[music playing] Please welcome the CEO
of Amazon Web Services, Andy Jassy. [applause] Thank you, and welcome
to the ninth annual AWS re:Invent and the first virtual one. Now, I think we all wish we were
together in Las Vegas right now but we’ve got the next best thing,
we’re live, so I’m really excited
to be with you. I am going to share a few words
about re:Invent, it’s a little bit
different this year. So one thing that’s similar is
we have a lot of people congregated. We have well over 400,000 people who have already registered
for re:Invent. We will be at 500,000
by the end of the week, and this year is going to be
different than prior years. re:Invent is going to be over
three weeks instead of one week, and then we have five keynotes,
so, the one you have today. Thursday there will be
a partner keynote. Next Tuesday will be our first ever
machine learning keynote with Swami. Then, next Thursday,
you will have Peter DeSantis telling you more about
our infrastructure and then, Werner will do the anchor keynote
the third week. So we’ve got a lot
in store for you. re:Invent is first and foremost
an education and learning conference
and this year will be no different. We have well over
500 technical sessions and while I think we would
all like to be together, I think one nice part
about it being virtual is that you don’t have that session
collision that you sometimes have, where there are two sessions
you want to go to at the same time or you have to travel.
They will all be online. Some will be scheduled.
Most of them will be on demand. So it will be much easier for you
to see everything you want. A lot for you. So I am going to give a quick update. But before I do
I just want to acknowledge what an unbelievably crazy last year this has been,
particularly these last nine months. COVID has been so difficult
for so many people. We have had so many people pass away. We’ve had so many people
lose their jobs. So many businesses are struggling.
It’s just been really difficult and I think if you live
in the United States it’s also pretty hard
to not be struck by the murders of Ahmaud Arbery,
Breonna Taylor and George Floyd and realize the sobering thought
that we have so far to go in this country
in how we treat black people. I think the reality is for
the last several hundred years the way we treated black people
in this country is disgraceful and something that has to change. We’re working on it at Amazon.
I know a lot of companies are but I think the important thing
for us all to realize is that this does not
get solved in a few months. It can’t be something that
we work on for a few months and then turn
our attention elsewhere. It’s going to take several years
of us working together but we need to do it. So a quick AWS update. The business is at a $46 billion
revenue run rate business in the last announcement that we made
growing 29% year-over-year. I think it’s also important to talk
about that year-over-year right because I think a lot of people
sometimes get confused and they try to compare different
companies’ year-over-year growth rate. The year-over-year
growth rate percentage only matters as it relates
to the base revenue. So you can have a higher
year-over-year growth rate but be growing at
a much lower absolute rate if you have a much lower
base of revenue. If you look at just AWS as an example
to grow to a $46 billion dollar revenue run rate with 29%
year-over-year growth, it meant we had to grow
an incremental $10 billion in the last 12 months to get there. That is much larger than you
will see elsewhere in the cloud. I think another data point to give
you an idea of how fast AWS is growing, it took us a 123 months,
a little over 10 years to grow to a $10 billion business. Then it took us only 13 months
to go from $10 to $20 – er, 23 months - to go from $10 to $20 billion,
13 months to go from $20 to $30 billion and then 12 months
to go from $30 to $40 billion. So the rate of growth in AWS
continues to accelerate. I thought this was also
an interesting slide which is: This shows
you the top enterprise IT companies out there
ten years ago based on revenue. And you can see on this list there’s some companies that don’t
really exist anymore in this forum but you will also see that AWS is nowhere
to be seen in this list. And then if you fast forward
just ten years in 2020 you can see that AWS is now
the fifth largest enterprise IT company in the world. Ahead of companies like SAP
and Oracle and of course, that growth is significantly driven
by the growth of cloud computing and the infrastructure technology space. And I think that you can see that with what’s happened
during the pandemic. And in the short term
in the first nine months or ten months of this thing
virtually every company in the world, including Amazon, has tried
to save money any way that they can. But what we have seen
and this happens a lot of times when you have a period
of discontinuity like a pandemic is that companies take a step back and they rethink what they are doing
and what they want to stop doing. And one of the things that we’ve seen
is that enterprises that we’ve been talking with
for many years about moving to the cloud
where there’s a lot of discussion and dipping the toes in the water,
but not real movement, so many of those enterprises
have gone from talking to having a real plan. And that I think is going to be
one of the biggest changes you'll see. See, I think when you look back
on the history of the cloud it will turn out that the pandemic accelerated
cloud adoption by several years. It's really hard to build a business
that sustains for a long period
of time - really, really hard. And this is a metric
that I think is interesting. If you look at the Fortune 500
just 50 years ago in 1970 you can see that only 83 companies
or 17% of them are still in the Fortune 500. If you look at just 20 years ago
the 2000 Fortune 500, only half of them
are still in that list. It is really hard to build a business that lasts successfully
for many years and to do it you’re going
to have to reinvent yourself and often you are going to have to
reinvent yourself multiple times over. And so, in the last nine months
I’ve thought a lot about reinvention and what it takes
to do reinvention well. And typically what you see is
the desperate kind of reinvention. You see companies that are
on the verge of falling apart or going bankrupt deciding
they have to reinvent themselves. And when you wait to that point
it’s a crapshoot whether you’re going
to be successful or not. It’s a little bit
like borrowing money. Everyone will tell you that you don’t
want to have to be borrowing money when the business is in bad shape because you may not get the rates
you want or you may not get money at all. You want to be reinventing
when you’re healthy. You want to be reinventing
all the time. And so we thought about
what are the keys to reinvention and some of it is building
the right reinvention culture and some of it is knowing
what technology is available to you and jumping on it
to make that reinvention happen. So I thought I would share with you
today what we see as some of the keys
to building that reinvention culture and then some of the things that we
see being reinvented as we speak. So what does it take to reinvent? And I am going to list eight keys
that I think are important if you want to build
the right reinvention culture. And the first is that
you have to have the leadership will to invent and reinvent,
and those terms sometimes they are a little different,
they are also similar. If you think about it,
people often say invention is inventing a new product or service
from whole cloth and reinvention
is reimagining an existing concept. But if you’re going to reinvent
and reimagine there’s a load
of invention in there. Just look at what Airbnb has done
in the hospitality space or what Peloton has done
in the exercise bicycle space or look at what Stripe has done
in the payment space. These are huge amounts of invention that has gone into
reimagining these spaces. And so if you’re going to be a leader
that’s going to reinvent you have got to be maniacal
and relentless and tenacious about getting to the truth. You have to know what competitors
are doing in your space. You have to know what your customers
think about your product and where you sit
relatively speaking. You have to know what’s working
and what’s not working. And you will always have a lot
of people inside the company who will try and obfuscate
that data from you. Sometimes they think
they are doing you a favor and sometimes it’s for
self-preservation reasons but it’s hard to get at that data
and you have to be relentless about it. You have to challenge people. Often people who know a lot more
about a subject than you do but you have got to get to the truth. And then when you realize that
there’s something you have to reinvent and change
you have to have the courage to pick the company up
and force them to change and move. And part of that is sometimes
acknowledging that you can’t fight gravity. If you step back and have conviction
that something is going to change because it’s a better experience
for customers, it is going to change. Whether you want it to or not, whether it’s convenient for you or not,
it is going to change and there are a lot
of examples of this. I think if you look at what
Reed Hastings and Netflix did several years ago where they cannibalized
their own DVD rental business because they saw where it was
headed with streaming, I think that turned out to be
a pretty good decision for them. I think if you look at Amazon
in the late 90s we had this owned inventory
retail business which meant we bought all this
product from publishers and from distributors.
We stored them in our warehouses and then we shipped them
to customers. And what we started seeing was
these companies like eBay and Half.com that were actually
offering third-party sellers’ products and they were shipping
the products to customers. And we had this huge animated
debate inside the company on whether or not
we should support that. And the reasons that
we were concerned about it were we just didn’t believe anybody was going to take care of
customers the same way that we did. And then also
the whole culture was set up to be this owned inventory business. People worried, well if we worked
with third-party sellers, how would our publishers
or distributors feel? So it was a very hard decision
but ultimately, we decided to build a marketplace
and offer third-party sellers. And we did it because we know
that you cannot fight gravity. It was better for customers.
It provided them better selection and it gave them more assortment
on price. Now that also turned out
to be a good decision for us because we sell more
than a half of our retail products through third-party sellers. But you have got to realize that
if something’s going to happen it is going to happen regardless
of whether you want it to or not. You’re much better off
cannibalizing yourself than having someone
do it to you and chasing it. The third thing you have got
to make sure of is that you have talent
that’s hungry to invent. Now this seems fairly obvious. Everybody says I have talent
that wants to invent. But it’s not always true. A lot of people who have been
at the company for a long time are very comfortable
doing things the way they have been
doing them for a long time. Have you ever noticed it’s often when
you have new blood in the company that they are leading
the transformation? And that’s not because existing
people can’t lead reinvention. It's just that you’re asking them
to reinvent something they built. It’s hard to rip up something
you spent a lot of time and energy and dedication doing. And it means you have got
to learn new skills and it means that you have got
to actually be curious about getting trained
on other technologies. Sometimes that’s true
sometimes that not true… I will tell you a quick story. There’s a CIO in a pharma life
sciences company that I’ve known a long time. I have a huge amount
of respect for him. And I went to see him
a few years ago and I was talking to him
for about 30 minutes about why I thought they should be
using the cloud more meaningfully. They were barely using
the cloud at that point. And he listened to me
and when I was done he said, “Look, I agree with everything
you just said, Andy, I agree that we could be inventing
at a much faster clip but that will be
the job of the next CIO. It will not happen on my watch.” And that’s what happened.
It took him a year or two. He retired, they hired a new CIO,
that CIO said: “What are we doing?” And they’ve significantly
moved to AWS in the cloud. But they lost two, three,
four years of inventing on behalf of their customers and you have to make sure that
you have got builders who are curious about learning who are excited about
leaning forward and inventing and reinventing
their customer experience. Now, you want builders
and talent that’s hungry to invent but you want to make sure
you guard against the opposite which is that you have people
who actually solve problems. That you want people
to solve problems for customers as opposed to solving problems
because they like the technology and they think it’s cool.
And you see this a fair bit. You know if you look in
the enterprise technology space there are some providers
who are competitor focused. They look at what
their competitors are doing and they try to fast follow
and one up them. We have a competitor like that across
the lake from us here in Washington. Then you have a number of other
providers who are product focused and they say look, it’s great
that you have an idea on a product Mr and Mrs Customer
but leave that to the experts. And that’s the group that you have
got to be careful about because they often are building
things that they think are cool as opposed to what really solves
the problems for customers. At AWS, we’re customer focused. What we build is driven
by what you tell us matters to you. And even if you can’t articulate
a feature we’ll try to read between the lines, understand
what you’re trying to build, and invent on your behalf. And so if you think
about over the years of AWS, we have built a lot of technology
that we believe is pretty cool, you know pretty ground-breaking
stuff, S3 and EC2 and RDS and Aurora
and SageMaker and Redshift. I mean a whole host of technology
but we never built it because we thought it was cool. We built it because we knew
it would enable you to build new experiences
and change your business. You’ve got to make sure that
your scarce resource of engineers are working on problems
that really matter to your customers. The fifth thing is speed. Speed disproportionately matters
at every stage of your business and in every sized company. And I think that a number of leaders
at enterprises have resigned themselves
that they have to move slowly. It’s just the nature
of how big they are. It’s the nature of their culture. They have engineering teams
that tell them “Hey look, this is too risky.
This is too big a lift.” Sometimes you control those teams
into trying something and the first sign of a problem they throw up their arms
and say “See.” Speed is not preordained.
Speed is a choice. You can make this choice
and you’ve got to set up a culture that has urgency
and that actually wants to experiment because you can’t flip a switch
and suddenly get speed. It doesn’t work like that. You’ve got
to build muscle to get speed. You’ve got to be doing it
all the time. And there is going to be time… Frankly I think that time
is happening right now. It happens a lot more frequently
than most companies realize. But there is going to be
seminal moments where if you don’t have the ability to have speed you will not be able
to reinvent when you need to. Now one of the enemies of speed
is complexity and you have to make sure that you don’t over complexify
what you’re doing. When companies decide
to make transformations and big shifts a huge plethora
of companies descend on them and providers descend on them and tell them all the ways that you
have got to use their products: “You need to use us for this
even if you’re using these people for these three things, use
for these two.” This company says: “Use us for this.” They don’t deal with the complexity
that you have to deal with in managing all those different
technologies and capabilities. The reality is for companies that
are making big transformations and shifts, it is much easier
to be successful if you predominantly
choose a partner and you learn how to do it
and you get momentum and you get success and you get
real results for the company. Then later on if you want to layer
on complexity and more providers,
you should go for it. But it’s not a great way
to start to a reinvention to have too much complexity upfront. And then one of the ways to help you
avoid some of that complexity is making sure
that you use the platform that has the most capabilities
and the broadest set of tools. Now I have played a round of golf with somebody who used a 5-iron
for every shot in the round. It’s doable. It was not pretty
and it was not very effective either. I don’t recommend it. And when you think about the cloud,
because all the services you only pay for as you consume them -
you don’t pay for it upfront - Why would you possibly go
with a platform that has a fraction
of the functionality of a leader? If you go with the platform
that has the most capabilities and gives you the right tools
for the job you need to do, it not only makes it easier
for you to migrate all your
existing applications, but also to enable your builders
to build anything they can imagine. And there’s nobody who is close
to the capabilities across the cloud infrastructure
technology platform as AWS. And you can see that whether
you are talking about compute or storage or database
or analytics or machine learning or the edge or IoT or robotics,
in every one of these categories you get a lot more functionality
in AWS than anywhere else. The eighth key is something that
really wraps all of this together, which is that the leadership team has
to build aggressive top-down goals that force the organization
to move faster than it organically otherwise would. And there are lots of examples
of this. I would like to talk about GE,
where about ten years ago their CIO decided that they were going to move
50 applications to AWS in 30 days. And her entire team told her
what a terrible idea this was and she listened to them
and she said, “We’re doing it anyway.”
They got to about 42 in 30 days. But in the process they figured out
their security and governance model. They figured out how to operate
in the cloud and they had success
which built momentum and the ideas came flowing
in such that she could set that second big top-down goal
to move 9000 applications to AWS in a few years. Capital One did the same thing where they set
this big audacious goal top-down goal: They were going to reinvent
their consumer digital banking platform on AWS.
And then that was on their way to moving
everything to the cloud in AWS. Setting an aggressive top-down goal
forces the organization to understand that they are not going to be able to dip their toe in the water
for a number of years. That you mean business
and you’re going to make this change and setting up
the right mechanisms to inspect whether you are getting
the right progress, and if not,
getting the issues on the table so you can solve them
is really important. Now, you will notice that most
of these keys, all of these keys really, except for maybe one,
I mentioned are not technical. They are really about leadership. And so you’ve got to make sure
that you embrace these types of keys to build a reinvention culture
like you can. It’s very doable
but you have to embrace them. My first speaker today
has embraced these keys and it’s really remarkable to see
how she has led her company to start reinventing on top of AWS. It’s my privilege to welcome to
the stage the CIO of JP Morgan Chase, Lori Beer. [applause] Thanks Andy, it’s great to be here.
JPMorgan Chase serves customers and clients from around the world
from individuals and their local communities
to corporations and governments. We have an over 200-year history
built on trust. A foundation that allows us to build
simple and intuitive experiences. Those solutions which are
increasingly technologically driven represent the complete spectrum
of financial services. From lending to banking, markets
to advisory, and everything in between. And everything we do we do
at tremendous scale. We have $28 trillion in assets
under custody and process $6 trillion of payments daily through our Consumer
and Community Bank. We have a relationship
with 50% of US households and we serve 54 million
active digital customers. We have long realized that while our
relationships and financial expertise are paramount it is technology that
continues to help us differentiate. With a 200-year history we’ve been
dealing with technological change since the time
of Thomas Edison literally. The firm financed Edison’s efforts
to invent the first light bulb and our former Wall
Street Headquarters was the first office in Manhattan
to draw on Edison Electricity. We were the first bank
to offer ATMs and our industry
embraced early enterprise computing to process transactions. We were quick to offer online
banking services during the rise of the internet
and our industry is increasingly driven
by mobile computing today. Simply put, we’ve been reinventing the financial services
industry for decades. But it’s different today.
The pace of change has accelerated. Innovation is more rapid
and cloud platforms are disrupting business models daily. So we decided to completely rethink our environment to embrace
a true modernization effort across 250 thousand employees,
35,000 developers, 6,000 applications
and 450 petabytes of data. We asked ourselves how do we leverage
our history of innovation and institutional know-how in order
to evolve our business for today’s technology revolution? We set the direction
at the top of the house with business
and technology leadership agreeing on a key vision.
We will have the best tech talent. We will own our destiny
in a hybrid cloud world. And we will work with leading
technology companies like AWS to deliver unique innovation. We’re in the midst of a truly
one-of-a-kind transformation. Leveraging AWS and modern
engineering practices like refactoring our applications
to be cloud native, leveraging more advanced
AI and analytics than we ever have before,
and doing all of this securely. AWS
is helping us along this journey. We’ve benefitted from their breadth
of services and capabilities. Understanding of the enterprise
and willingness to innovate and collaborate
on key strategic initiatives. We work together
on a unique approach holding actual hackathons
with AWS and JPMorgan Chase engineers to decompose
applications and migrate workloads. This process helped us
uncover problems, gain institutional knowledge
and enhance our collaboration. We’ve developed repeatable blueprints helping developers
architect modern applications in safe repeatable environments using the depth
of AWS services. We are a systemically
important institution serving clients through some of
the most turbulent events in history. Working with AWS allows us to scale
massive volumes like Amazon EMR for trading analytics or AWS Lambda
and Amazon Elastic Kubernetes Service for risk calculations
so we can innovate to stay
ahead of our competitors. Transforming yourself
takes a lot of work but we are creating
paved ways for developers and accelerating adoption. Our developer community is now
innovating at scale using AWS and we continue to migrate
critical workloads that can take advantage of the unique
capabilities of the platform. How can we transform our business
through AI and better analytics? As I mentioned earlier
we have customer relationships with half of all US households. The scale of our reach provides
an opportunity to use analytics to better service them. To do that we need
a modern AI platform that was secure and one that enabled
rapid experimentation. That’s why we built our firm-wide
AI platform OmniAI on AWS using Amazon SageMaker
for machine learning. SageMaker has helped us
create a platform to rapidly test and train machine
learning algorithms enabling us to run
more experiments with more data. These are real use cases
providing real value. As I speak, our data scientists
are leveraging our OmniAI platform to develop new capabilities. For example, incorporating
natural language processing to make client
interactions more personalized. Testing advanced machine
learning models to have a more
comprehensive view of risk. And performing
real-time coaching and recommendations
for call center agents so they can better
serve customers. We use SageMaker across
the model development lifecycle from data labeling
to model selection to experimentation
and model serving. Our success with
model development on AWS has influenced
our data management strategy. We’re now investing in cloud data warehousing technology
with Amazon Redshift to more effectively scale
our analytic capabilities in a modern environment. The combination of a scalable
AI platform and AWS’s Elastic Compute
environments will help us accelerate our efforts to infuse
analytics in everything we do. I will close with
what I started with. We are a business of trust. Everything we do is for the benefit of
our customers and clients and we deliver new innovations
in a safe and secure manner. This is why we devote
significant resources and collaborate with AWS to protect
and continuously improve the security
of our systems. Our adoption of cloud technology
is already paying off. We’re more agile. We’re more secure
and we’re more efficient. And through this journey
we will work with AWS to reinvent ourselves and build
financial services of the future. Thank you. [applause] Thank you, Lori. It is really an honor for us
to work with JPMorgan Chase. The partnership
has come a long way and we’ve just gotten started with
respect to what we can do together. So thank you very much. So I mentioned earlier that when
we think about the key to reinventing
it’s a combination of building the right reinvention culture and then also having
a set of technologies that you know are available to you
that you jump on to use to reinvent. So I thought I would spend
some time talking with you now
about some of the areas that we believe are actively
being reinvented that will allow you
to do this reinvention. And at the beginning
of each of these sections I am going to ask the customer
to say a few words about what they think about
this particular area and we’re going to start
with the SVP of Engineering at Snap, Jerry Hunter,
who will talk about Compute. [music] All right we’re going to try
and shoot this on Snap. I remember when we first
started doing Compute - that was the biggest thing
to hit tech. It disrupted so many parts
of the industry and I really enjoyed
being a part of it. Joining Snap where the company
is cloud native and taking advantage
of every innovation including things like
Graviton2 and Serverless. Every time a new innovation comes out,
we are one of the first to adopt it because it lowers cost
and improves performance. You might think that
there isn’t much left to reinvent when it comes to Compute
but the innovations just keep coming. [applause] Thanks, Jerry, we really appreciate
the partnership with Snap and we appreciate you
continuing to move to AWS. And Jerry knows what he is talking
about as it relates to Compute. He spent a lot of time
at Sun in the early years and then he was at AWS
where he ran infrastructure which are our data centers
and our network and our hardware before going to Snap
to run their engineering. So he’s kind of seen the rise
and change in Compute. He’s absolutely right that
Compute is continuing to be reinvented as we speak. There’s three major modes
of Compute that we see. Instances, which is
the traditional way that people have run
Compute - particularly when they want to get
all the resources on a box for their application. And then smaller units of Compute
like containers, where people build on these
smaller microservices because it lets them move faster
and be more portable. And then event-driven serverless
computing, when they don’t want to worry
about servers or clusters at all. And these three modes of Compute
are here to stay. They are going to be here
for a long period of time. Let’s start with instances.
If you look at Amazon EC2, which is our service
that vends instances to you we not only have the broadest
array of instances but we also have the most powerful
instances within those families. These are things like the fastest
networking instances where we have 100 gigabits per
second in all our recent generations and up to 400 gigabytes
per second in our P4d’s. We have the largest high memory
instances at 24 terabytes for SAP use cases
in our high memory instances. We have the largest local storage
instances with our D3en instances which are launched today
with up to 336 terabytes. We have the most powerful machine learning training instances
with our P4d’s. We have the most powerful machine
learning inference instances with our Inf1 instances. We have the best price
performance graphics instances with our G4ad instances
which are coming in a week. We’re the only ones
who give you a macOS instance type which we just launched last night
with our macOS EC2 instances that lets Apple’s millions
of developers now leverage the cloud
much more easily. We’re the only provider
that gives you the ability to run instances
with Intel, AMD, and ARM chips. It’s a very different set of
capabilities than anybody else has. And we’re also iterating
and innovating in a much faster clip. And people often ask us -
they say well, how are you innovating
at such a rapid clip right now? And there are really two reasons. The first is what we’ve done
with something called Nitro. We spend five years rebuilding
our virtualization layer and our Compute platform,
and we launched in 2017. What we did with Nitro was we took
the virtualization of the security, Networking, and storage
off of the main server chip and we put that
on our own Nitro chips. And what that did for customers
was it meant that you got all of the CPU
to run your instances so you get performance that’s
indistinguishable from bare metal but at a lower price. And it also meant you got
a stronger security posture because it used to be that
when you have to troubleshoot VM’s you had to worry about
whether somebody would do something to the main server
with your instances. But because the security now
is on that separate Nitro chip you don’t have to worry about that. What it did for us was, because
we broke out a bunch of those pieces into separable cards or chips, it meant we didn’t have
to make changes in making all these different
things change and evolve in lockstep, which allows us to innovate
at a much faster clip. Now what that means for you
is that you get instances now instead of every two to three years -
you get innovative brand now change-the-game types
of instances in months. That’s a big difference. The second thing that has happened
that has allowed us to innovate at such a rapid clip
is what we have done with chips. And we have a deep relationship
with both Intel and AMD and we will
for the foreseeable future. However, we realized a few years ago that if we wanted to continue to push
the envelope on price performance - which you have asked us to do - we knew we were going to have
to develop some of our own chips. And so we acquired a company called
Annapurna who were sophisticated and very experienced chip
designers and builders - and we put them to work. And we tried to pick chips that will
allow you to get the most done - have the most impact
for you business. And so what we started with was
we started with generalized Compute and we had the team build a chip
on top of ARM that we called Graviton. And so Graviton first manifested
itself in these A1 instances we had and they were really
for scale-out workloads, things like Web tier workloads
and things of that sort. And people loved them. They used them a lot quicker
than we ever imagined and they said: “Gosh if you would make the chip
more powerful so we could use it for more of our workloads that would
really change the game for us.” And so that’s what the team
did with their second version of the Graviton chip,
which we call Graviton2, which gives customers 40%
better price performance than the most recent generations of the x86 processors
from other providers. And those manifest themselves on C6g
instances and M6g instances and R6g
instances and T4g instances, 40% better price performance,
that is a big deal. Think about what you can do
for your business if you have 40% better price performance
on your Compute. And customers have loved this. We have a large number of customers
who are already using it from Nielsen to Netflix to Snap
and Lyft and NextRoll, and we are announcing today
a brand new Graviton2 instance which is going to be
the C6gn instance, which is our Compute heavy
and networking heavy instance with a 100 gigabits per second that will be coming
in the next week or two. We are not close to being done investing
and inventing with Graviton. People are seeing a big difference.
Let me give you a couple of examples. Here's a company called Honeycomb.io
and this is a blog that they wrote. And Honeycomb.io is a company that offers a comprehensive
debugging tool for development teams and they moved over
to the M6g instances. And you can see they are saving 40%
on price performance for their Compute and using a lot fewer instances
than they were using before. Or if you look at NextRoll, which
is a leader in digital advertising, they are saving 50%
on price performance. This is big deal and so Graviton
is saving people a lot of money. Very excited about it. We have a lot more
coming in this space. Then we asked ourselves what else
can we apply our chip team to that would solve problems that we know
are big growth areas for customers? And what we looked at
was we took machine learning. That’s the other area
that’s growing unbelievably quickly. And what you often see
with machine learning is that people talk
about the training because we’re still in the relatively
early stages of machine learning so, so many people are just
getting their models trained. And there’s a lot of effort
on machine learning training. But if you have models
that work at scale, you know, if you take for instance
Alexa, where it’s a big old machine learning model
that we have to train periodically. But we’re spitting out
predictions and inferences by the millions every hour. And so it turns out that 90% of
your cost is not in the training. It’s in the inference
of the predictions. And nobody was focused to try
to help customers save money and be efficient there. So we built an inference-focused chip
which we announced last year at re:Invent called Inferentia. And if you look at the number
of customers using Inferentia that also was really swelling.
But it's interesting to see - if you want to take a big
use case just look at Alexa. Alexa today has moved 80% of
their predictions to Inferentia. They are getting 30% better cost
and 25% better latency than they got from their prior chip.
That’s a big deal. Now while we focused on inference we haven’t forgotten
about machine learning training. And we continue
to iterate there as well. Of course we have these P4 instances
which are the most powerful machine learning training instances
in the cloud, but people understandably
still want to find ways to be cost effective
when they are training models. So today I have two announcements
to share with you that aim to help there. The first is that we will offer
next year, in the first half of the year, Habana Gaudi-based
Amazon EC2 instances. [applause] And so this is a partnership
between AWS and Intel and it will use Gaudi accelerators
that will provide 40% better price performance than the best performing
GPU instances we have today. It will work with all the main
machine-learning frameworks PyTorch as well as TensorFlow, and it will be available
in the first half of the year. Now like with generalized Compute
we know that if we want to keep pushing that price
performance envelope on machine learning training we’re going to have
to invest in our own chips as well. And so I am excited to announce
as well AWS Trainium, which is our machine
learning chip that’s custom designed by AWS to deliver the most
cost-effective training in the cloud. [applause] So Trainium will be even more cost
effective than the Habana chip I mentioned. It will support
all the major frameworks - TensorFlow and PyTorch and MXNet.
You will get to use the same Neuron SDK that our Inferentia
customers use. So if you use
Inferentia for inference it will be easy
to also get going on our machine learning chip - on Trainium.
It will be available both as an EC2 instance
as well as in SageMaker and that’s coming the second
half of 2021. So when you look
at the unmatched array of instances that you have in AWS coupled with the relentless
innovation in chips, and in the virtualization layer,
and the Compute platform, you’re now getting
reinvented instances every few months instead of every few years
which is a big deal. Now one of the things that’s really
interesting that we’ve seen in the last several years is that people are moving to smaller
and smaller units of Compute. And what I mean - there really
are containers and serverless. And so I will start with containers.
People really love containers. It allows them to build
on these smaller chunks of Compute which lets them move faster
as well as be more portable. And if you look at the growth
in containers in computing it’s pretty astounding. And the vast majority of it continues
to run in the cloud on top of AWS. Of the containers that run in
the cloud, about two thirds run on AWS - and that’s because while most other
providers have one containers Offering - typically
a managed Kubernetes offering - AWS has three. If it turns out
that what you value most is using the open source
Kubernetes framework, then we have our
Elastic Kubernetes Service, EKS. If it turns out what you value most
is having the deepest integration with the rest
of the AWS platform, you use our Elastic Container
Service, or ECS, which we can do because since we control it
we can make sure that everything launches integrated
with ECS right from the get-go. If what you value most in containers
is running containers without having to worry about servers
or clusters, then you use AWS Fargate,
which is our serverless container offering which nobody else
has anything like. And so one of the things
that was interesting when we launched containers
and we launched these offerings - we wondered once we had
a managed Kubernetes service would people use the other offerings? Or do Kubernetes
have so much resonance that people would only use that? And what we found is that all three
of these container offerings continue to grow like a weed.
Unbelievably fast. If you look at ECS we have over
100,000 active customers using it. We have billions of Compute hours
on EKS run every week on AWS. And if you look, most of the net-new
container customers in AWS start with Fargate
because it’s so easy to get going. And we actually have
a lot of customers who use two or even
three of these container offerings because different teams
have different preferences and have different use cases. You want the right tool
for the right job. You don’t want one tool
to rule the world, because there are lots of different
teams and preferences and use cases. So if you look at the problems that people who use containers
are trying to solve today they say it’s wonderful
that you have these three offerings. It gives me so much more choice. I can use different offerings
for different use cases. However, I still have
a lot of my containers that I need to run on-premises as I am making this transition
to the cloud. And so people really wanted
to have the same management and deployment mechanisms
that they have in AWS also on-premises - and customers
have asked us to work on this. And so I am excited to announce
two new things to you. The first is the announcement
of ECS Anywhere, which lets you run ECS
in your own data center. [applause] So ECS Anywhere allows you
to have all the same AWS-style APIs and cluster
configuration management pieces on-premises
that you have in the cloud. So it makes it easy as you’re making
this transition, to be able to run - if you’re running ECS on AWS
you can run it on-premises as well. It works with all of your
on-premises infrastructure. So not surprisingly when ECS
customers hear that we’ve got ECS Anywhere they say, “Well,
what about Kubernetes?” So I am also excited
to announce Amazon EKS Anywhere, which lets you run EKS
in your own data center. [applause] So again, just like with ECS Anywhere, EKS Anywhere
lets you run EKS in your data centers on-premises alongside
with what you’re doing in AWS, it works again with all of your
on-premises infrastructure. And it’s interesting when we’ve
talked privately to customers about EKS Anywhere,
they have been very excited and they’ve said,
“Well, I know that both ECS Anywhere
and EKS Anywhere are coming in 2021, but I want to actually get started
getting ready for EKS Anywhere." And so what we’re also
announcing today is that we’re going to open source the EKS Kubernetes distribution
to you so that you can start using
that on-premises. It will be exactly the same
as what we do with EKS. We will make all the same patches
and updates so you can actually
be starting to transition as you get ready for EKS Anywhere. So I think our container customers
are going to be excited by ECS and EKS Anywhere. But what we’re also starting to see
in addition to huge amounts of containers growth and adoption
is that more and more customers are using event
driven serverless computing. And we pioneered this concept
a few years ago with Lambda. And the problem that
we were trying to solve was that we had some customers
who said: “Look, I have certain workloads
when something happens, it triggers needing
to spin up some Compute and it forces to spin up
EC2 instances and multiple availability zones
for fault tolerance, and then they only really run these
jobs for a few hundred milliseconds, but I have to leave them
all up all the time because I don’t know
when the jobs are going to come in.” And they wanted us
to solve that problem and that’s why we built Lambda. Lambda lets you set a trigger
with a few lines of code. It spins up Compute.
We spin it up for you. We spin it up in a fault tolerant way, and then we spin it back down
when you are no longer running jobs and then we bill you in increments
of 100 milliseconds. By the way, I am going to announce
right now that we are changing the increments
in which we’re billing from 100
milliseconds to 1 millisecond, which means for a number
of these workloads customers will be able to save up to 70%. Customers have really loved
this event-driven computing model. They say, “Look I love this.
I don’t want to actually just do it for these use cases where I spin up
a little bit of Compute. I actually would like to have
my architecture for applications run that way.” And so we started actually
building other services that allowed you
to run serverless end to end. These are services like API Gateway
and EventBridge and Step Functions. And then we added triggers
in a lot of our AWS services so that you could actually
trigger serverless actions from those services.
We have them now in 140 AWS services, which is seven times more
than you will find anywhere else. And just to give you a sense
of the type of momentum that you are seeing with serverless
right now, if you look inside Amazon and you look at all the new
applications that were built in 2020, half of them are using Lambda
as their Compute engine. That’s incredible growth. Hundreds of thousands of customers
now are using Lambda. And so Lambda is growing
really quickly. Serverless is growing really quickly.
Containers are growing really quickly, and not surprisingly
a lot of customers are using both containers
and serverless together. And they’ve said, “I really wish that
you would make it easier for us to run these two smaller units
of Compute together than you make it today.” And they asked for a couple
of different things. The first thing they said was
“Look, our usage and adoption in containers is a little bit
earlier than it is for serverless” - in part because it looked
a little bit more like instances and the tools were set up
a little bit better than they were
for serverless early on. But we’re growing like gangbusters
on serverless and Lambda. And so they said,
“Look, we’ve invested so much time and energy
into these container images that we deploy from,
why can’t you just make it so we can deploy Lambda functions
from these container images?” So I am excited to announce today
the launch of Lambda Container Support, which lets you build
Lambda-based applications using existing container
development workflows. [applause] So now you can
package code dependencies as any Docker container image
or any Open Container Initiative
compatible container image, or really any third-party
based container image, or something that AWS has maintained
as a base container image. It totally changes your ability
to deploy Lambda functions along with the tools
that you have invested on containers. So I think customers
are going to find this very handy. Another big challenge
they asked us to try to help solve is this issue of trying
to manage the deployment of the smaller units of Compute
where you end up in these situations where you have all of
these microservices that have to be deployed together
that comprise an application. And it’s actually quite difficult
to do if you think about it -
it’s different than instances. With instances, typically you
build it as a single block of code. You have code templates you use
like CloudFormation to provision the infrastructure as
code, or services like CodePipeline that do the CI/CD for you or you
do monitoring with CloudWatch. Once it’s set up it doesn’t change
that much. And code is usually maintained as kind of a single release,
so it stays pretty coordinated. And there are tools today that
make this pretty straightforward. But if you look at containers
and serverless, these apps are assembled
from a number of much smaller parts that together comprise
an application. It’s actually hard, you know, if you look
at each of these microservices they have their own code templates and they have
their own CI/CD pipelines. They have their own monitoring and
most are maintained by separate teams and it means that there’s all
these changes happening all the time
from all these different teams. And it’s quite difficult
to coordinate these and keep them consistent -
and it impacts all sorts of things including quality and security. And so there really isn’t anything
out there that helps customers manage this deployment challenge
in a pervasive way. And so this is something that
our team has thought a lot about. And I am excited to announce
the launch of AWS Proton, which is the first fully
managed deployment service for container
and serverless applications. [applause] So this is a game changer
for managing the deployment of microservices.
Here’s how it works. A central platform team
or anybody central to an application will build a stack. And a stack is really a file
that includes templates that use code to define and configure
AWS services used in a microservice, including identity,
and including monitoring. It also includes a CI/CD pipeline template that defines
the compilation of the code and the testing
and the deployment process. And it also includes a Proton schema
that indicates parameters for the developers that they can
add, things like memory allocation, Or a Docker file,
or something like that. Basically, everything that’s needed
to deploy a microservice except the actual application code. Then the platform -
the central platform team - will publish the stack
to the Proton console - pretty often they’re going
to publish lots of stacks because there’s so many different
use cases for microservices. And then when a developer is ready
to deploy their code they will pick the template
that best suits their use case, plug in the parameters they want,
and hit “Deploy”. And Proton will do all the rest. It provisions the AWS services
specified in the stack using the parameters provided. It pushes the code through the CI/CD
pipeline that compiles and tests and deploys the code to AWS services. and then sets up all
the monitoring and alarms. Proton also lists in the console all the downstream dependencies
in a stack. So if the central engineering team makes some kind of change
to that stack, they know all the downstream
microservices teams that need to make those changes and can alert them and can track
whether or not they made it. This is a game changer with regards
to deploying containers and serverless apps.
We’re very excited about this and I think our customers
are going to be as well. So think about ten years ago,
if you flash back ten years ago, the CPU and GPU providers effectively
didn’t have any competition and it meant that you were able
to get new instances every two to three years or so.
Look at today. Look at what Graviton
is doing with 40% better price performance on your most
recent generations of x86 processors. It totally changes
what you can get done Compute-wise. Or look at what’s happening
with Inferentia and some of the machine learning training chips
that I talked about earlier. You are going to get new instances
that allow you to reinvent your business every few months
now instead of every few years. Ten years ago, we weren’t really
talking very much about containers or serverless at all.
Look how fast those are growing. And if you think about it, if you
think about these container offerings and how many you have
available on AWS and the ability now to be able to
manage them the same way both in AWS as well as on-premises, and you think about what’s
happening with serverless and you think about the tools now
that are allowing you to use these new smaller units
of Compute and to be able to deploy
much more easily, there is an incredible amount
of reinvention happening in Compute. We are not close to being done
reinventing in Compute. So you can expect
a lot more to come. Now Compute is obviously
being reinvented, but so is data and data stores
in a very big way. And to share some thoughts about
what he is seeing in this space, and he has been there since
the very start of the cloud, it’s my pleasure to welcome the CEO
and Founder of SmugMug, Don MacAskill. [music] The launch of Amazon S3 in 2006 was
a seminal moment for data storage. This invention totally changed
our trajectory at SmugMug. Today we use almost everything
AWS has to offer. They’ve continued to add more and
more data stores and analytics tools that have let us
really pick the best tools for the best job
to serve our customers. Tools like Athena, Redshift,
Elasticsearch, Kinesis, and DynamoDB, which allows us to scale
and serve photos at billions of requests per day.
The cloud has totally reinvented how we store, secure, analyze,
and share data at a scale that we couldn’t have
even imagined 14 years ago. [applause] Thank you, Don. SmugMug was the very first
big AWS customer and I remember
in the early days of AWS - this would have been March 2006 -
shortly after we launched S3. Basically, our only sales-person
at the time, Rudy Valdez, wrote us an email and he said,
“I just got a call from SmugMug and they are going to put
6 terabytes of data in S3.” And we said,
“Wait, terabytes or gigabytes?” And he said, “No, terabytes.”
We said, “With a T or with a G?” He said, “No, no, with a T.”
We just couldn’t believe it. But SmugMug and Don have seen
the rise of the cloud and the change in data
and data stores. And in fact have been a big piece
of informing what we’ve built. They’ve given us great feedback and
we really appreciate the partnership. And Don is correct in explaining
that data and data stores is radically being reinvented
as we speak. And if you think about it today with the way that the cloud has made
storage so much less expensive, and then Compute to do something with
that storage so much less expensive, it’s astounding how much data
is being created and stored today. Just a couple of data points: Analysts
say that in every hour today we’re creating more data than what we did
in an entire year 20 years ago. Or they predict that
in the next three years there will be more data created than
in the prior 30 years combined. This is an incredible amount
of data growth and the old tools and the old data stores that existed
in the last 20-30 years are not going to cut it
to be able to handle this. Every single type of data store
is being reinvented and will be reinvented
multiple times over. Let me give you a few examples.
Let’s start with block storage. So if you look at block storage, it’s a foundational and pervasive
type of storage used in computing. Unlike object storage which has
metadata that governs the access and the classification of it. Block storage has its storage
or its data split into evenly sized blocks that just have an address
and nothing else. And since it doesn’t have
that metadata it means that the reads and the writes
and access to them are much faster. And it’s why people use block storage
with virtually every EC2 use case. It’s also why that probably
the very most animated debate that we had as an AWS team
before we launched AWS was: Could we launch EC2
with just direct-attached storage that was ephemeral? Or could we not launch EC2 until we
had high performance block storage? And it was a very animated debate,
lots of opinions and we ultimately kind of figured out that it was going to take us
two more years to build to a high-performance block store
and we didn’t wait to give you EC2. So we launched EC2 with just
that direct-attached storage. But we hustled like heck
to get to building a block store and we launched our Amazon
Elastic Block Store, or EBS, in 2008. It was the first high performance
scalable block store in the cloud and it gave you an easy ability
to provision what storage you needed and what IOPS you needed
and what throughput you needed. And then you could adjust
as you saw fit. In 2014, we built the current version
of our general-purpose volume. It was called GP2
which is what the vast majority of EBS workloads run on top of. Every
imaginable workload you can imagine. The feedback that we’ve gotten
the last year or two from customers is that, “We love GP2.
But if we had a wish list there’s a couple of things
that we would like from you.” We’d like one, of course, we’d like
the cost per gigabyte to be less and then we want
to be able to sometimes scale throughput or scale IOPS without also having to scale
the storage with it, which is what GP2 asks you to do. So the team has been working on this
and I’m excited to announce today the new version of our general
purpose volumes, gp3, which allows you to have 20%
better cost per gigabyte as well as be able to provision IOPS and throughput separately
from storage. [applause] So the baseline performance of these
new gp3 volumes is 3,000 IOPS and 125 megabytes per second,
but you can burst that and scale that up to a peak
of 1,000 megabytes per second, which is four times that of GP2. And you’ll see that customers
will be able to run many more of their demanding workloads
on gp3s and they even work for GP2. Yet, there are certain types
of use cases where you need a lot more IOPS
than you get with gp3s. Cluster databases
are a good example of this. And that’s why we built
our Provisioned IOPS or io2 volumes for EBS which give you four times
the amount of IOPS of a gp3 volume. And you can, of course
if you start with gp3s, and it turns out you think
you need io2s, you can use
our elastic volumes feature to seamlessly move that volume. But it’s also true, while people
run their most demanding IOPS applications
on top of io2s, that there are still workloads
that are even more demanding. Some of the most
demanding Oracle databases are SAP HANA databases and offerings where people need in the neighborhood
of 200,000 IOPS or 3,000 to 4,000 megabytes
per second of throughput. And it’s true that you can stripe
together a number of these io2 volumes to give you more collective
IOPS of throughput but the more that you have
to stripe together, the harder it is to manage
and to keep consistent and to get the performance you want.
And so what customers have said is, “Look, what you’re forcing me to do
for these most demanding workloads, is you’re forcing me to use
these storage area networks, or SANs, which really they’re
a bunch of clustered discs with networking attached to them. You’re forcing me
to run these SANs on-premises and I don’t want to have to run these
SANs because they’re expensive.” If you get a good deal on a SAN
it’s about $100,000, but then when you factor in support
and maintenance, and maintaining it across
multiple data centers, you get to $200,000 pretty quickly and if it turns out you exceed
the capacity of a SAN, you have to buy an increment of another couple
hundred thousand dollars. So people don’t like the cost,
but it’s also hard to manage. You have to update
and maintain the software, you have to do the same thing
with the hardware, you have to manage it
in data centers. You have to make sure
it’s fault-tolerant. The customers have said,
“Look, you’ve left me with no option because there is no SAN
in the cloud.” Until now. I’m excited to announce
io2 Block Express, which is the first
SAN built for the cloud. [applause] So Block Express volumes
give you up to 256,000 IOPS, 4,000 megabytes per second,
64 terabytes of storage capacity. That’s 4X the dimensions of the io2s
on every single one of those. That is massive for a single volume. Nobody has anything close to that
in the cloud and what it means is that you now get
the performance of SANs in the cloud, but without the headaches
around cost and management. You just spin up
a Block Express volume, we manage it and we’ll back it up
and maintain it for you. We’ll replicate it across an AZ. You can use the EBS snapshot
capability to auto lifecycle, a policy to back it up to S3,
and if you need more capacity, you just spin up another
Block Express volume at a much lower cost than trying to do it
at $200,000 a clip. So, we will add additional
SAN features in 2021, things like multi-attach
and IO fencing and make elastic volumes
work with it. And this is pretty exciting. There was a lot of very complicated,
sophisticated, innovative engineering
to give you Block Express, but it’s a huge gamechanger for your
most demanding applications that you want to run in EBS.
So block stores are being reinvented. What about databases? Databases as you all know are right
in the middle of every application and they’re hard to manage. You have to set them up,
you have to tune them, you have to patch them,
you have to make sure you get the right fault
tolerance and performance. And it’s why companies have
so many database professionals that they have to hire. It’s also why we built
our relational database service, which is our managed
relational database service which we launched about ten years ago
and has been wildly popular. But if you look at, despite
the growth of things like RDS, it’s still true that the overwhelming
majority of relational databases live on-premises
and they live largely with these old-guard
commercial grade database providers named Oracle
and Microsoft with SQL Server. And this is an unhappy place
for customers. We’ve talked about this
for several years. This is why you’re seeing
so much movement. But it’s unhappy because
those offerings are expensive, they’re proprietary, they have
high amounts of lock-in and those companies
have punitive licensing terms where they’re willing to audit you
and if they find any discrepancy, they extract more money out of you. And these companies
also have no qualms about changing the licensing
terms mid-stream on you. If you just look at what Microsoft
did with SQL Server in the last year or two. They basically changed the terms
so you can’t use your SQL Server licenses anywhere
but Microsoft’s cloud. Is that good for customers?
Hell no. Is that good for Microsoft? I think they think so.
I think it’s short-term thinking because our experience
over the fullness of time is that customers flee companies
the first chance they get when they feel like
they’re being abused. And this is something
that customers are fed up with and they’re sick of -
and it’s why they’re moving as fast as they can to these open engines
like MySQL and Postgres. But to get the type of performance you get in
a commercial grade database in these open engines,
you can do it, but it’s hard work. We’ve done a lot of it at Amazon
and our customers asked us to fix that problem for them
and that’s why we built Aurora. And Aurora has 100% compatible
versions with MySQL and Postgres. It has several times
better performance than those community grade versions. It has at least the same fault
tolerance and durability and availability as the commercial grade databases,
but at one tenth of the cost. This is why customers
have been flocking to Aurora as quickly as they have. It’s the fastest growing service
in the history of AWS. It has been since its launch, and you see that we have over
100,000 customers now using Aurora. These are companies like Airbnb
and AstraZeneca and BP, and Capital One, and Fannie Mae,
and Petco, and Verizon, and Volkswagen. One of the great things
though about having a service that people love so much
and is growing so fast is that you get a lot of feedback on what else people would love you
to build and that is fuel for us. Please keep it coming.
That is how we choose what to build. And a lot of our Aurora customers
said, “Look, we love Aurora. If we had a big ask it would be that
we’d like to be able to run Aurora taking advantage
of that serverless architecture where we didn’t have to think
about servers and clusters.” And so that’s why we built
Aurora Serverless, which is really
an auto-configuration for Aurora which allows you to set up
an Aurora database and then when you need to scale up
because of capacity, we’ll scale it up in 5 to 50 seconds, usually doubling the capacity
each time you need more capacity. And we have thousands of customers
who’ve been using Aurora Serverless primarily for dev and test workloads. And they said,
“Look, we want to run… We love it. We want to run it
for production workloads, but we need some things from Aurora, what we’d normally get from Aurora
if we’re going to do that.” And they said, “We need Multi-AZ.
We need Read Replicas and then when we need
to scale up capacity-wise, we need it to happen instantaneously.
We can’t wait 5 to 50 seconds, and we only want to scale up
in the precise increment that we need to scale up.” So the team went away and started
working on that this year. I’m excited to announce the launch
of Amazon Aurora Serverless v2, which allows you to scale to hundreds
of thousands of transactions in a fraction of a second. [applause] And so Aurora Serverless v2
totally changes the game for you with serverless
as it relates to Aurora. You can scale up as big
as you need to instantaneously. It only scales you up in the precise
increments that you need to. So if you’re using
Aurora Serverless v2, you can save up to 90% versus
provisioning Aurora for the peak. It adds in a lot of the Aurora
capabilities people want - and Multi-AZ and Global Database and Read Replicas
and Backtrack and Parallel Query - and it really makes
Aurora Serverless v2 ideal for virtually every Aurora workload. MySQL is available for you now
with Aurora Serverless v2 and Postgres will be available
in the first half of 2021. Now I ask you, how many of these
old-guard commercial grade database companies would build
something like Serverless v2 that’s clearly going to take
a meaningful amount of revenue away from their core
database offering? I wouldn’t hold your breath.
I think the answer is none of them. They’re just not built that way. But we have a different way
of thinking about our business, which is that we’re trying to build
a set of relationships and a business
that outlasts all of us. And the best way we know of doing that is listening
to what customers care about. And if we can help you build more,
more efficiently, more effectively, change your experience
for the long-term, even if it means short-term pain
for us or less revenue for us, we’re willing to do it because we’re
in this with you for the long haul. I think that’s one of the reasons
why customers trust AWS in general and have trusted us in the database
space over the last number of years. If you look at it,
just in the last few years, we’ve had more than
350,000 databases migrate to AWS using our Database Migration Service and the pattern that usually follows,
they say, “Look, we used the Database Migration
Service to move our database data. We used your Schema Conversion
Tool to convert the schema, but there’s a third area
that’s making this harder than we wish it were,
that we want your help with, and that’s trying to figure out
what to do with the application code that is tied
to that proprietary database. And so customers have asked us, “Can you do something
to make this easier for us because we want to move
these workloads to Aurora?” And especially with the way
they’ve watched Microsoft get more punitive and more
constrained and more aggressive with their licensing, they want help. So the team has been working on this
for a little bit more than a year and I’m excited to announce today the launch of Babelfish
for Amazon Aurora Postgres, which lets you run SQL Server
applications on Aurora Postgres with little
to no code changes. [applause] So Babelfish is a new translation
capability that lets you run SQL Server
applications on Aurora Postgres. And what Babelfish
does is it understands Microsoft SQL Server’s T-SQL dialect
and it creates a tabular Data Stream or a TDS endpoint
for your app to connect to, and it understands Microsoft’s
proprietary schemas. And what it means for you is that now
you can use the Database Migration Service to move your database data,
the Schema Conversion Tool to move your schema or convert
your schema, and then you can use Babelfish to update
your application configuration to point to Aurora Postgres
instead of SQL Server and you get to shed those expensive
and constraining SQL Server licenses. Because Aurora now is able
to understand both T-SQL and Postgres, you can write
application functionality in Postgres to run side by side with your legacy
SQL Server code. And so customers that we’ve spoken
to privately about this, to say they’re excited would be
one of the more large understatements I could make. Very excited.
And they were so excited about this that as we were preparing
for re:Invent and discussing it, we realized that this was probably
bigger than just Aurora Postgres. That people really wanted the freedom to move away from
these proprietary databases and to Postgres which is where
most people move from these. And so we decided that
we’re going to open source Babelfish. And so I’m excited
to announce Babelfish for Postgres, which is an open source project. [applause] So Babelfish for Postgres will use
the permissive Apache 2.0 license. It means that you can modify
or tweak or distribute in whatever fashion you see fit. All the work and planning
is going to be done on GitHub so you have transparency of
what’s happening with the project. So this is a huge enabler
for customers to move away from these
frustrated, old-guard proprietary databases
to the open engines of Postgres. You can sign up for Babelfish
for Aurora Postgres today and then you’ll be able to sign up
for the open source project in 2021. So block stores are being reinvented, relational databases
are being reinvented. Heck, actually all of databases
are being reinvented. If you think about it, we’ve been talking about this
for the last few years. That world that you lived in
for 20 or 30 years where you used a relational database
for every single workload, that time has come and gone. If you have data in volumes that are
gigabytes and sometimes terabytes, you can get away with using
a relational database for everything. It’s not ideal,
but you can get away with it. But in a world where you’re now
dealing with terabytes of data and petabytes of data,
and sometimes exabytes of data, a relational database
doesn’t make sense. It’s too expensive,
it’s too complicated, and it doesn’t perform as well
as purpose-built databases that do a particular workload
or use case extremely well. And that’s what we’ve been
working on the last several years. We have built seven of these
purpose-built databases, more than you’ll find anywhere else
by a fair amount, that allow you to have
the right tool for the right job. So, I’ll give you some examples.
If you’re a company like Lyft and you have millions of geolocation
and driver combinations, you don’t want a relational database. It’s too complex, it’s too expensive,
it’s not performing. You want a high throughput,
low latency, key-value store like DynamoDB,
or if you’re a company like Peloton, you want to show your dashboards
in microseconds, you want an in-memory database
like ElastiCache. Or if you’re a company like Nike
and you’re trying to connect all these different graphs
of information that have relationships
attached to them, you want a graph database
like Neptune. Or if you’re doing work
at the edge with IoT where the data’s coming
in a timestream format, you want a time series database
called Timestream, which we have. If you want to run managed Mongo,
if you want to run managed Cassandra, you want those databases that allow you to have the right tool
for the right job. I think you’re seeing
the same exact thing happening with purpose-built
analytics stores. So if it turns out
that you want to do querying directly on your data lake,
or in S3, you use Athena. If it turns out that you want
to actually process vast amounts of unstructured data
across dynamically scalable clusters using popular distributor frameworks
like Spark or Hadoop or Presto, you use EMR. If you want to do large-scale
analytics on log data for operations you use
our Elasticsearch service. If you want to do real-time
processing and streaming data, you use Kinesis,
and if you have structured data where you need super-fast
querying results, you want something like
a data warehouse; You want Redshift, which was the first
data warehouse built for the cloud - continues to be the largest data
warehouse in the cloud - and they continue to innovate
at a rapid rate. Last year at re:Invent you saw
the RA3 instances, which separates storage from
the compute, which people have loved. We’re just around the corner from
the general availability of AQUA which moves the compute
to the storage and will give you
10x better query capabilities in terms of speed than anywhere else. You want these
purpose-built databases and these purpose-built
analytics stores. And we see customer flocking to them. Now, it’s actually brought up a
really interesting challenge and question for customers which is: Most companies
either have a data lake or will build a data lake to take all
that data from disparate silos and move it together, so you have one place
where you can do your analytics and your machine learning from.
And most of those are built on S3. We have tens of thousands
of data lakes built in S3, more than you’ll find anywhere else
because of the security and reliability and governance
and compliance capabilities, and the broad features
and the flexible and cost-effective performance of S3. So,you can have
these data lakes that you centralize all your data, so you can run analytics
and machine learning from, but as we just talked about,
you’re increasingly seeing customers using more and more
of these purpose-built data stores. And so customers say,
“Well, I want my data to be able to move back and forth
between these different stores because it’s very useful to take
some of these views that I have and use them in other spots.” And while we have capabilities
in many of these services that let people move them
back and forth, it’s really not easy enough
for people to do, such that they do it
in a pervasive way. So customers really, really want more freedom
for their data to move around. Our team’s worked on this
for the better part of a year, and I’m excited to announce today
the launch of AWS Glue Elastic Views, which lets you easily
build materialized views that automatically can find
and replicate data across multiple data stores. [applause] So, Glue is AWS’s ETL service,
and you can write SQL or you can use our Glue studio
visual tool to do extracting
and loading and orchestration. And we recently a few weeks ago
launched something called DataBrew to make it easy to clean
and normalize data with no coding. But Elastic Views is different. What it lets you do
is it lets you write a little of SQL to create a virtual table
or a materialized view of the data that you want to copy and move
and combine from source data store to a target data store. Well in the old days you’d have
to figure out a way to write code and make sure you can get
all the DynamoDB attributes and figure out how to move that,
and move it in sync, keep it up-to-date
as the attributes changed and new ones came in
or things got adjusted in any way. All of that muck is taken away
by Elastic Views. And so what Elastic Views does is
it allows you to set up a materialized view
to copy that data and move that data
from one of those source data stores to a target data store
and then it manages all of the dependencies
of those steps as they move. And if something changes
in the source data store, Elastic Views takes that
and automatically changes it in the target store in seconds.
If it turns out, for whatever reason, the data structure changes
from one of the data stores, Elastic Views will alert the person
that created the materialized view and allow them to adjust that. This is a huge game changer
in being able to move data. It takes a lot of the work
that people had to do – and frankly just found it
so much work that they rarely did. You know, when you have
the ability to move data and to have purpose-built data stores
and have a data lake, but also move that data easily
from data store to data store, there is a lot of power in giving
that freedom of movement of data and this is going to be
a big game changer for customers, having Elastic Views.
We’re very excited to give it to you. A company that I think is doing
one of the most revolutionary and transformational
things around today and that’s also using data and
computing in very innovative ways – is what’s happening at a company
called Boom, which is trying to build the first
supersonic airplane in 60 years. And to share with you
how they’re doing that and what they’re doing with data
and compute on top of AWS, it’s my pleasure to welcome
the founder and CEO of Boom, Blake Scholl. [applause] [music playing] [music] So what does a revolution
in high-performance computing have to do with a revolution
in high-performance aircraft? What does cloud computing have to do with how we fly
through actual clouds? And why now is
a software engineer building the next commercial
aircraft company? Air travel is integral
to our modern lives, yet aviation is a domain
starving for step-change innovation. The last big new thing
was the jet airliner invented more than six decades ago. Tokyo to Seattle has been a nine-hour
flight for going on 60 years. At Boom, we are guided
by one fundamental mission – to make the world
dramatically more accessible. And by the end of the decade,
millions of everyday travelers will enjoy the benefits
of supersonic flight aboard Overture, an airliner
twice as fast as any flying today. That means that Tokyo will be just
four-and-a-half hours from Seattle, and London
just three-and-a-half from New York. Speed unlocks new possibilities
for human relationships and business connections. That’s why major aerospace players
such as Rolls Royce, Japan Airlines, and the United States Airforce
are among Boom’s partners. All of this is possible
thanks to advances in computing that enable
a startup company to spark a revolution in speed. Twenty years ago, I hit two
life milestones as a new graduate. I started my first job here
at Amazon as a software engineer, and I started taking flying lessons just a few miles down the street
at Boeing Field. Around the same time, Amazon was
building the fundamental web services that would later become AWS, charting a new future of
computing across all industries. Little did I know
that these things would intersect so powerfully in my future. I never could have predicted
that the innovations in cloud computing happening at Amazon and AWS would fundamentally change
how we all fly. Well fast forward to today
and Boom is designing and building the world’s fastest
and most sustainable airliner, and we’ve just announced that
we’re going all in with AWS. Why? Well it turns out that high
performance computing is key to this
new era of airplane design. And AWS is the leading
cloud provider, allowing us to leverage a wide range
of capabilities and services. Plus, Amazon’s relentless focus
on the customer means that some of
the best minds in cloud computing are helping Boom innovate
and get to market faster. AWS levels
the playing field in aerospace, allowing a startup company
to develop what previously only big companies
or governments could do. So this is how airplanes were
designed before the age of computing. Engineers worked with drafting paper
and slide rules. They built scale models
to test in wind tunnels leading to a process of iteration
that was slow and costly. Today, to design faster airplanes, we need the fastest computers. And
computational methods leveraging AWS save us literally years
of schedule and millions of dollars. Moreover, because we can now test
many designs quickly and inexpensively,
we can deliver a better airplane. In October, we rolled out
XB-1, history’s first independently developed
supersonic jet. To design XB-1, we leveraged EC2
to stand up HPC clusters often with more than 500 cores -
hundreds of possible airplane designs flew through virtual
wind tunnel tests encompassing
thousands of flight scenarios. Because AWS allowed us to run many
hundreds of these simulations concurrently, we achieved a sixfold increase
in team product productivity. Simply put, without AWS today we would be looking at a sketch
of a future airplane concept, not an assembled jet because years
of design work would still remain. Since airplanes are amongst
the most complex machines ever created by humanity,
Boom will generate petabytes of data as we design and develop
our Overture airliner. Already we are transferring
525 terabytes of XB-1 design and test data to AWS. Because they let us
put compute next to data, we can run models across our dataset
gaining actionable insights. For example, we’re using machine
learning to calibrate simulations to wind tunnel results,
accelerating model convergence and allowing us to deliver
a more optimized aircraft. All in all, we have used
53 million core hours in AWS to design and test XB-1. And in the same manner we expect
to use over 100 million core hours as we finalize the design
of our Overture airliner. Because our pursuit of speed is about
making earth more accessible, we’re taking great care
to build an environmentally and socially responsible
supersonic jet, and I am proud that Overture
will be 100% carbon neutral from day one thanks to its use
of alternative fuels, which means that supersonic flight is going to be more
affordable than ever before. Well, great revolutions
and fundamental technologies enable benefits that are difficult
to predict or even imagine. Think of how many industries AWS has already transformed,
and today supersonic flight is one of those surprising
benefits of computing. Just as AWS is reinventing computing,
at Boom we are reinventing travel. So what further breakthroughs
will be sparked by a revolution on how we fly? At Boom our vision
is one of accessibility, of new possibilities
unlocked in the world around us. By the end of the decade
your flights will be cut in half. So, what will you do when Australia
is as accessible as Hawaii is today? And your flight is completely
carbon neutral? How might you transform
an industry? What new people could you
come to call friends? But Overture is merely
the first step towards our vision
of a supersonic future. Because I dream of a day where you
can go anywhere in the world in four hours for just $100. Where the fastest flight
is also the most affordable. Boom is making supersonic
flight mainstream and AWS is helping us deliver
on that promise. I am so excited to see
what you will invent when more of the world
is within your reach. Thank you. [applause] Thank you, Blake.
I’ll tell you what I'm going to do when Australia
is as accessible as Hawaii – I’m going to go to Australia more.
It’s incredible what Boom is doing. We often talk at Amazon
about thinking big. That is thinking big. It’s very
impressive, it’s very exciting. It could really change life
for all of us and we’re honored
to be partnering with Boom. You know, I remember a time
about 25 years ago or so, people thought
that search was boring, and there was really
nothing left to invent. And I remember a time
maybe 15 years ago where people felt like technology
and infrastructure was boring, and there wasn't much left to invent
and those turned out to be wrong. And I think oftentimes people
think of data stores and databases as being boring
and what is there left to invent. And I hope you can see
the answer’s a lot. You see that block stores
are being reinvented. You see that relational databases
are being reinvented. You see that purpose-built data
stores are being reinvented. And then the movement of that data
between those stores that frees up the power of what you can build
with Elastic Views… huge amounts of reinvention
happening with data stores. Now hand-in-hand with data
is what has been happening with machine learning
and the adoption of it. And to start us off on this section
I’m going to transition it over to the CTO of Intuit,
Marianna Tessel. [music playing] Intuit is an early adopter of AI, starting the journey
over a decade ago. We see tremendous opportunity
in applying AI to revolutionize our business
and to benefit our customers. And given the potential,
this is just the beginning. Working together with AWS,
we developed a robust ML platform that empowers our engineers
to incorporate AI into our products. We use AWS tools
for model development, training, and hosting, and integrate
our own capabilities for orchestration
and feature engineering. This has been game changing. Together with AWS
we made great strides in driving AI/ML
innovation with speed, helping us deliver
smarter products faster to more than
50 million consumers, small businesses, and self-employed
customers around the world. [applause] Thank you, Marianna. It is really interesting
when we talked about earlier about being able to build
the reinvention culture. Intuit has done that.
And you can see it, not only with how they have
been more sophisticated and more leaning forward
with respect to using the cloud for their technology
infrastructure, but also with respect
to what they are doing with machine learning where they’re well
ahead of most companies. So people have been talking about
machine learning for over 20 years. In the cloud,
with cost structure around compute and then the amazing amount
of capacity you have, has made machine learning
much more practical. And while more
and more companies - amazing progress - have started
using machine learning, make no mistake about it, it is very early in the history
of machine learning and almost everything
is continually being reinvented. There’s loads of examples. I’ll just give you
one simple one here. So this is the top frameworks
that have been used in new machine learning scientific publications
in the last five years, which is often a leading
indicator of what people use. And you heard us
talk a few years ago when TensorFlow was so dominant
in the frameworks used that the one constant we see
in machine learning is change. And you can see that. Look at how PyTorch has caught
TensorFlow over the last few years, because it’s much easier to use
and to get started with. And when we talk to machine
learning practitioners, 90% of them tell us
they use more than one framework and 60% use more than two. And so clearly you can see
that machine learning is in the very early stages. Frameworks was one of the most
stable things a few years ago, but you can see it’s changing
very substantially. Now we have a lot more machine
learning capability than anybody else
by a fair bit, and it’s one of the reasons
why last year in my keynote the section on machine
learning was 75 minutes. I’m not going to do that this year.
For the first time, we have broken out a machine learning keynote that
Swami’s going to do next Tuesday. I’m going to leave a lot
of the goodies for him. But I have a few
that I’m going to share and I’m going to really
frame them in the context of the asks we get
from our machine learning customers. The first ask we get is they say, “Look, we want to have
the right tools for expert machine learning practitioners,” and that typically involves
chips and frameworks and this is the group
of people by the way who operate at that bottom
layer of the machine learning stack. They’re comfortable
building and training and tuning and deploying machine
learning models of scale. And we talked earlier
about all the areas we’re trying to help
our expert machine learning practitioners with chips
around inference and machine
learning training, but I also think it’s important
to think about the frameworks. We took an approach
a few years ago where we said we’re going to support
every single one of the major machine learning frameworks and we built…
that was a unique approach. Everybody else was just
focused on TensorFlow. And what we did was
we built separable teams. We have one team that’s just focused
on how to optimize TensorFlow in AWS, one team that’s just optimized
on how to run PyTorch in AWS, and one that’s just focused
on MXNet in AWS. And that’s why you get
the best performance in all those frameworks on AWS. I usually show you
benchmarks at this point. I’m going to leave that
to Swami for next week. But we believe that we’re not done
seeing new frameworks pop up that you’re going
to care about. And the commitment you have from us
is that we will continue to support every single
one of those major frameworks because we know that expert
machine learning practitioners want the flexibility
to build however they see fit. The second ask we get is, “Look, the reality is there aren’t
that many expert machine learning practitioners
in the cloud and there aren’t
that many in the world, and those that exist tend to live at
the big technology companies.” So, if we want machine
learning to be as expansive as we all believe it should be, you’ve got to make it easier
for everyday developers and data scientists to be able
to use machine learning. And that’s why we built
a few years ago SageMaker, which sits at that middle
layer of the machine learning stack. And SageMaker was a step-level
change in the ease with which you can build,
train, tune, and deploy
a machine learning model. We have tens of thousands
of customers who are standardizing on top
of SageMaker and using it. And these are companies
like 3M and ADP and Cerner and Intuit and GE and Snap
and the NFL and T-Mobile and Vanguard,
just a broad group. And one of the things
that people like most about SageMaker is they see how quickly
SageMaker’s continuing to iterate. This is the second year in a row
that the SageMaker team has added over 50 new features
in the last 12 months. That’s like one a week.
That’s really unusual. And we haven’t stopped. If you think about last year
at re:Invent, I announced in my keynote
the launch of SageMaker Studio which was the first
Integrated Development Environment, or IDE,
for machine learning. And we gave you Notebooks, which were easy to create
in one click and share. And Debugger, which made it easy
to debug your models. And Experiments, which saved all
the artifacts of your experiments so you could kind of figure out
what happened and share them. And Model Monitor, so you could tell
if you had model drift, and Autopilot which looked at automatically
what was in a CSV file and created a new model for you, machine learning model
automatically with transparency and how it was created,
so you could pick it up when you were done seeing it,
if you wanted to evolve it yourself. That was a huge amount
of new innovation last year and customers have loved
SageMaker Studio. And they’re using it
in a really expansive way. And what we often do, again we have services
that have this amount of traction. We get a lot of gratuitous feedback
from customers which we love, but we also constantly
ask customers, “What else can we build for you
that would make your life easier?” And the topic that seems to come up
first and foremost almost every time is, “How can you make doing data
preparation for machine learning much easier?” And data preparation is hard.
If you think about it, to build a model you’ve got
to collect all this data from different sources
that come in different formats. None of it’s normalized
which you need for the models. And when you’re building
these models, you need to look at all
these different dimensions, what they call machine
learning features. I’ll take something
like a real estate app. If you want to predict
the prices of real estate, you need to look at features
when you build a model - like how many bedrooms or how many bathrooms
or how many square feet is the house or what are other houses
on that street selling for? Does it have a Starbucks
within five miles? These are all features that you need
that arrive in different formats that the model cannot understand, where you have to convert
these features into the right format that the model can understand. This is what is called
feature engineering. And then there are also,
by the way, times where you want to take
two different features and combine them, or two or more
different features and combine them. Take something like,
in a real estate example, you want a house pricing index where
you’re combining different features that make
the model more efficient. Converting these models,
converting these features, combining features themselves – that type of feature engineering
is really hard and it takes a lot of time.
You’ve got to write queries and code to get all that data
from various sources. You have to convert the data
to the format the algorithms can use. You sometimes want to actually
combine the features. You want to prototype
these features to see if your feature engineering worked or not before you actually apply
those transformations everywhere. And then you’ve got to apply
the transformations and make sure you don't have
missing data or outliers. It's just a lot of work. And people said there must be
an easier way, which is why I’m excited
to announce today the launch
of Amazon SageMaker Data Wrangler, which is the fastest way
to prepare data for machine learning. [applause] So Data Wrangler is a total
game changer with respect to speed of doing
machine learning data preparation. The way it works
is you point Data Wrangler at the appropriate AWS data store
or third-party data store and then Data Wrangler has over
300 built-in conversions and transformations
that, through machine Learning, will automatically
recognize the type of data coming in and suggest the right transformation
for you to make that you can apply. You can of course
do your own thing as well. It also makes it much easier
in the Data Wrangler console to combine features to build those
composite features I was mentioning. You can preview very easily these
transformations in SageMaker Studio, and then if you like what you see,
you simply apply that transformation to the whole dataset
and Data Wrangler manages all the infrastructure
and all that work under the covers. It’s a total game changer
in the time that it takes to do data
preparation for machine learning. Now, because you spend so much time
on the data prep as well as on
the feature engineering, not surprisingly you want to use
these features in lots of other models, so you need a place
to store these features and make them easily accessible. And this turns out
to be a hard problem. Sometimes you might step back
and say, “Well why is this a hard problem?
Why don't you just store them in S3?” The problem is, you may be able
to do that if you wanted to store as an object maybe
a simple set of features that are mapped to one model, but features are hardly ever mapped
to just one model. They’re usually mapped to lots of
models because they are highly useful, and you did all the work
to get the features in a state where the model can understand them. And then sometimes you’ve got
subsets in that set of features that want to be
their own sets of features. And then you’ve got multiple people
who want to access those features and the different sets
with multiple models, and pretty quickly
it becomes complicated and hard to keep track of, which is why people
often try to keep track of this in an email or spreadsheets
or sometimes build a clunky UI that takes time
to really work very well. You also need to… these same features
need to be used to train models and then also to make predictions. And they’re really
different use cases. When you’re training a model, you’ll use all the data
in a particular feature to be able to get the best
model possible to make predictions. But when you're making predictions,
you often will take just maybe the last five data
points in that particular feature, but they have to be stored
the same way and they’ve got to be accessible. When you’re actually using these
features for inference and predictions,
you need really low latency because you're trying
to make predictions in close to real time
in your applications. So it’s really hard to do all
these things in a generalized store. And so that’s why we have built
and I’m excited to announce today SageMaker Feature Store,
which is a new repository that makes it easier to store, update,
and share machine learning features. [applause] And so the Feature Store, it is a purpose-built feature store
that’s accessible in SageMaker Studio and it makes it much simpler
to name, organize, find, and share features with teams because we’ve built
a purpose-built store for features. It turns out it makes it really easy
for features to actually be accessed
either for training or for inference. Even though they’re different
use cases, we’ve built a store
that makes that simple. And because Feature Store
is located in SageMaker close to where your machine
learning models are running, you get really low latency
on your inference and prediction. So this is another big deal
for people that are trying to build machine
learning models. Now I think you can tell from
understanding what happens in machine learning or looking at data
preparation or Feature Store, that machine learning has a lot
of things that have to happen sequentially
or sometimes in parallel, but really lend themselves well
to orchestration and to automation. And this is true in normal code as well
when you’re building applications. It’s why they’ve built these
CI/CD pipelines. But in machine learning there
is no CI/CD. None of them exist pervasively.
People have tried to build their own. They’ve tried to do it homegrown.
It hasn’t been that scalable, it hasn’t worked the way
they wanted it to. And customers want an easier way
to do this. And so I’m excited to announce today the launch of
Amazon SageMaker Pipelines, which is the first purpose-built
easy-to-use CI/CD service for machine learning. [applause] And so with Pipelines you can
quickly create machine learning workflows
with our Python SDK. And then you can automate
a bunch of the steps from a number of the things you have to do on data preparation
in Data Wrangler - to moving the data from
Data Wrangler to the Feature Store to some of the activities you want
to take once in the Feature Store, to training, to tuning,
to hosting your models. You can build all of these things
in a workflow that happens automatically. And then Pipelines manages
all of those dependencies between each of the steps
and orchestrates the flow so you can have any number of these
automated steps in a Pipeline workflow. We also give you Pipelines
with preconfigured templates for building and deploying so that you don't have
to create these from scratch. You can either use those verbatim,
you can use those as a base and then customize how you see fit, or you can actually
build them from scratch. You can do all of those things
in Pipelines. And then Pipelines automatically
keeps track of all of the artifacts and then keeps an audit trail
of all the changes so it's easy to troubleshoot
if you need to do that. So SageMaker has completely
changed the game for everyday developers and data
scientists in being able to build, train, tune, and deploy machine
learning models. And people have flocked to SageMaker, not only because
there's nothing else like it, but also because of the relentless
iteration and innovation that you continue to see us
applying into SageMaker. It’s not just the launch
of the service. You can look at what we did
with SageMaker Studio last year. You can look at what
we’ve done this year, at least what you’ve heard so far –
there’ll be more. And what we’ve done around data
preparation and Feature Stores and the first CI/CD
in machine learning. We’re not close to being done
innovating here. The third ask that we often get
is customers say, “I would like to be able to use
your machine learning, but I don't want to have
to be responsible for creating the model myself.
I want to send you data. I want to use models that
you have built and trained on data and get the answers back via APIs.” And that’s what we think of
as this top layer of the machine learning stack.
We’ve a lot of services in this area. For people that want to look
at an object and say, “What’s in this?” or a video and say,
“What’s in this?” we have an object - a video recognition service -
called Rekognition. We have services that allow you
to go from text to speech, to transcribe audio to text,
to translate that transcribed audio, to do natural language processing
on all that translated, transcribed audio
so you don't have to read it all and know what’s happening.
It lets you do OCR, but also OCR++ and being able to pull out data
from tables and formats that don’t usually
come out in OCR. We give you the ability
to build chatbots with Lex. We give you the ability
to do internal enterprise search with Kendra,
which we launched last year. There was a bunch of Amazon
capabilities that you asked us to expose as machine
learning services which we have. So we have a personalization
machine learning service and a forecasting service
and a fraud detection service, and a code inspection service
called CodeGuru. And it’s interesting. As you start using
these models more and more, there's that point in every company
where you think about, “How much can I trust
these predictions and these answers to use
in my applications?” And there are certain use cases where
the ramifications are low enough that you can take more chances. Oftentimes in applications that are
doing translation or transcription, if you see good results early on, you’re going to really trust it
as your major input. But there are other use cases where
the ramifications of getting it wrong are significant enough where you’re
going to wait a little bit, and these are services
like facial recognition or forecasting your inventory
or the quality of your code. You’re going to look and make sure
that you like the predictions, that they’re actually high quality,
and you’re going to use those inputs really as one input
in several inputs of a decision. So it’s really machine
aided in that case. And what you’re going to wait for is you’re going to wait
for the predictions to become more
and more commonplace that you say, “Yes, that’s the answer,”
and you want to see that that provider is going to continue
to invest in that service so you know it’s going to be robust
and you can really rely on it and cut the cord
from what you were doing before. Let’s look at an example of this.
Let’s look at CodeGuru. CodeGuru is a service
we launched last year, which is a machine learning service that lets you automate
code reviews to tell if you have any kind of code
that you’ve written that we think is going to lead
to a problem. And it also allows you to identify
your most expensive line of code. By the way, we have over
120,000 apps at Amazon that are using that Profiler part to find the most expensive
line of code, which not only is helping them
find operational bottlenecks, but also saving them
a significant amount of money. And so customers
who are trying to assess, “Can I use…
can I rely on CodeGuru as a major input in how I think about my code?” they say, “Well what’s
your commitment? I understand you’ve got a profiler
or you have a code reviewer, but how about more
programming languages?” So the team launched Python
and a few other languages. Then they said, “Well how about
something around security?” So the team launched
a security detector feature that lets you know
whether or not you have code that would lead
to some kind of security issue. And then people said,
“Well, that’s great. You have so much information about
the way that my application operates and the way that all kinds
of applications operate on AWS. Why can’t you build a service
using machine learning that allows us to predict when there are going
to be operational problems?” And so we went away
and thought about that. And I’m excited to launch
a new service today called Amazon DevOps Guru, which is a new service
that uses machine learning to identify operational issues
long before they impact customers. [applause] And so DevOps Guru makes it
much easier for developers to anticipate operational issues
before it bites them. And so using machine
learning informed by years of Amazon and AWS
operational experience in code, what we do in the service
is we identify potential issues with missing or misconfigured alarms
or things that are… resources that are approaching
resource limits or big changes that could cause outages
or under-provision capacity or overutilization
of databases or memory leaks. When we find one of those that we
think could lead to a problem, we will notify you either
by an SNS notification or a CloudWatch event
or third-party tools like Slack. And we’ll also recommend
the remediation. And so when you look
at these top-level services and you find them super useful,
as you go through this journey that you will go through
as a company in determining when you can trust them
as either the major or a major
input in your applications, you’re always going to want
to get a lot of experience
running them at scale and see the predictions
being more and more accurate, and then also making sure that
the provider that you’re betting on for that service is continuing
to invest in that service and make it easier and easier
for you to use that as that input. And I think that’s what people
are starting to see with CodeGuru. The fourth ask that we get is from
customers who say the following: “It is awesome that you have
so much machine learning capability,
more than I could find elsewhere, but I would like to somehow
benefit from machine learning without having to even know
I’m doing machine learning.” Perhaps my favorite
business school professor, and also perhaps the best
innovation writer of our time, was a fellow named Clay Christensen -
and Clay just passed away. But what Clay used to always
talk about was he used to always say that customers hire products
and services for a job. And in this case what customers
are telling us is, “I don’t want to hire you
for machine learning. I want to hire you
for a particular job and if you use machine
learning to get it done, great, that’s fine either way. But I want to hire you
to get that job done.” And let me give you
a good example of that. If you look at business intelligence,
or BI, of course it started with these kinds of older forms
of this, like Oracle OBIEE, and then really over
the last few years Tableau has done a great job
reinventing this space and building beautiful
visualizations and dashboards. And we’re now at a stage
where BI is being reinvented again. And what people really want -
is they really want their BI service to be serverless, so they don’t have to manage
all that infrastructure. They want the ability to be able
to embed all their dashboards and visualizations
everywhere easily, flexibly. And then increasingly,
they want to be able to use all kinds of machine
learning and natural language. Really what they want
is natural language, whether machine
learning fuels it or not, that allow them to understand
the results more easily than having to do all the work
to glean the insights themselves. And that’s why we built
Amazon QuickSight back in 2016. QuickSight is serverless,
it’s very scalable. It has, not just all the features
you need in a BI service, but it also does embedding better
than you’ll find elsewhere. And then we’ve started
to invest in machine learning that will allow you
to have experiences that get things done
in a much more natural way. So for instance, we have this feature
in QuickSight called Autonarratives and what it means is that when
you actually do a query and you get the results, you don’t have to actually
always do the work to figure out what the results mean. We will provide natural language
what we think the key insight is. Now we do that through machine
learning, but nobody who uses QuickSight has to know anything
about machine learning. And customers
really love that feature. And they’ve said, “I wish you could
use that type of capability in other things
that we have to do in our BI usage.” So for instance, I don’t like that
our users have to know which databases or data stores to get the data from,
or how the data’s structured or how to ask a question
in a certain way or… I want them just to be able to type
into a search bar a natural language question
and get answers. And this is a hard problem. A lot of companies have tried
to take a shot at solving this, usually with natural language query, and they just haven’t been able
to get good results. And it’s really hard
to build the right dataset because there’s millions of questions
and pretty quickly you find you didn’t have all the questions.
You’ve got to go back to IT and they have to build it
when they get a chance. It just hasn’t worked out until now. So I’m excited to announce
the launch of Amazon QuickSight Q, which allows you to ask Q
any question in natural language and get answers in seconds. [applause] And so the way Q works is you
just type in a question in natural language in Q
in the search bar. You can ask a question like, “Give me the trailing 12 month
sales of product X,” or, “Compare product X sales to product
Y sales the last 12 weeks,” and then you get
an answer in seconds. That’s all you have to do.
You don’t have to know the tables, you don’t have to know
the data stores. You don’t have to figure out how
to ask the question in the right way. You just ask it the way you speak. And Q will do auto filling
and spellchecking so that it makes it even easier
to get those queries written. Q uses very sophisticated
deep learning, natural language processing,
schema understanding, and semantic parsing of SQL code generation.
It is not simple. However, customers don’t need to know
anything about that machine learning. For them, the experience is: I type
in natural words the way I would speak
and I get answers in seconds. By the way, we’ve trained those
models over millions of datapoints and we’ve also trained them over all
these different vertical domains, virtually every one you can imagine. This is going to completely
change the BI experience. So again, customers didn’t hire us
to do machine learning here. They hired us to make it easier
to ask the questions that they want business intelligence on,
to get answers quickly, whether we
use machine learning or not - and that’s what we try
to do here with Q. So it’s pretty amazing how fast
machine learning is continuing to evolve. And although we have over
100,000 customers who are using our machine
learning services, a lot more than you’ll find
anywhere else, and a much broader array of machine learning services
than you’ll find elsewhere, we understand that it is still
very early days in machine learning and we have a lot to invent
in all four of these areas that we get asked regularly from,
from customers. Now we’ve talked about
the reinvention of compute and the reinvention of data stores and the reinvention
of machine learning, and I think
that if you look more broadly, if you’re wondering are
horizontal application areas being reinvented or are vertical
industry segments being reinvented, the answer is absolutely yes.
And to kick us off on this section and to share how they’re
reinventing broadcasting, it’s my privilege
to bring to the stage the CTO and President of Digital
for Fox, Paul Cheesbrough. [music playing] Through our partnership with AWS, we’ve combined
best-of-breed infrastructure with leading-edge
media operations. Production teams who are
more distributed than ever before will be able to deliver
uncompressed video through and from the cloud
with full redundancy, which is an industry first. We’ll be able to produce
live events with less latency, increased reliability, and
more efficiently than ever before. This collaboration has resulted in the development
of entirely new services. Working with AWS, we’ve reinvented
how we produce and distribute content to our consumers across all platforms
and devices using the cloud. We’ve transformed our
existing operation but, more importantly, we’ve laid
the tracks for the future in a way that will help us innovate
and adapt for many years to come. [applause] Thank you, Paul. We love
working with Paul and his team. They are unafraid of leaning
forward and inventing. They have built
a real reinvention culture, and it’s remarkable
to see the way that they are reinventing
broadcasting as we speak. Very honored
to be partnered with them. One of the questions
that we get asked most from venture capital companies
and private equity companies is, “Which areas do you think
are going to be reinvented? Which application areas? Which vertical business segments?
What’s going to get reinvented?” And the answer we give
is often unsatisfying, because in our opinion
the answer is, “All of them”. And that’s true in every single area.
I think you’re seeing it happen today and those that you’re not
seeing it happen in yet, you will see it
in the next few years. And so let’s start with
horizontal application areas. And let’s start with an example.
Let’s look at call centers. So if you look
at call center solutions, the last 20-30 years people have not
liked these solutions very much. They’ve been hard to set up,
they required expensive consultants and lots of hours, if you need to change anything
you require a complicated code, they were hard to scale up and down,
they are expensive, and they are missing the two
most transformational technology advances of the last 15 years,
which is cloud and machine learning. Now, we learned
this first-hand at Amazon, where we had to build
our own call center solution to handle our retail business, and we had a lot of AWS
customers who said, “Why don’t you just generalize that
and expose it to us as a service?” And that’s why we built
Amazon Connect, which is our call center service,
which we built in 2017. And people love this service. It’s one of the fastest growing
services in the history of AWS, and the reason they love it is
that it’s easy to set up, it takes minutes instead
of outside consultants and lots of money,
it’s easy to scale up or down agents, it’s much more cost effective, because you only pay
for what you consume, and you are only paying for
when your agents are interacting with customers. There is no infrastructure
for you to deploy, and then it’s built right from
the get-go on top of the cloud and with a lot of machine
learning and AI embedded, so you can do things
like chat bots and IVR and transcription of audio
and sentiment analysis without having to worry
about the machine learning at all. And Connect is growing
really rapidly. You see thousands and thousands
of customers are using it, companies like Capital One
and Intuit, John Hancock, and GE and Petco
and Barclays and State Farm. And what’s also interesting is, during the pandemic
over 5,000 new companies, new Connect customers,
have started using the service where they spun up call centers
remotely to help them deal with the fact that all their customer
service agents were now remote. And so this is another
one of those services that’s growing
unbelievably quickly, and people are excited about
and giving us all sorts of input on what else they would like
to see us solve with the service. And so I am going to share a few
of their asks across three categories and five different feature ideas. And the first is really around
how can we make it easier for our customer service agents to have the right information
about products and customers right in the moment, in real-time,
when they are dealing with customers? And the problem is
an interesting challenge. When customers call in about
something, a service or product, you have all this information
that lives in all these different databases, some on-premises and first-party
and some of which are third-party, which you have your agents
try to access and toggle between different databases
and find the right information, which is slow
or they don’t do it all. The same thing happens
when you have customers who interact with you
across lots of different silos in your business,
and different databases, either first-party or third-party.
And so customers said, “Can you make it much easier for us
to have the right product information and customer information
for our agents to have that fast
holistic service?” So I have two new features
to announce here for Connect to help with this problem.
The first is Amazon Connect Wisdom, which is a new capability
that uses machine learning to deliver agents
the product and services information they need to solve issues
in real-time. [applause] So Wisdom has all these
built-in connectors to relevant knowledge
repositories, either your own
or there’s some third-party ones. We’ll start with having connectors
to Salesforce and ServiceNow, but there will be more coming. And then what Wisdom lets you do
is that, as a call is happening, Wisdom is using machine
learning to listen to that call transcription
from Contact Lens and then detect issues
that are happening and put the right information
in front of agents. So, for instance, if you have
a call and Wisdom hears ‘arrived broken’, it will search
all the relevant data repositories for the information
that you need around what to do when you have a product
that arrived broken. So the agents have it right there
in front of them automatically, right in their console.
That’s a game changer. And if it turns out that you don’t
get the right recommendation, or you want to get
more information, agents can also just type
in natural language questions into the Wisdom search bar which pulls all the information
across all those databases. So this totally changes
what information you have available to you, product
and service-wise, for your agents. And then, for trying
to figure out how to provide a more holistic
customer experience, and pull together all
the customer profile information, I am excited to share
a second feature, which is the launch of
Amazon Connect Customer Profiles, which gives agents a unified
profile of each customer to provide a more personalized
service during a call. [applause] So here’s how
Customer Profiles works. You will point Customers Profiles
at your internal databases, where again, a third-party database
will launch Customers Profiles with your being able to access
your information in Salesforce, in Zendesk, ServiceNow, and Marketo.
And then, when the call comes, Customer Profiles connects
that phone number or contact ID with a customer ID that’s used consistently across
all those data stores. And so, Customer Profiles knows
how to ingest and normalize that data across customer
contact history, or e-commerce status,
or order management, or marketing communications
you’ve sent, or sales or CRM information. And they know how to display it
in a concerted organized way so all that appears in front
of an agent’s screen. And so, take a simple example. Let’s say you are a hotel company
that’s using Connect, and let’s say you have
a customer call in who’s unhappy about a stay
they had the prior night. You might handle
that call differently if you also knew
that that same customer called you earlier in the week
for their company to ask for a quote to do a five-day off-site
at that hotel. You might have
a different conversation. Today, agents don’t have
access to that information. With Customer Profiles, all that information is there
for agents to use at the same time, and changes how they can have
holistic customer interactions. Very, very useful. The next category of asks
we get around Connect is, “How can I make it easier that
when I have a customer contact or call that’s going off the rails, how can I intercept that
before the call is done and before I have a bad experience
for the brand?” And you may remember that last year
we launched a service in Connect called Contact Lens,
which is a call analytics service. And what Contact Lens does
is it stores all your calls in S3, and then it transcribes it to text, and then it indexes and tags
that information so it makes it easier
to search and find, and then it gives people,
agents or managers, the ability to have full-text
transcriptions of those calls, to be able to understand
when there was a negative sentiment, or when there were long
lapses in conversation or people were raising their voice
or talking over each other. And so it totally changes
your knowledge and information about how these contacts went. And customers
have loved Contact Lens, and they have used it
very pervasively. But a number of our customers and the supervisors
of these customers have said, “Look, I wish I had the chance
to know in real-time when a call was going awry, because I could either coach
that agent in real-time, or have it move to me
to avoid any brand harm.” And so the team has worked
on this for the last year, and I am excited to announce
the launch of Real-Time Contact Lens, which identifies issues
in real time to impact customer interactions
during the call itself. [applause] And so Real-Time Contact Lens
uses more developed machine learning to do
natural language processing on audio calls
in real time, instead of waiting a few minutes
after the call ends and the data and the speech
is transcribed to text. They do it in real time.
This is not easy to do, which is why it took
the team a year to do it. And it’s a good example
of not only being able to leverage AWS’s machine learning expertise,
but again, being able to use machine learning where it’s working for you
under the covers, where you don’t actually
have to worry about it at all. The job you’re trying to get done is you want to be able to have
an impact on calls in real time. And so what managers do who are using
Real-Time Contact Lens, is they specify terms
where they want to be alerted. So they may say,
“If, in the transcript I hear ‘bad experience’ or ‘unhappy’
or ‘never using you again’…” When that criteria is met, it sends an alert in the Contact Lens
dashboard to supervisors and they can either choose
to coach agents real-time, or have the call
transferred to them. And if the call
is transferred to them, they also get the real-time
transcription of that call right in front of them, so they don’t have to ask customers
all the same questions, which is frustrating
for those customers. So again,
a really big deal for people who are doing customer service. If you can know when a call
is going awry in real time, and help that customer
have a better experience and not do harm to your brand,
that’s a big deal. The third category that we get people
asking us for help in is, “How can I further optimize
my agent’s time in various areas?” And the two areas that they asked are,
first, they said, “About half of my time for agents
is spent on doing tasks, tasks outside of calls”. You might have to file
an insurance claim for a customer. You may need to contact
the customer for a status change. Half of their time
is spent on these tasks. These tasks live everywhere. They are not organized
in any one central place. It’s hard to prioritize them.
They are kept on pieces of paper, that’s why oftentimes
people lose them. This is a big problem. The other thing that people ask
to optimize agent’s time is they say, “At the beginning of a lot of these
calls, our agents have to spend time manually authenticating customers and
asking all these manual questions, and it wastes a lot of time
for agents, it wastes a lot of time
for customers. Can you help with that?” And so we have two solutions here
to help you with these problems. The first is the launch
of Amazon Connect Tasks, which automates, tracks, and manages
tasks for contact center agents. [applause] And so Tasks makes it much, much easier
to be able to manage tasks. Now you have one central place where
an agent to see all their tasks, managers and supervisors
get that same view, managers can easily assign
the tasks there, they can choose to prioritize and where they want what agent
to get each task done. They can actually decide
to assign agents’ tasks based on how busy they are and whether they are
on the phone or not. And then agents have everything
right in front of them. And the nice thing also is
that both agents and supervisors can automate
a number of elements of tasks to make it even faster for them
to get these things done. So this is very useful. This is going to change
the efficiency of agents when they are not on calls.
People are very excited about this. And then, for the call
authentication issue, I am excited to share with you
the launch of Amazon Connect Voice ID, which is real-time caller
authentication using machine learning powered voice analysis. [applause] And so the way that voice ID works
is that if you’re using it, you’ll ask whether customers
want to opt into it, so they can have a better chance
of avoiding fraud and also
so they can save time in their calls. They opt in,
and they say a few sentences and we build effectively a voice
print for each customer. And then, when a caller calls in, we will just allow them
to start talking. They’ll start talking about
what they’re calling in for and what their problem is,
and within a few seconds Voice ID will return to you
who they think that customer is, and with a confidence level
of that prediction. Each company will set
their own threshold, but if it meets that threshold,
you just continue having that call, you know there is no chance of fraud, you don’t waste all that time
on the manual questions. And if it’s under that threshold
that the company sets, then either you can choose
to do manual verification, or transfer that call to a fraud
specialist, but again, it totally changes the efficiency
and the productivity both for agents as well as for customers
who are calling in. And so if you think about it,
call centers and call center solutions existed
in a certain way for a long time. And even though customers
didn’t like them, what happened was they left
this area ripe for reinvention. And when you have an opportunity
to reimagine an area in an experience like a call center, and you find a way to make it easy
to get started and easy to scale, and much more cost effective,
and you build capabilities that put all the relevant information
about products and customers
in front of agents so they can handle
those in real time, and when you allow agents
to be more productive with their tasks outside of calls, and you allow them to actually
start calls much quicker, and then you allow them to use
machine learning under the covers to do things like have the ability
to impact calls in real time before there is any harm done
to a brand - when you do all
those types of things, you can completely transform
any horizontal application space as we have with Connect. Every single one of these
horizontal application spaces has the ability
to be reinvented and will be reinvented
in the next number of years. How about vertical business segments?
Which of these are being reinvented? And the answer is, all of them. And I am going to give you
a few examples. If you look in the auto space. This may be the space
that has the most elements of reinvention
happening simultaneously, but you have
the electrification of cars, which is partially driven
by the environment and partially driven by companies that came up
right from the get-go, electric only companies like what
Rivian has done on top of AWS. You have companies who are working
and building autonomous vehicles. We do a lot of work
with BMW and Lyft on these, but virtually every company
is working on this. The car gives you a chance to have
all these new connected experiences that weren’t really thought
of many years ago. We’ve been working with companies
like Toyota on this, but even today you may
or may not have seen the announcement we just made
with BlackBerry, where they have built a new
intelligent vehicle data platform that they call IVY,
on top of AWS, where they basically
took their QNX software that does sensor management and that runs on
175 million vehicles on the road, and then what they’ve done
is they’ve helped automakers better access,
process, analyze, and share data from those connected vehicle sensors
to build brand new experiences for car owners
or drivers or passengers. So auto is transforming
in a rapid way. The same thing with healthcare. Healthcare companies will tell you
that the Holy Grail for them is to be able to take all these
disparate pieces of information that live in all
these different formats so they have
a 360-degree view of customers. And they are not there yet, although we’ll have
more to say on this and to help with this
in the next couple of weeks. But they’re not there yet.
But they are making progress. And they’re using machine
learning to solve problems that they couldn’t solve before. So a good example is look at what
Moderna has done in the last year or so,
the last nine months. They’ve built an entire digital
manufacturing suite on top of AWS to sequence their most recent
COVID-19 candidate that they just submitted
and it has 94% effectiveness. And they did it on AWS
in 42 days, instead of the typical
20 months it takes. Just think about that. Think about what it changes
with regard to what’s possible in healthcare
if you’re able to take actions that are so momentous like that much
more quickly and much more easily. You see the same type of invention
happening in media and entertainment. I think most people here know that the way
that we have started streaming and allowing consumers
to use video content has radically changed
in the last five years, and if you look at all
the major streaming services, Netflix and Disney+ and Hulu
and Prime Video, they all use AWS to do that, but you are also seeing
really audacious new inventions like what Paul was talking about
with Fox and ViacomCBS are trying the same type of thing,
which is completely reinventing how they do broadcast
and the ease and speed with which they can spin up
brand new channels for people. And then you see
the same thing happening in industrial
manufacturing, where they are using edge
and the data to rethink all their design
processes on the production line, thinking about how to change
their supply chains, how to change their products. And a company that’s
a great example of this, which is totally reinventing
themselves on top of AWS, is my next guest. It’s a privilege to welcome
the President and CEO of Carrier, Dave Gitlin. [applause] Thank you, Andy.
I am honored to be with you today. Like Amazon, Carrier is a company
founded on innovation. Willis Carrier invented
modern air conditioning at the age of 25 back in 1902,
and it changed the way we live. Innovation is in our DNA
and it is embedded in what we call ‘the Carrier way’. With a laser focus on customers,
Innovation, and agility, we introduce more than
100 new products a year. And that kind of relentless customer
focus and quick execution is what Amazon calls Day One.
We literally just had our Day One when we became
a public company on April 3rd, following our spin
from United Technologies. And it has been so energizing. We think of ourselves
as a 100-year old start-up. And our focus is clear. To be the world leader in healthy,
safe, and sustainable building and cold chain solutions. On the building side,
COVID has shined a light on the criticality
of health and safety, and we are leading the way
to ensure healthy buildings and safe indoor air environments. But today I am going to focus
on our other ecosystem, and that’s the cold chain and
our transformative alliance with AWS. So let me start
with a problem statement. The safe distribution of many foods
and pharmaceuticals requires uninterrupted cooling from inception
to destination. Maintaining precise
cooling conditions is more difficult
than it sounds, because foods and pharmaceuticals
can travel thousands of miles through a complex
refrigerated distribution network called the cold chain.
So consider a banana. Bananas are the most
consumed food on earth and are extremely
temperature sensitive. Once harvested,
their clock starts ticking. Within hours,
bananas must be cooled and maintained between
56 degrees and 58 degrees Fahrenheit, with relative humidity
between 90 and 95%. Even a one degree
temperature variation during that refrigerated journey
can cause damage which doesn’t show up
until arrival at the supermarket. So consider the two
to three-week journey of a banana from, say, Ecuador
to the United States. It starts at the farm. There, bananas are packed
into a non-cooled cardboard box. They are then put into a series
of refrigerated environments starting with a cold room
on the farm to a truck, to an ocean container,
shipped over the ocean, then onto a trailer
to a distribution center, onto another truck
to a ripening facility, to yet another truck
to the grocery store. Each of these hand-offs from one
refrigerated environment to another, occurs at outside temperature,
and there is no margin for error. And today, there is
no central repository for the temperature data
for the entire journey. So now it’s not surprising
to learn that one third of the world’s food
produced every day is never consumed, due to waste, loss, or damage.
Imagine that. Food waste costs the global economy
nearly $1 trillion. But here’s what’s
even more tragic. One in nine people
go to bed hungry every night. If we could eliminate all of
the cold chain challenges, we’d preserve enough food
to feed 950 million people. That’s more than all of the hungry
people in the world today. Think about the impact on the planet. Consider all of the resources
and resulting emissions that go into producing food
that never gets consumed. If food waste were a country, it would be the third largest
emitter of greenhouse gases. It’s tragic on a human level,
an environmental level, and a financial level. And the same problem exists
with pharmaceuticals. They must be kept cold
within very strict standards from the time the active
pharmaceutical ingredient is first manufactured, through a series
of hand-offs spanning ocean, air, warehouses, trailers,
vans, freezers, and then temperature-sensitive
medications must then be administered within hours
of hitting room temperature. Today, we lose $35 billion
in biopharma each year from temperature-controlled
logistics failures. It’s a challenging problem
in normal times, but an unprecedented challenge
with the COVID vaccine. So now let’s talk solutions.
We at Carrier are a global leader in providing equipment
that cools containers, trucks, trailers, temporary
remote storage, and cabinets. We also provide sensors
that track temperature, location, and cargo condition. AWS has the world’s most widely
adopted cloud platform with the broadest
and deepest set of services. Together we are going to transform
how perishable food and pharmaceuticals are protected
through a new offering that we’ve developed, called Lynx. You can think of Lynx
as an end-to-end digital connective platform
for the cold chain. We will use AWS’s IoT,
Analytics, and ML services to help customers
reduce food and medicine loss, optimize supply chain logistics, and enhance
environmental sustainability. Where data was previously siloed, we’ll now be able
to aggregate temperature data across the products’
entire journey onto a data lake. Say, from the time the banana
is first cooled on the farm to when it arrives
in the supermarket. And we can use AWS
ML capabilities to provide faster, smarter, actionable intelligence
to better manage the cold chain and prevent breakpoints.
With AWS ML services, for example,
we can learn that truck doors should be not be left open longer
than half an hour when loading or unloading cargo. With AWS ML services we can determine if a container
has to be pre-cooled to 56 degrees and then use position location
to initiate the pre-cooling process. We can also use weather data
to detect and avoid
transportation delays, and then use updated arrival times
to alert the supermarket to a delay, so they can then adjust
their produce displays. The data can also be used
for equipment health, to remotely diagnose
a compressor issue while the container is at sea,
or better yet, use AWS IoT and analytics to help anticipate
when a unit will fail, so we can then perform
preventative maintenance. And we can also optimize logistics
by combining truck positioning data with traffic and weather to help
a distribution company improve fuel consumption
and fleet up-time. All of this, helping food
and pharmaceuticals get to where they are needed
when they are needed and in the conditions
that they are needed. So exciting times for us
here at Carrier. With AWS we will be addressing
profound issues like hunger, climate change,
and safe vaccine distribution, and we could not be more excited
to be on this journey with our friends at AWS.
Thank you. [applause] Thank you, Dave. It is really impressive
what Carrier is doing, and really important,
particularly at this time. And as we said earlier,
the keys to reinventing is a combination of building
the right reinvention culture and then knowing what technology
is available to you to make that reinvention
change and using it. And that’s exactly
what Carrier is leveraging. It’s very important
and very impressive. So manufacturing and industrial
companies are a group of customers
who have said to us, “We know that we could change
our customer experience and how we operate
our plants so significantly using
machine learning, but we just don’t often
have the equipment or the talent
to make that happen, and I really wish
AWS would help us.” And so let me give you
a couple of examples. If you look at machine data, there are a lot of industrial
companies who know that if they could do
predictive maintenance better, they could save a lot
of money and time. And what manufacturing
companies will tell you is that it’s always
much less expensive to fix something
before it breaks, not to mention saving the money
of the downtime in the plant. But it’s actually
not that easy to tell. Most people at a plant
will tell you that you can hear it or you can feel it through
the vibration before you can see it. But a lot of companies
either don’t have sensors, or they are not modern
and powerful sensors, or not consistent
and they don’t know how to take that data
from the sensors and send it to the cloud, and they don’t know how
to build machine learning models. And our manufacturing companies
that we work with have said, “Could you just solve this? Can you build
an end-to-end solution?” So I am excited to announce today
the launch of Amazon Monitron, which is an end-to-end solution
for equipment monitoring. [applause] And so Monitron gives
customers sensors, a gateway device
to send the data to AWS. We build custom machine
learning models for you in a mobile app with a UI
so you can tell what’s happening. And all our customers do is you mount
the sensors to your equipment, you start sending the data
through the gateway device, and then, as we take the data in,
we build the machine learning model that looks at what ‘normal’ looks
like for you on sound or vibration, and then as you continue
to stream that data to us, we will use the model to show you
where there are anomalies and send that back to you
in the mobile app, so you can tell where you might
need to do predictive maintenance. That’s a big deal. That makes it much,
much easier for companies to do. Now, there are other companies
who say, “Look, I have modern sensors
that I’m fine with. I’m also okay taking the data
from those sensors and sending it to AWS, but I don’t want to build
the machine learning models. I just want to send you
the data, use your models, have the predictions come back
to me through the API.” That third layer
of that machine learning stack. And so we have something
for this group of customers as well to announce today, which is the launch of
Amazon Lookout for Equipment, which shows anomaly detection
for industrial machinery. [applause] And so with Lookout for Equipment
you just send the data to S3, or we have a service
called IoT SiteWise that lets you send
your machine data in a structured way
for your analytics. You send the data to AWS, we will assess sound,
vibration, temperature, and we’ll again build a model
of what normal looks like, and as we see anomalies
we’ll send them to you via the API, so that you can do
predictive maintenance. These are game changers
for industrial companies that want to be doing predictive
maintenance and saving money. And the second problem they asked
to help with was computer vision. And if you think about it,
there are a lot of split decisions that you’ve got to make
in facilities, on your production lines, or even in interactions
between people when you’re trying to be
socially distanced, where you just
don’t have the time to send that information to the cloud
and get an answer back. You need to make
that decision in real time. So what these
industrial companies need, and often try to employ
in some fashion, are cameras - these smart cameras that allow them
to do streaming video. But the problem is, most of
the smart cameras out there today are just not powerful enough
to run sophisticated computer vision models at the edge.
And most companies you talk to, they don’t want to rip out all their
cameras that they have installed, but they know they need
to give help to those cameras to be able to do computer vision
in a sophisticated way. They asked us if we would
try to help with that. And so I am excited
to announce today the launch of the AWS
Panorama Appliance, which is a new hardware appliance that allows organizations
to add computer vision to existing on-premises
smart cameras. [applause] So here’s how it works. You simply plug
in the Panorama Appliance and connect it
to your network, and Panorama starts to recognize
and pick up video streams from your other cameras
in the facility. The Panorama Appliance
can accept streams of up to 20 concurrent streams
and operate on those. If you need to have
more concurrently, you can buy more
Panorama Appliances. And then we have pre-built
models inside Panorama that do computer vision for you
and that we’ve optimized by industry. So we’ve got them
in manufacturing, in construction, retail,
and a host of others. You can, of course, choose not
to use the pre-built models and build your own in SageMaker, and then just deploy those
to Panorama. And then Panorama
also integrates seamlessly with the rest of the AWS IoT
and machine learning services, where, if you actually
want to send that data, not for real-time actions, but to do large-scale analytics
on what’s happening in the plant, you can send them to us
through Panorama and you can use it
in the rest of your AWS Regions. Now, this is pretty exciting, and people are pretty excited
about the possibility of having real computer vision there,
but they also have told us that, “Look, we’re going to buy the next
generation of smart cameras, and those smart camera
manufacturers have told us we want to actually
embed something that allows us to run more powerful
computer vision models in there,” so we are also providing
a brand new AWS Panorama SDK which enables hardware vendors
to build new cameras that run more meaningful computer
vision models at the edge. [applause] And so what this will do is, if you are a company
that runs cameras, and you are building
the next generation of cameras, you will be able to use this SDK
and the APIs associated with it. We’ve done all this work
to optimize the parameters around memory and latency
so you can actually, in the camera, fit more powerful models
in a more constrained space, and it’s going to change
what’s available for companies as they’re building
smart cameras moving forward. If you’re an industrial company, and you’ve built a culture
that’s able to reinvent, if you use these tools
I just talked about, along with a host of others
that AWS provides, you can totally reinvent
what you are doing in your industrial
manufacturing company, much like Carrier has. And the reality is that this is true
and available to every single company in every vertical business segment.
That reinvention is there for you. If you’re ready, we have a lot for you,
with a lot more coming. So the final area of reinvention
that I’m going to talk about today is really around
hybrid infrastructure, and I am going to yield the floor to the Head of Infrastructure
at Riot Games, Zach Blitz. [music playing] With Outposts, AWS
gave us a unique solution to ensure a level playing
field for our players, and streamlined our deployments
using the same tools and APIs on- premises and in the cloud.
We rolled out Valorant fast and we are continuing
to reinvent how we design and
deploy our games to provide our players with the best
possible game experience. We shipped AWS
Outposts to new colos quickly, enabling a rapid deployment
of game servers, standardized on a single build,
test, and production pipeline. And best of all, using AWS Outposts, we reduced latency
for players by 10-20 milliseconds, minimizing peeker’s advantage and creating a level playing
field for all players. [applause] Thank you, Zach.
I appreciate it, Zach. It’s very impressive
to continue to see the way that Riot is innovating
for their players, and it’s also impressive to see
how they’re using AWS in Regions as well as on-premises
in a seamless and consistent way. When you think about the term
‘hybrid’ and ‘hybrid infrastructure’,
I think a lot of people believe that this term
and these solutions are pretty set. But in our very strong opinion, both the definition of the term
and the solutions themselves, are innovating and being
reinvented really quickly. When people ask,
“What’s hybrid?”, you know, typically, people view it
as a mix of modes. And people define it early on
as a combination of cloud alongside on-premises data centers. And one of the reasons
it was defined that way is the people
who popularized this term were on on-premises
infrastructure product providers, and they wanted to ride along
with the momentum of the cloud. So that’s how
the definition started off, which was cloud
and on-premises data centers. And it led to all
this breathless debate about whether this was going
to turn out to be a binary situation. Would you only use the cloud
or only use on-premises? And we probably contributed
a little bit to that confusion, because we stated
our then very strong belief and now even stronger belief,
that the vast majority of companies, in the fullness of time,
will not have their own data centers. And those that do will have
much smaller footprints. But we always thought
that was going to happen in the fullness of time,
not this year. And we knew that it would
take several years and so we just spent
a lot of our time in the early years of AWS
on the cloud-specific pieces, and then building sensible bridges
back to on-premises data centers. And these are things
like Virtual Private Clouds, or VPCs, or Direct Connect,
or Storage Gateway. But in the meantime,
you had a number of companies who tried to jump on
owning what hybrid was, and built all of these
hugely-hyped capabilities that were supposed to be
hybrid capabilities that never lived up to the hype,
and really never got any traction. And so we were watching
this happening, and we went back to first principles
and we asked ourselves, “So, wait a second.
What really is hybrid?” So it’s cloud and on-premises.
What does on-premises mean? Is it just on-premises
data centers? What about a restaurant?
Is that on-premises? What about an agricultural field?
Is that on-premises? If those are on-premises,
they have very different requirements than on-premises data centers, and so we think
of hybrid infrastructure as including the cloud,
along with various other edge nodes, on-premises data centers
being one of them. But there are several of them. And the way that customers
have told us they want to consume our hybrid
offering is with the same APIs, the same control plane,
the same tools, and the same hardware that they are
used to using in AWS Regions. Effectively, they want us
to distribute AWS to these various edge nodes. So we reimagined
for ourselves what hybrid was, and we started to build solutions that picked off
the biggest use cases, but in a way that worked for
customers both short and long-term. And so we started with,
customers said, “Look, I want to be able
to use the same tools I’ve used to manage
my infrastructure on-premises for the last number of years.
I want to use that in the cloud. I want to use it to manage
my cloud deployment.” Most of the world, at this point,
has virtualized on top of VMware. So we started working
with Pat Gelsinger and the VMware team
on how to build a new offering, which is called
VMware Cloud on AWS, that allows customers
to use those VMware tools they have been using
for many years to manage their
on-premises infrastructure, and manage their infrastructure
on AWS. And this is a very
unusual collaboration. There’s no other managed service that VMware runs
with another cloud provider. There is none that have
the functionality and capability of this
VMware Cloud on AWS. It’s not just that
both VMware and AWS have its engineering teams and its product teams
closely tied at the hip, but also our field teams
and our partner teams work together with customers.
It’s a very unusual partnership. And it’s gaining a lot of momentum
and a lot of steam. And you can see there
are a lot of customers, whether you are talking about
S&P Global or PennyMac, or Johnson & Johnson or Phillips
or Palantir, Scottish Government, Lots of customers are using
VMware Cloud on AWS. You also see the growth, almost double the amount
of nodes year over year. IDC just had a report
that showed over five years you get a 500%
return on investment. There’s a lot of momentum
in VMware Cloud on AWS and it’s really, really handy, as you’re moving from on-premises
infrastructure to the cloud. Then a lot of our customers said, “Well, that is awesome,
that is very useful, but what about when I need
to keep workloads on-premises for the foreseeable future? Maybe they need to be close
to a factory or something, that lives near my data centers,
but I want to use AWS there right on-premises.”
So there are a number of companies that have tried
to solve this solution, none of which have gotten
any traction. And we tried to go to school
on what wasn’t working. And what customers
hated about those solutions was that it didn’t have the same
APIs or tools or control plane or hardware.
They were totally different. There was too much work,
and they weren’t doing it because it was too much friction. And so we changed how we thought
about what people wanted there, and we changed our mindset
to realize what they really wanted was they wanted us to distribute AWS
to the on-premises locations and nodes. And so that’s why we built Outposts, which we announced two years ago
and launched last year, which effectively rolls
in racks of AWS services. You’ve got compute, storage,
database and analytics, and soon machine learning.
It’s fully managed. We deliver it, we install it,
we’ll do all the maintenance. And it comes with the same APIs,
the same control plane, the same tools, and the same hardware. People are very excited
about Outposts. In a short amount of time with
thousands of customers using them, these are customers like Philips
and Volkswagen and Erikson and Cisco, Lockheed Martin, T-Systems,
and Toyota. Just loads of companies
who are using these services, and using Outposts
and very excited about it. And so customers said,
“While we love Outposts, and it really worked well on
on-premises data centers, but they are big racks.
What if I want to use Outposts in a much smaller space
where I can’t afford to have racks?... I don’t have the space.
I need servers.” And so I am excited to announce today
two new formats of Outposts, smaller Outposts that let you run AWS infrastructure
in locations with less space. [applause] And so instead of these big rack
solutions, these are server solutions
for Outposts. So the first size
is what we call 1U size, which is 1¾ inches tall,
the size of a pizza box. It’s a 40th of the size
of the bigger rack Outposts that we launched a year ago. And then we have a 2U size
which is 3½ inches tall, which is almost like two pizza boxes
stacked, and these two smaller Outposts formats
have the same functionality as Outposts, just for a smaller space.
And now it means that restaurants or hospitals,
or retail stores or factories can use Outposts to have AWS
distributed to that edge. Then customers said,
“Okay, that’s really cool. I can have Outposts on-premises data centers
in a much smaller space. How about the use case where I need AWS distributed
in major metropolitan areas where it may not be cost-effective
either for you at AWS or me, the customer,
to have a data center there, but I’d be willing to pay
a little extra in exchange for my most
demanding low-latency applications being able to reside
in major metropolitan areas?” And that’s why we built and announced
last year AWS Local Zones. We started with the first one in
Los Angeles, aimed at the film makers and graphics
renderers and gaming companies. And I am excited to announce today we have three new local zones
that are launching today. In Boston, in Houston, and in Miami, and then 12 more in the United States
in 2021, in Atlanta, Chicago, Dallas, Denver, Kansas City,
Las Vegas, Minneapolis, New York, Philadelphia, Phoenix,
Portland, and Seattle. [applause] I don’t know if you’re applauding because you’re excited
about the Local Zones, or that I remembered all 12 of those.
But we’re excited about it and we think it’s going
to help you deploy AWS as metropolitan city edges. “What about when I actually need
AWS distributed to the edge where there’s no connectivity,
or where the terrain is so rugged that I need something that can
be banged around a little bit?” This might be in an
agricultural field or an oil field, or a military battlefield. And for this, we built
our Snow Family of products. So, Snowball Edge, and Snowcone,
which are different sizes, but both are hardware appliances that you can bring
to that disconnected edge. It stores data that’s running off
the different assets that you have with the edge.
It has Compute on it, so you can do some analytics
and processing on it, and then, if at any point you want
to actually detach that appliance and send it back to us to have it
adjusted in our data centers so that you can do
large-scale analytics, you can do that as well.
“How about distributing AWS if I want to build
mobile applications and I want to take advantage of 5G?”
And people are very excited about 5G, because the latency
and the power of it give you a chance to build sub-10
millisecond latency applications that can do things
like smart manufacturing or autonomous vehicles
or various things in games. But any application that wants
to do anything interesting is going to need compute,
it’s going to need storage, it’s going to need infrastructure. And the problem is,
for mobile applications to leverage that infrastructure
which is typically AWS, they have to go from the device
to the mobile network, to the local aggregation location, to the regional aggregation location,
to the internet, to AWS and back. And that’s not 10 milliseconds,
that’s seconds. And so customers wanted
a way to change that, and that’s why
we announced last year the launch of what we built,
called AWS Wavelength, which extends AWS
infrastructure to the 5G edge, so now you only have to go
from the device to the natural 5G aggregation site,
and AWS has racks of Outposts right there for you
to do your infrastructure. We launched this
last year with Verizon, who’s been an amazing partner. We already have eight US cities
that we’ve launched, with more coming in the coming weeks.
We are launching with KDDI in Tokyo, and SK Telecom in South Korea
in the next few weeks, and then with Vodafone
in London in early 2021. And one of the problems
that we also thought about and saw that the customers
asked about was they said, “Look, there are all these
different telecom providers. They all have
different semantics. I don’t want to have
to learn all those. Can you build an abstraction so I’m just writing
to a Wavelength zone, and then you do all
the normalizing under the covers?”, which is what we’ve done. So there are a lot of ways
of being able to bring the AWS experience
to customers wherever they are. And so when you go back
to asking “What is hybrid?” It’s not just cloud
and on-premises data centers. It’s cloud along
with various edge nodes, on-premises data centers being
one of them, with several others. And we think most of this computing
will end up in the cloud over time, which will be like the hub
just given the cost and the agility and functionality and productivity
advantages for builders. But several other workloads
will reside where it makes most sense. You will have on-premises
data centers when you’re doing this transition
from on-premises to the cloud, where they need to be close
to something that lives
near an on-premises data center. You’ll have them in various smaller
venues where you want them, in a restaurant or in a hospital,
or in a factory. You’ll have them
in major metropolitan areas where you have your most
demanding low-latency workloads, where you are willing to pay
a little bit extra to have that low latency. You’ll be able to have them
in the disconnected edge, and you’ll be able to have AWS
as well when you’re building 5G mobile applications
that need to sit at a 5G mobile edge. And people will want this hybrid
experience to be delivered by AWS by distributing AWS to these
edge nodes with the same APIs, the same control plane,
the same tools, and the same hardware
they get in AWS Regions. That’s where we believe
hybrid is heading, and how we’re trying
to enable it. So I’m going to close
with a song lyric, and because of COVID and the way
we did this virtually, we didn’t have the band this year, but I’m going to use a song lyric
to try and bring us home. And the lyric says,
“I wish I could Google my ending. Someone give me reassurance,
answers, anything will do.” And this is from one of my
very favorite song writers, and what she was talking about was really the uncertainty
of being a young adult, and what’s going to happen. But I think that same message
is really applicable to companies, and what’s going to happen
in the future. If you’re a missionary
and you’re focused on building a lasting company versus
making a quick buck and being a mercenary, you know how hard it is
to build a sustainable business. So many things can derail you. New technology, new competitors,
losing leaders, losing key contributors,
regulation, pandemics. There are a lot of things.
And like for a young person wondering what the future
holds for them, the same is true for companies.
It’s daunting. And I think
the same counsel applies. You can’t control for every change
and every development, but you can build the capacity
to get to the truth, to make changes
when they’re needed, to have people around you
who want to help you change and can help you make the change,
stay focused on what matters most, to move fast when speed is required,
which is more often than you realize, and then to be aware
of what’s available to you and what’s changing around you,
so you can reinvent who you are and what your customer
experience will be. That’s what you need for reinvention,
and in my opinion, companies who aren’t
already reinventing themselves in some meaningful way are unwinding,
whether they realize it or not. The good news though is
that invention and reinvention is very doable, if you’re intentional
and focused on it. And we’ll be here every step
of the way to help you do it. I want to thank you for listening.
I hope everybody stays safe, and I hope you have
a great few weeks of re:Invent. Thank you very much. [music playing] [applause - cheering]