Amazon CEO: $220 Billion Spending Won't Stop 2027 Compute From Selling Out
Jassy says that even with $220 billion in capital spending this year, Amazon deliberately leaves some demand unfulfilled—2026 capacity already sold out, most of 2027 sold out, 2028 also selling a large chunk. The money was well spent.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Computing capacity for 2027 is already mostly sold out
Jassy says even with Amazon's $220 billion capital spending this year, the company deliberately leaves some meaningful demand unfulfilled: 2026 capacity already sold out, most of 2027 sold out, 2028 also selling a large chunk. This means continued capex is not wasteful but a rational choice under real scarcity, standing in direct contrast to market worries about ‘AI bubbles and insufficient capital.’
— Andy JassyAWS AI revenue is growing hundreds of times faster than AWS itself grew
In its first three years, AWS achieved annual revenue of only $60 million, but in the first three years of the generative AI wave, AWS alone generated over $25 billion in annual AI revenue—according to Jassy himself, ‘hundreds of times’ faster. He uses this comparison to illustrate the real scale of this wave of AI demand: not marketing narrative, but growth velocity the company has never seen before.
— Andy JassyThe $220 billion capex works because data center and server spending follow two entirely different cycles
Jassy breaks down Amazon's capex cash cycles: data centers need roughly two years of advance construction but have useful lives of 30 years and only need to be built once; servers and networking equipment only need a few months lead time but have payback periods of about three years and update cycles of five to six years, with most contracts running five years. Because data centers lead and servers lag, the company's free cash flow comes under pressure during rapid expansion; it swings positive only once the data centers are filled with servers and start generating returns.
— Andy JassyThe chip team acquired in 2015 became Amazon's most profitable acquisition
Jassy calls Annapurna Labs probably Amazon's best-ever acquisition, though it was not originally made to bet on AI chips. It began because virtualization, networking, security, and storage functions on EC2 servers were consuming too much compute, so the company wanted to offload them to chips and leave servers nearly 100% for customers. Through this team, Amazon went on to build the Graviton CPU chip (used by 99% of the top 1,000 EC2 customers, with ~40% better price-performance than x86) and the Trainium AI accelerator (with 30-40% better price-performance than comparable processors).
— Andy JassyInference cost is still an order of magnitude too expensive—the real bottleneck
Jassy says inference costs have dropped in recent years but remain ‘an order of magnitude’ too high. That is the key bottleneck preventing large-scale enterprise adoption of AI accelerators: if the cost to rewrite existing applications with AI accelerators is too wide a gap, enterprises need very high incremental returns to justify it; only when inference cost drops does that threshold fall, and applications start broadly using inference. Horowitz cited an example on the spot: his own team has one employee whose AI research alone ran up $2 million in inference charges.
— Andy JassySix people in 76 days did what once took 40 people a year
The Bedrock team rewrote the inference engine with 6 engineers in 76 days—work that would likely have taken 40 people a year in the past. Jassy says experiences like this are reshaping how Amazon thinks about organization: in the future, fewer people will do the same work, teams will be much smaller, and members need to work across functions because models already bring people to 80% completion. Whether established large companies can match this speed decides whether they'll get lapped by AI-native startups.
— Andy JassyModels find vulnerabilities too fast—patches must shrink from weeks to minutes
Jassy believes the most underestimated risk of this AI wave is not sci-fi scenarios like ‘evil robots’ but the immediate cybersecurity problem: models discover vastly more vulnerabilities than humans ever did. At that pace, applications in the future must be able to respond to security threats in minutes—but today most companies take a long time just getting from vulnerability discovery to classification to patch development to deployment. His conclusion: if an application has not been redesigned for agents to complete updates in minutes, enterprises have no defense in the new world.
— Andy JassyRedoing Alexa is harder than optimizing it, but refusing to redo it means being left behind
Horowitz divides AI-era strategy into three types: cutting costs, creating something unprecedented, and completely tearing down an existing successful product to rebuild it—Alexa is the third. Jassy says Alexa's original ‘brain’ design was cutting-edge at the time, but generative AI changed what the ‘brain’ can do entirely, so it must be completely rewritten, even if that means abandoning years of accumulated team experience and hard-won lessons. He believes this judgment applies equally to Prime Video, Fire TV, fresh grocery retail, and all existing businesses—not optimization, but total reimagining, or in three to five years you will not be the winner.
— Andy JassyIn their own words · checked verbatim
We always leave a certain amount for on demand, but 26 is obviously sold out. 27 is largely sold out. A good chunk of 28 is sold out.
Andy Jassy0:00
the financial markets are saying things like, oh, there's not enough money in the world to finance these things.
Ben Horowitz9:33
In the first three years of this generative AI wave, our business, or just our AI business in AWS is over $25 billion in run rate. So it's several hundred times faster.
Andy Jassy12:41
Is there, like, any business pressure to go, well, maybe we should buy out the old guys and sell them to the new guys because they'll pay us whatever.
Ben Horowitz22:57
it's one of the very best acquisitions we've done at amazon in the history of the company if not the best
Andy Jassy28:12
But it's still an order of magnitude too expensive, in my opinion, for it to be used as pervasively as possible.
Andy Jassy32:18
the Bedrock team rebuilt your inference engine with six engineers in 76 days. And that would have taken 40 people a year.
Ben Horowitz36:30
we are living in an era of three of the biggest transformations in our lifetime, and maybe ever, between AI, robotics, and space development.
Andy Jassy56:15
Figures
| 2026 capex | ~$220 billion | 0:00 |
| AWS annual revenue | $169 billion, 36.7% YoY growth | 4:16 |
| AWS AI business annual revenue (first three years of generative AI wave) | over $25 billion | 12:41 |
| AWS annual revenue in first three years | ~$60 million | 12:41 |
| Amazon total revenue vs. 2021 | ~80% growth | 2:05 |
| AWS business scale vs. 2021 | grew nearly 3x | 2:05 |
| Investment in Anthropic | ~$18 billion | 33:22 |
| Investment in OpenAI | ~$50 billion | 33:22 |
| Graviton chip price-performance advantage | ~40% higher than x86 processors | 31:17 |
| One employee's inference research spending over a period | ~$2 million | 33:22 |
Glossary
- Barbell demand
- Concentrated at both ends: AI labs and hit applications consume compute ravenously; enterprise end mostly uses it for cost-cutting; large swath of enterprise applications have not ramped yet.
- Stateful runtime environment
- An AI agent runtime that can remember contextual state; Jassy sees this as the standard architecture for future agents.
- Nitro
- Offload virtualization, networking, security, and other functions to a dedicated chip, leaving servers nearly all their compute for customer workloads.
- Annapurna Labs
- The chip team Amazon acquired in 2015, which created the Graviton and Trainium chips.
How to listen
Investors and founders tracking AI infrastructure capex and cloud competition, and executives wanting to understand how large organizations are being reshaped by agents.
Around 53:10, the section on how young people use AI for learning offers generic advice with low informational value and can be skipped.