Amazon CEO Andy Jassy on the $220 Billion AI Buildout

Amazon expects to spend about $220 billion on capital expenditure in 2026—roughly $600 million per day by the interviewer’s calculation—and Andy Jassy said it could spend more while still leaving meaningful AI demand unmet. AWS has sold out its capacity for 2026, is largely sold out for 2027, and is partly sold out for 2028, although Amazon keeps some capacity available for on-demand use. Jassy framed the buildout as a sound medium- to long-term investment, not a guarantee that every current demand forecast will materialize.

Jassy described AI demand as “barbelled”: AI labs and a few fast-growing applications use large amounts of compute, while many enterprises are adopting AI chiefly to reduce costs and improve productivity. Most enterprise production workloads do not yet use inference pervasively, but he expects AI to spread through existing applications over the next five to seven years. Data centers, power, and memory will constrain growth for several years. Amazon plans to expand capacity for both AI and core AWS customers rather than meet AI demand by squeezing existing customers, while balancing scarce high-bandwidth memory for AI accelerators against DRAM needs in its core business.

The investment has different timing and risk profiles. Amazon needs to build data centers about two years before they are needed, but can delay buying servers and networking equipment until a few months ahead and defer those purchases if demand fails to arrive. Data centers can be useful for about 30 years; servers and networking investments typically break even in about three years and last five to six years, while most customer deals run about five years. Building ahead therefore weighs on free cash flow before capacity starts earning revenue. Jassy also argued that the United States cannot afford to fall behind China in setting AI’s leadership and norms.

Lower inference costs are central to expanding AI use, but Jassy said they remain about an order of magnitude too high for AI to be deployed as widely as possible. Amazon developed Trainium to improve price-performance, which he said is about 30–40% better than comparable processors. Cheaper inference could make it more attractive for enterprises to rewrite CPU-based applications for AI accelerators. Amazon’s custom-chip work began with Annapurna Labs and a network interface card designed to offload virtualization, networking, security, and storage work from EC2 servers; the card is now in every Amazon server. That chip expertise also produced Graviton CPUs and AI accelerators. Jassy said 99% of Amazon’s top 1,000 EC2 customers use Graviton pervasively, and that it offers about 40% better price-performance than other x86 processors.

Jassy said AWS’s AI business exceeded a $25 billion annual revenue run rate in its first three years, compared with about $60 million for AWS after its first three years. AWS grew 36.7% year over year in the latest quarter to a $169 billion annual revenue run rate. Since 2021, Amazon’s total revenue has grown about 80%, AWS is nearly three times larger, and advertising revenue is about two and a half times larger. Amazon has invested in Anthropic and OpenAI, which Jassy said reflects his view of AI’s transformative potential and confidence in both companies. Anthropic runs on Amazon Bedrock and has worked with Amazon on Trainium; OpenAI models are available through Bedrock, and the companies are developing a stateful runtime product that Jassy expects agents to use.

Jassy said AI models can help people work beyond their usual specialties and bring them about 80% of the way through a task, but judgment, critical reasoning, ingenuity, and experimentation remain necessary. He expects established companies to match the pace of agent-native startups and said slow movers will be more vulnerable; he also expects companies to rewrite almost all their code so agents can update it within constraints. He sees cybersecurity vulnerabilities as a more immediate concern than speculative catastrophic AI risks: models can find vulnerabilities people missed, and applications may need to patch threats within minutes. AWS’s Continuum identifies vulnerabilities and builds patches, while models paired with automation could broaden penetration testing despite limited security staffing.

Amazon is working on substantial AI-driven changes to Alexa, retail, Prime Video, Fire TV, and Prime Music, though Jassy said the company should not expect to get everything right immediately. He said data-center concerns should account for the fact that Amazon’s backup generators run about 10 hours annually for maintenance, and he challenged estimates that assume continuous operation. He attributed energy infrastructure pressure in part to grid underinvestment and said golf courses use 200 times more water than all of Amazon’s data centers. Amazon has paid about $3 billion in taxes in St. Joseph County, Indiana, and committed another $1 billion over five years through Built Together for communities with data centers.