AWS posted its fastest growth in 18 quarters as Amazon raised AI infrastructure spending to $220 billion, citing memory costs and saying demand already stretches into 2028.
Amazon says it will spend even more on AI infrastructure this year – and still won’t have enough capacity to satisfy customer demand.
The company on Thursday raised expected 2026 cash capital expenditures to approximately $220 billion, up from about $200 billion, saying higher memory costs pushed spending above earlier expectations. Despite the increase, CEO Andy Jassy said Amazon expects AI capacity to remain constrained through 2027.
“We will still not have enough capacity to meet all the demand we have in 2026,” Jassy said during Amazon’s second-quarter earnings call. “I believe this dynamic will also be true in 2027, too. In fact, the demand we already have for 2028 is striking.”
The comments came as Amazon reported one of AWS’s strongest quarters in years. AWS revenue climbed 36.7% year over year to $42.2 billion, its fastest growth in 18 quarters, while operating income rose to $16.6 billion. Amazon said AWS now operates at a $169 billion annualized revenue run rate, with both its AI business and custom silicon business surpassing $25 billion annual revenue run rates.
For data center operators, Amazon’s earnings suggest the next bottleneck isn’t finding customers – it’s delivering capacity. The company isn’t slowing construction because demand is weakening. Instead, it’s spending more to secure memory and other AI components while racing to add power, servers, and data center capacity fast enough to meet commitments that already extend into 2028. Steven Dickens, CEO and principal analyst at HyperFrame Research, said the results reflect Amazon’s renewed emphasis on infrastructure.
“With the recent announcements around focusing away from model development and a focus on model hosting, the company is back to what it does best, namely infrastructure, and that focus is showing in the results,” Dickens said.
Demand Still Exceeds Buildout #
Amazon’s updated spending forecast underscores how rapidly hyperscalers continue expanding AI infrastructure.
Rather than reflecting a broader construction push, Jassy said the increase from roughly $200 billion to $220 billion primarily reflects higher memory costs. Amazon also highlighted “resource and supply volatility, including for memory chips,” among the risks facing its business.
Sid Nag, president and chief research officer at Tekonyx, said the earnings reinforce that the AI industry remains focused on deploying physical infrastructure rather than optimizing existing capacity.
“Amazon’s commentary confirms we are still in the infrastructure deployment phase of AI, not the optimization phase,” Nag said. “When AWS says demand already extends into 2028 while capacity remains constrained through 2027, it signals that compute, power, networking, and data center construction have become the primary bottlenecks, not customer interest.”
Nag said the increase in spending also highlights a shift in where AI infrastructure costs are emerging.
“The increase wasn’t driven by building more data centers, but by the rising cost of equipping them, particularly high-bandwidth memory, which has become one of the most constrained and expensive components in AI infrastructure,” he said. “The next phase of AI investment will be shaped as much by silicon and memory economics as by concrete, steel, and power.”
Holger Mueller, vice president and principal analyst at Constellation Research, said Amazon’s results also underscore the competitive race among hyperscalers to add AI capacity.
“We are in the gold rush era – you need to build it to sell,” Mueller said. “All vendors are trying to assemble the investment needed for more capacity. Amazon is no exception.”
The company disclosed AWS’s backlog reached $496 billion, growing at a triple-digit rate year over year, giving Amazon unusual visibility into future infrastructure demand. Jassy also said Amazon remains on pace to double its power capacity by the end of 2027 compared with 2025, while noting much of its planned capacity for 2027 has already been reserved by customers.
Why Amazon Is Comfortable Spending $220 Billion #
Jassy devoted a significant portion of the earnings call to explaining why Amazon believes the unprecedented investment will generate attractive long-term returns.
He separated AI infrastructure spending into two categories: long-lived data centers and shorter-lived servers and networking equipment.
Data centers require capital roughly two years before they begin generating revenue but remain productive for more than 30 years. Servers and networking equipment typically reach break-even in less than three years, after which they generate substantial free cash flow before replacement.
“As we get a few years out and the revenue growth outpaces the incremental CapEx growth ... the resulting revenue, free cash flow, and return on invested capital is very compelling,” Jassy said.
He added that Amazon now believes AWS could “very possibly” become a trillion-dollar annual revenue business.
Mueller said Amazon also faces a challenge that differs from some hyperscaler rivals because it must continue funding the capital needs of its retail business while sustaining record AI infrastructure investment.
AI Is Expanding Core Cloud Demand #
Amazon also argued that AI workloads are increasing demand for conventional cloud infrastructure rather than replacing it.
Jassy said reinforcement learning, post-training workloads, agent orchestration, storage and vector databases all rely heavily on traditional compute and storage services. AWS’s Graviton processors have benefited as customers deploy more AI applications.
“We’re seeing strong growth across both AI and non-AI,” Jassy said. “Growth in one is driving growth in the other.”
Chief Financial Officer Brian Olsavsky echoed that view, saying enterprises migrating traditional workloads are increasingly expanding their core cloud consumption alongside AI deployments.
Mueller said that trend aligns with what he’s seeing across enterprise deployments, where AI often handles intent and orchestration while conventional cloud infrastructure and databases execute much of the underlying work.
Trainium Adoption Broadens #
Amazon also highlighted growing adoption of its custom AI silicon.
Jassy said Anthropic and OpenAI have made multi-year, multi-gigawatt commitments to Trainium, while an expanding list of AI startups and enterprise customers have adopted the accelerator.
He added that Amazon is exploring selling Trainium chips outside AWS after receiving interest from customers that want to deploy the processors in third-party data centers.
“We’re actively having those conversations and exploring, and I expect there’s a real chance we’ll do that in the future,” Jassy said.
Matt Kimball, vice president and principal analyst at Moor Insights & Strategy, said the multi-year, multi-gigawatt commitments from Anthropic and OpenAI represent one of the strongest endorsements yet for Amazon's AI accelerator.
“Two of the frontier labs putting multi-year, multi-gigawatt commitments behind Trainium is about the strongest validation a piece of silicon can get,” Kimball said. “This training scale is where accelerators are truly tested – they either hold up, or they don’t.”
Kimball said the commitments also challenge the perception that Nvidia and AMD have effectively locked up the AI training market.
On Amazon’s plans to potentially sell Trainium outside AWS, Kimball said the larger question isn’t whether the company can sell chips and racks, but whether it wants to become a merchant silicon vendor.
“Selling chips and racks means public roadmaps, field engineering, lifecycle support, system integration, and channel,” he said. “Those are go-to-market motions other vendors have spent decades building.”