(calm swelling music) (dramatic upbeat music) - [Announcer] Please
welcome the CEO of AWS, Adam Selipsky. (upbeat music) - Good morning and welcome
to the 12th Annual re:Invent. I am so happy to be here, with more than 50,000
of you here in Las Vegas and more than 300,000
watching around the world. Thank you, thank you
everybody for joining. (audience applauding and cheering) That's right, thank you. (audience applauding and cheering) Now, we've always done this event just a little bit differently, and re:Invent remains above
all a learning conference, a learning event, and the opportunities to learn
and grow here are everywhere. We've got more than 2200 sessions and you can connect with an
incredible array of partners in the expo, all around. And of course, there are many,
many opportunities to connect with fellow members of the AWS community. So we've also got a great
slate of keynotes for you. Yesterday, last night,
hopefully a lot of you saw Peter diving into all of our
innovation around serverless. And tomorrow we've got Swami
on AI and machine learning, as well as Ruba's partner keynote. And then of course Werner
caps off the week on Thursday. But I'm extremely excited
about what we have in store for you right now here today. So let's get started. So one of the things
I love about re:Invent is the sheer variety of the customers that I get to meet. I mean folks from every
industry, every region, every use case, and there's
just nothing like it anywhere. And innovators all over the
world are relying on AWS to drive their businesses, including leading enterprises
from across every industry. Just to name a few, in financial services, we're
working with JP Morgan Chase, we're working with Goldman Sachs, NASDAQ, Fidelity, Allianz, Itau, BBVA, HSBC, Capital One,
S&P Global, Nubank and more. In healthcare, we're working
with Pfizer, Gilead, Merck, Novo Nordisk, Roche, Moderna, Illumina, Medtronic, 3M Health
Information Systems and more. In automotive, we're
working with everybody from BMW to Stellantis,
Honda, Toyota, VW Group, Nissan, Hyundai, Ford Motor
Company, Mercedes-Benz, and of course more. So many amazing customers and partners and still just a smattering of the leaders we're working with to shape the futures of these industries and all of the customers,
of course, that they serve. One customer partner we've
been working closely with for a number of years is Salesforce. So in 2016, Salesforce chose AWS as their primary cloud provider, and today Salesforce uses
AWS compute, storage, data services, many others, to power key services
like Salesforce data cloud and many others. And yesterday we announced that we're significantly
expanding our partnership to bring new capabilities
to our mutual customers. Salesforce is gonna be
significantly expanding. It's already, you know,
very large use of AWS. AWS will be using more
Salesforce technologies, and we're making it easier for developers to quickly build and deploy
generative AI applications, by bringing together Amazon Bedrock with Salesforce's Einstein platform. We're also helping to
unify data management with our new zero-ETL integrations between Salesforce data
cloud and AWS storage, compute and data service,
data and analytics services, so the customers can get
even better insights faster. And we're also building
tighter integrations between Service Cloud from
Salesforce and Amazon Connect. Also, for the first time, Salesforce is gonna be
putting its products on the AWS marketplace, making it even easier for
AWS customers to find, to buy and to manage Salesforce offerings. So with this expanded partnership, AWS and Salesforce customers
are gonna get new powerful ways to innovate and build applications, but it's not just global enterprises. The most promising up and comers also continue to choose AWS as well. According to PitchBook, there are more than a thousand
unicorns, or startups, with over a billion dollars in valuation. And the vast majority of unicorns, over 80%, choose AWS. Like Wiz, a cloud security
company, and great AWS partner. Wiz is the world's largest
cybersecurity unicorn and fastest SaaS company to
reach a $10 billion valuation. Really amazing. Or Ultima Genomics, one of Fast Company's most
innovative companies in biotech. And Ultima is revolutionizing
DNA sequencing by reducing the cost
of sequencing sixfold. Or Plaid. Plaid's, a data startup
that powers fintech tools like Venmo, Acorns, and Robinhood that millions of people are
using to manage their finances. Customers of every size,
every industry, every region, organizations you might never imagine would rely on the cloud,
are innovating on AWS. I just recently heard
about Lowden Guitars, a high-end, expert guitar maker, located in a tiny village in Ireland. And they're using AWS technology to create an electronic passport program that's going to ease the passage of their very special guitars, for musicians as they
travel through customs. I mean all sorts of use
cases, it's amazing. Enterprises and startups, small and medium sized
businesses, universities, community colleges, government
agencies, nonprofit. The cloud is for anyone. And our customers are
solving big problems, serving critical needs, and dreaming up the
world's next big thing. They count on us to be
secure, to be reliable, to innovate rapidly, to
delight their customers and enable new ways to
grow their businesses. All on AWS. So why is all this happening on AWS? We are relentless about working backwards from our customer's needs
and from their pain points. And of course, we were the first, by about five to seven years, to have this broadest and
deepest set of capabilities that we still have today, and we're the most secure
and the most reliable. But we also think differently
about our customers' problems and their potential. And this has led us to
reinvent continuously, to push through the barriers of what people thought was possible, so you can do the same. Reinventing is in our DNA and it continues to drive us every day. After all, this is how
cloud computing came to be. In the early days of
operating 'Amazon.com', we experienced firsthand how
hard, how expensive it was to provision and manage infrastructure, and how much of the undifferentiated
heavy lifting there was that distracted really talented teams from really innovating. So we set out to rethink IT infrastructure completely with AWS, so everyone, even a kid
in a college dorm room, could have access to the
same powerful technology available to the largest and most sophisticated organizations. All on demand, all secure, all reliable and all cost effective. Now, folks got pretty excited about this, as they moved into the cloud, some even hosted data
center closing parties and posted videos of
shutting down server rooms. It was really, really fun. At least it was fun for us. So we reinvented infrastructure
from the ground up to optimize for scale, for
security, for reliability, from routers to cooling systems,
to networking protocols. And because we look at
these things differently, our global infrastructure
was fundamentally distinct from that of other cloud providers. And that is still true today. For example, the AWS global infrastructure spans 32 geographic
regions around the world, and we've announced plans
for for five or so more. Each and every region consists of three or more
availability zones, or AZs. And no other cloud provider provides that in every single geographic region. And this is critical. Within each region, the AZs
are physically separated by a significant distance up to 60 miles, or a hundred kilometers, and then they're interconnected
with redundant fiber network that provides single
digit millisecond latency. Now each AZ is at least one
fully discreet data center with its own redundant power, water, networking and connectivity. So if there's a utility failure,
unexpected traffic spike, human mistake, or even a natural disaster, the region remains operational
and so do your applications. Now, others would have you think that all clouds are the
same, but it's just not true. I mean, imagine if you
had a region supported by a single data center, or if you thought your
provider had multiple AZs, in France for example,
and you were resilient, but it turned out that they're actually in the same location. I mean, one incident like a water leak followed by a fire could take
to honor an entire region for days. (speaking french) Having the broadest and the deepest set of capabilities matters. Like having three times
the number of data centers compared to the next
largest cloud provider, which makes a huge difference
in capacity and availability. And having 60% more services
and over 40% more features than the next closest cloud provider. So you can innovate without constraints. We reinvent so that you can reinvent. Take storage, for example. So we certainly didn't invent storage, but we reinvented how people consume and scale storage in the cloud with our very first
cloud service, Amazon S3, simple storage service. So back in the day, storage
was primarily disk arrays, which could take weeks to provision and obviously had finite capacity. So S3 provided highly durable, highly performant object storage, that customers could scale
instantly and nearly infinitely. And I remember talking about
S3 at our very first trade show in 2006 on the day that it launched, and folks had a hard
time believing the cost. I mean, it was so dramatically lower than what they were used to. But we kept going. We added a whole lot of storage tiers, including S3 Glacier Deep Archive, which delivers archive storage at less than a 10th of
a cent per gigabyte. But why should you have to even figure out which storage tier to use? So we built S3 intelligent
tiering to optimize storage by automatically moving data to the most effective access tier when access patterns change. Intelligent tiering has
already saved AWS customers, you ready? Over $2 billion. But it's just storage, right? (audience applauding)
I mean... Here's to the savings. But it's storage, right, I mean, surely this can't be an area that's ready for more reinvention? Well, recently we've
been rethinking storage for your most frequently accessed and most latency sensitive workloads. Workloads that need high
performance analytics processing, like financial trading
analytics, real time advertising, fraud detection, and
machine learning training. Now, these workloads access data up to millions of times
in just a few minutes and do require single
digit millisecond latency. And so today, what happens
is customers move data often from S3 to custom caching solutions and then run analytics at these speeds, but managing multiple
storage infrastructures and APIs has a lot of complexity. So we looked hard at what
we could do to add to S3. Okay, are you ready for the
first announcement of the day? (audience cheering) Okay, so today I'm excited to announce Amazon S3 Express One Zone, a new S3 storage class, purpose built... (audience applauding and cheering) Purpose built to offer
the highest performance and lowest latency cloud object storage for your most frequently accessed data. S3 Express One Zone uses
purpose-built hardware and software to accelerate
data processing. It also gives you the option
to actually choose your AZ for the first time, and you can bring your
frequently accessed data next to your high
performance compute resources to minimize latency. So S3 Express One Zone supports millions of requests per minute with consistent single
digit millisecond latency, the fastest object storage in the cloud, and up to 10 times faster
than S3 standard storage. Now, the faster data
processing lets you save up to 60% on your compute costs as well. For workloads that require the
high performance analytics. And the data access
costs also are 50% lower than S3 standard. So you can run these workloads
at much, much lower cost. So for example, with S3 Express One Zone, Pinterest has observed over
10 times faster write speeds while reducing their total cost by 40% for its machine learning powered visual inspiration engine. Now this is enabling them
to do faster experimentation and personalization, improving
experiences for their users. So 17 years ago, S3 reinvented storage, by launching the first
cloud service for AWS. And with S3 Express one zone,
you can see S3 continuing to change how developers use storage. The same simple interface, same low cost, and now even faster. Let's look at another
example of reinvention, general purpose computing. So we realized almost 10 years ago that if we wanted to
continue to push the envelope on price performance for
all of your workloads, we had to reinvent
general purpose computing for the cloud era, all the
way down to the silicon. And in 2018, we became the
first major cloud provider to develop our own general
purpose compute processors, when we announced GRAVITON, our server processor chip. And customers loved the cost
savings that it delivered for scale out workloads like microservices and web applications, but asked for more. So we brought you GRAVITON2, which improved performance sevenfold for a wide variety of workloads,
and we didn't stop there. Our current generation GRAVITON3 delivers up to 25% better
compute performance compared to GRAVITON2, and the best price performance in EC2. Importantly, it also uses 60% less energy for the same level of performance. And this is so important
to so many customers today, and that's compared to
comparable EC2 instances. The energy efficiency is only
gonna become more important. Today we've got over 150
GRAVITON based instances across the EC2 portfolio, and more than 50,000 customers, including all of the
top 100 EC2 customers, use GRAIVITON based instances to realize price performance benefits. For example, SAP, they've
partnered with AWS to actually power SAP
HANA Cloud with GRAVITON. And with GRAVITON, SAP is seen up to 35%
better price performance for analytics workloads and aims to reduce their carbon footprint, their carbon impact, by an estimated 45%. Super impressive. We are relentless about innovation and we continue to push
the envelope further. So today I'm excited to announce the latest generation of
our GRAVITON processor, AWS GRAVITON4. (audience cheering and applauding) GRAVITON4 is the most powerful and the most energy efficient chip that we have ever built,
with 50% more cores and 75% more memory
bandwidth than GRAVITON3. GRAVITON4 chips are 30% faster
on average than GRAVITON3, and perform even better
for certain workloads, like 40% faster for database applications, and 45% faster for Java applications. We were the first to develop and offer our own server processors. We're now on our fourth
generation in just five years. Other cloud providers
have not even delivered on their first server processors yet. We're also announcing the preview of our first instance based
on GRAVITON4, R8g instances. R8g is part of our memory
optimized instance family. They're designed to deliver
fast, fast performance for workloads of process,
large data sets and memory, like databases or real
time big data analytics. R8g instances provide the
best price performance and energy efficiency for
memory intensive workloads. And there are many, many more
GRAVITON4 instances coming. Highly reliable regions
with availability zones, durable, instantly
scalable, low cost storage, leading price performance
with custom design chips. These are just a few examples
of how we're reinventing to constantly deliver what you need to drive your infrastructure, your innovation, and your businesses. And for nearly two decades, we have been so intent on providing with the broadest and
deepest functionality with the fastest pace of innovation, the most reliable, the
most secure cloud services. And of course the largest and
the most diverse community of customers and partners. And these advantages are always relevant, but no more so than today because they're also the foundation for reinventing with generative AI. Now here too, you're
gonna need the broadest and deepest capabilities,
the best price performance, and the security and
the privacy you trust, you need at all. Amazon's been innovating
with AI for decades. We use it across the company, to optimize things like our
company wide supply chain, to create better retail
search capabilities, to deliver new experiences
like Amazon Alexa, or our 'just walkout' shopping technology. GenAI is the next step in
artificial intelligence, and it's going to
reinvent every application that we interact with at work and at home. We're ready to help you
reinvent with generative AI because again, we think
differently about what it takes to meet your needs, and
reinvent again and again. And because we're intimately
familiar with what it means to reinvent ourselves, we
understand how to help you with your reinventions. From startups to enterprises, organizations of all
sizes are getting going with generative AI. They wanna take the momentum
that they're building with early experimentation and turn it into real
world productivity gains and innovate, innovate, innovate. So we think about generative AI as having actually three
macro layers, if you will, of a stack, and they're
all equally important. And we are investing in all three of them. So the bottom layer is used
to train foundation models, and large language models,
'FMs' we'll call them, and to run these models in production. The middle layer provides
access to the LLMs and other foundation models that you need and all the tools that you need to build and scale generative AI
applications with them. And then at the top layer,
we've got applications that are built using Fms, so
that you can take advantage of generative AI quickly, without even needing any
specialized knowledge. So let's look at the bottom
layer, infrastructure. So with the foundation models, there are two main types of workloads, training and inference. Trainings to create and prove Fms, by learning patterns from
large amounts of training data. And inference uses that to
run the models in production, to generate an output such
as text, images or video. And these workloads
consume massive amounts of compute power. So to make GenAI use cases
economical and feasible, you need to run your training. You need to run your inference
on credibly performant, cost-effective infrastructure
that's purpose built for ML and AI. Now GPUs are the chips that
can perform a high volume of mathematical
calculations simultaneously, making them popular for
lots of different workloads, from ML, simulations, 3D rendering. And we've been collaborating
with NVIDIA for over 13 years to bring GPUs to the cloud, building compute instances that support a wide range of use
cases including graphics, gaming, HPC, machine learning, and now of course demanding
generative AI workloads. In 2010, AWS was the first to
bring NVIDIA GPUs to the cloud with CG1, or the cluster GPU instance. Then we were the first
to offer NVIDIA V100 GPUs with our P3 instances. And then we were again, the first to offer the
A100 tensor core GPUs with our P4 instances in production. And earlier this year, yes, we were the first major cloud provider to bring NVIDIA H100 GPUs to market with our P5 instances. And the P5s are providing
amazing performance, for training they're
up to four times faster and 40% less expensive than P4s. Now having the best chips
is great and necessary, but to deliver the next
level of performance you need more than just the best GPUs. You also need really high
performance clusters of servers that are running these GPUs. Our GPU instances can be
deployed in EC2 ultra clusters, and they're interconnected with up to 3,200 gigabits per second of networking through our elastic
fabric adapter, or EFA. And this enables customers
to scale up to 20,000 GPUs in a single cluster, providing up to 20 exaflops
of aggregate compute capacity, which is equivalent to a supercomputer. And all of our GPU instances that are released in the last six years are based on our
breakthrough nitro system, which reinvented virtualization
by offloading storage and networking to specialized chips, so that all of the service
compute is dedicated to running your workloads. And this is not something that other cloud providers
can offer, again. Now with Nitro, this delivers performance
indistinguishable from bare metal, but at a lower price. And the Nitro system also
provides enhanced security that continuously monitors and protects and verifies the instance
hardware and firmware are all robust. We've worked closely with NVIDIA
to integrate our networking and our virtualization
capabilities with their chips, to bring our GPU instances to customers. And today I'm thrilled to announce that we are expanding our
partnership with NVIDIA, with more innovations
that are gonna deliver the most advanced infrastructure for generative AI workloads with GPUs. And I'd like to invite Jensen Huang, founder and CEO of NVIDIA,
here on stage with us to tell us more. (audience applauding and cheering) (upbeat dramatic music) Jensen, so good to see you. - I'm thrilled to be here. - Yeah, thank you so much
for joining us today. It's amazing. So, as I've just discussed, so we've got a long partnership, AWS, NVIDIA have been doing
this a long time together. I'm so excited about how
the teams are collaborating. You and I were talking
about that backstage, collaboration's amazing. It's been happening for a lot of years, and I'm just thrilled to
deepen the partnership. - Thank you, it's great to be here. I'm thrilled to be here to celebrate the amazing work of our two teams. And to help announce
our big new partnership. You know, we've been working
together for a very long time. In fact, AWS was the world's first cloud to recognize the importance
of GPU accelerated computing, and you put the world's
first GPUs in the cloud, a long time ago. And since then, this is
an amazing statistic, in the last several years alone, just for the Ampere and Hopper generation, we've put up, deployed
2 million GPUs in AWS, that's three zettaflops,
3000 exascale supercomputers. - [Audience Member] Whoop. - Yeah, exactly. Most nations will be happy with
one exascale supercomputer, AWS has 3000 of them, and
we are just getting started. I'm delighted to see that we
are going to announce today a deployment of a whole
new family of GPUs, the L4, the L40S, and the brand new H200. The H200, this is really an amazing thing, the combination between the brand new TensorRT-LLM optimizing
compilers for generative AI, and H200 improves the
throughput of inference, large language model
inference, by a factor of four, reducing the cost in just
one year by a factor of four. Now our collaboration starts
with, of course, great GPUs, and we're ramping so incredibly fast, each quarter we're
standing up and deploying more than one zettaflops more in AWS, one zeta flops more each quarter. It's an unbelievable number. And all of this is of
course, made possible because of the accelerated
computing libraries. Our two teams stood up a whole
bunch of new infrastructure, but we're also bringing
our most popular libraries, the NVIDIA AI stack, our NeMo LLM, large language model framework, the Retriever inference model for RAGs. Our BioNeMo, digital biology, large language model foundation models, Isaac Sim on Omniverse for robotics. This is something that Amazon Robotics uses to simulate its robots, and it's incredible
warehouses that you guys do. Just all of these stacks,
software stacks and libraries, we're gonna integrate into AWS. So our two teams have
been really, really busy. - Busy indeed. And that was a lot of
zettas and everything. (Jensen laughing) Yeah. So, as you said, the collaboration's
gonna enable developers to access amazing technology, and they're gonna need it to
innovate with generative AI. Now, one of the big
announcements, of course, is that AWS is gonna be, you
know, the first cloud provider, to bring the latest NVIDIA
GH200, Grace Hopper super chips, with a new multi-node NVLink
to the cloud, first place. And maybe you could tell us
a little bit more about that and what makes that computing
platform so powerful, and how we're working
together and introduce it. - Well, we're both really
passionate about Arm processors. And the reason why Arm is so incredible is because we can mold
it to exactly the type of computing needs that we have. It's incredibly low energy,
it's incredibly cost effective. And so Grace Hopper, which is GH200, connects two revolutionary
processors together in a really unique way. It connects them together using
a chip to chip interconnect called NVLink, at one terabytes per second. And it's connected in a coherent way, so that the GPU could access
all of the CPU's memory, the CPU can access all
of the GPU's memory. And so the two of these
processors could work hand in hand in a really fast way. The second thing that we did
was we invented a new way of extending NVLink to
a very large domain. So now 32 Grace Hoppers could be connected by a brand new NVLink switch, and that becomes one unit. And with the Nitro, AWS Nitro, that becomes basically one
giant virtual GPU instance. I mean, you gotta imagine
this, you've got 32 H200s, incredible horsepower,
in one virtual instance because of AWS Nitro. And then we connect it with AWS EFA, your incredibly fast networking. All of these units now can
create into a ultra cluster, an AWS ultra cluster. So I can't wait to see
all this come together. - I mean, it's really an
incredible combination. I mean, how customers
are gonna use this stuff, one can only imagine. I know that the GH200s are
really gonna supercharge what customers are doing. It's gonna be available, of course in EC2 instances coming soon. Now, in the world of
generative AI, you know, a lot of companies are
looking to onboard AI into their business, and it's great to see the infrastructure, but it extends to the
software, the services, and all of the other
workflows that they have. So this leads of course to the second big announcement we have today. So partnering AWS NVIDIA to bring the NVIDIA DGX Cloud to AWS. And I'd love to hear a little bit about how the DGX Cloud's an important
platform for customers. - Well, first of all, DGX
Cloud is NVIDIA's AI factory. This is how our researchers advance AI. We use AI to do neural graphics. The way we do computer graphics today is impossible without AI. We use AI to advance,
we use our AI factories to advance our digital biology models. Our large language models,
use it for robotics, for self-driving cars. We use it to simulate Earth-2, a digital twin of the earth to
predict weather and climate. And so DGX Cloud is
really important to us. We are incredibly excited to
build the largest AI factory NVIDIA has ever built. We're gonna announce inside our company, we call it Project Ceiba. Ceiba, you all probably know, since you're at a AWS conference, is the largest, most
magnificent tree in the Amazon. We call it Project Ceiba. Ceiba is gonna be sixteen thousand... Sixteen thousand three
hundred and eighty four GPUs, connected into one giant AI supercomputer. This is utterly incredible. We will be able to
reduce the training time of the largest language models,
the next generation MoEs, these large, extremely large
Mixture-of-Experts models, and be able to train it
in just half of the time. Essentially reducing the cost of training, in just one year, in half. Now we're gonna be able
to train much, much larger multimodal MoE models, these next generation
large language models. These 16,000 GPUs will be 65 exaflops. It's like 65 exascale
supercomputers in one. And so I can't wait for
us to stand this up. Our AI researcher's chomping at the bit. Now of course, this is also
a place where we collaborate with partners and customers who need to build custom AI models. You know, it's great to be able
to use off-the-shelf models, and there's gonna be a whole
bunch of off-the-shelf models in public clouds. It's gonna be available in
software platform companies, SaaS companies, for example, Salesforce that you mentioned earlier. They'll have all kinds of
off the shelf co-pilots and generative AI models. But a lot of companies need to build their own proprietary models. And so we set up the AI factory so that we could partner with them, to help them create their custom AIs. And now it'll run all on AWS,
we'll develop a model on AWS, collaborate with our customers on AWS, connect them all to your
services and storage services, security services, and all kinds of other
generative AI services, and then deploy it all in AWS. So we're gonna be able to do
this very first time in AWS. - It's all amazing. I mean, you guys have
been doing amazing things. You know, we've been working hard, to see us come together even more tightly. I mean, it's just gonna
bring incredible benefits to customers who are really gonna need it. But we appreciate the collaboration, appreciate working with you and look forward to a lot
more of that going forward. - Thank you, Adam.
- All right, thank you, Jensen. - Have a great show, everybody. (dramatic music) (audience applauding) - It's been a great partnership. We've really appreciated it. Now let's talk about capacity, or actually getting access to
all of that amazing compute. So customers who wanna
build and train models, often need large amounts of
reliable clustered capacity. It has to be together, or in clusters. But they don't necessarily
need the same amount of it all the time. For example, customers training a model, I mean, they might stop to
evaluate, use new training data, or make other changes, and then they want to do more training. So what this means is that
they often need, you know, short term cluster capacity. And it's something that really no one, no cloud has been addressing. Until just a few weeks
ago when we announced EC2 Capacity Blocks for ML. And EC2 capacity blocks for
the first consumption model in the industry, that enables any customer to reserve the highly sought after GPU capacity to run for short duration
machine learning workloads. And this eliminates the need
to hold onto the capacity, hold to capacity that they're using, just so they can be certain
that they can use it later, if they need to use it later. Now, EC2 capacity blocks are
deployed in ultra clusters, interconnected with our EFA networking that we talked about this morning. And that enables customers to
scale up to hundreds of GPUs in a single cluster
with this new offering. So with the capacity
blocks, customers can plan for their ML workload
deployments with certainty, and they know that they're
gonna have the GPU capacity when they need it, and
only when they need it. Now support for the latest generation, high performance GPUs, you know, combined with our
innovations, virtualization, networking, cluster management, these are reasons why
AWS is the best place to run GPUs in the world. But to drive innovation
and the broad choices that our customers want in ML and AI, we realized a few years
ago that again, here, another place we'd have to innovate all the way down to the silicon, just like we did with GRAVITON for general purpose computing. And that's why we built
Trainium and Inferentia. So Trainium is our purpose-built chip for training machine learning workloads, machine learning models. And Inferentia is our
chip optimized for running inference on those models. So earlier this year, we already announced the second generation of our Inferentia chip, Inferentia2. And because we optimized for inference, the two instances, which
powered by Inferentia2, offered the best, lowest
cost inference in EC2, while also delivering up to
four times higher throughput and 10 times lower latency
compared to the Inf1 instances. Customers like Adobe,
Deutsche Telekom, Leonardo AI, they've all seen, you
know, great early results. And they are, you know, deploying their generative
AI models at scale with Inferentia2. We've all seen a lot of
interest in Trainium as well. Our Trn1 instances, powered by Trainium, have been deployed by a lot of customers, including Anthropic, partners
like Databricks and Ricoh. And our own Amazon search
team uses Trn1 instances to train large scale deep learning models. They're taking advantage of
Trainium's high performance, scale reliability and low cost. But as customers continue to
train bigger and bigger models on larger and larger data sets, we realized we needed to keep pushing, pushing on price performance. We're never done there. And today I'm really excited to announce AWS Trainum2, our second generation. (audience applauding) It's our second generation
ship that's purpose built for high performance training
of foundation models. Trainium2 is designed to deliver up to four times faster performance compared to our first generation chips. And that makes it ideal for
training foundation models with hundreds of billions, or
even trillions of parameters. Trainium2 is gonna power
the next generation of EC2 ultra clusters, and will deliver up to 65
exaflops of aggregate compute. And we expect the first
Trainium2 based instances are gonna be available next year, and we are really excited to see what customers are gonna do with them. Meanwhile, a lot of other cloud providers are still just talking
about their own ML chips. We've also made a lot of exciting progress in building the software tool chain that supports Trainium and Inferentia. So AWS Neuron is our
software development kit that helps customers get maximum
performance from ML chips. Neuron supports machine
learning frame frameworks like Tensor Flow, PyTorch, with JAX support coming really soon. And so customers can use
their existing knowledge to build training and inference pipelines with just a few lines of code. And Neuron supports the
most popular models, including 93 of the top 100 and counting. Importantly, you also need the right tools to help train and deploy your models. And this is why we have Amazon SageMaker, our managed service that makes it easy for developers to train, tune, build, manage machine learning
and foundation models. In the six years since
we launched SageMaker, we've introduced a lot
of powerful innovations, like automatic model tuning,
distributed training, flexible model deployment,
tools for ML ops, and built-in features like responsible AI. SageMaker has been instrumental
in democratizing ML for tens of thousands of
customers, including Stability AI, AI21 Labs, Thomson Reuters, AstraZeneca, 3M Health Information
Systems, Ferrari and more. And SageMaker can train models with up to billions of parameters. For example, TII has trained the 180 billion parameter Falcon FM, which is the world's largest
publicly available model on SageMaker, using 3000 GPUs, and petabytes and petabytes
of data stored in S3. And you can also deploy this
top rank model using SageMaker. We've also been working
closely with Hugging Face to support their models on SageMaker, and have collaborated to create the Hugging Face AWS
deep learning containers, to help accelerate training
and deployment of FMs using SageMaker, Trainium and Inferentia. So if you're building your own models, AWS is relentlessly focused
on everything you need, the best chips, most
advanced virtualization, powerful petabit scale
networking capabilities, hyperscale clustering and the
right tools to help you build. But of course, an enormous
number of organizations just want to be able to access the most powerful models out there. They wanna get started quickly, experimenting with different FMs, and testing different use cases, and to meet the enormous
opportunity head on. But they also have questions like, which model should I even use? And how do I decide which
are best for my application? How can I move quickly to build and deploy generative AI applications, and how can I keep my
data secure and private? That's why we're investing in that middle layer of the stack. Now we know that many of
you need it to be easier to access a powerful and
diverse set of models, LLMs, other foundation models, and then to quickly build
applications with them, all the time maintaining
your security and privacy. And that's why we built Amazon Bedrock. So Bedrock is the easiest
way to build and scale generative AI applications for LLMs and other foundation models. And customers in every industry
are already using Bedrock to reinvent their user experiences, their products and their processes. And to bring AI into the
heart of their businesses. So why Bedrock? So first, you enjoy the the
broadest choice of models, many of which are available first, or in some cases only on Bedrock. You can add your own
business context quickly, easily and privately with
the broadest selection of customization options, and you get enterprise
grade security and privacy because we designed it
that way from day one. And the customer excitement's
been overwhelming. We made the service generally
available in September, and now over 10,000 customers are using Bedrock around the world, across virtually every industry. Adidas is enabling developers
to get quick accurate answers to deep technical questions. Carrier is using Bedrock to
combine historical actions, predictive analytics, systems alerts, and data trends to
provide recommendations. And these are helping customers reduce their energy consumption
and cut carbon emissions. NASDAQ is using Bedrock to automate investigative workflows on suspicious transactions, and to strengthen their
anti financial crime and surveillance capabilities. So many more, including
Bridgewater Associates, Clariant, Cox Automotive, GoDaddy, LexisNexis, Merck, Omnicom Group,
are all using Bedrock. They're all using it to build
GenAI applications today. But it's still early days. Everyone's moving fast,
experimenting, learning, and at the same time, the
generative AI technology itself is evolving quickly, with new developments
practically every day. And customers are finding that different models actually work better for different use cases, or on different sets of data. Some models are great for summarization, others are good for
reasoning and integration, and still others have
awesome language support. And then there's image generation, search use cases and more. All coming from both proprietary models, and models that are publicly
available to everybody. I mean, things are moving so fast. And in that type of environment, the ability to adapt is the
most valuable capability that you can have. There's not going to be
one model to rule them all. And there's certainly not
going to be one company providing the models that everybody uses. So you don't want a cloud provider who's beholden primarily
to one model provider. You need to be trying
out different models. You need to be able to
switch between them rapidly, even combining them
with the same use cases. And you need a real
choice of model providers, as you decide who's got
the best technology, but also who has the dependability that you need in a business partner. I think the events of the past 10 days have made that very clear. We've been consistent
about this need for choice for the whole history of AWS, and that's the approach
that we clearly laid out when we started talking about
our generative AI strategy almost a year ago. And that's why we continue to innovate, to make building and
moving between a range of foundation models
as easy as an API call. Today, Bedrock provides access
to a wide range of models, from leaders like AI21,
Anthropic, Cohere, Stability AI, as well as Amazon's own
Amazon Titan models. But we didn't stop there. We were the first to make
Meta's Llama 2 model available through a fully managed API. And we also recently announced our expanded collaboration with Anthropic. So Anthropics, a longtime AWS customer and one of the world's leading
foundation model providers. Their model is Claude, and they got others as well, excel at a wide range of tasks from sophisticated dialogue and
creative content generation, to complex reasoning and
detailed instructions. So as part of the collaboration, Anthropic is gonna use
Trainium and Inferentia to train future generations
of their models. And Bedrock customers are
gonna have early access, not available in other
clouds to unique features for model customization and
fine tuning capabilities. So to dig in on what we're
doing with Anthropic, I'd like to bring out our partner, Dario Amodei, who's the CEO
and Co-founder of Anthropic. Let's give him a hand. (upbeat dramatic music) Great to see you, all right. Oh, we have chairs, let's sit. Dario, thanks for being here. - Thanks for having me.
- Yeah. Earlier we talked about how Claude is available on Bedrock, but before we dig into that, let's just talk a little bit
about what Anthropic is about, and what makes Claude so unique. - Yes. So Anthropic was founded on the idea that we can make AI models
that are safe, reliable, and steerable while still
being highly capable. The founders of Anthropic
were a set of people who worked at OpenAI for several years. We developed ideas like GPT-3, reinforcement learning
from human feedback, scaling laws for language
models, some of the key ideas behind the current generative AI boom. Seven of us left and founded Anthropic. Our foundation model Claude, in addition to being safe and reliable, excels at a lot of knowledge work tasks like content generation,
summarization, Q and A, complex reasoning, things like that. We have an industry leading context window of 200K tokens, that's
about 150,000 words, which makes it well suited for areas like legal, finance, insurance, coding, things like that. And about half of the Fortune 500 are using, or are testing Claude. - So amazing progress. So just a few months ago, obviously Amazon and Anthropic announced kind of an enhanced
strategic collaboration. Anthropic named AWS as
your primary cloud provider for mission critical workloads, and you're gonna be training, and deploying future generations of Claude and foundation models on AWS. So just tell us a little bit
about, from your perspective, what led up to this and how that relationship
between Anthropic and AWS grew. - Yeah, so, Anthropic's use of AWS goes back to our early history. Back, you know, all the way
in the midst of time, in 2021. - Way back.
- We trained- - [Adam] I mean who can
even remember that, right? - We trained our first model on AWS, and so we're excited now
to say that, you know, you guys are our primary cloud provider for mission critical workloads. I would say the partnership
has three components. First is the compute side,
which I just discussed, which we've been doing, you know, throughout our history. The second is on the
provision to customer side. We're very excited to
hear about Amazon Bedrock. We really recognize the
value that we could provide to customers working together. And the third side is the hardware. We're currently working with you guys to optimize Trainium and
Inferentia for our use cases, and hopefully for other use cases. And we really believe that we can achieve that four x faster than the
first generation Trainium, that you just announced, you know, if we optimize these things together. And so putting together
these three pieces, layers of the stack, really
allows us to provide things to customers, together, that
couldn't be provided before. Things like customization of
models, use of proprietary data and unique fine tuning features. - Well, it's a... The respect we have for you
guys is just tremendous. I mean, truly world-class expertise for the Anthropic team. I've really enjoyed working together, and you know, the models
are just so powerful. So obviously we are aiming to, and I think are in the
process of improving incredible array of customer experiences through this collaboration. And of course, our own Amazon
developers also have access to Anthropic models on Bedrock
and throughout the company. And we're not only working to
provide customers with this, but we're giving them early access to unique features that they're
not gonna be able to get anywhere else. And we're out there together, with our generative AI innovation center, working with customers in the field, helping them to actually get targeted. And that's allowing us, of course, to bring this expertise, and the domain specific expertise that they need to customize the models. And as we've been out there together, and it's been hot and heavy, particularly the past,
you know, few weeks, but before that as well, what are some of the most
interesting use cases that you've seen? - Yeah, so I'll give three examples across different industries. There's, there's many others, but this should give a sense of the range. And, you know, we've seen
a lot of organic adoption. One example in the biomedical space, we're happy to announce
we're working with Pfizer. I'm a former biologist
myself, so I'm really excited for any way AI can advance medical science and, you know, bring lifesaving drugs to people around the world. In the legal space, we made
a custom fine tuned model, deployed it through you
guys for LexisNexis, to help with their legal search, you know, in a way
that's kind of fact-based and produced citations. And in the finance space we're working with Bridgewater Associates to develop an investment
analyst assistant. Beyond that, you know, just some examples, we've engaged with seven outta 10 of some of the largest
players in the insurance and banking industries, and, you know, could tell similar stories about other industries. So, you know, we're excited
to see new use cases and where all of this goes. - That's incredible momentum. Just the financial services
example to be working with that many leaders. So just last week, you're
not standing still, right? So, just last week you
announced Claude 2.1, so that's got a whole
lot of new advancements and key capabilities. Maybe you can tell us a
little bit about that. - Yeah, so Claude 2.1 really,
you know, is a release that's targeted at enterprises. So a few different features. One is we've doubled our context
window to 200,000 tokens. Tokens is a term of art in the AI space. What it amounts to is about 150,000 words, which is a relatively long book. And so use cases like
reading financial statements, reading the entire long book, and one of my favorites, putting multiple documents
in the context window and asking the model to
compare them or integrate them. A second feature is that
we've reduced the rate of so-called hallucinations,
which is when models say things that aren't true by about two x. We've seen across the industry, a core obstacle to deployment
in the enterprise world is people are worried that models will say things that aren't true. No one has perfectly solved this problem, but we are really leading the way in getting more and more reductions. And then we also have some beta features. We're introducing system prompts,
tool use as a beta feature and a developer workbench. Again, all tools targeted at people building for enterprises. We're excited to see all the use cases that emerge from this. - I'm speechless. That is such incredible progress. You talked about the
reduction for hallucination, really important topic. You know, just more broadly,
I think we share a belief that AI should be, you know,
safe, secure, responsible and benefit society. And we've been engaged a lot of groups and organizations that are
promoting the development and deployment of responsible AI. In fact, I keep bumping into Dario around the world at some of
these various gatherings, which is nice. So I know this is a
particular passion of yours. And can you tell us a
little bit more about Anthropics work in this area? - Yeah, I mean, you know,
obviously this is something we were founded on and
that we care about deeply. You know, I would
categorize into two areas, where we've done work in this area. One is the way we design our own models, and the other is the way
that we try to set an example for the ecosystem, and bring others along, so that ultimately everyone
becomes safer and more reliable. So large fraction on the
first one, a large fraction of our research org works on safety. We've put a lot of work into, you know, making our models hard to break, making them secure against
misuse and harmful use. There was a recent study
out of Carnegie Mellon, where people tried to adversarially
break different models, and it found that they were able to successfully break Claude
10 times less often than they were able to break
the competitor's model. So, there's a real difference here. (audience applauding) Second, we've put a lot of
effort into interpretability seen inside the models. There's no commercial
deployments on that so far, but it's been a long
arc, and I am hopeful, that relatively soon for both compliance and for understanding what models do, and therefore reducing bad
behavior from the models, we'll be able to make a lot of progress. On the wider ecosystem, you know, we always try to
create a race to the top, do something, and then competitors, other players in the
ecosystem follow along and do the same thing. An example of this is a
responsible scaling policy, where we front loaded
safety investment ahead of model scaling, a kind of
pre-deployment testing regime, like what's done for cars and airplanes. And in the policy space
we've continually advocated for governments to build up their ability to evaluate and measure AI systems. If we're not able to
evaluate and measure things, we'll get either no regulation or poorly informed regulation. We need to all work together to make sure that we can develop quickly, responsibly, but in a way that's secure. - I just love the concept
of the race to the top. It's so positive. And what you're doing is so amazing. I mean, we've said it the whole 17 years that AWS has been providing
services that, you know, so much of what gives us
energy is just working with the coolest, most cutting
edge, you know, startups and entrepreneurs in the world. And there is just no better
example than what's happening in generative AI now and
what Anthropics doing. So it's great to be
partnering together so deeply and really appreciate you
joining me here on stage today. - Yeah, excited for
everything we can do together. - Thank you. All right. (upbeat dramatic music) Alrighty, so amazing. Thank you so much again, Dario, for coming out and joining
us, really appreciate it. So the innovation around
generative AI models is explosive and we really look forward to working with all of our Bedrock
partners for years to come. Incredibly innovative space. And we're also excited about
the large language models and other foundation models that we're building ourselves at Amazon. So our Amazon Titan models were created and pre-trained by AWS. And we are building
our own models in order to bring actually a few
benefits to customers. So we do have a lot of expertise here. I mean, Amazon has been
working on AI for 25 years, got a long track record. We use the technology across our business, and we've learned a lot about
building and deploying models. And as a result we can provide
a whole range of models that offer really powerful capabilities, and also great economics to support a whole bunch of use cases. And we carefully choose
how we train our models and the data that we use to do so. And we're actually gonna
indemnify customers against claims that our models, or their outputs, infringe on anybody's copyright. Now they're already multiple
models in the Titan family. We've got Titan Text Light, which is ideal for text-based
tasks like summarization. We've also got Titan Text Express, which is great for more
advanced general language tasks, like copy editing, copywriting. And we've also got the
Titan text embeddings model, for search use cases or personalization. And we're gonna continue adding
models to the Titan family. And if you're interested in that, then I highly recommend you tune in to Swami's keynote tomorrow. A little word up. Now all of the foundation models that I've been talking about,
they're all extremely capable and they all enable a lot
of great applications. But to really drive your
business, you need generative AI that understands your business,
understands your customers, understands your
products, your operations. And that's only possible
when you use your own data to improve and customize the models. The models have to become unique to you. So Amazon Bedrock is giving you all of the purpose-built
capabilities that you need to customize the AI
applications with your own data. So with a few clicks you can
go from generic applications, to ones that are highly
specific and highly tailored. So one important technique
is called fine tuning. Now with fine tuning you point the models to examples of data in S3, that you've labeled, labeled to provide context. For example, you might point to proposals, and you've labeled some of
those proposals as strong. So in this way the model learns from you, learns from your data, what
matters the most to you. And then Bedrock makes a
copy of the base models, it trains it, it creates a private fine tuned model, so you get tailored responses. So fine tuning is available in Amazon Titan Text Lite and Express. And today it's also generally available for Cohere Command Lite and Meta Llama 2. And fine tunings coming
to Anthropic Claude soon. (audience applauding) So a second technique for
tailoring foundation models for your business, is called retrieval
augmented generation, or RAG. Another one of the terms
you see out there a lot. So RAG allows you to
customize a model's responses when you want the model
to consider new knowledge, or up-to-date information. So Bedrock does this for you by calling and bringing back your own data, your data from multiple sources, including document repositories, databases and other APIs and more. So for example, the model might use RAG to retrieve search results from our Amazon OpenSearch Service, or documents from S3. In September we introduced
a RAG capability, called Knowledge Bases, that connects FM to your
proprietary data sources to supplement your prompts
with more information. And today knowledge basis
is generally available. (audience applauding) All right, thank you. Alright, now, a third way
that you can customize models inside of Bedrock is a technique called
continued pre-training. So continued pre-training is a technique that uses large amounts of unlabeled data, before you fine tune
into the labeled data. Here's unlabeled data, like the raw text of internal reports, financial plans, or research
results that you have. And that's gonna be used to
improve the model's knowledge and its reasoning capabilities, using things in your own specific domain. So continued pre-training is of course available in Bedrock today. Now all of these techniques, they're all powerful tools in
making generative AI targeted, making it truly useful for your business. And Bedrock makes it really easy to combine these techniques together. So a great example of a
customer using Bedrock to customize their
generative AI applications, is Delta Airlines. So Delta's actually modernizing its customer service
capability with generative AI, built on Amazon Bedrock. The new tool is gonna be
available to answer questions in a more relevant and
conversational style, by accessing Delta's travel policies, realtime flight schedules,
customer re-booking options, and airport conditions. So for example, it could
respond to simple questions like, how many bags can I
check on a Delta flight? And it can also do more
complex questions like, can I carry my cat with me
in the cabin to Jamaica? Of course, what else are
you gonna do with your cat? Delta's been testing with different foundation
models in Bedrock, including Anthropic Claude
2.0, Claude Instant, and a fine tuned Amazon
Titan Text Express model. And in this way they
can use the right model to answer different customer questions. It's really powerful how they can combine the most relevant information needed to answer any specific question. Now with the new customer service tool, Delta Airlines is aiming
to provide a delightful and conversational
self-service experience, and it's gonna improve
customer satisfaction. So everyone wants more relevant
generative AI applications, right? And Bedrock's making it a lot easier to customize those applications
with your own data. But at the end of the day, you want FMs to do more than just provide useful and targeted information. Ultimately, what you're
gonna want is to use FMs to take action, to
actually get things done. And we wanted you to
be able to use Bedrock to complete actions like booking travel, filing insurance claims,
deploying software, maybe ordering replacement parts. And this usually requires orchestration between multiple systems that
operate in your organization. And that's why a few months ago we introduced Agents for Bedrock. And today I'm happy to announce that this powerful feature
is today generally available. (audience applauding) Now with Agents, GenAI
applications can execute multi-step tasks across company
systems and data sources from answering customer questions about product availability, to taking a sales order and more. And building with Agents is really easy. First, you use Bedrock Setup Wizard to select the appropriate model. Then you give the model basic instructions to help the model
understand how to respond, like you are a friendly
customer service agent who helps customers return items. Next, you specify the lambda functions that execute the API calls
to retrieve the information that's relevant to the task. For example, return policies,
or inventory levels. And finally, you select the data sources. Then you just select 'create agent' and Bedrock configures your
agent and it's now ready to run. When the agent receives a request, it uses its reasoning capabilities to analyze and to plan the task. And the model works out
the sequence of steps, the information needed, the APIs to call, and when to call them. Then it executes the plan, taking all the steps that it
needs to gather the information and then complete the task. All in the background. No need to engineer prompt, no need to train the FMs. No need to manually connect systems. With agents, generative AI is becoming an integral part of your business, making it easy to take
actions and get tasks done. No heavy lifting required. So we've talked about how
Bedrock's making it easy to choose amongst leading models. And also how it gives you
so many different ways to customize your
generative AI applications with your own data. And how it can help
combine the power of models into your organization
systems, your datas, and your API to actually take action. But you won't actually use
any of this in a serious way if it isn't secure and private. I mean, you demand the same
rock solid reliability, the same security for generative AI, that you do for any workload. You really need this to have
nothing more and nothing less than everything you expect
from an AWS service. It's just like S3, EC2, RDS, Bedrock has got to have
enterprise grade security features to make your data safe and private. And that's what it has. Bedrock customers can trust that their data remains protected. No customer data is gonna be used to train or improve the
original base models. When you tune a model, we make
a private copy of that model, we put it in a secure
container and it goes nowhere. Your data's never exposed
to the public internet, never leaves the AWS network. It's securely transferred
through your VPC. It's encrypted in transit and in rest. And Bedrock enforce the
same AWS access controls you have with any of our other services. Bedrocks also supports a
variety of regulatory standards. It's HIPAA eligible, it can be
used in compliance with GDPR. We've added a bunch of other security and governance capabilities, including integrations with CloudWatch, to track usage and metrics, CloudTrail, to monitor API activity and troubleshoot activity. And today Bedrock is SOC compliant. Now this helps ensure that customers can have strong controls
and policies in place, and helps it be used in every business. Now beyond security, we need generative AI to be deployed in a safe, trustworthy, and responsible fashion, like we were talking
about a few minutes ago. And that's our North Star. And the capabilities that make GenAI such a promising tool for innovation, also do increase the potential for misuse. We've got to find ways to unlock generative AI's full potential while mitigating the risks. Now dealing with this
challenge is gonna require unprecedented collaboration in a truly multi-stakeholder effort, across technology
companies, policy makers, community groups, scientific
communities, and academics. We've been actively participating
in a lot of the groups that have come together
to discuss these issues. Earlier this year, I
joined President Biden and other industry readers like Dario, as he announced that we've made a series of voluntary commitments to
promoting the safe, the secure, and the transparent
development of AI technology. And just this month I joined
UK Prime Minister Sunak for his AI safety summit, to discuss new approaches
including the formation of their new AI safety institute. Now, an important component
of responsible AI, is promoting the interactions
between your consumers and the applications that are safe, that avoid harmful outputs, and that stay within your
company's guidelines. And the easiest way to do this is to actually place limits on what information the
models can and can't return. And we've been working hard here. And today we're announcing
Guardrails for Amazon Bedrock. (audience applauding) Now this is a new
capability that helps you easily safeguard your
generative AI applications with your responsible AI policies. To create a guardrail Bedrock provides configuration wizards. You can enter and enter a
natural language description of the topics that you
want the model to avoid. Guardrails can be used with
any of the foundation models accessible via Bedrock, including the custom
models that you create through your own fine tuning. And you can also use
Guardrails with Agents. So now you can have a
consistent level of protection across all of your GenAI
development activities. For example, a bank could
configure an online assistant to refrain from providing
investment advice. Or to prevent inappropriate content, an e-commerce site could ensure
that its online assistant doesn't use hate speech or insults. Or a utility company could remove personally identifiable
information, or PII, from a customer service call summary. So we're approaching the
whole concept of generative AI in a fundamentally different way, because we understand what it takes to reinvent how you're going
to build with this technology. Just as you've used AWS to reinvent how you've built and innovated for years, and ultimately to transform
how your organization operates. Whether you're migrating
enterprise applications, building the next hot startup, or deploying your generative AI strategy, our approach remains the same. Customer obsession, innovation,
and a long-term view of what joint success really means. To tell us more about how
her company is built on AWS and is now reinventing with generative AI, please welcome to the stage Lidia Fonseca, the Chief Digital and
Technical Officer of Pfizer. (audience applauding and cheering) (bright uplifting music) - Hello everybody. (audience applauding and cheering) I'd like to start by asking, how many patients do you think
Pfizer treated last year? You may be thinking in the millions, but in fact we treated
1.3 billion patients with our medicines and vaccines. That's one out of six
humans on the planet. A truly humbling accomplishment. And we're grateful to be
recognized as an innovative and ethical company, listed in Fortune's World's
Most Admired Companies, Forbes' World's Best Employers and Time Magazine's
World's Best Companies. Pfizer is experiencing an exciting period of innovation and growth, with the launch of 19 medicines
and vaccines in 18 months. This has never been done, by us, or any other pharma company. It is an incredibly ambitious goal. And we are well on our way,
with 13 already launched. Digital data and AI are
critical to our success. And what differentiates Pfizer is that we apply this
across all business units, our 36 plants and the 149
markets that we serve, quickly and at scale, to bring medicines to patients faster. Our success today rests on the groundwork we laid for technology and AI to flourish. Centralizing our data, creating standard platforms,
cultivating strong talent, and building a secure foundation, all to innovate for maximum impact. We could not have achieved
this tremendous reach without our close relationship with AWS. So let me share a bit
about our journey together. In 2019, Pfizer and AWS created a pioneering scientific data cloud that aggregates multimodal data from hundreds of laboratory instruments, empowering our scientists to search all historical
molecule and compound data in real time, compared to weeks or months in our prior fragmented environment. And the scientific data
cloud accelerates analysis and computational research. And with AI algorithms that
help us identify and design the most promising new molecules. Building on that success in 2021, Pfizer embarked on one of
our boldest initiatives, going from 10% of our
core in the cloud to 80%, moving 12,000 applications
and 8,000 servers in 42 weeks. One of the fastest migrations
for a company our size. This move to AWS saved
us $47 million a year and helped us exit three data centers, reducing 4,700 tons of CO2 emissions, or the equivalent energy use
by a thousand homes in a year. But very importantly, it enabled
us to innovate with speed and at scale. For example, we reduced time
to set up computing capacity from months to hours, speeding up data generation
for drug submissions by 75%. And when Covid hit, AWS was one of the first
to offer their help, scaling us into the tens of thousands of additional cores in the cloud. Accelerating manufacturing
and clinical trials. When we ran computationally
intensive analysis to understand how to
manufacture our Covid vaccine, AWS jumped in with additional
CPUs for us to move fast. And when we needed to submit
data to the FDA in under a day, AWS increased our capacity, enabling Pfizer to move
at the speed of science. In fact, it took just 269 days, from when we announced
our plans to develop a COVID-19 vaccine with BioEn Tech, to the day that we received the FDA's emergency use authorization. This normally takes eight to 10 years. (audience applauding) Thank you. With the world waiting,
we needed to manufacture and distribute the vaccine
as quickly as possible. We held firm to our
mantra, science will win. Well, science did win and
digital helped us do it faster. Pfizer's industry leading
digital operations center allows colleagues to see
production status across plants and resolve issues in real time, yielding a 20% throughput increase, our mRNA prediction algorithm yielded 20,000 more
vaccine doses per batch. Imagine the lives impacted. To give you an idea of the scale, prior to Covid, Pfizer produced 220 million vaccine doses in total and scaled to 4 billion
of Comirnaty in 2022. This is now at the heart
of how our plants operate and we're applying these innovations to our other medicines and vaccines. More recently, AWS helped us supply AI to generate alerts about potential
supply chain disruptions, prompting us to reroute shipments to continue delivering medicines
to clinical trial sites, ahead of Hurricane Ian's landfall. Digital and AI make it possible for Pfizer to bring medicines to
patients around the globe faster than ever before. Now let's look at what's next. As you heard, Pfizer
has been harnessing AI to drive innovation and
productivity on a global scale. More recently, we are
leveraging generative AI, which is estimated to
deliver annual cost savings of 750 million, to a billion
dollars in the near term. Real tangible value. Using AWS cloud services, Pfizer quickly deployed VOX, our internal generative AI platform, allowing colleagues to
access large language models available in Amazon Bedrock and SageMaker. AWS's breadth of services and the variety of LLMs in
Bedrock means we can select the best tools for use cases
in R and D, manufacturing, marketing and more. Enabling Pfizer and AWS to prototype 17 different use cases
in a matter of weeks. AI and generative AI will help us identify new oncology targets, a process that is largely manual today. With AI, we can search
and collate relevant data and scientific content
from many more sources in a fraction of the time. And algorithms generate and validate potential targets to improve our scientific success. With Bedrock, we can scan the marketplace for companies and assets, automating the creation
of company profiles, to prioritize potential acquisitions. In manufacturing, Bedrock takes the optimal
process parameters to identify what we call the golden batch, and uses generative AI to detect anomalies and recommend actions to
our operators in real time, aiming to improve cycle time by 25%. AWS's agile culture aligns well with Pfizer's light speed way of working. We look forward to
building on our learnings, to understand not just how AI
can transform our daily work, but how they help us disrupt the industry. To close, we're excited to accelerate our battle against cancer, with Pfizer's proposed
acquisition of Seagen, a biotech focused on cancer therapies. We plan to bring Seagen's leading edge antibody drug conjugate technology to more patients with cancer faster. You see, we hope to achieve with cancer what we were able to achieve with Covid. Our work with the AWS
will help us continue this great momentum, fueling our ambition to change a billion lives every year, not just in a pandemic year. Thank you AWS for helping
Pfizer bring breakthroughs to patients faster, because in the battle against disease, time is life. Thank you all. (audience applauding and cheering) - One out of every six
people on the planet are impacted by Pfizer. I mean, that is just so
impressive, so impressive. Lidia means so much to us,
honestly that you've trusted AWS to help deliver all these capabilities, these breakthroughs that
are changing people's lives. Thank you very much. - And thank you, Adam. AWS has just been a amazing partner. - Thank you. We look forward to more.
- Thank you everybody. (dramatic music) - So organizations across
the globe are using AWS for agility, for speed, and for the scale that they
need to drive their futures. But this requires employees
with top-notch cloud skills. And we know that not enough
people have that cloud expertise and that there's a large skill
gap that's gotta be closed. So generative AI is of course the next incredibly exciting new opportunity, but it's only gonna add to that gap. And so we're investing to
help provide the cloud skills that are gonna be needed across
the world for years to come. AWS is committed to training
29 million people for free with cloud computing
skills by the year 2025. And we're well on our way with
21 million already trained. (audience applauding) And we just recently launched
the AWS Cloud Institute, which is providing deep cloud expertise to help students launch
an entire cloud career. And we've also launched a hundred
different machine learning and AI courses and resources
to support innovation. And just over a week ago,
we announced AI Ready, which is our new commitment
to provide free AI skills to another 2 million people by 2025. And to help students around the world to learn the foundation skills and to prepare for a career in tech, we actually created an AI and machine learning scholarship fund. Let's meet some of the recipients. - [Vani] My name is Vani Agarwal. - [Jose] My name is Jose Tapia. - [Olympiah] My name is Olympiah Otieno. - [Jose] Without the AWS
AI and ML scholarship, I wouldn't have been able
to develop my own questions, know what tools are
available to solve them. - [Vani] The Nano degree
has definitely unlocked my interest in AI. I constantly try to go out there and talk to girls who are
confused about what to do in life. - [Olympiah] In Kenya
due to crop diseases, the rates at which the
yields have been dropping is devastating. So my colleagues and I have
made a mobile application. With AI, farmers prevent the
spread of the crop diseases. The fact that I was able
to go through the program, I now know that I am capable
of doing well in this field. (audience applauding) - I mean, this is why
we do this stuff, right? But we're actually lucky enough
to have Jose and Olympiah here with us today. So come on, let's give
'em a round of applause. (audience applauding) There we go. (audience cheering and applauding) To both of you, I'm so proud. I'm really proud of
what you guys have done. And I just really hope
the reinvent this week is just a great, great experience for you. Thank you so much for letting
us share your stories. And for everyone watching,
we wanna empower all of you to build and to have that kind of impact. And we see a huge opportunity to help that by infusing generative AI
into all of the systems that people are using
into their daily lives. We believe that generative AI
should help everyone at work, seamlessly with helpful,
relevant assistance, whether or not you know the first thing about foundation models, RAG or any of the rest. And that brings us to
the top of the stack. The applications that
you use that are powered by generative AI, leveraging
foundation models. We believe that generative AI, of course, has the potential over time to transform virtually every customer experience that we know. It can pour through the nooks
and crannies of a system and actually help you find data that you didn't even know existed. And then help you put it to optimal use. It can generate insights that
can save hours of manual work, and it can take projects that were slogs and make them a snap. Now, one important example is how generative AI can
make the task of writing code radically faster. And so last year we released
Amazon CodeWhisperer, which helps you build generative AI, which helps you build
applications, I should say, faster and more securely by
generating code suggestions in near real time. Now with CodeWhisperer, developers use a natural language prompt and then receive accurate code suggestions in every popular IDE. And for 15 of the most
popular programming languages. A CodeWhisperer gives
developers a powerful boost for general code examples. And of course, we made
sure that it's great for providing specific
suggestions for AWS Code. Customers, including Warner
Music Group, the Bundesliga and the Cigna group have
embraced CodeWhisperer to accelerate their
development capabilities. And Tata Consultancy Services, Accenture, and HCL Tech are making
CodeWhisperer available to tens of thousands of employees each. Now KONE, a global leader in the elevator and escalator industry. They actually believe that CodeWhisperer is a game changer for them. We want as many developers as
possible to reap the benefits. And so what we've done
is to make CodeWhisperer, to make the code suggestions in it free, free for individual use. Now the recommendations in CodeWhisperers are great for general coding tasks, but general coding suggestions can only get your developers so far. Imagine you've just hired a new developer. Even if they're world class, they're not going to do their best work until they understand your code base and your internal best practices. And AI power coding tools are similar. You can have a great tool,
but for the kind of speed and relevancy that makes
the most difference, your solution has to understand
the context of your code. And the best place to find that context, is inside of your internal
APIs, your libraries, and your packages. That's why a couple months
ago, we previewed a capability that allows CodeWhisperer to securely learn from
your internal code base, to provide much, much more
pertinent, customized, and useful code recommendations. And with customizations, CodeWhisperer is an expert on your code, and can provide much more
tailored suggestions, creating an understanding of
your internal SDKs, your APIs, your languages, and your classes. And of course, any
customizations are isolated. We never use customer's
content from CodeWhisperer to train any of our underlying models. Now to understand the impact
of this kind of customization, we worked with Persistent, a
digital engineering provider. Now, their studies showed that developers using CodeWhisperer
completed their tasks an additional 28% faster with
customizations than without, 28%. And AWS is the only major cloud provider that's currently authorizing offering a customization capability
that anyone can sign up for and use today. And we are really excited
about CodeWhisperer, about CodeWhisperer customizations, and how it's gonna help
you accelerate your work and increase your impact. But we're just getting started and we see a lot more potential
uses with generative AI. For example, think about
generative AI chat applications. So these days it seems
like just about everyone is experimenting with them, right? And what the early providers
in the space have done is really exciting and it's genuinely super
useful for consumers. But in a lot of ways these
applications don't really work at work. I mean, their general knowledge and their capabilities are great, but they don't know your company. They don't know your data, your customers, or your operations. And this limits how useful
their suggestions can be. They also don't know much
about who you are at work. They don't know your
role, your preferences, what information you use, and what you do and don't have access to. So critically, other providers
who have launched tools, they've launched them without data privacy and security capabilities, that virtually every enterprise requires. And so many CIOs actually banned the use of a lot of these most
popular chat assistants inside their organizations. And this has been well publicized. And just ask any chief
information security officer, or CISO, and they'll tell you, you can't really bolt on
security after the fact and expect it to work as well. It's much, much better to build security into the fundamental
design of the technology. So when we set out to build
generative AI applications, we knew we had to address these gaps. Had to be built in from the very start. And that's why today I'm really proud and excited to announce Amazon Q, a new type of generative
AI powered assistant, designed to work for you at work. (audience applauding and cheering) Q lets you answer questions quickly, with natural language interactions. You can easily chat, generate
content, take actions, that's all informed by an
understanding of your systems, your data repositories, your operations, and of course we know
how important rock solid security and privacy are, so that Q understands and
respects your existing identities, your roles and your permissions. If a user does not have permission to access something without Q, they cannot access it with Q either. And we've designed Q to meet enterprise customers' stringent
requirements from day one. And we never want business
customers content, we're never gonna use that content for Q, to train underlying models. I really believe this is
gonna be transformative. We want lots of different kinds of people, who do lots of different kinds of work to benefit from Amazon Q. Now, as you might expect,
we want Q to help developers and builders be more efficient, more knowledgeable, and more proficient. So for starters, Amazon Q
is your expert assistant for building on AWS. Now to supercharge work
for developers and IT pros, we've trained Amazon Q on 17 years worth of AWS knowledge, so it can
transform the way you think, optimize and operate applications
and workloads on AWS. And we put Amazon Q where you work. So it's ready to assist you
in the AWS Management console and documentation, in your
IDE via CodeWhisperer. And in your team chat rooms like Slack. You can chat with AWS Q to
explore AWS capabilities, learn unfamiliar technologies,
architect solutions. It's an expert in AWS
architected patterns, best practices and
solution implementations. For example, you could ask Amazon Q, how do I build a web application with AWS? What are my options? It's gonna answer with a
list of potential services that you could use, such as Amplify, Lambda or EC2. And then it's gonna offer reasons why you might consider each service. From there, you can further
narrow down your options through natural language, like which of these would be preferred if my app only needs to
run for a few seconds and only has very infrequent traffic? Lambda. Amazon Q's gonna take your
requirements into consideration and it's gonna provide you with the best possible recommendations. And once you've chosen a service, you can ask Amazon Q,
how do I get started? And Q's gonna give you
step-by-step instructions for configuring the solution, plus links to relevant information, to documentation and so on. And this is just the
beginning of how Amazon Q is gonna help you with your work. For example, if you encounter
an error with Lambda in the console, you just press the troubleshoot
with Amazon Q button. And Q is gonna research the error, and it's gonna suggest how to fix it. Amazon Q can also troubleshoot
your network connections. It can analyze your end-to-end
network configuration and help resolve
connectivity issues quickly. That is a huge time saver. Now what else can Q help us with? How about choosing the
optimal EC2 instance? So you can just tell
cue about your workload and get an accurate, quick and economical instance type recommendation. For example, which EC2 instance types provide the highest
performance for video encoding and transcoding workloads
supporting gaming use cases? And you get your
recommendation, it's that easy. This is all very cool, and we expect that Amazon
Q is gonna save customers so much time in architecting
and troubleshooting and optimizing your workloads. But that's not all the Q can do for you. Amazon Q is also gonna be
in the IDE for developers. If you're unfamiliar
with the code in the IDE, or you need to dig in into
how a specific function works, Amazon Q and CodeWhisperer is
gonna help you get up to speed by using its deep knowledge of AWS, and its understanding of your code bases. So you need to add tests, just ask Q, and it's gonna automatically
generate the test for you. And this is gonna make a huge
difference for developers. But we knew we could do even more to help you build even faster. For example, think about everything that goes into building a new feature. You have to understand
requirements, think through design, write all the code, add tests,
document the code and so on. And a file could spend hundreds, to even thousands of lines of code. And this is a daunting task. If you're using Amazon CodeWhisperer, it's gonna generate chunks
of new code for you, which is a real time saver, but that still leaves all the rest. So this sounds like a job for Amazon Q. Specifically Q's feature
development capability, which is gonna automate the
multi-step development task of adding new features
to your application. So instead of all that work,
now you just write a prompt and Q is gonna use its expertise, enriched with an understanding
of your project code, and it's gonna do all the
heavy lifting for you. So Amazon Q will create the draft plan, which you can then collaborate and iteratively improve
using natural language chat, till it's polished and ready to go. And then Q really kicks into gear. It's gonna implement the
plan across your files in your application, and you remain in control
using your expertise to review changes and ensure
the very best quality. So this is hours of work
that Q can do for you with just one good prompt. And this feature is available
today in Amazon CodeCatalyst and will soon be fully
available in the IDE. You will not currently
find an AI assistant from another major cloud provider that can do this for you today. But we're not done yet, let's keep going. We know that developers
spend a lot of time slogging through the muck that comes with maintenance and upgrades. Maintenance and upgrades, big deal, right? So take take language version
upgrades, for example. How many of you're using
older versions of Java because it's gonna take months, or even years to upgrade? Hands. Oh yeah, a lot of 'em. So migrations inevitably come with a lot of nuance and edge cases. And so you end up spending
a ton of time writing and rewriting the code
over and over again, and trying to get it to compile and work. And sometimes these programs take so long, that it feels like you're in
a migration infinity loop, just as one ends, the next begins. It's not hard to understand the temptation to just maintain status quo, but doing that means that you miss out on performance improvements. And it means much worse
that it can also open you up to potential security vulnerabilities. And that's why we built into Amazon Q, code transformation, which
helps you with upgrades to transform code in
a fraction of the time that it takes today. All a developer has to do- (audience applauding) All a developer has to do is ask Q to perform the code transformation. And it handles the rest. From identifying and upgrade and mandatory code package
depositories and frameworks, to replacing deprecated code, and incorporating security best practices. Q's even gonna run tests on
the upgraded application. So we've been using this application, this capability internally
and the results are stunning. So with Amazon Q transformation, a very small team of Amazon developers successfully upgraded
a thousand applications from Java 8 to Java 17 in just two days. Now I'm gonna say that again, 1000 application upgrades in two days. That's how long a single
application upgrade used to take. (audience applauding and cheering) I mean, this is months, months if not years of
developer time saved. A lot of very happy people
at Amazon over this, I promise you. So today, Amazon Q can work
its magic on Java upgrades, but soon you're gonna be able to use it to help migrate your .NET
workloads from Windows to Linux. And there are a lot of
applications out there stuck on Windows because
of the sheer effort required in making the migration. And this is an opportunity
for huge cost savings, on costly licensing fees, on
top of the inherent performance and the security benefits. Well, what about the other
people in your organization? All the folks in marketing, finance, HR, product management, and more. There's so much information spread across your organization's
documents, all of your data, all of your applications, and people in all of
these different roles. They struggle every day
to find the information that they need to then transform
it into making decisions and taking actions and
doing that fast enough to say competitive. Hmm, finding relevant information, making recommendations, taking action. That sounds familiar, right? Yep. Amazon Q is also your business expert. Q has connections to over 40
popular enterprise systems. So employees across the organization can ask complex questions
and get detailed, accurate, nuanced answers, that are
relevant to their role at work. And again, secure is secure,
and Q is secure and private. It respects your existing roles and the permissions that are
granted uniquely to each user. We're really excited about this, and I actually wanted to show you just a little bit of Amazon Q
for your business in action. So to put Q through its paces,
please welcome AWS's own Dr. Matt Wood to the stage. (upbeat dramatic music) - Thanks Adam, and good morning everyone. I'm excited to give you
a first look at Amazon Q, a new type of AI assistant, designed to be an expert in your business. With Amazon Q, you can
get quick, accurate, relevant answers to your most
pressing business questions all delivered securely and privately. Getting started with Q is simple
and takes just three steps. First, you configure Amazon Q, by connecting and
customizing Q with details of your own organization. Q connects to existing data
sources like S3, Salesforce, Microsoft, Google, Slack, and more, supporting over 40 popular
business applications and services right out of the box. Once connected Amazon Q starts indexing all of your data and content, learning everything there is
to know about your business. This includes understanding
the core concepts, product names, organization structure, all the details that make
your business unique. As well as indexing the
data from these sources, Q also uses generative AI to understand and capture the semantic information which makes your business unique. This additional semantic information is captured as vector embeddings, allowing Q to provide
highly relevant results, which are tailored to your
specific company and industry. Your data remains completely
under your control at all times. Amazon Q never shares it externally, or uses it to improve any
of the underlying models. And that's it. Surprise, there is no step three, just open Amazon Q in your
browser and away you go, with a fully customized secure assistant, who is an expert in your business. You communicate with Q through this friendly web application, designed for everyone in
your organization to use. Q knows who you are and
about your identity and role. And because Q understands
your business semantics, you can ask far more
detailed nuanced questions and get tailored answers
faster than ever before. For example, you can ask Q to analyze which product features customers are struggling with and recommend ways to improve them. As soon as you hit enter, Q gets to work on
understanding your question and preparing an answer. Q starts by creating a set of
input prompts automatically. Q then uses all of the
business context available to find relevant data,
information and documents and picks the best ones, before combining everything
together into a response. All in just a fraction of a second, using the power of generative AI. Amazon Q remains faithful
to your original sources, citing them in line, so that you can validate
the information easily. Q's experience might be magical, but there is no magical black
box just giving you answers. You can upload new Word documents, CSVs and other files on the fly to Q, to incorporate into its responses, letting you ask questions
about ad hoc data, which hasn't yet been added
to any corporate system. And Amazon Q was built with
privacy and security in mind from the ground up. So even if you were to seek information to which we don't have access, Amazon Q respects your
existing access controls, only returning information
you're authorized to see. Admins can also restrict sensitive topics, filtering out inappropriate
questions and answers where necessary. Finally, Amazon Q can take
actions on your behalf, through a set of configurable plugins. As an example, if you update, say, your training priorities, Q can automatically
create tickets in Jira, notify leaders in Slack, and update dashboards in ServiceNow. Q allows you to inspect
actions before they run, so you can review them and
add additional information. And after the action runs, Q will link to the
results for verification. So in summary, Amazon
Q brings AI assistant into the business world,
providing secure, relevant and actionable guidance tailored
entirely to your company. We can't wait for you to try Q and see how it will
help your organization. With that, back to Adam. Thanks a lot. (audience applauding) (dramatic upbeat music) - That was awesome. Amazon Q is gonna make a huge
difference for businesses. Employees are gonna love how Amazon Q is gonna help solve problems, discover and synthesize new information and lift folks out of monotonous
drag of repetitive tasks. And IT is gonna love how all of this comes with rigorous security and privacy. Now, I know at least
a few of you out there are wondering if Amazon
Q can access my data, does that mean it can help
with my business intelligence? And of course the answer
is a resounding yes. So we've been working for a while to make business
intelligence more accessible to people without BI expertise. We think anyone should
be able to ask questions of their data using natural language. And so Amazon QuickSight,
our ML powered BI service, now has Amazon Q features built in. First Amazon Q is gonna
help reduce the time that it takes business analysts to create dashboards and reports, from hours down to minutes, by letting them simply tell Amazon Q what they want to visualize. For example, regional sales by
product as a sankey diagram. Q comes right back with that diagram and you can add it to
your dashboard with ease. And you can tell Q to
further refine the visual. Say you wanna change the chart
to a vertical stack bar chart by month and color coded by region, your wish is Q's command. With Q, you can get
better dashboards faster. But even the best dashboards don't tell every angle of the story. For example, you can ask Amazon Q, where did we have the highest sales? And you can get a response
customized to your business data. You can ask a follow-up question, even as little as one word like "London." And because Q maintains the context of your current conversations, you can get insights
from your data quickly. Now, enabling people to
talk to Q in QuickSight is a big step forward in
helping your organization be even more data driven. But not everyone is
getting hands on with data. A lot of people consume
data through presentations and reports, and so so
much of you do at work is telling stories with data. And as your business expert Q is here to make that easier too. Let's say you wanna
create a monthly report on your business in North America. You can ask Q to write a monthly
report about the business and make recommendations for next month. And Q's gonna give you
visualization options. You choose the ones you want,
and then Q will format them. And in seconds you have a
beautiful and compelling story. And it's completely customizable. You can actually add more
context to this section by asking Q to add a
visual of the usage trend. And Q will generate a short summary, and then you can further
tweak it by making it longer, shorter, or reformatting it to bullets. Let's make it longer. Once we're happy with the story, we can securely share it with others. With Amazon Q and QuickSight, you now have an expert BI assistant, that's gonna make it easier
for business analysts, and it's make it easier for end users to get answers to end insights faster. And these new Q capabilities in QuickSight are available in preview today. (audience applauding) Now, a lot of organizations
are looking for solutions that are targeted to horizontal use cases, or specific lines of business. And AWS has applications
in these areas too. So our first specialized
line of business solution was Amazon Connect. So it's born in the cloud and of course Amazon Connect
is a scalable cost-effective, easy to use contact centers application that's reinvented customer service. Generative AI that knows your business is gonna have an incredible impact on the whole customer service experience, across so many different industries. Your customer service teams are at the heart of your businesses, delivering interactions
that people remember, for good and for bad. So getting this right really matters. Today, contact center
agents spend a lot of time gathering information from customers, to understand their questions, and then they spend even more time searching for the right answer. Now, connect has made
this a lot easier already with machine learning
transcription analytics, but we knew we could make it even better, for your agents and
certainly for the customers waiting on the other end
of the line for help. Enter Amazon Q and
Connect available today, and this is gonna give contact
center agents a major assist. Agents can chat with Q, directly inside of Amazon Connect to help them respond quickly
to customer questions. Live chat with Q with
fast answers is great, but what if you can make
the call itself the prompt? Now inside of Connect, Q is
actually there on the calls, assisting with proposed
responses with suggested actions, or with links to relevant
articles, no typing required. So contact centers,
supervisors and administrators are also gonna get a generative AI boost. Connect now automatically
creates post-call summaries, that supervisors can use
to track follow-up actions and to uncover agent
coaching opportunities. And for administrators,
Connect can set up chatbots and interactive voice responses, through simple natural language prompts, streamlining the process. Amazon Q is just the first of many specialized industry and
use case specific services that are gonna supercharge Amazon Q. So stay tuned. Amazon Q builder, your AWS expert, AWS Q business, your business expert, Amazon Q in QuickSight, your
business intelligent expert, and Amazon Q in Connect,
your contact center expert. All engineered to offer
assistance that remains faithful to the information that
you choose to feed it. And with with the built-in
security and privacy that you know and that
you count on from AWS. And this is just the start
of how we're gonna help to reinvent the future of work. We continue to innovate across
all layers of the stack, to bring you what you
need to harness the power of generative AI for your organization. Performant, cost-effective infrastructure, a new secure, private easy way to build and scale powerful new applications with foundational models. Generative AI applications
and capabilities that can be enriched with
the context of your business. And it's all on AWS,
built with our high bar, that gives you broad
and deep capabilities, choice and enterprise readiness. But it's so early in the game and we're incredibly excited about what we're gonna be able to do together. Now there's one more thing
that customers are gonna need for any of this to work. And that's data. Your data. No one else has it, no one else
can do exactly what you can. Your data's key to
setting your organization apart from the pack. And as we've discussed, generative AI has just made it so much
more relevant and impactful when you tailor your data. So let's hear from a longtime customer, about how they're using data to create first class
experiences for their customers. So please welcome to the
stage Stephan Durach, VP of Connected Company at BMW. (audience applauding) (keyboard keys clicking) (upbeat music) (upbeat dramatic music) (audience applauding) - Thank you. Hello everyone. It's fantastic to be here. I couldn't agree more, data is key. The automotive product complexity
is strongly increasing, and development speed is essential. Let me show you how BMW is mastering data, accelerating the development process, and delivering best in class
experience to our customers, anywhere and anytime. Across the entire product portfolio, from BMW to Mini and even Rolls Royce, BMW is providing game
changing vehicle entertainment and experience powered by AWS. We optimize our tech stack,
improve our workflow, and use data by leveraging
AWS cloud technology. Let me give you some examples. Cloud data hub, BMW Central Data Lake, built by BMW and AWS, hosts data from our
connected vehicle fleet. And while upholding the
strictest data privacy standards. On top of it, teams can build
their application and model, as well as deploy updates and improvement. Together we are exploring innovative use of machine learning,
including Amazon SageMaker. This automated driving platform is BMWs next generation of
driving assistant system on AWS. This automated driving platform is capable to handle several
hundred petabytes of data from our global test fleet. We are accelerating our
development, simulating and validation process dramatically, for our product launch in 2025. With our route optimized
charging services, we are providing real
time traffic calculation, based on driving style,
preference, traffic, or available change in
station and many more. We also use Amazon technology like Alexa, for our intelligent personal
assistant and Fire TV, to create unique customer
experience inside the vehicle. Our connected vehicle AI
platform is built on AWS, and forms the backbone of the largest connected vehicle fleet. Ready for some more
even eye-opening stats? Let's dive in. 20 years ago, BMW introduced
its first connected vehicle, and we come quite a long way since then. Now the BMW group has the
largest connected fleet, with more than 20 million
vehicles globally. Over the last three years, we migrated the connected
vehicle backend to AWS. By utilizing the global
presence of AWS in many regions. Worldwide, we are updating
more than 6 million vehicles over the air, on a regular basis. Those numbers are growing constantly. It requires to manage more than
12 billion requests per day, with over a thousand
different microservices. Ultimately, we are processing more than 110 terabytes of data traffic per day, with a remarkable reliability. Look at these numbers,
I'm always impressed. However, with our new
product architectures, this number will triple
in the next two years. Let's change perspective and take a look at a couple of numbers
in the development phase. More than 8,500 software
developers worldwide create first class code for
our vehicle and infrastructure. The BMW 7 series runs on over
500 million lines of codes. Our worldwide DevOps teams execute more than 110
software builds every day, by using more than 60,000 virtual CPUs. To keep up with our
demand, we use AWS services to scale up and accelerate development whenever it is necessary. Let's take a deeper
look at the development. Traditionally, BMW comes from
a hardware driven development. Every developer had to
do the development work on a physical test track. Today, we are using innovative technology to streamline our infotainment
development process. The new approach allows us to
write test functionalities, remotely to our global developer teams. An example is using Amazon
EC2 for virtual development and testing, in order to improve
cross-functional collaboration. Our next step is to extend this approach to build a fully virtual digital twin. Now let's talk about the product, and some exciting feature we
created for our Mini customers. On the picture, you can
see the mini interior. It's our way to emphasize
brand typical design. Take the impressive round OLED display in our new Mini family,
as a perfect example. Speech is rapidly becoming
the dominant modality for interacting with vehicles. We have a long history
integrating Alexa voice assistant. However, this year we are raising the bar. We will use Alexa technology to bring our intelligent
personal assistant to the next level. Now let's take a seat in
the back of our 7 series. One of our most exciting
feature is A BMW theater screen. It's the 30 inch screen
with an 8K resolution. It is spectacular, integrated in the car. It gives you a real cinematic experience. With Fire TV, our customers
can seamlessly enjoy their favorite series and movies. It's our strong collaboration
with the AWS and Amazon that unlocks new possibilities, enables us to think about
the car in the new dimension. So one last thing to give you
a glimpse into the future. This is the BMW Neue Klasse. This will launch in 2025. BMW has been always on the forefront of automotive innovation. The Neue Klasse will be a quantum leap in customer experience. It will take the driving
experience to a new level. As you can see, the Neue Klasse introduced a completely new interior. With the BMW panoramic vision, we developed a new head up
display to project information pillar to pillar, for the perfect driver
orientation end user experience. All of this emphasize, improve, BMW is mastering data, accelerating the development process and delivering best in class
experience to the customer, anywhere and anytime. This is absolutely exciting, and all of this is happening right now. AWS and the BMW group are
working closely together to shape this future. Thank you all. Goodbye. (speaking german) (audience applauding and cheering) - What an incredible data
platform you've built, and I mean the incredible
experience it's creating for your consumers. Just amazing. Thank you Stephan. - It was a pleasure.
- So great to see you, as always.
- Thank you. - Thank you. (upbeat dramatic music) So, as Stephan has just made so clear, success with AI, as well as with nearly every
other business priority relies on having a strong data strategy. And one question I hear from
customers frequently is, how should I think about data
in a generative AI world? Well, the answer is
that for generative AI, your data is your differentiator. Now, we've talked about
customization through Bedrock, CodeWhisperer, customizations, Amazon Q tailored for your business, and examples from Delta
Airlines and Pfizer, (indistinct) adding context to your generative AI applications with your own data is absolutely critical to providing value. Now, in each of these examples, the data used to customize the application has to be up to date. It had to be complete and accurate. It had to be discoverable, and it had to be available
when you needed it. So the first step in achieving all of this is to making sure your
data is in the cloud. Now, generative AI is
just one of many reasons that you want this. Having effective
cloud-based data strategy, is gonna increase your insight. It's gonna reduce your cost, and it's gonna accelerate innovation. To understand this, just think about how many different issues
there are to consider when trying to make the most of your data. Organizations have
different types of data, it comes from different origins, and it comes at various scales and speeds. Different teams all
across the organization are gonna work with different data in many, many different types of ways. And then how you want to use
this data is just as varied. You're storing it, you're
powering your applications, you're analyzing it, you're
generating predictions from it. There's no one tool or
solution that can do it all, unless you're willing to make
sacrifices like performance, cost, or capabilities. But you don't want to
make those sacrifices. So what you need is a
comprehensive set of tools, that accounts for the scale
and the variety of this data and the many, many purposes
for which you want to use it. You need to integrate it
and combine all that data that's spread all over to
create a better understanding of what's happening and
predict what will happen. And of course, you need governance, so that you can have the right balance of access and control
every step of the way. You need it all, if you miss even one aspect,
your ability to use that data is significantly hobbled. So in order to make sure
that you can store, analyze, and process the data, you've gotta have a comprehensive
set of data services. So you can get the price performance, you can get the speed, the flexibility, and the
capabilities for your use case. And that's why we have
built the most comprehensive set of cloud data services out there. To start we have Amazon S3, and customers use S3 to make hundreds of thousands of data lakes, where they can bring structured data together with unstructured data, and make it all available for analytics and machine learning. We also offer eight relational databases, including Amazon Aurora, the
relational database service that's built from the
ground up for the cloud and is still the fastest
growing service in AWS history. We also have eight
purpose-built databases, non-relational databases, including Amazon DynamoDB, which is used by over
a million developers, I'm sorry, a million customers, for its serverless, single
digit millisecond performance at any scale. And we also have the
broadest and deepest set of analytics services. And this includes our
data warehouse service, Amazon Redshift, which delivers up to six, or even seven times
better price performance, depending on your workload, than other cloud data warehouses. And we've got Amazon
EMR, to easily run Spark and other big data workloads. And Amazon OpenSearch Service, with tens of thousands
of active customers, and hundreds of thousands of
clusters under management. And that all processes,
trillions of requests per month. And many, many, many
more analytic services, I won't go through them all. But you need this breadth
and this depth of choice, so that you can never have
to compromise on cost, on performance, or on scale. And instead you can choose
the right tool for the job. Now, we've already talked
about AI and ML extensively, so I'm not gonna dwell on that here. But you need to be able
to choose the right tools to feed your AI and everything else that you're doing in your business. And the second thing that you need, as you move to put data at
the center of your business, is to break down your data silos. Most people have data spread
out all over the organization, and databases, analytic
tools, SaaS applications, spreadsheets and more. And to get the most value from your data, you need to be able to use it
all, no matter where it lives. And to do this today, most
people move data around, or try to connect the
various silos using a process called ETL, or extract,
transform and load. Yeah, I see some of you
flinched when I said that. 'Cause you know, right? You know exactly how bad
this process could be. It is terrible. First, developers have to design an ETL pipeline architecture. They have to decide where
to extract the data from, and it's often coming
from multiple sources. Then they have to write
code to transform the data and remove duplicates. They have to filter out
outliers, retrieve missing data, and identify corrupted data. After all that, you have to load your transformed data to its new destination, which once again means more code to write. Then if something changes, like a data engineer
swapping out a data table, or adding a new field,
you have to go back again and update all of your code once more. It is monotonous, it is tedious, and it is just a whole lot of unfun work. Which is why we introduced our
vision of a zero ETL future. And in this future you can
wave goodbye to the boring, frustrating process of ETL. Your data's automatically connected from the source to the destination. It's that easy. And you get your data ,where
you need it, when you need it. No more building and managing complex, brittle processes and pipelines. It's really exciting. But is it real? Well as a down payment, last year we announced a fully
managed zero EL integration between Aurora, MySQL and Redshift. Within seconds of data
being written into Aurora, you can use Redshift to do
near real time analytics on petabytes of data. You can enable this integration
with just a few clicks and zero ETL. It is incredibly easy. Now this first zero ETL integration is already done. But we want to deliver so much more. So I'm really excited today to be announcing three
more zero ETL integrations with Redshift. Let's get 'em all done at once. Aurora PostgreSQL, RDS
for MySQL and DynamoDB. These zero ETL integrations and Amazon Redshift are all
available on preview when? Today. (audience applauding) So that makes... Yeah, you don't like a muck, do you? So that's four of our
most popular relational and non-relational databases, together with the data warehouse that gives you the best price performance, without any of the ETL. Fantastic. But remember, we're not
done with this vision. So of course our next question was, well, where else could we
be removing the pain of ETL? Now, many of you need to use search on your transactional data, to enhance the experience
of your application. So for example, if you have
an e-commerce application and you wanna provide
personalized recommendations to your user, you'd use similarity search on transaction data that's gonna
be stored in your database. So to do this, currently you're
designing, you're building, you're managing complex
data pipeline architectures, from your database to a search engine. And tens of thousands of you
are building those pipelines to Amazon OpenSearch Service for its powerful search capabilities. I bet you can tell where
we're going with this. So today I'm really pleased to have unveiled one more
zero ETL integration, DynamoDB zero ETL integration
with OpenSearch Service, generally available today. (audience applauding) And this new integration makes it so easy to search DynamoDB data
with our OpenSearch Service. And it's just the first of
these database integrations to our OpenSearch Service that's gonna help you do more with data. So our new zero ETL integrations are bringing transactional data and analytics together
faster and more easily. And customers are really
excited about the idea of a zero ETL future. So you can be certain we're
gonna keep innovating here. Now, we've talked about all
of the tools that you need and how you can have
data where you need it without the painful ETL burden. Next, you wanna empower
people across the organization to use your data to its fullest. You want them wallowing in
your data, experimenting, asking questions, sparking ideas. You wanna give people access. But there's one problem, your data is incredibly valuable, so you want to grant access only to the right employees
at the right time. You need to keep it safe and secure. So you also need control. Access, control. These are opposites, right? No, you can have, in
fact, you must have both. And the key to making sure
that your data is safe, is governed, has access, is governance. Creating visibility into your data and controlling who can access your data and when they can access it. With robust governance, you're
gonna give people access to the data they need and only that data. And when you know what data
you have, you know where it is, you have guardrails and tools
in place to safely unlock it, that builds confidence. And then people within the organization develop trust in the data, you actually empower
innovation and experimentation, rather than stifling it. And so to help you with governance, we developed Amazon Data Zone. Data Zone is a data management service that helps you catalog, discover, share, and govern data across your organization. And we're getting great feedback from lots of different customers, including the financial
services leader in Brazil, Itau. And we're really pleased with the impact that Data Zone's having. And we're gonna continue to
innovate rapidly in this space. Today, Data Zone uses machine learning to automatically add
metadata to your catalog. But in order to find the
right data at the right time, customers still have to choose
to add business context, like explaining the data's purpose, or clarifying column descriptions. Now this helps employees
find the data that they need. It's really important because if you don't know the data exists, the question of access
and control becomes moot. But adding all this metadata is tedious. It's a manual process, well no longer, we're making your cataloging even smarter. And today I'm really pleased
to announce a new capability that automatically suggests
business descriptions. To add context for your data
set in just a few clicks. (audience applauding) Customers simply choose a data set and Data Zone will use generative AI to create business
descriptions for your data, using our train models, making
it easier to understand. These descriptions can
be reviewed and accepted, or even further clarified or edited for your exact preferences. Data Zone's also gonna
provide recommendations for how to use your data. This is a huge, huge time saver, and this capability is
gonna help employees to find the right data
for even faster insights. We're gonna continue to innovate, to help you use data governance
to set your data free. Whatever you wanna do in your business, you need an end-to-end data strategy that provides capabilities across the entire span of your data needs. You need to easily
connect data across silos. And you need your data to be governed, to be confident that
the right data is used in the right ways, by
only the right people. We're making AWS the best place for all of your data needs across all of these different areas. And when you have all this and when you're ready to make
data your differentiator, in AI and across your
business, it's amazing. We've talked a lot
about reinvention today. And here at AWS, the most important
reinvention of all is yours. To support you in all the ways that you wanna adapt and evolve, we continuously challenge ourselves. We try and think differently. We try and give you the capabilities that you've always wanted and even some maybe that
you've never imagined. Now, this isn't the work of a few months, or even a year, but we're okay with that. We're okay with that because
at Amazon we make bold bets. So third party selling on Amazon, Prime, even AWS itself was a big bet. I mean, it took a lot of years and a lot of capital to build out the on-demand infrastructure services, that simply became known as The Cloud. This long-term focus is a good thing. Being willing to think
differently and long-term is a good thing. It's the only predictable way to profoundly changed what's
possible for customers. Now there's another big
bet that Amazon's making. You might've heard of Project Kuiper. (audience cheering) Yeah, Kuiper. (audience applauding) Kuiper is building a constellation of thousands of low
earth orbit satellites. Why? Well, hundreds of millions
of people on earth lack fast or reliable, or
even any internet access. Project Kuiper's gonna help
close that digital divide, by delivering fast, affordable
broadband to a wide variety of customers operating places without reliable internet connections. In fact, just last month, the Kuiper team successfully
launched its first two proto satellites into orbit. It's just an incredible thing. I mean, they put satellites up there, they're working flawlessly. Everybody's really happy. Let's take a look. (upbeat dramatic music) (audience applauding and cheering) Maybe I should have stopped there, right? Wow. I mean, Kuiper is a global
high speed satellite network that's gonna provide coverages in places that have never had coverage before. I mean, imagine the power of delivering, in really hard to reach places, the ability to access the
internet with the same speed, the same reliability that most of us take for granted every day. Now, the possibilities for
consumers is for sure enormous, but so are the benefits to
businesses and governments. And there are all sorts of use cases, like renewable energy providers, that want to analyze in real time data from offshore wind farms. Or first responders who need
fast, reliable connections during natural disasters. Or organizations that are looking for connectivity backup in
case of an unplanned outage. Kuiper's gonna help to
make all of this possible. And AWS customers are already
planning for the opportunities that this opens up to
connect to their facilities, their sensors, their data, their networks. It's going to be amazing
for so many customers. But some customers of course, don't want their data running
over the public internet. And as you know, security is
always our number one priority. So today I'm really excited to announce, that in addition to public
internet connectivity, Kuiper's also gonna
provide enterprise ready, private connectivity services. (audience applauding) And now you're gonna be able to move data from virtually anywhere, over
private secure connections. And as you use these
connections with your data, you can use it inside
of the AWS cloud too. And early adopters are gonna
be able to start testing in the second half of 2024. So we make big bets like these, so that customers can do things that they couldn't imagine before. And you come on this journey with us because we are willing to take these kinds of audacious leaps. And the benefits have been enormous. The exciting challenge ahead
of all of us this next year is to look closely at our
work and to reimagine it, to reinvent it. To dream up experiences that
were previously impossible and make them real. We are so excited to be your
partner in this new world that we're building together. And I'm thrilled that we get
to explore it all this week. This morning has been incredible. I want to thank all of
my guests on stage today. And thank you to everybody watching. And of course, thank you
to the whole AWS team for working so hard to make this happen. (audience applauding) Now the team does have
one more little surprise up their collective sleeve. So I dunno if you remember
at the very beginning I mentioned Lowden guitars in Ireland, well we actually have
a Lowden guitar here, from our friends at Lowden, and we've invited a talented guitarist, a longtime AWS favorite,
Zach Person, to play us out. So we have got a great conference
here for you this week. Now let's get out there and reinvent. Thank you. (audience applauding) Take it away, Zach. (guitar strumming)
