# Annapurna gave AWS a path to custom chips and lower inference costs

Markdown copy of https://realalpha.app/s/annapurna-gave-aws-a-path-to-custom-chips-and-lower/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Andy Jassy](https://realalpha.app/people/andy-jassy/) · Andy Jassy is CEO of Amazon. · Founded and led AWS, then became Amazon's CEO in 2021, steering its $220 billion AI buildout.
- Host: [Ben Horowitz](https://realalpha.app/people/ben-horowitz/) · Co-founder, a16z · Co-founded a16z and Opsware, sold to HP for $1.6 billion; wrote The Hard Thing About Hard Things.
- Host: [Erik Torenberg](https://realalpha.app/people/erik-torenberg/) · General partner, Andreessen Horowitz · Helped build Product Hunt; founded On Deck, Village Global and Turpentine.
- Show: a16z Podcast
- Episode: [Amazon CEO Andy Jassy on the $220 Billion AI Buildout](https://realalpha.app/v/-YMMRQ1qscI/), Oct 23, 2026
- Moment: 27:50 to 33:34 of the episode (6 min), playing on https://realalpha.app/s/annapurna-gave-aws-a-path-to-custom-chips-and-lower/; the episode from there: https://realalpha.app/v/-YMMRQ1qscI/?t=1670

## Quote

> “It's about **40% more price performance** than other x86 processors.”
> — [Andy Jassy](https://realalpha.app/people/andy-jassy/) at [31:06](https://realalpha.app/v/-YMMRQ1qscI/?t=1866) · a16z Podcast

## Takeaways

1. Annapurna first helped free more server capacity for customer workloads.
2. Graviton delivers about 40% more price-performance than x86 processors.
3. Jassy says inference costs must fall further for widespread enterprise use.

## About this segment

Annapurna began as a way to offload server functions, then built Graviton and the Trainium AI accelerator. Andy Jassy says custom chips improve price-performance and lower inference costs enough to broaden enterprise use.

From “Amazon CEO Andy Jassy on the $220 Billion AI Buildout”, Oct 23, 2026. [Watch the full episode](https://realalpha.app/v/-YMMRQ1qscI/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[28:01](https://realalpha.app/v/-YMMRQ1qscI/?t=1681) So, kind of speaking of that, you did something back in 2015 when you bought Annapurna Labs, which was probably before, you probably weren't really thinking, I mean, maybe you were, but about like giant AI demand in 2015 because there just wasn't much AI demand in 2015. So, kind of what drove that decision and then how has it kind of paid off in this new world? Yeah, I think, you know, the Annapurna team is one of the most talented teams and set of people I've ever had a chance to work with. And I, it's one of the very best acquisitions we've done at Amazon in the history of the company, if not the best. And it was, we foresaw some pieces, but I don't think that we, we did not make the Annapurna purchase for AI chips per se.

[28:51](https://realalpha.app/v/-YMMRQ1qscI/?t=1731) But, you know, when you get to a certain scale, you start to have access and visibility to patterns of customers at very large scale that you realize if you can vertically integrate and you could build certain capabilities yourself, you'd be able to provide more value for customers and, you know, and better economics for the business itself. So what Annapurna started as for us was just on the kind of core EC2 server, we felt like we were using too much of the server for virtualization and for networking and for security and for storage.

[29:33](https://realalpha.app/v/-YMMRQ1qscI/?t=1773) And we wanted to be able to use nearly 100% of the server for compute for customers. And so we had been working with Annapurna before we acquired them on this NIC that offloaded those components to the NIC to allow us to have as much of the server for compute as possible. And we really liked the team and we really liked the product that we collaborated with them on together. And we realized that was going to end up being in virtually every EC2 server, and it's turned out to be in every server we have.

[30:04](https://realalpha.app/v/-YMMRQ1qscI/?t=1804) So we originally bought it because that was going to be a core capability of our compute capabilities. And because the team was so talented at building chips, we felt like they could probably build other chips, too. And at that time, we had this very deep partnership, which we still have a deep partnership with this company, with Intel. But Intel had virtually all of the CPU share. And when you have a partner who has virtually all the share, they're not often thinking about how to drive price performance.

[30:36](https://realalpha.app/v/-YMMRQ1qscI/?t=1836) They're trying to think about other features that allow them to get more share and drive more revenue. By the way, it's totally understandable. These are some of the dynamics of how markets work. It's part of how the innovator's dilemma, you know, comes up. And so we had tried to push on price performance, and it was hard for us to get attention there. But we had customers who wanted to move more and more of their on-premises workloads to the cloud, and they wanted better price performance.

[31:02](https://realalpha.app/v/-YMMRQ1qscI/?t=1862) And so, the next thing we turned the Annapurna team loose on was to build a CPU chip. And they built Graviton, which 99% of our top 1,000 EC2 customers use it very pervasively today. It's about 40% more price performance than other x86 processors. And so that has been hugely impactful for our customers and for our business. And we just, a few years ago, we saw that same movie in the AI space where you really had one dominant provider who's very good.

[31:37](https://realalpha.app/v/-YMMRQ1qscI/?t=1897) And we had a deep partnership with, we still have a deep partnership with, we will for as long as I can foresee. But they have giant markets and lots of different products they're building and priorities. And we knew our customers needed better price performance. And that's why we spun out a separate team in Annapurna to build an AI accelerator. which we call Trainium, which is also about 30% to 40% more price performance than the other comparable processors.

[32:07](https://realalpha.app/v/-YMMRQ1qscI/?t=1927) And, you know, I think when you talk about that middle of the barbell we were just talking about, if you ask what are the things that will change that middle of the barbell to have more of those current enterprise production apps using inference, there are a couple things. You know, some of it is, you know, some of the talent is comfortable doing post-training and fine-tuning and some of the talent and organizational muscle that knows how to prioritize what to go after and how to do it. But a big, big chunk of it is the cost of inference. The cost of inference has come down quite a bit in the last few years, but it's still an order of magnitude too expensive, in my opinion, for it to be used as pervasively as possible.

[32:49](https://realalpha.app/v/-YMMRQ1qscI/?t=1969) Because if you're asking enterprises to rewrite their applications that use CPU today with AI accelerators, if the difference in those costs is very significant, the amount of incremental benefit they have to get from having the AI accelerators on there is a much higher bar to want to rewrite those apps. But if you can take the cost of inference down, that bar is much lower and all applications are going to be better and smarter and more capable with AI accelerators and inference.

[33:16](https://realalpha.app/v/-YMMRQ1qscI/?t=1996) So that is a really important part for customers to being able to use AI pervasively is driving the cost of inference down. And our chips business This is a big piece of that. Yeah.