Amazon CEO Andy Jassy on Why Amazon Is Investing $220 Billion in AI

Amazon expects to spend about $220 billion on capital expenditure in 2026, and Andy Jassy said demand is strong enough that even this investment will leave meaningful customer demand unmet. He said Amazon could spend more and still consider it a sound medium- to long-term use of capital. AWS capacity is sold out for 2026, largely sold out for 2027, and substantially committed for 2028. Jassy expects constraints to last several years as limited data-center space, power, and memory force trade-offs in how capacity is allocated.

Jassy described the investment cycle as a reason the spending does not immediately translate into cash generation. Amazon generally has good visibility into AWS demand before buying servers and networking equipment, which it purchases only a few months before deployment. Data centers must be funded about two years before servers can be installed and have useful lives of roughly 30 years. Servers and networking investments break even in about three years on average, last five to six years, and are generally supported by deals of about five years. Until facilities are operating and their capacity is monetized, construction weighs on free cash flow.

AI demand is a “barbell,” Jassy said: AI labs and a few fast-growing applications use substantial compute, while many enterprises are focused on cutting costs and improving productivity. Most enterprise workloads do not yet use inference pervasively, but he expects inference to become the largest absolute segment of AI usage, with many, if not most, existing enterprise applications using it broadly within five to seven years. Inference costs have fallen considerably, but Jassy estimates they remain about an order of magnitude too high for the technology to be as widespread as possible. Lower costs could enable more capable applications and make it worthwhile to move more CPU-based workloads to AI accelerators.

Amazon’s AI business passed a $25 billion annual revenue run rate within the first three years of the current AI wave, compared with about $60 million for AWS after its first three years. AWS grew 36.7% year over year last quarter to a $169 billion annual revenue run rate. Jassy tied Amazon’s chip development to improving customer price performance: Graviton, the company’s CPU, is used by 99% of its top 1,000 EC2 customers and delivers about 40% better price performance than other x86 processors. He said Trainium offers about 30% to 40% better price performance than comparable processors. AWS will not take capacity from core customers to serve AI customers willing to pay more; it will seek capacity for both groups on sustainable economic terms.

Jassy cast data-center construction as important infrastructure for services including banking, healthcare, shopping, education, and entertainment, comparing AI infrastructure growth with the U.S. highway and bridge buildout of the 1950s. He also argued that the United States needs to build data centers to avoid ceding leadership and the ability to shape industry norms to China. Inside companies, he expects AI to let people work beyond narrow formal roles and give teams a more advanced starting point, though teams may become smaller and some existing tasks will require fewer people. He said AI, robotics, and space will also create new jobs.

Jassy warned that companies using AI only for incremental gains could lose ground in customer experiences within three to five years. Amazon is rethinking Alexa, retail, Prime Video, Fire TV, and Prime Music for AI rather than relying only on gradual upgrades; he said Alexa’s original “brains” were insufficient for long-term relevance and had to be rewired. The company expects to iterate, and it remains unclear whether some redesigned experiences will be standalone apps or integrated into existing ones. Jassy described a four- or five-person team that agreed on an idea on Thursday and built a working prototype that was about 80% shippable by the weekend, arguing that established companies must learn to build at the pace of agent-native startups.

Jassy identified software vulnerabilities as a more immediate AI risk than catastrophic scenarios such as an evil botnet. AI models are finding more vulnerabilities than people had been finding, while many software systems and embedded devices were not built with strong security. He expects applications to detect threats and deploy fixes within minutes rather than wait through dependency-heavy processes. AWS customers of all sizes are concerned about security; AWS professional services helps them, and Continuum identifies vulnerabilities and builds patches. Jassy also said AI combined with automation can extend penetration testing across an organization’s applications, beyond what limited security teams can cover.