Amazon's Scale, Infrastructure and AI Future | WSJ

Venture Vibes
28 followers
#Amazon #AI #WSJ
David Zapolsky has witnessed Amazon’s growth since his arrival at the company in 1999. The chief global affairs and legal officer discusses AI data centers, nuclear energy and the effort to balance speed, safety and global regulation.
AI Generated·May contain errors
Video Summary
AI GeneratedDavid, a long-time Amazon executive, discusses the company's evolution from a small operation to a global giant. He details Amazon's strategic shift toward sustainable data center expansion, the integration of nuclear power to support AI demands, the development of a versatile AI ecosystem via AWS, and the deployment of a global satellite internet constellation to bridge the digital divide.
First morning of the show, and we got our,
what, the second slot?
Exciting, exciting stuff.
It's good to see all you here.
And I'm really grateful to have David here
in conversation with us.
There's a lot to talk about when it comes to Amazon
and global affairs, I would say.
Yeah, first of all, thanks for having me.
Great to be here.
And it's first thing in the morning,
so it may take some time to get warmed up.
But yeah, there's a lot going on.
Well, if we wanna warm up, we can start way back in 1999.
So you, one of the reasons I was excited
to talk to you today is that you've been at Amazon
from when it had about 7,500 or 7,600 employees.
Actually, like 2,500.
Oh, really?
Yeah, yeah.
Okay, amazing.
2,500 employees, and now you have how many?
One and a half million-ish.
Okay.
It's been an amazing story and an amazing transformation.
And, you know, having been there
since we were just a little bit more than a bookstore,
I joined in late 99,
so we were already into music and video,
and we'd just launched tools.
Like, that was the big thing, tools.
We had electronics and toys.
So, like, tools, like power screwdrivers.
Yeah, like power tools.
Yeah, it was, that was a pet project
of our then-short-lived president who we had,
who was with us at the time, who was a tools expert.
All right, so you got your 1999 self
and you got your 2026 self.
What does your 2026 self desperately wanna tell
that new guy, that new lawyer, fresh out of the firm
who's showing up and meeting this Jeff Bezos guy?
So, I wouldn't give advice to myself about Jeff
because, you know, I got to meet Jeff very early on.
The first weekend I was working at Amazon,
we were in the middle of the,
I started as the first and only
litigation and regulatory attorney at the time.
And so, we were in the middle
of the infamous one-click fight with Barnes & Noble.
And we'd sued Barnes & Noble over a one-click patent,
which was widely derided as being not patentable.
But we had a patent and we got a preliminary injunction
against Barnes & Noble for sort of copying the features
that Amazon was pioneering.
And so, Jeff was a witness in that.
And so, my first weekend there,
I was in Jeff's conference room.
And I remember, you know, prepping him for his deposition
and walking into the deposition room
and he looked at me and said, where'd you come from?
And, you know, and so I had the opportunity
to work with him and get to know him very early on.
So, I wouldn't give myself advice on working with Jeff.
He's an inspiration to work with
and I've enjoyed my relationship with him.
What I would say is the company's gonna get a lot bigger
than you think it is.
I remember seeing a senior leadership PowerPoint.
That's how old, like we still had PowerPoints at Amazon.
We haven't had them since 2003.
But like there was like a 2000 era PowerPoint
that was on my boss's desk and it said something like,
you know, the road to five billion.
I just thought, that's never gonna happen, that's crazy.
That's just bananas.
Valuation company?
No, no, like revenue.
Oh, revenue, revenue, yeah, yeah.
And we were at about 780, 800 million at the time.
I'm like, five billion, that's crazy.
So, you know, having just crossed, you know,
600, 700 billion last year, just keep thinking big.
Like the leadership principle about think big, it matters.
And you really have to think big
in order to stay ahead of where the company's going.
And the other thing I would have said
would be don't try to predict the future because you can't.
And I stopped doing that, you know,
three or four years into my Amazon tenure.
I just, people would say, what's gonna happen next?
Like I have no idea.
Like I won't predict the future
for more than like three to six months
because you just can't tell
how different the company's gonna be.
You know, when I took over as general counsel in 2012,
you know, I wouldn't have thought we'd have to build
an entire legal group in LA, in Culver City
to support, you know, a major entertainment business.
Nor would I have expected that I'd have to like
create a group that, you know,
that is expert in satellite law and space law
or autonomous vehicles or, you know,
we were already in the cloud at that point,
but no idea how big that would become, so.
So let's talk about some of the other things
you've had to grapple with over the period.
We're all thinking nowadays about data centers, right?
I mean, this is something we,
like many of you are probably wondering
when's he gonna ask about data centers?
Well, right now.
The fact is, you know, that it's not only something
that's suddenly, you know, occurring to many people,
like what's going on in my backyard?
It's something you guys have been building for decades.
And so I wanna ask you as, you know,
an experienced builder of data centers
and other giant buildings that no one really
get to see the inside of, what's that process?
Like you show up and you go,
that field is gonna be a giant building.
Well, I think I'll start by saying,
I think the industry as a whole can do a lot better
in talking about what data centers are,
what the value they can bring to communities
where they're cited.
And I think that whole data center industry
kind of grew up in a culture of secrecy.
And so we're having to really push to-
Why is that?
I think in part because it's super important
to keep them secure.
And so early on, you know,
if nobody knows where your data centers are,
they're gonna be more secure.
And so I think there's a mentality of like,
we don't really want to be very public about them
because they could be vulnerable to attack,
to, you know, sabotage.
So I think that's part of it.
But I think we certainly have completely pivoted
away from that because we realize
we're gonna have to be much more transparent
and we are at this point.
One of the problems in operating in secrecy
is that if there are narratives that get out there
that are incorrect, it's harder to set people straight.
And so there's a lot of incorrect narratives
out there about water.
Most people don't realize that,
at least the data centers that we use,
you know, they don't use water to cool
like over 90% of the time.
If you really care about water,
you should shut down your golf courses
because there are many, many, many times
the usage of water than data centers.
In many cases, we don't use water at all to cool.
It really depends on the climate
and where the data centers are.
And where we do use water, we use reclaimed water,
non-potable water, and Amazon has gone further
and made a water-positive pledge to be.
2030.
By 2030 to return more clean water
to communities where we have data centers than we use.
We're 75% of the way there.
And, you know, that means, in some places,
that means we're not using water at all.
In a lot of places, that means we're investing
in water reclamation efforts that's community-led.
We've done that in Virginia and in eastern Oregon.
The issue, I think, that kind of clouds that
is that there are power centers nearby
that are using an awful lot of water,
and that is not always part of the tally
that you guys, because you don't own the power plant.
Well, in some cases, we, you know,
we're the major consumer of power plant.
We have over 700 wind and solar facilities
around the world, and over 28 countries alone.
We generate over 40 gigawatts of renewable energy.
So we have some input into that.
We, you know, we offset, as of a few years ago,
we offset all of our operations with renewable energy.
And so, you know, we're working cooperatively
with the utilities that, you know,
that are in the communities where our centers are
to make sure that the water impact
of what we do is mitigated, and that, you know,
ultimately, we're gonna be working with communities
to ensure that we don't have a negative impact
at all on community water.
Do you still have a net zero carbon pledge for 2040?
Absolutely.
The climate pledge, emblazoned on the side
of Climate Pledge Arena, which is a few blocks from my house,
go Kraken, is very much in force.
We, in fact, this is Climate Week, as well,
as everybody in the room, we're consistent
with our pledge to have over 100,000 by 2030.
And we have over 700 other companies
who have signed on to the climate pledge
to work with us and together,
some of whom are our biggest commercial competitors,
to trade best practices, information
on how we can all move towards a carbon neutral world.
So how does nuclear play into all this?
I mean, we hear a lot about it.
We know that it is one of the power sources
that has somehow won over, like,
both sides of the political spectrum
because of its greenness and because, I guess,
nukes are cool enough for, I don't know, whatever.
But somehow, like, Jane Fonda
is not making any noise anymore.
Where are we really at with nukes?
So I'm old enough to remember the no nukes days,
the no nukes concert, remember that?
Jane Fonda and Ted Turner. Bruce Springsteen.
Yeah, Neil Young.
And there was a lot of fear and certainty
and doubt in the late 70s and early 80s
about nuclear.
And there were also bad designs.
Yeah, correct.
But here we are, 50-ish years later,
and we're looking at places like France,
where 80% of their grid is nuclear
and it seems to work for them.
And it is an incredibly important strategic advantage
that France has, that they are 80% nuclear.
And so I think the science has gotten better.
The way we design these plants has gotten better.
And you're seeing a recognition of that
in the desire to include nuclear in the portfolio
of non-carbon energy sources that we need to invest in.
And so for us, that means we're working with,
in Pennsylvania, to site a data center
right next to an existing full-scale nuclear plant,
not Three Mile Island.
They had a branding issue.
Is it one that's been continuously operating
or was it one that was?
I think it's been continuously operating.
I think it's in Berwick, Pennsylvania.
And so that's underway.
We have an investment, a significant investment
in a Washington state company
to develop small modular reactors.
And that's in Eastern Washington.
That's the timeline I'm most curious about.
I've heard-
Pennsylvania has these small modular reactors,
but we don't have them yet.
I've heard anywhere from like 2030 to 2035
as being when we can roll it out at scale.
I had a lawmaker tell me last week I was in D.C.
that they feel like it's a bit like fusion, right?
It's always 10 years away.
I'm much more of an optimist.
And I think the fact that we've invested,
I think 500 million in this X energy project
in Eastern Washington really shows
we're putting our money where our hopes are.
And I think that's especially exciting
because one of the big barriers
to bringing on more energy generation
is the transmission system,
which is obviously out of date.
70% of the transmission system
was built more than 25 years ago.
Permitting for new transmission is kind of a nightmare.
We may see one of our most important policy objectives.
One of the reasons I was in D.C. last week
is to look to continue the momentum
toward possible permitting reform at the federal level.
We might be in a moment where that's possible,
although we felt that way two years ago
and we didn't get there.
So transmission is super important.
And obviously if you can build a facility that's small
right next to what you need to have energy go to,
then you reduce the load.
You could theoretically power your whole operation
with your own nuclear reactor
and sell excess power back or give
if you want to be charitable
to the place that's hosting your data center.
But even more important,
you're not reliant on the transmission infrastructure
to get it from where you're generating it
to where you need it to go.
And so that said,
we're also working with utilities
to pay the cost of transmission upgrades
where we know we're gonna need that
in order to avoid affecting rates of consumers.
We thought that the rate payer protection pledge
that the White House put forward earlier this year
was a useful one and we signed on to that.
And you do that by working with utilities
and working with communities
to make sure that if there's transmission infrastructure
that needs to be built as a result of a new large load,
that we're paying that cost.
We're paying the cost of new generation, obviously.
As an example, we actually had to get
a regulatory exception in Virginia
to allow us to build a transmission facility
that would enable some of our large loads
to serve data centers upfront
instead of going through the utility
because we wanted to make sure
that the cost didn't get included
in the utility's rate base.
And they did grant that permission.
So I think there's lots of tactical creative ways
to work with utilities and communities
to mitigate any impacts on consumers.
So before we move on to AI,
I just wanna ask one more.
In terms of the new power that you're adding,
how much of that is carbon emission power?
Because that's been the issue, right?
This ramp up of AI, the power demands,
no one has these alternatives in enough places
and mature enough in development.
Right, I mean, there's gonna be places
where we're gonna need to rely on natural gas,
at least as a bridge.
And natural gas, by the way, is a lot cleaner
than sort of coal grid.
But again, it really depends on where you are in the world,
where you are, and this is not limited.
All these issues, one thing I've learned,
I haven't learned more about this space
than any other subject in the last three years.
They didn't teach it in law school?
They didn't, no.
And it looks, I mean, there are similar problems
all over the world.
Everybody's experience, you go to Ireland,
you go to Singapore, it's the same story.
And so we have to be creative and we have to be flexible
in terms of working out how to bring on
new renewable generation.
I started working, leading the public policy team
about four years ago, and on one of my first trips
out to Singapore, I was like,
well, why don't we just build more offshore wind?
And they all kind of looked at me like,
I didn't have a head or something.
We're on the equator, there isn't any wind.
There isn't enough wind to make it commercially viable.
So we're learning.
Yeah, that's interesting, I hadn't thought about it.
Now I'm gonna think about that all day.
All right, so let's talk about AI.
I mean, there's, I mean, I'm not gonna ask you
what your P-Doom is, but everyone probably knows
that expression now after last week.
We went from being like, oh, AI's taking jobs
to AI's gonna kill us all, right?
I mean, and nervous laughter ripples through the room
in every room we're in now, because we're like,
is that real?
We don't know if it's real.
And of course, we can all guess our own,
I give it a 10% chance of destroying, you know.
But the truth is, the powerlessness that most people feel
is because all of the decisions are in the hands
of big tech companies right now, including Amazon.
And so the question is, like, how much of that discussion
is happening with this rapid rollout of AI?
Like, what are you guys talking about?
Has the conversation shifted?
I know you were probably talking about it
way before we were, but like, what is going on right now
behind the scenes that you can share with us?
Well, you know, it's not new, and it's not rocket science,
even though it seems like it.
Like, we've actually been an advocate
for responsible regulation of AI for years.
I've been talking from stages like this
for at least three years, about how we need to look at,
you know, what are the risks, work backward
from the worst case risks, and have some reasonable
guardrails, regulatory guardrails, and make them
interoperable between different parts of the world,
make them standards.
And you know, the headline in today is that, you know,
OpenAI and Anthropic are calling for global standards.
Yes, duh.
Like, you know, we've been saying that for a long time.
And so, you know, there's risks involved
in any new technology.
It is critical that companies exploring new technologies,
whether it be nuclear or AI, like, that when they release
products to the world, that they're safe,
and that they're tested, and that they're vetted,
and that we use the best science and the best
knowledge that we have to ensure that they're safe.
It's a little bit like, you know, racing so hard,
you know, releasing cars without remembering
to put the brakes on, like, okay, it's up to the companies
to put the brakes on the cars before they release
the next model car.
And so, you know, we don't think there's a trade-off,
there should be a trade-off between speed and safety,
but we need to make sure that when we launch new products,
that they're safe, and we've done everything we can
to ensure their safety, so.
But when I say that it's in the company's hands,
I'm not just saying because you're the ones driving it,
I'm also saying because we don't see both, you know,
a federal regulatory, like, push,
or a global cooperation push.
And, you know, I mean, this is like,
that's the frustration I'd say.
I think there is more cooperation than people think.
I mean, the AI sort of safety institute is still a thing.
I mean, it was known as a different,
under a different name under the last administration,
but I think that's being built up.
You know, we're members of the Frontier Model Forum,
which is a group of companies that are trying to build
and building frontier models that trade information
about how we're testing these models,
how we're making sure they're safe.
And so there is a fair amount of collaboration
that's happening both at the policy level,
and more importantly, at the science level
in forums like that.
So we're working hard to make sure that connectivity exists
and is robust, and that those conversations
are taking place.
And I think that sort of frontline cooperation,
you know, may get lost a little bit
in some of the headline rhetoric
that comes from public statements
that some of the leaders like to make.
Yes, super intelligence and all that.
The question is, when are we going to see
something that signals that the government,
that the global community is taking this
as seriously as some people?
Well, I don't know is the answer to that question.
We've been talking about it for a while.
Maybe it requires the kind of scary rhetoric
that we now, that has entered the debate
in order to get governments to sit up and take notice.
As somebody who sits in industry,
I don't think we ought to be waiting,
when we're not waiting for governments to regulate
in order to make sure that the products
that we offer and services are safe.
Like that's actually the industry's responsibility
first and foremost.
It would be great if we could get as much cooperation
among different companies, different scientists,
and governments to help in that effort.
But I don't think there's a sense like,
well, there shouldn't be a sense that we can't,
you know, we can't put the brakes on the car
until there's regulation requiring brakes.
You know, like that's the wrong order of things.
And I don't think it's actually how the sort of thoughtful
people in this industry are operating.
Right, okay, well, so let's pull back a little
from the doomsday scenario
and just go more to the day-to-day.
I mean, you guys are, your business,
let's break it down.
You have, you're selling, you know, through AWS,
you sell access to other companies' models.
You are an investor in Anthropic, OpenAI,
and others as well.
And then you're making your own models.
So that kind of makes you like the Amazon store
of AI models, if you would.
Well, you know, who would have thought?
But we're also, you know, we're also building,
you know, the infrastructure to train models
and to run the rest of our economy,
our services from advanced manufacturing
to the electrification and automation
of our transportation system,
to the movies that you watch on Amazon Prime or Netflix.
You know, all of that is powered by this infrastructure,
which, you know, the build-out of which I kind of liken,
everybody has their own favorite model.
I like the interstate highway build-out.
Took 20 years, created a generation's worth
of great construction jobs.
It was critical to, not just to commerce
in the United States and the ability to travel,
but to national security,
which was one of the big reasons why it was.
So I think the sort of infrastructure build-out,
which will also be used in the adoption of AI
and the training of AI, is akin to that.
Biggest difference being it's all being financed
by private capital instead of the government,
which I think is a huge opportunity
for the government and for communities.
So you want the government to help fund it?
I'm not saying that.
No, I'm saying we want the government to let us do it.
Right, right, right, right.
Right, to inform permitting.
There are parts of the country where you really,
you can't build a new renewable energy plant.
You just can't.
It's too hard.
So we're at that level of the stack.
We're also using AI to build outward-facing applications.
So things like our shopping assistant,
the feature that allows small, medium-sized businesses
to come to Amazon, use the tools,
create absolute premier content advertising
without the need for an expensive studio,
and feature it as they sell their products.
And obviously the most recent consumer-facing
and enterprise-facing innovation would be Amazon Quick,
which is a desktop assistant
that uses a bunch of AI models in the background,
but allows you to hook to your email,
to your Slack, to everything.
I don't think we've played around with that yet.
You should play around with it.
I use it, I mean, I think it's pretty incredible.
It really is a day-to-day assistant
that lives on my desktop.
And now it's integrated with the phone,
which we didn't have initially, which was super annoying.
But now it is very useful.
In terms of AWS clients
who are taking advantage of the AI offerings,
what are you seeing in terms of what they're buying
and what they're buying it for?
Who, are they gravitating towards certain types of models?
Are they looking at lighter weight, cost-effective?
I mean, what are the business decisions?
I mean, I think it's all of the above.
One of the really great features about Bedrock,
which is the main premier feature
where enterprises within the cloud,
within the secure cloud environment,
can use any number of models.
I think there's over 100 models,
from the big frontier models to open weight models
to our own models to smaller ones,
because they don't all cost the same.
And it lets enterprise customers experiment,
mix and match, decide what kind of model they need
for which workloads.
It's not at all clear that you need Mythos
for every single, to process email in your back room.
You could use a lot of a cheaper one.
So we're seeing a whole variety of innovation.
I think one of the most exciting things for me,
as I've watched this space grow,
is Amazon's gonna do its best
to innovate in new applications.
But if you think about the number of enterprise customers,
and it's over 100,000,
who are using these models, experimenting with them,
I think the biggest interesting innovations
are gonna come from them,
when they apply these things to their business
and the way they operate their organization.
And are you able to meet the demand?
Is demand still off the charts?
You hear, I mean, this is part of why the stock market
tends to be buoyant these days.
I don't know, I haven't checked it today.
But the question is whether or not
we're gonna see any slackening of this demand curve.
I think the demand for inference,
which is really the use of these models,
is continuing to expand quickly.
Bedrock is one of our most successful launches.
Do you have to limit how much you give customers
or charge a premium because the demand is so high?
I don't know for sure.
I don't think so on the inference side.
I know that there's tremendous demand
from the model builders for chips and training.
And I know that that's a constrained resource, for sure.
And where that balances out
in terms of the resources we have to devote
to users of Bedrock, I'm not positive.
Right.
Okay, so let's talk about outer space.
So how soon till Amazon delivers space internet
to catch you up?
There's a constellation of satellites
that is trying to rival Starlink.
You're at around 400 satellites in the air now.
We need about 3,200 for our generation one
to cover most of the earth.
The way this works is that if you get a critical number,
and 400 is about right,
you can provide service to a certain latitude band
all around the world, the earth.
Uzbekistan?
I think that's what you guys just signed a deal with.
We did.
A year ago, I was here for UNGA,
and my first meeting of the first day,
I was still kind of adjusting to the time zone,
was with the president of Kazakhstan
to announce a deal with the Kazakh Telecom.
Another sort of thing I never thought I would do
26 years ago when I started,
but apparently now, and it was like,
they filmed the whole meeting,
and the paparazzi were outside.
Apparently I'm now huge in Kazakhstan, but.
I mean, no, we're building,
we can serve a certain latitude band,
and as we get more of our satellites up,
we can serve more areas.
I think we're still on track to turn on
the first tranche of service later this year.
At least that's what our, that's what Panos told me.
And that'll be really exciting,
because the service is amazing.
I've seen it in action.
At the moment, it's significantly faster
than the nearest competitor.
It's enterprise-ready.
It's gonna have some synergies with the AWS infrastructure
for enterprises that want to use it.
It'll serve everyone from consumers to schools
to hospitals to large enterprises.
Yeah, how much does this provide
kind of a proprietary backhaul for you and AWS?
In other words, if you've got data centers all over,
and you want them talking to each other,
and there are places where you can't.
Yeah, I think there's lots of interesting applications
to merge those functionalities.
We also have satellite-to-satellite communication,
which is an upgrade made possible
by the most recent technology,
and that's actually attracting a lot of interest
from government customers as well.
And so there's a lot of possibility there.
It's very exciting.
And for me, the most exciting thing
is when we're able to bring service
to places that really haven't seen
that type of high-speed internet service before.
So earlier this year, I was in South Africa.
We signed in.
One of our approaches that differentiates us,
I think, from some of the competition
is we're really willing to work with local partners,
with local telecoms.
I mean, you have to,
because you're already doing other business
in those towns, in those areas.
Yeah, although we don't,
that's not something Amazon is normally known for,
you know, sort of working with partners
or outsourcing to partners.
But we signed a deal with Herotel in South Africa,
which means they're gonna offer the service.
It's gonna be branded them.
We're sort of powered by Amazon.
They'll own the customer relationship and pricing.
But as a result, we're gonna be able
to serve rural communities, farmers,
areas that haven't had access to internet
that would enable them to do things
like remote learning or telehealth
or really sort of high-end applications.
And so, you know, that's pretty exciting
and we're the first provider there.
Specifically because even if you have
a satellite overhead somewhere,
you still need permission to broadcast
and have that radio relationship
with the ground in that country.
Right, you need the right license.
It's not like you can just be like,
oh, we have satellites,
we don't have to talk to these folks.
You need the right licenses and you need permission.
In some cases, you need ground infrastructure to enable it.
So, you know, all of that's, I think,
on track in South Africa
and in many other places in the world.
And, you know, hopefully, the northern United States,
coincidentally, including Seattle, who knew,
would, you know, should be part of the first.
The first.
That is a coincidence.
Who knew? Okay.
And I think in some cases,
you have to use SpaceX rockets to get your satellites up,
which is kind of a funny thing
since SpaceX obviously owns the only leading competitor.
We are, right, it's true.
We're, you know, working with every launch provider
that can provide meaningful functionality
to help get our constellation in orbit.
And that means working with Ariane Espace in France.
We're working with SpaceX.
SpaceX has been a terrific partner
in terms of, you know, launch reliability and functionality.
Working with ULA and, you know.
And then Blue Origin.
And we'll be working with Blue Origin
as soon as they can.
Jeff Bezos's rocket company.
Yeah, very separate.
Very separate, but it'd be nice.
It'd be a synergy, right?
You know, mostly, we're just looking at,
like, getting our satellites in orbit.
Yes, absolutely.
Well, you have a deadline, right?
You have to meet.
Well, we met the, we're fine with a deadline.
You know, we sort of got enough built and got enough,
because we have 1,000 satellites on the ground ready to go.
So it's really, launch capacity is a constrained resource.
And we have, I think, we have rights to,
we've purchased rights to over 100 launches
in the next couple of years
in order to enable us to move more quickly
to get global coverage.
Well, it's hard to get things into space,
which doesn't change no matter how good
our tech on Earth gets.
You still gotta get those rockets.
Although it's getting cheaper.
And getting cheaper in part because of the efforts
of SpaceX and Blue Origin.
And, you know, the advent of reusable rockets.
And that's the design.
That was, talking to Jeff 20 years ago,
that was his vision.
He wants to make the cost of space go down
so that there can be much more innovation,
much more activity in space,
much like the internet was when he innovated Amazon.com.
