Amazon Raises 2026 AI Spending to $220 Billion and Still Won't Have Enough
Amazon told investors on its July 30 earnings call that it now expects to spend roughly $220 billion on capital expenditures in 2026, up from the $200 billion it had guided to previously, citing rising memory chip costs. Amazon's stock jumped more than 9% in after-hours trading on the news, driven less by the spending increase itself than by what came with it: AWS posted $42.2 billion in second-quarter revenue, up 37% year over year and its fastest growth rate in 18 quarters.
CEO Andy Jassy did not frame the extra spending as a sign Amazon is catching up. He framed it as still not enough. "We will still not have enough capacity to meet all the demand we have in 2026," he told analysts, adding that he expects the same to be true in 2027, and that contracted demand already stretches into 2028. AWS's backlog, customer commitments representing future revenue, grew to $496 billion, expanding at a triple-digit percentage rate year over year.
Where the money is actually going
Jassy said the jump from $200 billion to $220 billion traces mainly to higher costs for high-bandwidth memory rather than a broader construction push, a distinction two industry analysts picked up on independently. Sid Nag of Tekonyx said the increase shows AI infrastructure spending is still in a deployment phase rather than an optimization one, since compute, power, and memory, not customer demand, are the binding constraint. Steven Dickens of HyperFrame Research linked the results to Amazon's recent pivot away from chasing a frontier model of its own and back toward what he called its core strength: infrastructure.
AWS now runs at a $169 billion annualized revenue rate, a figure Jassy noted would rank 24th on the Fortune 500 if AWS were spun out as its own company. Operating income for the unit rose to $16.6 billion, up 64% from a year earlier, and margins expanded to 39.4% from 32.9%. Across Amazon as a whole, net income hit $62.6 billion for the quarter, though that figure includes $53.4 billion in non-operating gains tied to Amazon's investment stake in Anthropic. Free cash flow, meanwhile, flipped to negative $7.6 billion from a positive $18.2 billion a year ago, a swing Amazon attributed to a $66.1 billion year-over-year jump in equipment purchases for AI infrastructure.
Betting the house on Trainium too
Jassy used part of the call to defend the spending to investors directly, splitting AI infrastructure into two buckets: data centers, which take about two years to start generating revenue but can run productively for 30 years, and servers and networking gear, which break even in under three years and then throw off cash until they need replacing. He said Anthropic and OpenAI have both made multi-year, multi-gigawatt commitments to Amazon's custom Trainium chips, a detail Matt Kimball of Moor Insights & Strategy called one of the strongest third-party endorsements Trainium could get, given that both companies could have leaned entirely on Nvidia or AMD instead. Amazon said it's now considering selling Trainium chips to customers who want to run them in their own data centers outside AWS, a move that would put it in more direct competition with the chipmakers it currently relies on for everything else.
Jassy also pushed back on the idea that AI spending is cannibalizing Amazon's traditional cloud business, arguing that reinforcement learning, agent orchestration, and vector databases are all driving demand for conventional compute and storage alongside the AI workloads themselves. Whether that read holds is likely to get tested again in three months, when Amazon reports how much of the newly booked 2028 demand actually converts into revenue against a $220 billion number that, by Jassy's own account, may not be the final figure this year either.
Sources: Fortune · Data Center Knowledge · Washington Times