J.P. Morgan Presents ACQUIRED at AWS re:Invent (ACT100)

Please welcome to the stage co-hosts of the Acquired podcast, Ben Gilbert and David Rosenthal.

Hello, Reinvent.

We like to say on the show, welcome to Acquired Live at Reinvent.

At reinvent.

I'm Ben Gilbert.

I'm David Rosenthal.

And we are your hosts.

Well, uh, this is fun.

This is a little bit different than my basement studio.

Yeah. In my backyard.

So, for folks who listen to Acquired, which I'm assuming if you're here is is you, uh, what we normally do is these 4hour deep dives into hundred-year-old

companies. You can listen in any podcast

companies. You can listen in any podcast player. You know, you put in your

player. You know, you put in your headphones, you basically don't see us.

You go on your long run, you're doing your dishes, you're on your commute, and we're telling you a story over, you know, a hundred years of a company.

And today we have the uh sort of indulgence for David and I of talking about the present, of talking about the technology era that we are living in and getting to talk through that with the

protagonists of some of the most important companies in the world here on this stage.

Yes. So, we have four really incredible guests here to join us today. We've got

Max Newer, who is the global co-head of JP Morgan Payments. We have Greg Peters, the co-CEO of Netflix.

Heard of it?

Heard of it. [laughter]

We have Arvin Shinovas, the co-founder and CEO of Perplexity, which I use all the time.

And then, this is funny. I I still haven't figured out the right way to introduce our final guest.

Our final host. guest host because he's both Yes. Yes. Our guest and our host

both Yes. Yes. Our guest and our host Matt Garmin, the CEO of AWS.

Yes.

It is going to be just an incredible morning and afternoon here today. Thank

you all for joining us. And huge, huge thank you to Amazon, to Amazon Web Services, to JP Morgan. They have been incredible partners to us over the years

at Acquired. I'm sure to all of you in

at Acquired. I'm sure to all of you in the room. planning today, getting these

the room. planning today, getting these four incredible people here today to have great conversations with all of us.

So, with that, let's get started. We'll

throw it to a video with JP Morgan.

JP Morgan payments is known in the industry for innovation. There's so many key trends that we are not only participating [music] in but driving.

Every country has an rapidly developing ecosystem around payments. [music] For

us, it's critical to develop those linkages to develop those products by country and deliver to our clients who may be very large multinationals operating in dozens of countries. Cuz in

JP Morgan Payments, we ultimately want to help our clients run their business.

Please welcome to the stage the global co-head of JP Morgan Payments, Max Neerchin, [applause] Max. Great to see you.

Max. Great to see you.

All right, so we have to pretend we don't know each other well and like we haven't been hanging out all the last few days at reinvent and doing JP Morgan meetups and everything and what? Hanging out at Radio City, hanging out at Chase Center.

We just get to see you in the most amazing venues around the world. Chase

Center in San Francisco, Radio City in New York, this beautiful old Italian theater in Las Vegas.

It's a testament to our partnership.

We've done already so many things together and you know, you've been such amazing partners. So great to be here

amazing partners. So great to be here with you.

Thank you. So David and I know the JP Morgan payments story and I suspect most people in this room don't and they're kind of like wait there's this big you

know enormous fintech company inside of the world's largest bank uh America's largest bank not the world's largest bank large bank how did this happen [snorts]

what is JP Morgan payments so we ultimately work with companies of all sizes to manage their money and move their money So think about all of us. We

have a bank account. We have cash there and then we make payments out of it. We

do this for companies of all sizes. And

that means it's a bit more complicated because companies operate in many countries. They have different

countries. They have different currencies. You have an entire finance

currencies. You have an entire finance department that needs access. So that is the service um that we provide to to our clients. And we do that at a massive

clients. And we do that at a massive scale. We process about 12 trillion

scale. We process about 12 trillion dollars on a normal day up to 15 trillion dollars on on peak days. And

that means that the world's GDP flows through our pipes basically every 10 days.

And um we do that also as a at a massive precision. So our efficiency and

precision. So our efficiency and precision rate is 99.9999%.

which is basically the same level as space flight or nuclear power stations operate at. That is that is the service

operate at. That is that is the service that we provide to our clients to make sure that their their money is safe and and that we can do all of this that they

need. Do payroll, buy other companies,

need. Do payroll, buy other companies, pay suppliers, and do all of this in more than 150 countries and accept payments from customers. I

mean principally and accept payments from customers.

Great point. We just passed Black Friday. Um, Cyber Monday was a huge

Friday. Um, Cyber Monday was a huge event probably for all of us here, also for the payments industry. Everything

went well. And if you bought something in the last week or so, likely we processed it and um, it's an really important part of our business.

So, how how did it's an incredible business. How did this come to be part

business. How did this come to be part of JP Morgan though? I mean the uh this usually does not exist within a bank.

[laughter] Uh payments companies are usually separate. Uh this is the result

usually separate. Uh this is the result of a whole bunch of strategic decisions over the years by Jamie by the leadership team to build up this this technology business within the bank.

Right. It is a business that really was built over decades and we have invested over a long period of time to build out the technology um to deliver all these

services because ultimately exactly to your point this is a technology business that needs to work very reliably and constantly renew and innovate. We also

did a few acquisitions um over the years. So for example, our merchant

years. So for example, our merchant acquiring business that was an acquisition probably 10 years ago and we also work with the ecosystem um the fintech ecosystem in particular very

actively. So we do not build everything

actively. So we do not build everything just in house but we have a network of over 100 partners that help us deliver all these services and innovate for our clients.

So one of the themes that we're going to be talking with all the guests today about is investing in the future. You

talked about the incredible precision and and being a giant incumbent. I mean,

this is an 18 plus billion dollar business on its own, just JP Morgan payments that exists within the bank.

How do you think about allocating that next incremental dollar on innovation when you kind of have to protect this really durable important franchise?

Investments are absolutely key. One, it

is a very very competitive environment.

So you need to invest to provide cutting edge solutions. But at the same time the

edge solutions. But at the same time the payments world itself is also evolving so fast. It's scaling. So you need to

so fast. It's scaling. So you need to make sure you keep the infrastructure fresh. And there we operate like any

fresh. And there we operate like any other technology company. We have a very rigid process around where we invest, how we invest.

We absolutely continuously invest in good times and in bad times because you do need this continuous growth in the platform to make sure that you deliver

solutions. And by the way, we also make

solutions. And by the way, we also make mistakes. So you do want to continue to

mistakes. So you do want to continue to invest so that you have an an a cutting edge solution out there. And we think about investments for the future in

three horizons. Most of our investment

three horizons. Most of our investment budget actually goes into Horizon 1, which is scaling the existing solutions, moving to the cloud. We have a fantastic

partnership with AWS um to make all of this work, enhancing the payment methods we offer. So that is probably the bulk of our investments.

Then we have a second horizon. This is

where we build new products, where we go to new countries, where we expand our network. Think about real-time payments.

network. Think about real-time payments.

Think about opening up new countries in the Middle East. That is there. And then

the third part which is probably 10 20% of the budget is reserved for real innovation because we are absolutely here to shape the future of payments.

And this is where our innovation in blockchain technology, biometrics, AI, etc. fall. And all of this together um

etc. fall. And all of this together um continues the the buildout of the platform. It's funny. Uh we wanted to

platform. It's funny. Uh we wanted to ask you about blockchain which um you know in some ways it feels like this is 2025. This is the wrong year to be

2025. This is the wrong year to be talking about blockchain should be but it's actually despite everything that's happened in the crypto markets over the past few years. It's become a big part

of fintech and JP Morgan payments. Tell

us about what's happening with blockchain and why you guys have continued to invest in it over the past couple years.

Blockchain as you say is a very very hot topic in payments at the moment. Um

actually we have been at it already for more than a decade. So we have seen all the ups and downs um of the crypto space and and the blockchain space. Most

recently we launched our JP Morgan coin which is a tokenized deposit on the public chain. So think about it as a

public chain. So think about it as a bank deposit where you can move money using the public chain.

Is it a stable coin? It is similar to a stable coin that it leverages the blockchain to move money, but it is different given that it's a true bank

deposit, which means it's insured.

Yeah.

Which means that [snorts] you can pay interest and it also means that um it's fully integrated into the JP Morgan infrastructure and you have the full backing of the JP Morgan balance sheet behind it

and it's not a speculative asset, right?

It's not a speculative asset. It's

literally the same as putting your bank, your your cash into into JP Morgan. What

we like about the tokenized deposit is that it truly combines the benefits of both the banking world as well as the

crypto digital world. It allows all the great characteristics of a traditional deposit, but then you can move money

truly 365 24/7. And you also have this feature of programmability where a CFO or a treasurer can say this is how we want to move money. You can put in some

rules and it allows them to manage their entire treasury operation a lot more efficiently.

And does it reduce expense also of changing currencies versus traditional foreign exchange?

You still need to exchange money quite in the same way because ultimately most clients still want to move in and out of fiat currency at some point.

The real interesting part is we see tremendous growth there. So we already process probably a few billion dollars every day through the tokenized deposits.

Wow.

And the growth is truly exponential.

I mean this is uh we interviewed Jamie uh over the summer uh in Radio City uh Jamie Diamond which was amazing. Um

that JP Morgan is launching JP Morgan coin in 2025. If you if you had told that to Jamie when you took over the payments business a few years ago that

JP Morgan would be launching JP JPMCoin uh what would his reaction have been then actually so I worked with Jamie now for a long time he would have listened he

would have made me make the case as we did and I think there's a truly differentiated view between the blockchain technology that makes the

payment system a lot more efficient and can really help create the benefits that we talked about and some of the cryptocurrencies where um you know you can have different views in terms of how

they adopt and since this is truly focused on the blockchain in a very controlled manner I think the whole company is behind it. All right. What I

gota ask one more question. What did

cause the institutional flip uh to stop saying, "Hey, we hate Bitcoin and it's evil to embracing it in the last year."

I think it was definitely um becoming more nuanced in the messaging. Digital

assets, digital currencies is not all Bitcoin and Bitcoin is not the same as digital assets and digital currencies.

And I think they were often mashed together. Yeah. But with the whole

together. Yeah. But with the whole ecosystem maturing, I think everyone now sees the different aspects and I I think the the answer are just more nuanced now.

Makes sense. Um let's flip to probably what I'm sure is on um everybody's mind here and reinvent certainly is on.

We're 12 minutes in and we haven't said AI yet.

Yeah. [laughter]

Well, I think Max said AI would you know it's one of the key three areas. We

haven't said AI. We haven't said AI.

Yes. Uh talk to us about AI. I mean you uh you move 12 trillion dollars a day.

Um how how are you using is is AI making an impact today on that? AI is next to digital currency is probably the second

big mega trend in payment that is really the talk of the industry and in my view will actually have long-term a lot more profound impact and change the way we

operate and for us having cutting edge AI solutions is a question of survival in in the business over the long run. So it is front and

center and I deal with it every day running the business.

There are some use cases that we pursue for AI inside the company. Um for

example in our operations group we use AI to run investigations sanction screening more efficiently and it's more traditional AI think machine learning

and we've seen really good impact already. So our volumes have increased

already. So our volumes have increased double digits over the last um five six years and the headcount only two 3%. So

you get true operating leverage. We're

excited about this. This will continue.

Same engineering productivity. Our

engineers you know use AI tools extensively and it has absolutely made them more productive better output less errors and given how big our agenda is

that is really really exciting. And then

it's also helping our our sales team become a lot more targeted. You get

signals, you know, when to reach out to clients. So there's a lot of internal

clients. So there's a lot of internal use cases that we pursue.

For clientf facing, we also have um a lot going on. I think fraud and protecting our clients money is probably one of the key areas uh where where we

use AI. We have a large payments model

use AI. We have a large payments model that truly now moves from traditional AI to new AI. So we are no longer just

using fraud labels and um assisted learning to more predictive models unleashed on the 12 trillion worth of transactions and hundreds of millions of

bank accounts that we see that truly creates superior fraud signals. So we

have seen 30 40% improvement. I mean

this is if you can move the needle on fraud that's massive impact to profitability for you and your customers right it's it's massive impact and if I speak to our clients fraud and keeping their

money safe is probably the number one topic that they care about they love all the discussion about AI blockchain they say please make sure the fraudsters don't come about our money and and AI is making that harder than ever

because the fraudsters are getting correct it's a true arms race and um by middle of the year we had already seen more fraud attempts on our clients money

than in all of last year. So you can definitely see how this is becoming a very very very prominent challenge.

So the AI investment is really helping us safeguard our clients here. Um and

then there's also solutions um where we help our clients run their business better. For a CFO, it's for example a

better. For a CFO, it's for example a really hard task to forecast their cash flows, which ultimately tells them how much liquidity they need to hold. So AI

can make this a much less manual task, much more precise, and ultimately help them save liquidity and working capital.

So there's a lot basically in every aspect of payments that you can think about.

Yeah. I mean, your your biggest customers are merchants. Uh so

predicting their cash I mean we've studied so many um merchants retail companies on acquired like the the the difference between a great enduring company and not is is often you cash

flow management it's working capital management it's can you have a negative cash flow cycle and especially with more and more payments happening real time and in the

moment you need to take quick decisions in nanoseconds is this fraud suspicious or not so the bar is also raising from a

consumer expectation standpoint which makes AI and the ability it offers even more important. So, one thing that I

more important. So, one thing that I have had a tough time doing in this AI era is reading beyond the headlines when there's a a hot topic and trying to look

under it and say, "Okay, how much usage is there actually of that topic?" And

one of the buzzy terms right now is agentic commerce. And I'm curious to

agentic commerce. And I'm curious to between your your clients and all the different companies that you power um what are examples where agentic commerce is really happening and there's a lot of activity there versus a lot of people

showing cool mockups and saying agentic commerce is the future dot dot dot.

It's a great question and aentic commerce definitely meets the criteria of a lot of noise at the moment. Even a

year ago, barely anybody was speaking about Agentic Commerce and today you see multiple announcements every day.

We're going to talk about it with Matt later.

Exactly. We look at Agentic Commerce as a transaction truly end to end from consumer identifying what they want to buy to the checkout to the post-purchase

experience. And agents are already

experience. And agents are already pretty prominent across the full full value chain. They are often embedded in

value chain. They are often embedded in merchants or marketplaces as platforms. You have agents embedded in the various AI platforms. I mean, Perplexity is also

coming. Um, you have consumer agents

coming. Um, you have consumer agents coming up. So, there's a lot of activity

coming up. So, there's a lot of activity in the space. Most of the impact has been so far on the search side of the um of the customer experience where people

use them to to to to find what they want to buy. But more and more you see also

to buy. But more and more you see also checkout and actually customer experience and customer decisions taken over by agents. I do think there's still

quite a lot of open questions. Um first

of all, what are the use cases? Probably

if you buy toilet paper, you're happy for an agent to do it. If you want to buy a nice pair of shoes or a nice dress, you probably still want to do it yourself. Then how do you deal again

yourself. Then how do you deal again with fraud liability? What do you do if a truck with 20,000 rolls of toilet paper is outside your house ready for delivery? You know who is now

delivery? You know who is now responsible?

Actually happened to be with bananas. I

hit the wrong button.

It happened to you with bananas.

I thought I was ordering 10 bananas and I ordered 10 bushels last week. We have

a lot of bananas at my house right now.

Yeah. You ordered 10 bushels of bananas [laughter] on Amazon.

Amazon.

Imagine this.

Were you on Amazon business? Like what?

Okay. Okay. All right. [laughter]

So there is open question. I think also the business model a lot of um commerce today is funded by by um advertising and commercials you know how will this

evolve going forward so a lot going but I do think it will mature and then have an impact by the way another area we are very passionate about is agentic

treasury so this is where treasures and CFOs just give a payment file and then they say you optimize dear agent for

speed cost global reach and then identify for every transaction what's the best way to do it. It's also a great use case.

This is also I imagine an area where JP Morgan kind of like uniquely can create solutions for customer like hey optimize my treasury. Uh there's everything you

my treasury. Uh there's everything you just said but there's also you know what portions of my treasury am I going to keep in what investments at what balance and like match that with predict my cash

flows and my cash flow needs like you know. Yeah. Correct. And we do have also

know. Yeah. Correct. And we do have also a very close partnership with our colleagues in the consumer bank to really look at the end to end journey from the consumer all the way to the

merchants and everyone in between to embed our solutions there.

Makes sense.

Amazing.

Well, Max, thank you so much for joining us today. That's a great place to leave

us today. That's a great place to leave it. Really appreciate you coming.

it. Really appreciate you coming.

Thank you so much, Ben and David. And

thanks also for the partnership that we have. Truly amazing.

have. Truly amazing.

It's truly amazing. Thank you, Max.

Thank you, [music] appreciate that. [applause]

appreciate that. [applause] [laughter] Uh that was fun. Um for our next segment of the morning here, we um have a

company, this is appropriate on so many levels, has reinvented itself. Heyo.

Heyo. Uh, at least three times, arguably four or five kind of from whole cloth.

Ship me a lot of envelopes. And

a lot of envelopes.

More envelopes.

Not a lot of envelopes. In fact, zero envelopes these days.

Um, and yet I consume more of their product than ever.

As do we all. As do we all.

So, uh, before we announce our next guest, let's watch a little video about the company Netflix.

This is the battle for hearts and minds.

Building a real atscale global streaming business is hard because you got to combine great tech product and great content.

We're going to be able to move more quickly than other streamers as we leverage pre-existing tech and data science assets and expertise.

This is furious action that broadens the Netflix offering.

I've been training for this all my life.

We hope that you know that ultimately reinforces our value as the most important service for your entertainment needs that how many times a day do you see that logo?

I mean once but like I I do I get the warm fuzzies seeing the the rainbow Netflix. You're like so amped up for

Netflix. You're like so amped up for what's about to come next. Well, please

welcome to the stage the co-CEO of Netflix, Greg Peters.

[applause] How's it going?

Great.

Good to see you.

How are you?

I like your setup here.

Thank you. Thank you. Well, we uh maybe inspired by some, you know, Netflix shows here.

I was going to say, is this an audition?

Are we should we talk later? [laughter]

We're We're all ears.

You're open. Okay. All right. Good.

That's good to know.

Uh I did see Nobody Wants This up there and I was wondering, did you throw that in because we're kind of doing a podcast here as an homage?

I was going to say, you know, it all fits together as a universe of entertainment that's connected, you know, bouncing from one to the other.

So, it does. It does.

it does. It does.

Well, this this feels really poetic. I

don't know how many people out there are like David and I and watched 12-year-old, 13-year-old reinvent keynotes, but in the very first keynote, Andy Jassie first reinvent

interviewed Reed Hastings as a case study for the cloud for large companies moving to the cloud which at the time was complete heresy was provocative yeah

AWS was for startups it was for Y combinator companies and here's Rey Hastings well and even then I think it was this idea that you would lose control over your infrastructure and you know and your data and all these other things

were out there so it was definitely it was pushing the edge uh but you know so we started that process in 2008 8 actually a couple years before that keynote. Um and just to to step back and

keynote. Um and just to to step back and say over a decade and a half, you know, what has happened both with Netflix and with AWS is pretty amazing. We were

doing a million hours a month back then, a billion at the um the time of that 2012 reinvent and you know now we're doing 50 billion a quarterish every quarter kind of thing. So growing up

together has just been amazing.

Wow. Wow. That's that's incredible. Um

well, as we as we teed up in the intro, um you know, appropriate enough for reinvent, we have to ask you, you know, Netflix, unlike so many other companies, not just

reinvent, but like you've killed [laughter] what you were as a company in the past and gone to a a whole different thing. There was the red envelopes to

thing. There was the red envelopes to streaming. There was we will never have

streaming. There was we will never have ad supported to actually we really are into ad supported. there was we won't have sports. We have NFL games on

have sports. We have NFL games on Christmas. Now,

Christmas. Now, how do you make these decisions? When do

you decide that we're not just going to do little big change?

We're making a big change.

Well, I think um you know, most of that is driven by the desire to grow the business and you know, serve the customer more. So, it's not probably

customer more. So, it's not probably different than I think what anybody else has experienced. And you know, we we

has experienced. And you know, we we really have felt that focus um and being limited in what you did for a long period of time, especially as there's this big macro shift in how entertainment was going to be

distributed was the way to win. And I

think that served us quite well for, you know, a period of time, but then you also say, okay, now we need to think about what we're doing. Um that's more and it's interesting because you I think you described them as big shifts and

there were big shifts in strategy.

That's definitely true. But I think, you know, even some of those there's good examples where we've taken sort of an evolutionary or iterative approach towards how we entered those spaces because we generally come from the perspective that when we're going to do

something new, we don't know what we're doing. We have not figured out how to be

doing. We have not figured out how to be good at that thing. And you know, I always I think it's funny because people say when you're going doing something new, you just hire all the people that know how to go do that. And certainly

you can go do that and that's a a useful approach to to learning quickly. But I I think it underestimates the the need to essentially build institutional capability that understands how to go do something like you can hire a bunch of

expertise in but that expertise has to learn how to do it in the context of the business that we are currently operating especially if you're doing something brand new. I mean, it's not like when

brand new. I mean, it's not like when you decided to do streaming through Silver Light to IE browsers launched.

I've loved the fact you're going way back to Silver. You were one of the very few people that I was very frustrated because I had a Mac. Uh I remember um it was so novel.

Mac. Uh I remember um it was so novel.

It's not like there were a lot of people you could go hire that had that know-how. There is a

know-how. There is a So, and there's there's multiple examples here, right? So, I think it's a it's a um you know, so you talk about like you're doing something completely new. We were building from scratch. I

new. We were building from scratch. I

remember even like talking about like how are we defining the standards that will be instituted you know for the streaming universe etc. So that's a that's one example where you're where you're building it as you go. I think

you know ads is another really interesting one where you know it's not like this is uh you know a new space right people have been doing it and doing it well for a long period of time and so we you know thought okay we're

going to get into it we'll do it in an iterative approach you know partner to get an ad stack at the at the get-go and then build our own and learn through that process but even this I think is a is a bit of a new version of ads because

we're trying to bring essentially what has been amazing about digital advertising that you know folks like Google and Meta have been doing for a long time and are great at, you know, great targeting, great relevance, all

those things, measurement, but then also bring it into a premium, you know, TV like ad experience, right? And so that's a whole new thing as well.

Yeah. Like we'd be terrified of programmatic ads on acquired like we're so brand protective. I would never run a random commercial against it. Netflix is

premium like that. I mean, you guys have to really be thoughtful about Yeah. And I think and that's some of the

Yeah. And I think and that's some of the differentiation that we have, right?

Because when you know you think of brands uh when they what can they get from Netflix that they can't get from maybe some of those other advertisers it is that you know that that premium placement the security that they have around brand and increasingly what we're

trying to do also is um be able to do things that TV would never been been able to do or not really to do at scale which is bring brandaw aare brand appropriate brand sensitive creative

into title sensitive title aare creative and mash those two worlds up in a way that everyone feels you know super excited about which you know used to be quite an involved process and take a lot of time but we're we're seeing you can

actually accelerate that and so that sort of you know um plays on that differentiation or mag do it programmatically well it's a I mean we're I think you we're going to walk towards that way so this and this is maybe the path that we

see leveraging some of these new tools to get to which is you start by taking what was an existing call it 18month process heavy humans involved right and see if you can get that process to like

three months right and then you start to really build more iterations and then more data and then you ask the question.

Okay, is there a certain set of situations or class of advertisers where you could actually make it, you know, truly automated, right? And so I think, you know, we're not there yet, but I think that would be the path we would seek to go.

So this is the story behind ads. Give us

the story behind, hey, we're never going to do live. We're never going to do sports. And I I

sports. And I I How do you walk your way into well the NFL?

Yeah. I mean, so live um you know, one of the big differentiators we had at the beginning was on demand, right? And so

this whole idea that we are going to, you know, counterposition relative to traditional entertainment because you know we don't have the same constraints and so you know when you're in that early stage you really focus on the

counterpositioning and live doesn't you know that you know it obviously runs counter to that strategic now at some point you become the it's like you know Hamilton Helmer like [laughter] Hamilton you love Hamilton he's and

shout out to Hamilton Helmer if you don't know him read his book it's it's great um so you know when you are at a certain scale scale though and you've become not so much where you're

counterpositioning but you are for an increasing number of people the place that they show up to be entertained then you get to start to say okay now we think we can layer on some of these more traditional things and compete in that

way so that's you know where we you know and live maybe the other thing too on this is that you know I I think about the entertainment ecosystem sort of as increasingly moving to two poles right

we get highly personalized highly niche experiences that has tremendous value because you're speaking very specifically to like you know what I want but then that creates a space that you know I think um and a need for this

other poll which is you know how are we all having a shared experience like what what what are we all doing you know in a single moment where we know that there's millions tens of millions maybe hundreds of millions of people around the planet

that are all doing the same thing so that we can talk about it and feel the power of that shared experience and that's where I think live you know and some of the things we've been doing there come into play I was going to save this for later but

uh Did you have any idea that K-pop Demon Hunters was going to be one of those things? [laughter]

things? [laughter] We thought it was great. We thought we thought it was an amazing movie. Uh but

I would say there there's these this extra layer that you get to where something magic happens where I think it, you know, it it enters into the Zeit case and sort of builds on itself and it

becomes one of those events where now, you know, whether it's social media or your kids Halloween costumes, you know, that you cannot, you know, escape. I

mean Ben Thompson is writing about K-pop demon which that's a yeah and for him to write about an animated show like that is amazing. So

is amazing. So um well in that vein one of the things we really wanted to ask you is if you look at the industry of Hollywood 10 years ago and you look at the

industry of Hollywood today just like the value chain how it works it's completely different arguably completely different but most folks in the audience are not from the entertainment industry. So, would

you be willing to give us a a a brief primer on here's how you make a movie or a TV show 10 years ago and here's how it actually works today?

Yeah, I may sort of run super counter to your thesis here. Okay, great. Great.

Because I would say I don't think it's actually that fundamentally different and I know and I don't want to understate the differences that are involved in changing company dynamics and things like that. So, that's that's fair. If you step way back, right, and

fair. If you step way back, right, and take a long-term view, you'd say, what are what are the what are the sort of core functions or properties of this one is you got to have a great story. You

got to figure out a way to sort of produce that story in a compelling way for audiences, you know, leveraging the latest technology that exists out there.

And then you have to find a way to connect that story with a bunch of people to watch it, who are going to enjoy it. And you could, you know, I

enjoy it. And you could, you know, I think we could have, you could have used the words that you used around, you know, several shifts, whether it's like adding sound to movies, talkies, right?

Or television, tele remember television that was like, you know, gigantic and then, you know, VCRs and DVD and so on and so on. So, you know, I mean, do I think we have a we have a very

compelling, very effective distribution model in streaming. Now, we're now able to build

streaming. Now, we're now able to build essentially a global entertainment network. And that means that creators

network. And that means that creators when they have that story, right, that that's really compelling. We give them the tools to tell that story in an amazing way, sometimes in ways that they never could have before because we're upleveling their production

capabilities. But then we also then

capabilities. But then we also then connect that story with the largest audience that that show could have ever possibly have seen, right? I mean, you know, we think about whether it's Casa de Papel or Squid Game. I these shows

coming from countries and languages that have found audiences that are global audiences. That rarely rarely happened

audiences. That rarely rarely happened before, if ever. And so so in in my mind it's it's you know it's a maybe a quantitative difference if you will but qualitatively the processes feels very

similar to me.

Um I'll ask it at a sharper way. It used

to be that the distributors didn't like to own all the risk and all the upside.

And now it seems like Netflix and its peers want the risk and the upside. And

you see less of these like actors or studios or getting big backends on things. Does that feel like a fair

things. Does that feel like a fair characterization of what happened?

Well, I think the the reality is there's multiple models that are in play today, right? And so you still have, you know,

right? And so you still have, you know, all the way from folks who will literally produce, you know, something on their own. They will budget the whole thing and they will say then I want to go see who I can sell this to from a distributor all the way to where you

know folks like us and our peers get involved at the very beginning and say no, we will take the risk and we will support that. So one there's just a

support that. So one there's just a diversity of approaches out there which I think is very healthy too for the ecosystem because you know we all know you know when you when you have an idea and you're trying to get that idea out

there you know having many varieties and paths to yes is a very powerful thing I think to allow that to happen. So you

know there are some I you know shifts in that process but I think those shifts are also consistent with you know some of the what we see as like the value that we deliver and what we hope to offer which is again about this if we're

excited about an idea and we feel really strong about the fact like if you know if we can pull this off with the creative partners that we've got we think we can find a really big audience for this thing. So then it's you know appropriate for say like we're going to enable this to happen. We're going to

lean in to make sure that that idea doesn't die because it can't find funding in some other way. Um and then of course you know we want to participate in the upside of that as well and so then we get to these structures

right I mean it is interesting of course Hamilton Helmer has already come up once in this uh for for anyone who doesn't know u Netflix was was Hamilton author of seven powers the great strategy

framework we use on every episode of acquired Netflix was like the uh like like the test baby like the perfect company for developing the seven powers when I and I don't know if you know this

too but we actually had Hamilton come in and he taught essentially a strategy course for all of our directors and above because we we so believed in this this the because again in our model we

sort of try and distribute the decision-m out right and so having this really very strong strategic capability embedded throughout the company was important so that little on this particular let's talk about

scale scale economies this is like Netflix is the ultimate scale scale economies where the more the subscriber base grows the more theoretically Netflix should be

willing to pay for any given piece of content or to create any given piece of content because there's a wider base of people to amortize it over. And then if you kind of play out that flywheel, it's like, well, you get the best content,

you market it to the most people, you get more subscribers, and then you just can kind of like grow faster and pay more for content than anybody else.

Where does that fall down? Because it

sounds like utopian, like once you start that flywheel, you just run.

What's hard about that for you guys?

Well, I think the maybe the one component that's important to layer on top of that is that um not every title essentially has a global audience. And

so that sort of global scaling effect and increasing leverage it it it doesn't really work across the entire uh you know population. You know we're so at

know population. You know we're so at this point we're you know approaching you know getting close to a billion people who show up seeking to be entertained from us. they have a very wide range of tastes and so increasingly

what we have to do is think about you know how do we serve that that wide range of tastes and not you know when we produce a show we think we're going to find a pretty big audience for it um and that's the goal obviously but you know

we also really are anchored in what we call local for local which is you you can't start the idea with you know you're going to make a Squid game that's amazing you know for Koreans but then it's going to be globally you know

hugely popular it just doesn't work that way you get to really inauthentic and sort of lame stories as a result. So you

start with super um local thinking about the local needs and then some of those work globally but you're not really is there like a point in the in the development process where you're like this started local but we think it might be global

you know some the creative execs have idea that this is going to go but then but you know like I mean mostly you get surprised by how big that those can get more than banking on it I would say but it just gets to this idea that you know

you really are serving sort of sections of audiences at the end you know at the end Maybe now that you are at this giant scale that you're at, uh how do you

decide when a regional investment is worth it versus something that you can amortize across all users?

Uh it's a it's a I'm not surprising this question comes up all the time. Uh so I think it's one of the trickiest things that we that we try and balance. Uh you

know not surprisingly we come from the perspective of there's great leverage in building global scaled you know approaches. Um but also what I have been

approaches. Um but also what I have been you know what I'm constantly reminded of is that actually serving that wide variety of people is a very local game right and so you try and do it at multiple layers right you're building

core central stuff that you get great leverage off of you know we do a single investment almost some of the leverage scale benefits you talked about on content are the same for fixed cost investment on engineering and product advancements right so you get that same benefit there then there's some of that

you're building called central systems that have an adapt adaptation locally so think about like payments rails right you know you do a central payments rails. You're using a lot of great logic

rails. You're using a lot of great logic there, but then you're actually interfacing with, you know, many, many different payments providers to actually make that work. But then ultimately, it's sort of individual personalization

at the end of the day, right? Which is

you're bringing, you know, this vast library of content um uh and and we think about a personalization from not only the title perspective, but how do we present that title in a way that we think is going to connect with the user?

And increasingly now we're moving into personalization at the UI composition and module level. Right? So that allows us to think about not only you know what's in the UI but the form and

function of the UI and sort of say hey you know when you show up for us what can we give you that specifically meets your needs and that might be different um you know time of day or day of week right so Thursday night versus Sunday.

And I've heard you say so I'm sure everyone in the room is familiar with in the last year Netflix rolled out a new UI.

Yep. Hidden in that UI is the fact that you are not prefetching all the titles.

You are watching my behavior on the UI and then in real time fetching the right content for me based on my behavior.

That's right. So we used to premputee overnight essentially in trough when you didn't have you know the operational systems um sort of at peak. We would

premputee those recommendations but it doesn't allow us to do what exactly you're describing which is react in session dynamically based on the signals that we're getting. So we had to rearchitect everything right we had to

re rearchitect the whole uh recommendation systems how those are pipelined into the UI how the UI gets composed page construction all those things but it has a whole bunch of benefits because it allows the dynamic

response but it now allows us to think about the construction of the UI um based on need state or or use case and we can use those signals to adapt to those things.

That is so cool. I'm now going to be like very deliberate about where I'm sort of dwelling as I'm scrolling through like how much of a trailer do I let play or you you can do that or you can just let it fly. Here's an interesting twist on

it fly. Here's an interesting twist on this too because um you know it's it it takes implicit signals which are valuable and stuff like that but it also allows you to give um very simple

explicit queuing behaviors right where I think you'll see more and more well we'll give you because you know we sort of there's a vast universe of things that you could need right but ultimately you we sort of know that that universe

is centered around maybe four or five core use cases and so if we can find shorthand ways that you can understand where we say we know you're in one of these states give us a hint about what you need from us right now. You can probably

accelerate that process and get us give us give us a clue about exactly what you need.

Is this one of we were talking a little bit downstairs before um in my opinion at least I'm curious you as a a heavier user than me if you would agree. Um you

have the simplest UI of all the major streaming and entertainment platforms out there.

You created the current streaming UI that then everyone else copied.

But but I think even you know you look at like um we have young children. We

spent a lot of time on your competitor that has a mouse as its mascot. Uh

[laughter] uh that is a very um complicated UI. Uh

well, when they smash two streaming services together, it became complicated.

Yeah.

Yeah. Uh or or or even YouTube, you know, a different like, you know, it's actually quite uh busy when you open it up. Um is this part of, you know, the

up. Um is this part of, you know, the ability to dynamically generate UI? Is

this pro part part of the philosophy of simplicity? Well, we've we've always

simplicity? Well, we've we've always been sort of really anchored in um a simple UI philosophy and we we think that there's a general trend in companies um because they just want to serve more people, right? And there's

more features that you can add and you know some product manager says like I know that would be somebody's going to use that. But we we generally think that

use that. But we we generally think that there's a tendency to over complicate and add and so we always try to have this um very strong sense that every um feature has to earn its weight, right?

It's got to be delivering value commensurate to its pixel allocation, its engineering maintenance budget, whatever you know metric you think. So

we call this scraping the barnacles, right? Where you inevitably build things

right? Where you inevitably build things in where you think they're going to be big and maybe they're not or they're big at a time and then they seek, you know, they don't they change over a period of time.

At one point I star rated thousands of movies on Netflix.

It was your repository, right? Well,

actually it you know and those aren't totally gone because essentially we mapped them into the new model. So we

didn't lose that data but we did find that a more simple model believe it or not and even it seems so trivial right going from stars to just you know three thumb states essentially but even that

reduction changes how users engage in the amount of um signal that you get because just that that much simplification. So, you

know, we really try and be, you know, uh, disciplined about pulling stuff back when it when it's not meeting that criteria anymore. So, we can allocate

criteria anymore. So, we can allocate those resources to something else that'll do better.

There's a cool This is not This is not on your iPad, so you will not see this.

This we were talking backstage. Um, I

think there's a really cool story that I had no idea about about how deep Netflix integrates with the partner ecosystem and also just a a cool story of

technology abstraction. Can you share

technology abstraction. Can you share with the audience how how a how many Netflix clients there are, how many these you need to build and two how that happens.

Yeah, when you say clients, we're talking about the devices that you use to stream to essentially endpoints. Um,

different kind of clients than JP Morgan.

Yeah, exactly. Yeah, different kind of clients. They're good clients though

clients. They're good clients though too. So um you know we're we're on you

too. So um you know we're we're on you know thousands and thousands of of types of devices literally you know we're at the point where we count in billions the number of you know actual physical instances of those things but the you

know we got to this because we sort of were talking about like how do you handle the diversity of this stuff and the actual answer is we don't really have a lot of diversity. We have written an abstraction layer um essentially an

SDK that we call the Netflix ready device program which we give to all of those device manufacturers. But really

the maybe the fun part you're into is we do this at the chip level. And so you know when a new chip tapes out that's targeted at um TVs or settop boxes or adapter products when that's you know by

the time that chip chip tapes out Netflix is already integrated onto it and the latest version of NDP is there.

So that just moves through the value chain. If you're a television

chain. If you're a television manufacturer, you're like the in the Netflix integration is already baked into the chip that you are buying from a MediaTek or a whatever.

That's right. So 10 years ago, you know, 12 years ago, the Netflix pre-integration was a feature essentially that when you were thinking about chip selection and what you were buying, you're like, "Oh, I just get the Netflix piece. I don't have to worry

Netflix piece. I don't have to worry about that anymore." Right now it's become ubiquitous. So essentially, I

become ubiquitous. So essentially, I think it's become a a must-have feature when you're selling chips into the ecosystem, which is so cool. It's a great business model piece of leverage once you have it too because then those chip

manufacturers need to do that in order to become valuable enough for their customers to buy them.

Yeah. I think and it's just and you know obviously what fuels it is the customer demand at the end of the day. That's the

thing that matters because the whole thing works because there's enough demand where a TV manufacturer says that's pretty important for me to have that feature and then that flows through the ecosystem that way.

Um, uh, we spoke about this a bit with, uh, with Max and it's going to be a recurring theme here and, um, maybe this is the right time to bring it up with

you. Agents and AI. Um, how are you guys

you. Agents and AI. Um, how are you guys thinking about this? Like, is is there a future that you're thinking about where agents and AI services for consumers are

the ones that are bringing them content to consume? Yeah, I I think of this more

to consume? Yeah, I I think of this more as maybe a software architecture question or or you know, first and foremost. So, if you think about where

foremost. So, if you think about where most of our systems have been architected before, there's um you know, hard-coded business logic that essentially is calling out to services

or whatever to to execute that logic and we're really pivoting more towards a model, an agentic model at the center of that. And so think about this as maybe

that. And so think about this as maybe like a member model that says I you know understand you know a a wide variety of data around how members behave and I can think about how to optimize that for a

certain set of goals. And so now that you that model sits at the center it calls out to all these services. Maybe

it orchestrates other models to do other things for it to achieve a bunch of different goals. But it it really

different goals. But it it really provides a tremendous amount of flexibility and adaptability to how you then extend those systems in new spaces.

So it thinks like um if you had a new content type, right, and you want to do personalization for a new content type, it used to be that that was a you know many month process to stand that up, get data in, tune the models, figure out

what's working, all that stuff. Now we

can do that, you know, very very effectively. Um it also allows I think

effectively. Um it also allows I think different use cases of how you do product development. So think about like

product development. So think about like non technical or less technical folks who might have an idea and say like okay um let's say after a football game, you mentioned live, right? you know, how do

we pivot folks off? There might be a specific need state to pivot people off of a football game um to another show that they might want to watch. Maybe

maybe it's not even a football game that we don't have, somebody else's football game because we notice a surge of traffic after the sporting events happen. Come on to the service. So that

happen. Come on to the service. So that

you know that product manager, let's say, or this merchandising person says, "Okay, how do we fill the football shaped hole in your in your heart right now?" Right? And before

now?" Right? And before

Thank you. Well, I I got to say as a huge football fan, thank you for quarterback and receiver coming out during the offseason. Like it it fills a lot.

That's a way to fill your Yeah. Exactly.

Exactly. So, but you know, doing that and flighting sort of a promotional model before would take a lot of work.

You think about, okay, what's the universe of titles? How do you filter that? Maybe you put that into a title

that? Maybe you put that into a title rank or a traditional, you know, sort of ML model to to rank those in a personalized way. But now, you can sort

personalized way. But now, you can sort of have that person literally say, "Okay, this is what I'm trying to go do." speak to that central model and

do." speak to that central model and have it sort of put together a bunch of stuff to try and solve that problem. And

then and then these ideas sort of um you know they they they live or die or they get you know sort of elevated in the UI based on their utility to any given user at the end of the day.

Netflix is famously on the cutting edge of machine learning.

I mean the Netflix algorithm, the Netflix have you challenge that's right.

Uh, have you deprecated a lot of the older models in this world of LLMs?

Well, I'd say we're we're in this interesting spot and this is a you talked about evolutionary versus sort of revolutionary changes, right? Where this

is an interesting model where I think there are moments where you really don't get the benefit of the new approach unless you really shift over, right?

cloud I think is another really good example when we were doing that process we talked about like from 2008 or so to shift over you know you unless you're fully on the new model you really aren't

reaping all the benefits and we call this this awkward period in the middle of Roman writing where like you have one foot on one horse one foot on another horse and that's like that's a place you don't want to be for too long you want to minimize that um and so this is a

situation where um you know we think about leapfrogging essentially the existing approaches and trying to shift over and usually this thing is really hard to go do because the existing approaches are they're highly refined,

they're very sophisticated, hardened, all the corner cases have been worked out. And then you go to a new approach,

out. And then you go to a new approach, you're like, "Wow, we're we're super excited about what this could offer."

But it's, you know, it's there's often times, you know, like a a tough transition in that approach where you're betting on the future and then you're saying, "We got to get through an awkward phase to get there." But that's how we think about sort of shifting off of more

uh when do you guys um unleash the the chaos monkey? Is that what it's called?

chaos monkey? Is that what it's called?

Yeah. Wait, at what point in the um you know new technology onboarding process do you unleash the the chaos monkey on it?

Well, I think what for us that's a continual practice, right? Because

that's how we ensure that we are resilient against a whole variety of things and we we really come from the philosophy that if you're not doing that constantly, you're just building in vulnerability and weakness. So, I'd say we're that's just that's an ongoing

and this is a this is a process that runs within Netflix that is constantly just breaking stuff.

At least I read this like 15 years ago.

I don't know how true it is today. No,

it's we still very much believe in it.

Whether it's chaos monkey which is called service level, you know, killing service levels or chaos Kong which it takes out whole availability zones, you know, and see if you're are you resilient to availability zone failover and things like that. It's a pretty it's

pretty crucial part of how we think about our re reliability schema. And I I would say our reliability, you know, metrics tend to demonstrate the value of that.

All right, this feels like this this like funny question I almost had to ask you. There's two points of view. There's

you. There's two points of view. There's

there was the streaming wars and they're over and Netflix won. And then there's this other perspective that's like there a lot of streaming services out there.

There's a lot of streaming services. A

lot of them have really matured. They

sort of consolidated into I don't know about formidable competitors, but vertically integrated competitors.

You've got YouTube. It is in some ways the most competitive media environment that has ever existed. How do you think about this?

I would say it is definitively the most competitive media environment that's ever existed. If you think about the the

ever existed. If you think about the the the range of choice that is available to any given, you know, person on the planet for what they can be entertained with, there's never been this range of options available and they're all I mean they're increasingly completely

friction-f free, right? You can like if you don't like what you're doing right now, you're a flip or a click away from doing something else. It's very very simple. So, um in my mind, it's

simple. So, um in my mind, it's incredibly competitive. We've expanded

incredibly competitive. We've expanded the types of entertainment that are available to people too, right?

Gaming.

Yeah. Games or you know, social media, short form. There's so much that's going

short form. There's so much that's going on. And so I think that's great for

on. And so I think that's great for consumers. It's, you know, it's

consumers. It's, you know, it's incredible. Um, but it's also, you know,

incredible. Um, but it's also, you know, it requires us, our peers, everybody to basically be constantly reinventing themselves, pushing themselves to deliver more because I, you know, if you

stand still in this environment, you will have your lunch eaten for sure.

Well, um, that feels like a great place to leave it. Greg, this has been such a blast. Thank you so much.

blast. Thank you so much.

My pleasure. Thanks for having me.

Really appreciate it. It's

been great.

Thank you. [applause]

Thanks for I watch so much Netflix. [laughter]

I was going to say USB is so happy right now. I feel like every time Ben and I

now. I feel like every time Ben and I talk uh you're like did you watch this?

Like if you listen to our carveouts, Ben's are always, you know, Netflix shows. Yes.

shows. Yes.

And um uh yeah, you must have so much more insight into like what happens in your evenings these days.

Yes.

Yeah.

All right. our next company. So the the JP Morgan and Netflix, these are old companies, some older than others, some older than others.

Uh neither born in the AI era, both using AI technologies, but not native.

Uh this next company, Perplexity, is the complete opposite. It is exclusively

complete opposite. It is exclusively formed in the AI era, an AI native company, and is allin. So we want to roll a quick video before we introduce Arvin.

Our company is reimagining the future of search by trying to take us from Kendall links to personalized answers that cut through the noise and gets to exactly

what we want.

Pull up the clip of Jensen demoing Perplexity Labs. I've pulled up a

Perplexity Labs. I've pulled up a YouTube video showing Jensen demoing Perplexity Labs. This is such a complex

Perplexity Labs. This is such a complex product to run and a hard problem to solve. Hence why we decided to go all in

solve. Hence why we decided to go all in on AWS.

Please welcome the co-founder and CEO of Perplexity, Arvin Shrinavas.

Great to see you again.

All right, Arvin. So, I I got to say I use Perplexity all the time. Many many

times every day. I've got the what is it? I something P my fingers. I I have

it? I something P my fingers. I I have the muscle memory. I have the keystroke.

I use it all the time. Um and it has replaced much of web search for me. I

still use traditional web search for some things, but perplexity has eaten a lot of it.

If we're sitting here 10 years from now, does traditional web search still exist?

Probably not. Um, there is a big distinction between consumer desire and consumer behavior. Um, Mark Andre

consumer behavior. Um, Mark Andre actually once said in his podcast with Lex Freedman that 10 links was always a hack in in the late 90s, of course. Um,

explain that. A hack to do what?

Well, ultimately the the desire of the consumers to get the answer. Like when

when you when you go and search pizza near me, pizza is the answer. Pizza near

me is the the search query and of course you show a bunch of links with so many different uh websites and you have to internally read all those things figure out like what you exactly want and all

that cognitive power you can now delegate that to an AI that actually would go and do this for you and you directly get the answer. So when you can

directly get the answer and if it's often accurate than not you start trusting it more and more and as a result you start asking more questions and then the behavior finally changes.

It's going to take a it's going to take a while, right? It's very hard to change behavior that's been happening for like more than two decades. But your question is like we sit here 5 to 10 years from

now like yes that's sufficient time for behavior to change and also like the younger generation is going to be you know faster in adopting this.

So I remember the first time I used perplexity two and a half years ago probably it's felt like forever but AI was not that long ago. uh it just immediately worked.

long ago. uh it just immediately worked.

It would it was an answer engine. It

would site its sources. It had that nice little UI at the top where you could see the thumbnails of each of the sources.

And it was one of these weird instances where I was like maybe I found this company really late in its life, but it seems like there's sort of instant product market fit. Did it feel that way to you or was there some wandering in

the woods to get there? So, um I think you guys said this in one of your episodes, like the earliest adopters of the internet were the nerds, and that's kind of why Amazon started selling

books. Uh that that that's kind of true

books. Uh that that that's kind of true in AI, too. The initial adopters were the tinkerers and um a lot of academics, people

interested in doing medical research, um academic research.

Yeah. Uh GBT W GBT1 was a thought experiment but GBT2 was a and even GBT3 was research access like there was no UI for it.

Exactly. Yeah. It was just people were just playing tinkering it on the playground the API playground. So uh the the first set of queries the first set of users within a couple of weeks of

launch were just um people saying oh I love using it for searching about papers on bioarchchival like I I love using it to make sense of like some of the research papers

and um and and so that made sense a citation format appealed to those people um I'm an academic by training my co-founder is also the an academic so

it's it's not surprising that initially we appeal to people like us but Then turns out our belief was that it's not just academics who would do research.

The whole concept of a citizen researcher like everybody's going to do research eventually. It's just that they

research eventually. It's just that they never had the tools to do that. They

never had the ability to go ask a question and get an instant answer. So

that's why they were never like actually asking a lot of questions.

So always we found that like if you give people a new superpower, they're just going to like love it and keep doing a lot more of it. Was the initial idea to create this full replacement for a web search or was it just that narrow thing

around helping researchers find answers in research?

So when we tried to hire our first founding engineer, I was trying to close him and he asked me after all the what year was this?

2022.

Okay.

Pre prej or post September even before chat months before.

Uh yeah. And um he was like okay like do you guys give health insurance and I never I never cared for my own health insurance. It was more important

insurance. It was more important insuring the company. Uh and then yeah of course you got to give health insurance and I don't want to bother VCs asking like what provider you should go

with. It sounds so embarrassing. Um and

with. It sounds so embarrassing. Um and

then obviously insurance is one of the most advertised keywords on on tin blue link search engines. So uh I thought it would be a it it should be possible to have a

feed the product. Yeah. Yeah.

Yeah. So we built it for ourselves. We

first built it without any web search integration. Just ping a model and give

integration. Just ping a model and give an answer in in a chat format and obviously we couldn't trust what it said and so we realized like you need that factual grounding

and I come from an academic background and my adviser first took a look at the first draft of the paper I wrote my PhD first time. He said, "There's so many

first time. He said, "There's so many sentences here that have no like peer-reviewed citations. You're just

peer-reviewed citations. You're just like saying whatever comes to your head." That's not how you write an

head." That's not how you write an academic paper. It's not a opinion

academic paper. It's not a opinion piece. You have to actually like

piece. You have to actually like grounded in facts. And I said, "Okay, what if we baked in that principle into a conversational system and we could actually trust and use it." And so we

built it for ourselves and then uh turns out like a lot more people wanted it.

whether it's going to perfectly be the end of 10 blue link search engines or be a complimentary thing. Um at that time we didn't care. We just thought it was useful and like let's get in the hands

of others and see what happens.

Right. So we're we're here at reinvent.

Uh so I want to take a question sort of in the infrastructure direction.

Perplexity is astonishingly fast. I mean

when I ask it something it it is it's a near instant answer instead of a like let me think about that and synthesize information and write something for you.

It's like here's information.

Why? How?

Well, uh it's glad I'm glad you say that because when I first couple of days before launching the product, uh I sent it to one of my uh angel investors. He

used it and said you should actually call it a submit job button instead of a hit query button because it's so slow.

[laughter] Takes seven seconds. All the concepts that you're so used to now where you stream the answer. You don't you don't wait for the full answer to be decoded on the back end and then put it on the

front end. You actually stream it token

front end. You actually stream it token by token or chunks by chunks. All these

were UX hacks we figured out very early on. Um because that way you the

on. Um because that way you the perceived latency is what matters. It's

not the real latency. If you actually measure the real latency in terms of time from first token to the last token, it's going to be pretty bad even even today. But what you want is the user to

today. But what you want is the user to already feel like they got the answer while they're reading it. So uh for more complex questions, it you need to have the model think for a while.

So we innovated on interaction patterns like streaming the intermediate steps. Y

which today is often referred to as chain of thought. um when when the model is not just doing answers but like doing a bit bit of actions stringing together

tool calls you want to show to the user what all is doing how hard it's working to actually get you the answer you want to show the user all that and I think the user gets some dopamine from watching the model in action so we

innovated on a lot of that um of course just core infrastructure how you put the links the URLs in the index how you cache them really effectively uh CDNs um

you know like obviously server side rendering um all these things that you can do to just make the app feel faster and of course you know like famously keep the homepage as minimal as possible

load time it's very essential um keep the cursor ready to type there should be no need to put the mouse in the search box just have the user just start typing all these subtle details matter

uh I have a question while we're on infrastructure um I have a question I realized that I really want to ask you we plan to ask it to Matt in a minute we we'll ask it to him too but uh which is

from a technical infrastructure and sort of like a stack perspective how do you build an AI company like you know before AI like I'm not technical I could tell you how you know it's like you spin up

EC2 you get S3 you get a database like I don't think I could tell you like how did you build perplexity like what is your stack well uh it's it's good that we live in a world where cloud computing is already

pretty mature um like unlike the past search engines, we didn't have to build our own data centers, servers, GPU racks, necessary, databases, storage,

you know, Kubernetes like all these things are much more mature like distributed computing is a lot more mature. I think you were doing this uh

mature. I think you were doing this uh with the Netflix Coco like it Netflix definitely benefited from having access to AWS. So the similar thing uh that we

to AWS. So the similar thing uh that we benefit from which is complexity is built on AWS but you also need models and AI like

correct. So one of the most

correct. So one of the most controversial decisions u in 2023 early 2023 that I made is don't don't actually

go and build your own model pre-training um wait for open source to land. This

was before even Llama 2 uh and before GPT4 after 3.5 almost all the VCs when I go and meet them they'll be like do you train your own model and it's like I no and then

they you could immediately say they they would like look down upon me and now everybody wants to like back application layer companies because they know that like this is so capex intensive um but I

I I saw through that uh there'll be like maybe three or four relevant players and everybody else is going to really like have a hard time and there'll be a bunch of companies that would open source models. I didn't bet that it would be

models. I didn't bet that it would be the Chinese foundation model labs. I

definitely bet that Meta would be uh an important player. But all these things

important player. But all these things turned out to be right like like the costs per token went down dramatically.

The capabilities expanded dramatically.

Open source started closing the gap and commoditization of the model layer helped an application layer company like ours to like route across different models. almost every 17 days we've

models. almost every 17 days we've brought a new model onto the product. Uh

like the f like a play on the famous Jeff Bezos quote, right? Like your

profitability is not the customer's concern. Like I always say internally

concern. Like I always say internally like the which model you use to answer the query is not the user's concern.

Like no one cares like as long as they get a really great answer that's accurate and fast and trustworthy that builds the foundation for asking the next question and like next task that's what really matters. And so we bet on

that and that turned out to be when did you start building the model router?

I think we we started post-raining it after llama 3. I would say that was when we thought it was decent enough that we could close the gap to 4.1 GPD 4.1 or

claude 3.5 at the time. But we I I always was comfortable with the fact that whatever model we post train with the current best open source model within 3 or 4 months that'll be a frontier model from a lab that would

beat it. And it's okay. It's totally

beat it. And it's okay. It's totally

okay for that. We still benefit from having better margins by serving our own models for easy queries.

So you you do your own post trading on perplexity propriet you post train an open source model and make that a perplexity proprietary model.

That's right. We post train it for accuracy grounding uh keeping it less verbose in terms of the answer length and making sure the footnotes are all correct. M

correct. M um and making sure context across multiple turns is uh correctly retained to calls um we have lots of queries across so

many different categories finance you know all these things so just calling the right data tools for that so in so you've raised billions of dollars at least I think billions you've raised a lot of money yeah

uh given you're not building your own frontier model from scratch what actually are the most expensive things in building perplexity in the business yeah So the like I said the user doesn't

care what model you use. So if there is an amazing model offered by a model lab that helps you just get dramatically better accuracy or new capabilities like

agentic actions on a browser uh you want to get that in the hands of people to collect the tokens and then figure out distillation later. So that inference

distillation later. So that inference costs us a good chunk of money. So we do that at the same time like we are very capital efficient uh compared to like model builders.

Yeah.

Hence why we like you say race one or two billion but not like 10 billion or 100 billion. Right. Right.

100 billion. Right. Right.

So there's a dramatic difference between that.

But yeah we don't anticipate our capital needs to be increasing over time like in fact I would say u the more experiences that we able to provide the faster we've

been able to like bring down the cost.

So you launched a browser in addition to this sort of perplexity search chatbot.

How do you talk about the the first product? What's the right product

product? What's the right product category?

Answer engine.

Answer engine.

Answer engine.

Uh now that it's been out for a while, what have you learned and what's different than what you hoped or thought with the browser?

With the browser.

With the browser. Yeah. And and and why launch the browser in the first place?

Yeah. So the most important realization for us was u the amount of questions people ask. So obviously a lot of even

people ask. So obviously a lot of even though that was not our intention we obviously got a lot of users who are using traditional search engines like Google to be asking questions on flexity and the number of questions they ask

increase as a result because this is a very new behavior. Um on comet the browser the number of questions people ask on comet compared to using perplexity on another browser increases

multiffold like 6x or something like that. U it's because the browser

that. U it's because the browser go back to our Google part two episode in Chrome and the toolbar. Yeah. Yeah.

Yeah.

Yeah.

The browser has context to whatever you're seeing. So instead of copy

you're seeing. So instead of copy pasting something and then going to an AI tool and then asking something about it, you can just pull the sidecar on voice mode or just type in your query and say hey help me help me make sense

of this or like select a piece of text tell uh I don't quite understand that actually can you do can you can you pull this up and just send it as an email to my colleague and and schedule a meeting

on that to discuss further all these strung together orchestration of actions and research that is contextual to what you're seeing. These are questions that

you're seeing. These are questions that you could not ask before on a standalone tool that's not personalized to you. It

works the same way for everybody. But

now with the help of Comet, you have a personal assistant. I actually see it

personal assistant. I actually see it more as an the the way to build a real assistant that has agency and context that you have.

And so it's able to like do actions for you and and and understand you well. The

problem of memory in AI. Imagine an AI having access to all the tabs that you're h you you use your browser history your past questions your calendar your email that's the only way

to build a really useful assistant and so that was our motivation with comet do you uh try to think of a way ask this in a way you can answer could you imagine a

future where um you know you follow the Google playbook and there's more products and services that you would launch like do do you want to build out a whole suite or like how do you think about a strategy of of doing that versus Not yeah actually like you guys did this

three three three-part series on Google and like how you talked about all these ecosystem of products they built. So I

think there's only one other company that's done this really well which is like bite dance I guess right like um and I've always thought like why is there not a bite dance of America like like why not like build a lot of

different products together and and connect them well and and and in here it's like you could say Google Apple Amazon they've all done it really well.

So I think there are certain but Google's really done it you know at a scale that nobody else has nobody else has 9 billion user products exactly so I think I think it's certainly doable it takes time you got

to play the long game like when you when you do it at the stage we are earn which is 3 years old everyone's going to be like [laughter] uh oh look at this company like they haven't even had like 1 billion user

product yet and they're already out there building the next set of things but it's the the rules that if you're planning for the 10year journey like you obviously We want to build a tight ecosystem. So um our our our the way we

ecosystem. So um our our our the way we think about it is we don't want to launch yet another product for the sake of it. U you comet was not yet another

of it. U you comet was not yet another browser. We we thought we think it's the

browser. We we thought we think it's the only way to build a really good personal assistant. So whatever helps us build

assistant. So whatever helps us build that truly general purpose assistant that helps you and is really personalized to you um and helps you

make your decisions. If another product can help us do that better than our current two products, we'll absolutely do it.

So, there's a lot of founders in the audience today, and I'm curious.

Yeah, you're only a three-year-old company, but you're sort of this like successful incumbent AI company. What

have you learned about building an AI company that you want to share with people that is different than the classic entrepreneurship advice that was given in prior eras?

So um I would say like you know move fast and break things or like like have a long-term vision. All these are like all the usual um advice you read

everywhere. Um

everywhere. Um the one thing that I keep going back to as far as entrepreneurship goes is the single biggest determining factor is your grit. Like you just have to be

your grit. Like you just have to be extremely determined. um nothing else

extremely determined. um nothing else matters like like the true test of an entrepreneur is just your grit because things are not always going to go up.

You're going to have like obstacles.

You're going to have moments of doubt and every great company has gone through like difficult moments. You know,

there's this famous story where Amazon stock almost went down to nothing during the com bust and then it went back because they figured out AWS. So, I

think all these things well and retail for a long time and because Exactly. But yeah, I mean Amazon

because Exactly. But yeah, I mean Amazon did a couple convertible debt offerings um too. Yeah. Uh right after going public

too. Yeah. Uh right after going public uh I think and like that hadn't happened like we wouldn't be here today. Yeah.

Yeah. I think I think that that that's probably my biggest learning like just get get back to the whiteboard and think think through the problems from first principles. And by the way I listen I I

principles. And by the way I listen I I listen to your podcasts pretty much all of it. You have done like such

of it. You have done like such incredible work.

Oh thank you. Thank you.

And so uh you're a busy dude so thank you. No, no,

actually I I use it a lot for uh my own like strategy like thinking though the one thing I I wish I could do is I could just have Comet listen to you guys and then tell me exactly what to take because [laughter]

hey we have a business model here now.

Maybe we could talk about this.

Sometimes like the the episode you did on Rolex is like five hours and man I want to learn everything. But

Arvin, we've posted transcripts on our website. I'm absolutely certain that

website. I'm absolutely certain that those are those have been sucked into the reason I like I like I like listening to your stuff is um yeah there there's a lot of the um depth you go

into is really good. So that that that's kind of how I deal with my my own questions like comment perplexity and [laughter] well that is a great place to leave it.

Thank you so much.

Beautiful ask.

[applause] Grit.

Grit. It's It's uh I I really expected Arvin to say like, you know, all the things that always applied still apply, but here's something else that apply.

You know, you have to raise a billion dollars. You have to do Yeah. Same. Same

dollars. You have to do Yeah. Same. Same

as ever.

Yep. same as ever. Um well, for our final segment, we have um the reason we are all here uh and an incredible conversation that we've been looking

forward to for months. So, before we bring Matt out, let's watch a little video on AWS.

What drives us every day is giving you all the freedom to invent. There's a

bunch of launches that AWS is very excited about. We're witnessing an

excited about. We're witnessing an explosion of invention with AI. It is

incredible to feel the energy [music] as you walk through the halls here in Las Vegas. Dive deep into the details and

Vegas. Dive deep into the details and start inventing. And we're just getting

start inventing. And we're just getting started.

Please welcome the CEO of Amazon Web Services, Matt Garmin.

[applause] How you going?

Good to see you. Thanks for being here.

Nice to see you. Yeah.

Well, uh, hope you hope you've been too busy this week.

No, just been hanging out. I was out by the pool.

Ah, thanks.

Let's see. I've seen you on Jim [laughter] Kramer. I've seen you on

[laughter] Kramer. I've seen you on TBPN. Um, no poolside in either of them.

TBPN. Um, no poolside in either of them.

No. No.

How many of these [clears throat] media things have you done so far this week?

Uh, I've done a bunch, but it's awesome.

It's uh it's such a great opportunity.

Uh it's so awesome to have all of our customers here. Uh and and as I think I

customers here. Uh and and as I think I just said in the the the thing um on the video and and hopefully you guys have experienced the the energy and excitement that we have when you get this many people together, not just to learn about AWS, but they learn about

our partners, they learn from each other. Uh it's a cool opportunity then.

other. Uh it's a cool opportunity then.

Um it's been fun to build and this is our 14th year. I remember the first year here, I think we had two or three thousand people and I thought it was crazy that two or three thousand people would get together to talk about our

technology and uh we've grown since then.

When you had told somebody back in the dot era Yeah.

that in 2025 Amazon [laughter] would be convening 60,000 developers in Las Vegas. [laughter]

Yeah. like

uh it's a lot of books or something. Yeah.

What's the I distinctly remember when we launched AWS uh uh we we went to you know I can't even remember what conference it was but like we had one of these like booths right that we were saying like hey we're starting a new

thing called Amazon web services people are like why would I buy computers from a book seller like they were very confused for for a while and we eventually got them over that but yeah

all right so I wanted to start today uh we've been talking a lot about AI I wanted to ask you if you could honestly put yourself in the headsp space of summer 2022 before the chat GPT moment.

And if I had come up to you and said, "Hey, Matt, um, we're going to talk on stage at the 2025 reinvent." Uh, what do you think we'll be talking about? What

would your honest answer have been?

Uh, yeah, it probably wasn't going to be AI and agents necessarily. um you know but uh but I will say it it may have been AI because I think we were at the time even in 2022 like if you go back

another 10 years maybe not but in 2022 we were very excited about the potential of AI and saw where it was going to go.

Do I think it was going to be as it is right now? Of course not. Um and I don't

right now? Of course not. Um and I don't think anyone um would have projected exactly that but um but it wouldn't have surprised me that it was a main focus of what we were doing. We already had SageMaker. We already had a bunch of the

SageMaker. We already had a bunch of the things that we were building. So, um, I wouldn't have been surprised that if it, in fact, I probably would have guessed it was bigger than it would have been in 2022, that it was a bigger focus for people today because we saw some of the

early machine learning benefits, if you will, of of uh, the technology, but nothing like it. You know, I think there's there's some other areas that I would have expected that have actually grown faster that haven't. Um, and

they'll get there, but I would have thought um, robotics would have grown faster than they have today. Actually,

it's gone a little bit slower um, than I would have thought. Um, in fact, I think uh I can't remember what we we actually launched a service that was early. It

was just ahead of its time um called RoboMaker at the time that didn't take off because it was just robotics weren't ready for it. But um but yeah, I think that's that's some of it. But I will say

it's it's you know, it's it's that and I um some of the the challenges that our customers are facing today are remarkably consistent. you know, it's um

remarkably consistent. you know, it's um you know, we we often say today less than 15 to 20% of workloads have actually moved to the cloud from onrem and still still and it look if you if you

looked at uh you know I don't if you saw on Monday we blew up a bunch of servers if you saw that video. Um, but that's all about like getting rid of tech debt and helping people migrate off of mainframes and migrate off of like that

is all the same. Like people back in 2022 definitely wanted to get off Windows and VMware and mainframes and and and they have a lot of them have and there's still a bunch there.

It's funny, you know, one of the principles um every time Ben and I make an episode we try and you know, we're only studying the greatest companies in the world. We try and take you know

the world. We try and take you know something from every company we study and just, you know, bring it to our little operation here. And uh one of the many that we've brought from Amazon and Jeff is what's not going to change and like you know we're going to spend a

lot of time talking about AI and agents and all that but like even if that hadn't happened people are still going to be moving to the cloud.

They are and years from now people still asked me actually I was it's funny some people were saying like hey you know what I have a bunch of these like enterprises and I think there's some analysts were saying this and they're like you know do you think next year

when you come here like everybody will have moved AI? I was like, "No, [laughter] no, like look, only 20% of people have cloud 20 years and everyone hasn't moved to the cloud yet." And for those people whose data is still locked

in a mainframe, like they're definitely not jumping in and getting value out of AI yet.

When is the last year you will be moving people to the cloud?

I don't [laughter] It's not going to be 2026. I'll tell you that. Like it's there is a long runway

that. Like it's there is a long runway ahead of us for sure. H um give us a sense I don't know how if you want to frame it in terms of like revenue or workloads or compute

how much of AWS right now are AI things AI applications versus yeah it's I I actually think that that question in some ways is getting harder and harder to answer over time right I can tell you that AI and bedrock is a

multi-billion dollar business I can tell you like we sell GPUs but like I I do think that that is a um in some ways kind of undersells the impact of it right let's say you have an application

like workday that has like AI assistants in it and agent workflows behind the scenes and and etc. Like is that an AI application? Does all of workday count?

application? Does all of workday count?

Do you just want the GPUs that they're running? Like what counts as AI? And so

running? Like what counts as AI? And so

we see um uh a massive, you know, if everybody's storing stuff in S3, right?

And some of it's vector embeddings, but some of it is referencing their objects and it's driving them to get more of their objects into S3. Is that an AI workload or not? Like I do I think um and actually I introduced this uh in my

keynote last year and I think it's more true than it's ever been that it really is this concept of inference is a building block there there were there are AI applications right arvin talked about perplexity that is an AI

application like pretty much but I think increasingly every application will just have AI in it there's not like a database application there's not like a storage I mean there are fright but there but but everything's just a database with a

crowd wrapper on top [laughter] but but you know like there's these are just components that you go to build interesting customer experiences or value or whatever. And when you think about Chase Payments, like is that a database application? No. It's like an

database application? No. It's like an interesting payments application that does really cool things. Is it going to have AI in it? Yeah. And is it making an AI application? I don't know. And so I I

AI application? I don't know. And so I I think that like today it's definitely a multi-billions um business for us. It's a huge tailwind to our business, but it's not just for

the quote unquote like model running, right? It's all of those pieces. It's

right? It's all of those pieces. It's

people are driving to the cloud faster to get value from the that AI. Um their

applications are getting more usage and so it needs more CPUs. We're using more graviton. All of those kind of pieces.

graviton. All of those kind of pieces.

And so uh there's a question we want to ask you. I think this this is the right time

you. I think this this is the right time to ask it. Um I apologize in advance.

It's not very well baked. So maybe you can help us get We're happy like we can we can workshop it here.

Yeah, we'll workshop it live.

That's right. Um, AI is the first cloudnative paradigm shift, I think.

Like, you can't have AI be an AI company and not be in the cloud. There's just no option. Like, yeah, sure, you could

option. Like, yeah, sure, you could install a model on your, you know, Mac Pro or something, but like you're not going to you're not going to build a company like that. It must be in the cloud.

Uh, how does AWS's like job change in that?

Like when you're working with AI companies, like are you a different or a a bigger part of that company than you are with like a SAS company that's like, "Hey, I'm we're on AWS as our infrastructure."

infrastructure." I don't know. Um I haven't thought about that. Uh I would say in my mind there

that. Uh I would say in my mind there are companies that are all in on AWS.

There are companies who have their like or or the cloud generally, but AWS, they have their kind of towin and then there's ones that are kind of moving the majority of their workloads there. And

there's probably like a a continuum of those, but um you know like I don't know that a AI native company uh is and their relationship with us is materially different than than Gregs and Netflix. I

mean they're they're 100% in on AWS.

They have been for many and and like when they need features like they talk to us like when they need functionality, when they need capability. Um they're

great partners. We learn a ton from them and that's true for AI companies and um and and Netflix and you know could Netflix go and build data centers and run all of that stuff like I guess so probably. But, you know, but but we we

probably. But, you know, but but we we have a great partnership and they haven't needed to.

But it's just like even, you know, when you look at the foundational model companies like Yeah, they need cloud.

Like they're not building their own, you know, like they need it.

Yeah. It's Yeah. I mean, they they uh I do think that there's a I don't know if they need it, by the way. Like I I think some of these large companies um um OpenAI has talked about going and building Stargate or whatever it is.

They can build data centers. Um now, you know, are they going to run them as efficiency? Is it a good use of their

efficiency? Is it a good use of their capital? Like all of those things are

capital? Like all of those things are all good questions that um I think lots of people ask and and um uh and and those are open questions but it's possible right there's nothing um that's impossible for them to do but I don't

know why they would right and they get a lot of benefits from the cloud they get a lot of benefits from the customers also being in the cloud right and that that um uh is uh really powerful I think particularly when you move away from um

consumer applications for consumer applications like you get a lot of benefits of scale you get a lot of benefits of the technology pieces when you get to enterprise applications You also get benefits because your customers are in that same cloud and you can

deploy those models into your VPC. You

can be in a secure environment and there is a lot of really great flywheel things that happen there where those enterprises trust AWS. We trust they trust us with their data security. They

trust us with a bunch of those things that we've built up over the last 20 years where then uh a new startup that you know has a new model that's exciting can put it on bedrock and just get some of that trust level which is uh which is cool for because it takes a while to

build that and access the enterprise.

That's right.

Yeah. Totally.

So, David asked a question earlier that I want to reframe for you. So, I used to be a web developer. I I haven't written code. I actually me and Claude wrote

code. I actually me and Claude wrote some code earlier this week, but I haven't [laughter] like sat down in a text editor and myself written you actually built it.

The good news is not many people are going to do that going forward.

That's true. [laughter]

Acquired has never had a um like piece of software really and you we now have like a software uh you know a cloud generated software pile of scripts that I run to do. Good.

Um, I'm not going to ask you where you deployed them, but I hope it was on.

Uh, they're not deployed.

Okay.

We'll talk afterwards. [laughter] We'll

get you we'll get you to account manager. Okay.

manager. Okay.

Uh, the um it used to be five years ago extremely obvious if you said, "How do you build a web app?" Like, it was a pretty settled

web app?" Like, it was a pretty settled frontier. There's very off-the-shelf

frontier. There's very off-the-shelf components. Everybody's sort of using

components. Everybody's sort of using the same paradigms. There's languages, there's frameworks. There's S3, and

there's frameworks. There's S3, and there's EC2. And David came up to me

there's EC2. And David came up to me earlier today. He's like, "Do you know

earlier today. He's like, "Do you know how to build an AI app or like an AI company?" I was like, "I don't actually

company?" I was like, "I don't actually know what the core building blocks are."

I mean, I know there's a these models.

Yeah.

Do you feel like all the cards just got thrown up in the air or is it actually It's uh I mean, yes and no. Like I I would say that there's a couple of pieces that I would say to that. I think

on the first part like uh I would challenge you to build like an interesting AI app without using like some sort of computed storage and databases. Like you're probably not,

databases. Like you're probably not, right? It turns out that all of those

right? It turns out that all of those things still need that because turns out that like you know when you have like the front end of a chatbot like CPUs are the most uh effective way of rendering that web page. Like it is that is true.

You're going to need a bunch of storage.

So so all of those components are the same and and and similar. And um I remember actually and I'll just like as an aside I remember when mobile first started getting big and people were like

we need to build mobile building blocks and then I was like are there different servers that you need for a mobile app?

And it turned out no actually it was like you know there's there's some like front-end development tools and things that were different but like not really like you still needed comput and storage and databases and now now in AI is different in that I do think and this

goes back to what I was just saying earlier I think inference is one of those new building blocks that was never a developer building block before. So

it's like the world invented a new Lego.

You had all your Legos and you could build whatever and they've now invented some new Lego that was never possible before and you're like woo that is a cool new Lego. I can build incredible things with that. I still need all the other Legos I had before, but now I can

take this one and build some super interesting things with it. Um, and

agents are kind of an extension of that where now you're going to be able to build completely new experiences. You're

you're still going to like it turns out like Lambda is actually a great execution engine for a lot of of agents workflows where you can call the model and do stuff, but you actually need a compute environment to go do that. Um,

and so that I think is how you would think about that. I think by the way though, there's a lot of things that do change, right? how you go to market

change, right? how you go to market might change. How you think about how

might change. How you think about how you um deliver a a user experience is going to change. I actually think by the way like and and this is one of the things that a lot of as I talk to a bunch of customers out there um org

structures are going to change like if you're starting a company I don't know that it's the exact org structure that we've used for the last 20 years to organize a development team like a lot of those things may actually change.

What what's different?

Um I think you can you can deliver with much smaller teams. So um you know a a huge effort that used to take dozens if not hundreds of people may take 10 people or five people or something like

that and you can deliver just as much and maybe faster.

We we were in an event with um uh the CF cursor a couple months ago and it's like it's a tiny team. They may have 100 people now. Maybe, maybe, maybe not

people now. Maybe, maybe, maybe not even.

I heard they did pass a 100. So, I think they passed 100. But yeah, like you just need tiny teams and even there like and you know, as we're finding it, you know, the the roles may change like you may not have like a traditional dev manager.

You may have actually a dev lead because the dev lead can actually dive in and actually help design some of the things you the whole org structure may just just change. I don't actually don't know

just change. I don't actually don't know the answers to that, but but with those capabilities and how you think about that software development life cycle, um I think a lot of those things will change. I mean, I got to imagine

change. I mean, I got to imagine there is no more exciting thing for Amazon and AWS than the invention of a new Lego, like a new Lego block. Yeah.

Like that's like a I mean, this happens every 20 or 30 years and like this is the the the biggest needle mover that could happen for your business.

Yeah. Well, it's it's both exciting for a business perspective, but I will tell you like it's exciting because it means you get to go build new things. Like

it's not just like polishing the old thing. Like part of why everybody like

thing. Like part of why everybody like why I think reinvent is so um popular with a bunch of people is that when we first launched it, we didn't build we didn't launch this as a marketing conference uh as like a traditional kind

of like and and those are popular and lots of people have those marketing conferences, but if you go and you you look in the hallways out there, right, it's like people on their laptops, they're out there doing work sessions.

They actually are doing hackathons. like

we're doing like this is meant to be a learning conference where people can learn how to new use new technologies and most of the sessions here are not given by Amazon people they're given by customers who want to share what they've learned and when there's a new building

block like developers eyes light up they're like woo what can I build with that and uh and so it's it's for me it's exciting for that like it is a great business opportunity don't don't get me wrong but like that ability to go build

some really cool new things like perplexity that was just that wasn't a possible thing that you would have built in 2022 and now it is it's pretty cool Yeah. Uh so you're one of the most

Yeah. Uh so you're one of the most important platforms in the world. The

job to be done of a platform is to build a set of capabilities that other people can build things on top of.

Yeah.

Sometimes that means building your own either demonstrations of those capabilities or just extending the layer a little bit and doing some of the applications yourself.

Yeah.

And some of the agents you announced earlier this week are an example of that where you sort of have a first or best-in-class version of it.

philosophically, how do you decide when should we do this and when should we leave it to customers?

Yeah, it's a it's a good question. And

by the way, there is no magic answer.

It's not like there's like a very clear cut and dry answer for what that is. Um

uh we always think that our um ecosystem of partners is an incredibly important and impactful thing for growing our business and for helping our customers.

There was a cool stat I saw earlier this week that partners usually drive about $7 of revenue for every dollar of revenue they drive for AWS. Um, just

like that's pretty awesome for them and it's like a huge opportunity for everyone. Um, and we think about how we

everyone. Um, and we think about how we deliver value up and down that entire level of the stack. And if you think about it's it is a funnel of the like number of opportunities, right? Because

if you think about that top layer that's driving $7 for every like every dollar that is just a massive massive business opportunity. And so there are thousands

opportunity. And so there are thousands or tens of thousands or hundreds of thousands or millions of applications at that level. Um we think we'll have a

that level. Um we think we'll have a couple of successful ones up there and and we do today, right? And we will continue to have a handful of those and and you mentioned we launched things like our frontier agents in the software

development space at that top layer, you know, and then and we think about where can we add differentiated value, but it is a funnel like you go down to the bottom there's cloud, right? like and

then you go up to this next layer and you we'll take uh whether it's models or databases or other kind of pieces or or analytics engines and things like that.

um where we do, you know, compete with some of our customers, but but in a very like positive way for both of us where if you land on MongoDB or you land on

Snowflake or you land on data bricks or or any of the the other partners at that level, um great as long as it's running on AWS like we love that tales I win.

Well, and we have and we but we have a lot of products in that space too, right? We have tons of customers who

right? We have tons of customers who love using open search. We have tons of customers who love using um Aurora and and and all of our stuff at that level and we compete at that space, right? and

but but there's a real kind of partnership there where everybody is growing huge really successful businesses at that level and there's dozens maybe hundreds of companies at that level and then you go to this top

level um and that's where there's there's millions and so that's where at that top level in particular you know I think we're going to build stuff at these first two levels like for most of those things and we'll compete with

people but there's not that many things at that level that just doesn't make sense for us to not have a native solution for and customers re always tell us this right customers are going to demand it they largely say like I want you to deliver something of that.

It's a good thing you have red shift.

Yeah. Um you know I think if you go to this top layer there's lots of things that I mean there's no way that we could possibly build all of them and so we build some where we think we have a

particular either a competitive advantage or there's something at Amazon or AWS where we have an expertise andor a different way of approaching the problem that we think delivers differentiated results. I think actually

differentiated results. I think actually one of the things that Amazon is particularly bad at is being a fast follower. Like we just when we try to

follower. Like we just when we try to copy someone, we're just bad at it. Like

and so we just don't. But if we can innovate and we can go build something different infamous Amazon products from history, but I'll I'll [laughter] we're just Yeah, we're we're not great

at it because we we actually when we should fast follow, we still have this inherent thing like we should do something different and then we usually mess it up. Um so we just don't we're just bad at that thing. But when we think first principles about solving a

customer problem and we can do it differently, take these kind of frontier agents where like we we know developers like we really know what they want. We

understand AI. We see some of the struggles that they're having with the first generation of AI development tools where kind of vibe coding is is good, but gets you to a place where it doesn't it's not structured. It doesn't help you

develop in an enterprise environment. It

it doesn't allow you to really have agents go and be autonomous and drive fast and and you know, we saw some of these observations. We like we think

these observations. We like we think that we have some ways to differentiate um and deliver that. And we have same thing about cloud operations, right?

We're thinking about a DevOps agent and that could go and autonomously go and and help you find problems in your infrastructure. We have a lot of data

infrastructure. We have a lot of data there. We have a lot of knowledge about

there. We have a lot of knowledge about what best practices are, how we can help you do that. Same with security. We we

we think we do some of the best security work in the world. We see more problems than anybody does. We have solutions across them. we think we can deliver a

across them. we think we can deliver a differentiated product in those spaces.

Um and uh and so that's when we go after it and we're not always right. We don't

have a 100% batting average, but that's kind of how we think about it.

I I like your three stages of the funnel. Um what stage do you think

funnel. Um what stage do you think agents are? I mean, they're not they're

agents are? I mean, they're not they're not the bottom. They're not the cloud.

Are they the millions or they I think agents uh so if you if you think about that um they're definitely not the bottom. I think uh agent core is that

bottom. I think uh agent core is that middle layer, right? And so the agent core is these building blocks where it's like you have a secure compute environment that you can run your agent in that gives you like an isolated

compute environment. Um you can do um

compute environment. Um you can do um gateways where it's like okay you have permissions and so your agents can go talk to other agents that can talk to tools that have fine grain permissions etc. So all these kind of building blocks to go build the agents on we kind

of think is somewhere in that that middle layer. still builds on our core

middle layer. still builds on our core building blocks, but it it really makes it much easier um and more secure to build like a wide range of of agents in a in a real enterprise world where you can have an audit log, you can have

controls, you can do a bunch of the things you want to do. Um and then the actual agents that people build are probably more at that application layer.

But I don't know, it's new. Who knows?

It might answer may change in a month.

You know, that was that was our next question. Uh

we're back here a year from now.

Yeah. What is the conversation about agents in 12 months?

Uh, you know, look, I think, uh, just like any technology, I think that there's still going to be a bunch of people trying to figure out how to how to work them into their environment. And

I think, um, almost for sure over the next year, you're going to see a ton of people getting a huge amount of value from agents. I do think that it is um

from agents. I do think that it is um when you kind of look at that value creation where um you think about content summarization, you think about content creation as opposed to like

actually going and accomplishing tasks like the the value to an enterprise is so much bigger for that second thing because it's it really explodes out to like not just like summarize what happened but actually go you know

process my insurance claims like that is a totally different um unlock from for value and um and I think we'll start to see some really, really, really

compelling um use cases that frankly I'm going to be really bad about predicting right now. But that's part of what I

right now. But that's part of what I love about AWS. It's actually one of the things we've always loved is we build these building blocks, we give this technology, and then we let the world have the creativity to go find some

really cool things for them. And I think next year we'll have some really awesome uh use cases to highlight. That's

it's it's interesting to think about the agents as a paradigm because you flash back 30 years ago, the web happens, we have websites, then mobile happens, we have apps.

AI happens and in 2022, we get this like these bare models with like a chat interface slapped on top.

Couple more years go by and now we have this notion of agents. What do you think the dominant paradigm is in the AI era?

Is is it agents? Is that going to be the sort of app of AI?

That's not really how I think about it, but uh that's an interesting question. I

think that's slightly different. Like

I'm not sure that those are uh equivalent paradigms necessarily.

They're like somewhat orthogonal because you have like like part of what was interesting about the web was it was like a different platform to actually go interact with and build stuff on top of and and there was the user interface was

different and like you could and so that was like you had to build a website, right? um when you had your phone like

right? um when you had your phone like you had to like how do you shrink it down? How do you actually have an app

down? How do you actually have an app that's like that's native and has capabilities but like the the application and and and a lot of the things you could do were mobile and things. So I think with agents it's more

things. So I think with agents it's more like a capability by the way someone may invent a new way of consuming those things which would be interesting and and uh I think would be some of those

shifts today. I think there in some ways

shifts today. I think there in some ways it's a complimentary technology like the chatbot that you example turns out that's a website. It's not like it's not like it's not a website, right? The

website was actually something different. It wasn't an application that

different. It wasn't an application that was on your computer. It was a website.

The mobile app was a different thing. It

was actually on your phone. These are

things that are some ways complimentary and supercharging the things that you already have. Doesn't mean someone won't

already have. Doesn't mean someone won't come up with it yet, but it and it may or may not, but in some ways it's like I assume that my mobile apps will have agents that are working for me behind the scenes. I assume that my websites

the scenes. I assume that my websites will have AI and agents built into them.

Like all of those kind of things. So, I

don't know. Someone may build something different, but they haven't yet.

Agent management still feels really messy to me. Like

that's true. I believe there will be some interesting applications around agent management for sure. Um, and that is a very unsolved problem. I think it's an interesting one that someone out there should go build um uh or think about

to put more structure around it. I mean

uh if I pull up a website, it's very obvious what the website is. If I pull up my phone, I I've got a UI that's my mobile app. And if I have agents acting

mobile app. And if I have agents acting on my behalf all throughout the world with this new compute paradigm, I don't know how to track or manage them. Yeah,

I I think that that is an interesting um but but uh at the same time um you know you have uh when you click the button on a website there's like a database query that's going and doing

work for you behind the scenes. You

don't know how to do that either. So

like maybe you don't have to I don't know like maybe that's the job of everybody else to kind of figure not everyone like the individual like whoever has the website it's their job to make sure that like your database query actually works. Sometimes you get

a crashed website and it didn't but uh um but yes otherwise that's [laughter] So, so maybe right, but I do think that there's an interesting um if there really are a bunch of these autonomous agents, which I do believe there going

to be, there is going to need to be some way of coordinating across them cuz in a world of billions of agents, like no one can like keep billions of things in their heads. And so, but in the same way

their heads. And so, but in the same way that there's billions of compute processes, like we have to come up with a good kind of router and and orchestration engine to to manage all of those. No person's going to manage

those. No person's going to manage those, of course.

For our next set of questions, we want to keep going with AI and it'll be related, but um we really want to ask about you know you too like you're in a very um very interesting position and and start with

with AI and transition here. You spend a lot of time with a lot of CEOs.

I'm sure they are asking your thought.

What are CEOs asking you and asking AWS about AI when they're coming to you with problems? Yeah, I think there's a

problems? Yeah, I think there's a couple. Um, uh, I do get fortunate, uh,

couple. Um, uh, I do get fortunate, uh, that that, um, you know, I think it's it's, uh, it's fortunate for us that cloud and AI has elevated to it is

actually like a CEO level thing, right?

Technology used to be something that CEOs question CEOs and CTOs and like CTO.

Sure. Yeah. But I bet CEOs are the ones CEOs, it's it is a CEO level question now, right? What is your cloud strategy?

now, right? What is your cloud strategy?

What's your AI strategy? Like those are CEO level questions, which by the way is great and cool for us. um and what they're thinking about um it's it's not too dissimilar than some of the things that we're talking about. I think one um

they're thinking about my they they see like they see kind of the trend and I think particularly if you go with established it's there's a little bit different right if you're the CEO of a AI startup company it's a little bit different right so but if you

take that if you're saying like enterprise side of like someone who established business um they are trying to balance the fact that they they realize there is a real risk that they

could get disrupted with this technology like there is a very real risk that 100-year-old enterprises like Chase could get disrupted if they don't lean forward and uh and figure out how to

stay ahead of this technology. Now, like

they still have the pole position, don't get me wrong. Like they're like they like if if you're a bank if you're a bank, you would definitely rather be JP Morgan than like a brand new startup, but the risk is there, right? And so

that risk is very real. And so I think a lot of people are thinking about um how do they but a lot of businesses out there, frankly, are not used to moving fast, right? they are they're they

fast, right? they are they're they actually got to where they are by being conservative, by making sure they didn't make big mistakes, right? And that was like and and so I get that if you're an

insurance company that's a 100-y old insurance company, you got to where you are by making sure you didn't like make dumb bets or or do something that we we just covered Coca-Cola as our most recent episode and um you know, I mean

Coca-Cola did amazing innovative things and then like you know the worst thing they ever did which turned into the best thing was New Coke. you know, you don't do new coke.

Coke. you know, you don't do new coke.

That's right. You just don't want to do you don't want to be the one that did New Coke. But they're also recognizing

New Coke. But they're also recognizing that like if they stick on that that path and they're still slow and conservative, like this may be the thing that kills them. So, I think a lot of people are thinking about how do they

make sure that their team has the right skills to be competitive in the in the world today with AI? How do they make sure that their data and that their infrastructure is set up in the right way? I think they're thinking about do

way? I think they're thinking about do they need to hire or train for new skill sets. They're thinking about do they

sets. They're thinking about do they need to organize differently so that they can have that they can move fast.

And you know you you mentioned earlier kind of thinking about like Jeff said like what are the things that don't change. I actually think that in today's

change. I actually think that in today's world one of the things you want to be thinking about which won't matter what the next technology is is how do you make sure you can move fast and be agile. Like those are skills that are

agile. Like those are skills that are not going to go out of style. Right.

There's never you're not going to get to a world where it's like man I wish we were less agile. Right. And [laughter]

particularly in today's world like that is man I wish stuff were more expensive on you. Exactly. It was not that kind.

you. Exactly. It was not that kind.

Right. Exactly. Never going to happen.

And so, you know, I do think that that's where a lot of the leaders are thinking about right now because they're thinking about look as a as a CEO, there's not that many things you can impact. You can

impact who your leadership team is. You

can act, you know, where the direction you're going is, what the structure of it is, kind of what some of the goals are. And so, that's what a lot of people

are. And so, that's what a lot of people are thinking about.

Can I ask the opposite question? So when

you go and you talk with your mentors and other CEOs and or members who who give you advice, what are the questions that you're asking them?

Yeah. Uh that's a good question too. Um

I do a lot of that because I'm only like and you know I think it's like you're two or something of me being a CEO. So

I'm I'm still learning. Um part of that is uh and I think everybody actually any any good CEO is hopefully or any good leader is hopefully spending a lot of their time learning. Um but uh you know

there's a lot that I don't know about. I

I love just picking up tips on how do you think about leading at scale? How do

you think about um rapid decision-m? How

do you think about pushing decision-m down to your teams and um and I think that's one of the ways that you do real rapid decision-m is don't make decisions and get your team enabled to make decisions. Um and then but but picking

decisions. Um and then but but picking up like how to do that and and what are the subtle things that you do, what are the areas that you don't do. And by the way, sometimes you get advice and you're like interesting, I'm not going to do that. Um, you're [laughter] like, you

that. Um, you're [laughter] like, you know, like I like I I I sat down with Jensen and he gave me some good advice and he told me what he did. I'm like,

I'm not going to have 70 directs. Like,

[laughter] it's not going to do that.

Like, okay, I appreciate that that works for you. I'm not going to do that. But

for you. I'm not going to do that. But

um you know but uh uh you know I think that is uh but but assimilating all that and then trying to figure out what works for you and your culture and your business and but it's incredibly valuable and um and I appreciate every

time uh a CEO or leader is willing to take some time and share what works for them and and have that conversation is just in so valuable because kind of taking assimilating all of those different data points is is incredibly

helpful because these jobs are hard and they're fast moving and and we all want to do well for our company and our customers. customers and our

customers. customers and our shareholders and um but you know trying to pick up tips is helpful for sure.

there. Um, you know, I don't know how many folks in the audience know this.

Um, we know this. Uh, you started as a dev manager intern intern for EC2, right?

Yeah. I was actually AWS I was I interned out of business school in 2005 uh for Andy Jasse and uh when AWS was an internal project, we were building it internally uh and I

came and and I worked on it. I was like, "Holy crap, this seems awesome."

[laughter] And uh and I came back basically effectively as the first product manager for AWS.

Wow.

And EC2 was the first product you worked on.

Uh that was actually a year later. So

the first year I was at AWS, I was a product manager for all of AWS cuz it was [laughter] the product. So I

was the AWS product. Actually, it's not true. There was one product.

true. There was one product.

I guess you're still the AWS product manager.

There was there was a product manager for S3 and then I was the rest of AWS.

Okay. This this nullifies my premise then. But [laughter]

then. But [laughter] Okay. Okay. Okay, audience where we were

Okay. Okay. Okay, audience where we were going with this is uh Dave and I met Matt a few years ago when you were leading sales go to market for AWS.

That was a different story. Like it was crazy how I ended up there too cuz I' been it had been 13 years that I'd been leading EC2 at the time and then one day Andy came to me and said hey do you want to lead sales and marketing? I literally

looked at him I said I don't know anything about sales and marketing.

Well that was that was the question.

[laughter] like you very clearly came out of a product and technology path and I was like this is the strangest person to be leading sales and then it kind of occurred to me this is deliciously

Amazonian like of course you're the person he was he literally and I I you know it was it was interesting and I said uh I don't know about sales or marketing he's like it's okay he's like you know a lot about AWS I was like yeah

I do that's true he's like and you you can operate at scale and I was like well I've hope I've hopefully learned to be able to operate at scale he's like you'll pick up the rest of it'll be fine uh and And the teams were awesome. I

will say like it wouldn't I don't know that it would worked at every company, but the teams were awesome. And I I'd sat down with some of my sales leaders and be like, "Can you please explain how commissions work to me?" And uh they

could have instantly flipped a bit and been like, "Nope, like this guy's name."

[laughter] Um but no, they're like they're like, "Thank you for asking."

Like great. We'd love like first of all, they did they did have like a a bit of a double take when I asked those questions, but then they walk it through it and then I go like do a bunch of research and make sure I understand exactly how it works or whatever. and

and you know, I'd find out that um I feel like half of the time I would ask dumb questions and they'd be like, "That's just a dumb question." But the other half the time they'd be like, "Actually, no one's ever asked that question before, but it's a really good question. I don't know why it works that

question. I don't know why it works that way." Like, "It doesn't have to work

way." Like, "It doesn't have to work that way." And and we would go find

that way." And and we would go find something interesting because I didn't actually know how it worked. Um uh and just listen to to what people said. So,

and then you sort of get to build it from first principles, too, rather than it's the way it's done.

That's right. How how did the sales or start within AWS? I mean Amazon Amazon is not a salesdriven company.

Uh yeah, it was um we uh we're a different kind of sales. Let's

put it. Oh yeah, we um we had call it when I started I think we when I started back full-time we had hired I can't remember if it was two I think two

business development people uh and um our goal was to like uh our stated goal at the time we had a lot of hubris back in the day of like what was possible or not but we assumed that AWS would always

just be self-service and that we would probably never need more than like 10 20 30 salespeople um uh and then it turned out that like enterprises just want to talk to a human. Um they they didn't

want to just swipe their credit card and just use compute and uh and and as we got into it and like we learned about what the business could be and uh and as we got into it um we saw you know we

built a different type of a sales team by the way. We wanted one that was incredibly customer focused. We wanted

we we hired people and said look you've got to be customer focused. We want you to go to our customers and tell them we want you we want you to proactively lower our customers bills, right? We

want you to go find ways that they're using AWS wastefully and cut their bills down. And uh

down. And uh this is not part of the enterprise software and Exactly. And people were like our

and Exactly. And people were like our sales people were like well that's kind of refreshing like okay awesome like I don't have to like hide that under the covers. And customers were like just

covers. And customers were like just floored that we would possibly think like as a as a technology company come and help them lower their bill. And

we've been doing it for 20 years now.

And um and we continue to do it. And

we're not always perfect at it. we find

other ways to to do it. But um but we're launching products, we're launching features, we're launching capabilities to try to help people because every time we saw where we would lower their bill, they would use more. It'd be like they free up budget to go launch a new

product which would drive their revenue, but then they would drive more. And so

it was uh it was a real positive flywheel effect for us. But it's a very different type of a of a um a sales support organization than had existed in the past.

All right. So now going back to this trajectory you've been on this all these different roles you've you've had you get in a time machine you go back 20 years you get to give yourself some advice

about the future save yourself some missteps or some frustration or what are you imparting on yourself?

Um that's a good question. Uh

well there's a ton of things that I would do differently but um but I'm also quite happy with where I am so like I wouldn't want to mess those things up.

There's a whole path dependency question.

Yeah exactly. I I wouldn't like like a sliding door situation where I somehow mess it up and but um but uh uh cuz I do like it's like and and I think all of like even when we have like the worst

times actually and and and I'll use this like like yes like what I love um if you go to the early days of AWS we had some like major outages where we like took down the internet for like 4 days.

Wasn't really that bad but but it like felt that way internally that that that's what we were doing. like we we take it so personally cuz customers trust us with their their operations and their data and um and uh would I have

loved to like avoided that entirely like yeah except we learned a ton of really valuable things for that and our services today are way better because we learned some of those lessons I wish we would have found cheaper less painful

ways to learn those lessons for sure but from a a business perspective um you know there's there's so you know I I kind of I it's a fun question to ask and

like it's really hard Because a lot of the bad things and bad choices you made or even bad hires that you made are are such good learning opportunities that if you learn from them, you have to learn from them by the way. So like you got to make sure you don't like make that

mistake again. But I don't know that you

mistake again. But I don't know that you could be in the same place or the products could be in the same place or the people could be in the same place if you didn't get to have those crappy.

I mean that's always the honest answer.

It's like we get asked that question sometimes and like we made some really dumb things. We changed the name of the

dumb things. We changed the name of the show at one point in time. Like we

Yeah, we interrupted the compounding unnecessarily. We did some really dumb

unnecessarily. We did some really dumb things, but like yeah, would I take But maybe you don't end up with like a pretty fantastic show. Like who knows?

Like you take the wrong path and um so it's hard to say, right?

Particularly when you end up in a pretty good spot. If you end up in a spot that

good spot. If you end up in a spot that like you're not happy with um and there's lots of things that I would like to be better by the way. So I think you're um you know one of the the things that's another Amazon kind of um

mentality is is thinking about like our c one of the things we love is our c customers are are perpetually displeased like they always want more. They always

want you to do things differently which is great. It means there's huge

is great. It means there's huge opportunity to always get better and we take that really seriously. Like we are never pleased with our performance. We

never think that we're doing um we're never growing fast enough. Our

operational performance anything below 100% is not good enough. Um, we always want to drive cost down and performance up and um, you know, and all that being said, I think um, you know, we've we've

built a a good business and and a pretty exciting um, customer community here.

All right, my favorite closing question recently.

What's something you've changed your mind on that you used to be really sure of and you're now very sure of the opposite?

Uh, I have two. Number one, uh, uh, back in 2013, I was really sure that the Seahawks were going to be a dynasty.

[laughter] I thought you might be this was going to be part of the That did not work out. Uh, but we're back. You know, we have a we have a good

back. You know, we have a we have a good team again. So, uh, I know we both live

team again. So, uh, I know we both live in Seattle.

Well, thanks to the San Francisco 49ers quarterback rehabilitation program. So,

you're welcome. Thank you. Uh but but in a more serious answer um I will say like um I remember I'll get the timing right but it was wrong but it was probably six or seven years ago. We had a leadership meeting across Amazon and I will tell

you one of the things we I was I was absolutely most worried about for the future of AWS and for Amazon was we were charting a path on where we thought the business could be and the growth and the

number of SDEs that we thought it was going to take to get there. And we

thought the world would not have enough SDEs for us. We were literally like, I'm going to have to hire a million SDS to deliver the code at the rate that we're doing and where we want to get to and whatever. And

we're like, the the world is not producing SDS at that rate. And I

thought that was going to be our single biggest constraint um to growth. And uh

I do not think that today. We still have we still have the need, by the way, for great SDS, but I actually think the growth in the AI world um is going to be ideas. Before it was like we had way

ideas. Before it was like we had way more ideas than we could possibly get to. And um and I think because you can

to. And um and I think because you can deliver things so fast, you're constrained. It's going to be great

constrained. It's going to be great ideas um and great things that you want to go after. And I would never have guessed that um 10 years ago.

Fascinating.

Well, that's where we're going to leave it.

Awesome.

Cool. All right. Thank you guys. Really

appreciate you doing this for us. So,

thank you.

Oh, our pleasure.

Thank you. Great to see you. Thanks a

lot. [applause]

[applause] Well, well, that about does it. We want to thank all of you so much for coming and hanging out with us today.

Giving us a couple hours during reinvent, which I know is precious, precious time. Thank you. Big thank you

precious time. Thank you. Big thank you to all of our guests, to Max, to Greg, to Arvin, and especially to Matt, our guest and host. Um, thank you to Hannah and Nick and Sherry and Michael and

Orion and the JP Morgan team and the Amazon teams for months. I mean, there was months of work that went into making this happen. Uh, would not have happened

this happen. Uh, would not have happened without all of them.

And, uh, for those of you who are joining uh, the meetup in an hour at 2 p.m., uh, we can't wait to hang out and

p.m., uh, we can't wait to hang out and meet as many as many acquired listeners as we can.

Yes. Yes. And finally, huge huge thank you just to all of AWS and all of JP Morgan. as we've said, they've been

Morgan. as we've said, they've been incredible partners to us in our journey over the last several years, and I hope to many of you, too. Um, we hope you all have a great rest of reinvent, a great

holidays ahead, and we'll see you next time.

[applause] [music] Who got the truth?

Is it? Is it? [singing] See you.