gigawatts ^[4]- which gives AWS forward visibility on its own chip roadmap that doesn't depend on Nvidia allocation.

The payoff shows up in the margin line. AWS operating margin expanded to 39.4% in Q2, up from 32.9% a year earlier ^[5], even as capital spending hit records - the opposite of what a simple story about spending pressuring margins would predict. Bloomberg Intelligence's Anurag Rana made the case that the chips run rate, not AWS's AI revenue figure, is the number worth watching: shifting inference and training workloads onto in-house Trainium and Graviton silicon lets AWS keep margin that would otherwise flow to Nvidia, and combined with AWS's enterprise-heavy customer base and its Anthropic partnership, gives it a durable structural edge over Azure that a pure GPU-reseller model wouldn't have.

Jassy used the earnings call to publicly revise upward what AWS's ceiling could look like: "We long believed AWS could become a few hundred billion-dollar revenue business and now believe it'll be at least double that, and very possibly be a trillion-dollar annual revenue business for us in time" ^[6]. That's not just rhetoric - it's backed by a backlog that reached $496 billion, up roughly 2.5 times year-over-year ^[5], and by concrete enterprise adoption stories like the PGA Tour, which migrated its ShotLink ball-tracking, media, and broadcast systems onto AWS and now runs agentic AI on Amazon Bedrock for automated content and betting-profile generation ^[7].

The market's reaction validated the reframe: Amazon shares jumped as much as 7 to 15% after the print, reportedly the stock's biggest move since 2012 ^[2], and at least 15 brokerages raised price targets, including Goldman Sachs to $375, JPMorgan to $365, and Barclays to $365 ^[5]. Market commentary picked up on a wider pattern beyond Amazon alone - AWS's roughly $169 billion annualized run rate and 37% growth sit alongside Microsoft Azure and Google Cloud posting their own accelerating run rates, suggesting the entire hyperscaler cloud market is compounding at once rather than AWS simply taking share from a shrinking pie.

Not every reaction to the print was celebratory. A recurring thread in investor discussion questioned whether hyperscaler capex accounting is becoming partly circular - one company invests in an AI lab, the lab spends a meaningful share of that capital back on that same company's cloud infrastructure, and the arrangement inflates both sides' reported growth without necessarily reflecting new end demand. A related pushback challenged Jassy's own break-even framing directly, arguing that the tidy 'servers pay back in under three years' math glosses over how long it actually takes newly built capacity to reach full utilization - if utilization ramps slower than capacity comes online, the effective payback period stretches well past the headline number.

A related and higher-profile version of this skepticism came from market commentator Ed Elson, who argued that a large share of Amazon's and Google's net income is now coming from paper gains on their equity stakes in OpenAI, Anthropic, and SpaceX rather than from operating profit - and that this makes both companies' current price-to-earnings multiples look artificially cheap. It's a useful counterweight to the bullish capex-defense narrative: the payback framework explains why AWS keeps building, but it doesn't resolve the separate question of whether the AI-driven earnings underpinning today's stock reaction are as durable as they look.