Amazon CEO: $220 Billion Still Can’t Meet AI Demand

Summary

• Amazon expects about $220 billion of 2026 CapEx and still cannot meet visible AI demand. Jassy says 2026 capacity is sold out, 2027 is largely sold out, and a good chunk of 2028 is sold out, with some capacity reserved for on-demand use. AWS’s AI business exceeded a $25 billion revenue run rate within three years: “we’re leaving a meaningful amount of demand on the table.”
• The infrastructure cash cycle creates near-term free cash flow pressure before capacity earns its return. Data centers require investment roughly two years before monetization and have about 30-year useful lives; servers and networking are purchased only months ahead, with better demand visibility. That equipment typically breaks even in about three years against five- to six-year useful lives and roughly five-year customer deals.
• Jassy sees enterprise production applications becoming AI’s largest absolute usage segment, with inference cost a major adoption hurdle. He thinks quite a few, if not most, existing enterprise applications will use inference pervasively when viewed five to seven years out, with adoption developing along the way. Inference remains “an order of magnitude too expensive,” in his opinion; Trainium’s claimed 30–40% price-performance advantage helps lower the hurdle for rewriting applications.
• AWS intends to protect existing customers despite AI labs’ willingness to pay for scarce capacity. Ben Horowitz presses the economic temptation to replace established workloads with higher-paying labs; Jassy answers, “We’re not gonna squeeze our customers.” Both AI and non-AI demand are growing rapidly, and he expects shortages across different components for several years, requiring choices between resources such as HBM and DRAM.
• Amazon’s lab investments deepen relationships spanning models, distribution, and custom chips. Jassy points to Anthropic’s Bedrock usage and Trainium collaboration, and OpenAI’s arrival on AWS after prior business arrangements prevented it. OpenAI also made a multi-year, multi-gigawatt Trainium commitment and collaborated on a stateful runtime: “You’re gonna need state.”
• AI changes both staffing requirements and the speed at which software must defend itself. The hosts cite Bedrock rebuilding its inference engine with six engineers in 76 days versus an estimated 40 people for a year, an experience Jassy calls eye-opening. He expects fewer humans doing today’s work and smaller teams, while model-discovered vulnerabilities will require applications capable of receiving agent-driven updates “in minutes.”
• Jassy thinks incremental AI improvements will very likely leave incumbents without winning experiences in three to five years. Amazon is reimagining Alexa, retail, Prime Video, Fire TV, and Prime Music, while remaining uncertain whether the results belong in new applications or existing products. The organizational obstacle is attachment to prior work; as Horowitz recalls his boss saying, after the first line of code, “it’s religious.”

Deep dive

1. Amazon’s fulfillment crisis unlocked a faster retail model

• Since 2021, Jassy says Amazon’s top line is about 80% higher, AWS is almost three times bigger, and advertising revenue is about two and a half times bigger. Prime Video has become “a large and profitable business.” But his early CEO experience centered on repairing retail economics after an extraordinary expansion.

• Amazon doubled a fulfillment network built over 25 years in about 18 months, then built a last-mile transportation network the size of UPS in about 20 months. Costs suffered. Jassy now calls that difficult starting point “a real blessing”: it forced reconsideration of assumptions, including regionalizing a flat national US fulfillment network, which shortened transportation distances, lowered costs, and accelerated delivery.

• The hosts’ challenge is how to tolerate those inefficiencies without weakening operating standards. Jassy says Amazon’s 16 leadership principles include tensions: pandemic customers and sellers needed capacity quickly, so speed prevailed over optimization. The safeguard was a “truth-telling culture,” direct discussion, and metrics exposing problems; the cleanup then distinguished small productivity improvements from architectural changes.

2. The $220 billion buildout has a long cash cycle and visible demand

• Jassy contrasts early AWS, which reached roughly a $60 million annual revenue run rate after three years, with AWS’s generative AI business exceeding $25 billion in run rate over its first three years. AWS overall grew 36.7% year over year in the latest quarter he cites, on a $169 billion annual revenue run rate.

• AWS was originally conceived to free Amazon’s product teams from infrastructure work: Jassy says they spent 70–80% of their time on non-differentiating infrastructure and only about 20% on product features that differentiated them. Internal demand for shared building-block services validated the idea before external launch; he says they did not anticipate the business growing externally as fast as it did.

• Horowitz asks how Amazon knows when the buildout becomes excessive, citing financial-market anxiety and CoreWeave’s debt and equity markets. Jassy’s honest non-answer: “I wish there was an equation.” Even about $220 billion of expected 2026 CapEx leaves meaningful demand unserved; he believes additional investment could make economic sense over the medium to long term.

• The financing logic separates buildings from equipment. Data centers require investment about two years before servers arrive and monetization begins, but have approximately 30-year useful lives. Servers and networking require commitments only a few months ahead, depending on the component. If demand does not materialize, Amazon can refrain from buying that equipment.

• Servers and networking tend to break even in about three years, with useful lives of five to six years; most deals run about five years. Building many sites simultaneously therefore creates free cash flow “headwinds in the early years,” before operating capacity and equipment payback catch up.

3. Jassy frames data centers as essential infrastructure with local obligations

• Asked about political opposition, Jassy starts with existing dependence: banking, healthcare, shopping, education, and other digital services already use data centers. He compares AI infrastructure construction with building highways and bridges in the 1950s, arguing the US “cannot afford not to build” or risk another country, such as China, establishing leadership and norms.

• His rebuttal to resource criticism includes the claim that golf courses consume 200 times more water than Amazon’s data centers—and more in two days than its data centers use in a year. On pollution, he says site registrations disclose emissions assuming backup generators run continuously at full capacity, whereas actual operation is about 0.1% of the year, roughly ten maintenance hours.

• On electricity prices, Jassy acknowledges that rates have risen in some locations and fallen in others. He blames inadequate grid investment, with AI triggering overdue expansion. Once capacity arrives, he says rates in most geographies will flatline or fall, pointing to Indiana, Georgia, and Mississippi.

• Communities must benefit too: Jassy cites about $3 billion in taxes paid in St. Joseph County, Indiana, and another $1 billion committed over five years through Built Together. His prescription is continued public-private cooperation, while conceding the industry can improve how it builds.

4. Enterprise production workloads are the missing middle of AI demand

• Jassy calls current demand “barbelled.” At one end sit labs consuming “gobs and gobs of compute” and a few runaway applications such as ChatGPT and Cloud Code. At the other sit enterprises using AI largely for productivity and cost avoidance: customer service, fraud detection, and business-process automation.

• The middle comprises existing enterprise production applications, the overwhelming majority of which, he says, still do not use inference pervasively. He thinks this becomes the largest absolute AI usage segment. The adoption rate is hard to predict, but looking five to seven years ahead, “quite a few, if not most” applications will use inference broadly; change will accumulate before then.

• Horowitz’s pushback — worth keeping: why not buy out established customers and resell scarce resources to labs willing to pay almost anything? Jassy says AI and non-AI businesses are both growing rapidly. Physical trade-offs include allocating memory production between HBM for accelerators and DRAM for core workloads, but “we’re not gonna squeeze our customers.”

• Jassy cites AWS proactively helping customers optimize costs during the pandemic despite reducing its own short- and medium-term revenue. Capacity expansion follows that same long-term partnership logic. With 2026 sold out, 2027 largely sold out, and a good chunk of 2028 sold out, he expects constraints in different components for several years, while reserving some on-demand capacity.

5. Custom silicon lowers the economic hurdle for enterprise inference

• The 2015 Annapurna Labs acquisition was not an AI-chip bet. Amazon initially wanted a NIC that offloaded virtualization, networking, security, and storage so nearly 100% of the server could support customer compute. That capability now sits in every server, Jassy says. Scale revealed opportunities for vertical integration, while the team’s talent suggested it could build other chips.

• Graviton followed frustration with improving CPU price-performance when Intel held virtually all the share. Jassy describes Intel’s incentives as understandable, but poorly aligned with customers seeking cheaper cloud migration. He says 99% of AWS’s top 1,000 EC2 customers use Graviton pervasively, with about 40% better price-performance than other x86 processors.

• Trainium addresses a similar accelerator-market dynamic, offering a claimed 30–40% price-performance improvement over comparable processors. The mechanism matters: enterprises considering replacing CPU applications with accelerator-based inference need greater incremental benefits when costs rise sharply. Inference remains “an order of magnitude too expensive,” in Jassy’s opinion; cheaper inference lowers the threshold for rewriting applications.

6. Lab investments connect financial conviction with operating partnerships

• Horowitz tentatively puts Amazon’s investments at about $18 billion in Anthropic and $50 billion in OpenAI, asking whether investing in customers resembles dot-com-era excess. Jassy points to the teams, technology, applications, and extraordinary revenue growth. He believes both will be among AI’s long-term successes, while stressing there will be many successful companies and not everyone will win.

• With Anthropic, Jassy emphasizes customers running its models through Bedrock and the company’s work on Trainium. He says both organizations learned substantially from the chip collaboration. The investment rests on that partnership alongside Amazon’s conviction about AI and Anthropic’s prospects.

• OpenAI addressed customers who wanted its models on AWS but could not access them because of prior business arrangements. The partnership includes Bedrock access, a jointly built stateful runtime, and a multi-year, multi-gigawatt Trainium commitment, like Anthropic’s. Jassy’s agent architecture call is categorical: “You’re gonna need state.”

7. Smaller teams need software that agents can repair in minutes

• The hosts cite Bedrock rebuilding its inference engine with six engineers in 76 days, versus an estimated 40 people working a year. Jassy calls the experience “very eye-opening” and says Amazon circulated it internally to challenge assumptions about specialization and team size.

• Jassy expects fewer humans doing today’s work, with smaller teams and broader functional ownership as models provide much of the initial work across disciplines. New jobs will emerge, including some not yet known. Established companies that adapt slowly will, in his opinion, become more vulnerable to startups already building with AI from inception.

• He predicts companies will have to rewrite almost all their code to support agents, with constraints governing updates. Cybersecurity supplies an urgent reason: models find vulnerabilities humans were not finding, while existing classification, patch development, and deployment dependencies take too long. Applications will need to respond “in minutes”; without agent-capable updates, he says, companies will have “no chance.”

• Horowitz calls the security threat “clear and present,” contrasting it with hypothetical catastrophes. Jassy says AWS is helping through professional services and Continuum, which identifies vulnerabilities and builds patches. Models plus automation can also test a broader application surface than scarce security staff; waiting for trouble is a bad approach when protecting consequential data.

8. Winning AI experiences require reinvention and human judgment

• Jassy argues that businesses treating AI primarily as an incremental improvement will very likely lack winning experiences three to five years from now. Alexa’s need for new “brains” was obvious, but implementation remained difficult because teams were proud of existing work and attached to hard-earned lessons about what did and did not work.

• Horowitz recalls his former boss’s warning: “everything is up for discussion until the first line of code is written, and then it’s religious.” Customers may also love the existing experience. Jassy compares the transition with infrastructure leaders resisting cloud adoption: compelling technology and economics still require substantial change management.

• Amazon has teams reimagining retail, Prime Video, Fire TV, and Prime Music. Jassy does not expect to get everything right initially. One unresolved question is whether successful ideas become entirely new applications or enter existing products incrementally, changing the overall experience over the next few years.

• His advice to young builders combines extensive AI use with continued critical reasoning. Business-school students once had to research, annotate, and analyze three cases daily; AI can now bring everyone to an advanced starting point. But assembling the pieces into strategy, products, and businesses still requires judgment and experimentation: he does not think AI will do that “for most people most of the time.”

9. Single-thread ownership supports Amazon’s wider ambitions

• Jassy identifies AI, robotics, and space development as three enormous transformations Amazon participates in simultaneously. Amazon Leo’s low-Earth-orbit constellation targets the roughly 500 million households worldwide without broadband; retail robotics, autonomous ride-hailing through Zoox, healthcare, and grocery extend the range of customer problems Amazon wants to address.

• His answer to spreading too thin is organizational: separate, decentralized, single-thread teams rather than the same leaders managing multiple businesses. Opportunities need capable leaders and builders, a differentiated idea, an underserved market, and enough potential scale to move the needle. Beyond those conditions, breadth depends on how much Amazon is willing to invest.