Hi, it's Andy Jassy. I'm back here at re:Invent today on Wednesday. Yesterday, I shared some areas that I thought were really exciting with respect to launches. We talked about Trainium2 chips and Trainium2 instances. We talked about our Aurora DSQL database and we also talked about our new Amazon Nova Frontier models. Today I'm gonna talk about three areas of launches that particularly interest me and that I think are gonna be interesting for you.
I'm gonna start with storage, and I think a lot of times it's easy to forget how core and integral storage is, but we're continuing to innovate in a really important way for you around storage. And let me give you an example. In S3 today, which is our object store, we have over a million data lakes. And increasingly, as companies are expanding the amount of data they have to do analytics and AI, you're seeing companies store objects in the Parquet format. It's just easier query. And then over time, what you're seeing is that companies are using Iceberg to extend the functionality and capabilities of these Parquet files. So it allows you to inserts and updates and schema changes and things of that sort. It also integrates really well. Iceberg integrates really well with query engines, analytics engines. However, if you're managing these Iceberg tables, it's a lot of work. You actually have to do table management, you have to do data compaction, you have to clean up the files, you have to do access control, and it just takes a lot of time. So what the S3 team launched is they launched a new bucket type in S3 called S3 tables. And what it effectively does is it makes Iceberg a first class primitive in S3. So it does all the table management for you, all the cleaning up of files, all the data compaction, all the access control, much, much easier for you. And it actually makes your queries, your table queries and Iceberg three times faster and 10 times more transactions per second. So it's much better performance versus managing those Iceberg tables than your own. So I think this is gonna be very, very helpful for our customers in storage and in analytics.
Then the second area I was gonna talk about was my excitement about what we're doing in Bedrock right now. So for those of you don't know, Bedrock is our service that has the broadest selection of leading third-party frontier of large language models and then has the best collection of features to make it easiest for you to build great generative AI applications. These are things like model evaluation and guardrails and integrating with RAG and taking agentic actions. And so Bedrock has a lot of traction and especially at this early stage of where people are building generative AI applications, where companies are still experimenting with lots of different model types and often multiple model types in the same application and multiple model sizes. Bedrock being able to seamlessly allow you to move back and forth between models and sizes is very compelling. And I think that's part of why it has lot traction. So the Bedrock team has announced a number of things over the last couple days. First, we've had, for a long time, anthropic models and LAMA models and stability models and Mistral models and Cohere and AI21. We added today Poolside and Luma models into Bedrock, both very popular, very performant models, as well as a new Bedrock Marketplace for models where we've added over a hundred new models into the Bedrock Marketplace. So you'll have, you already have the leading selection of models and leading models, now you have even broader selection of those models. Second thing they added was a feature called Distillation. And so this feature allows you to take the intelligence of a larger model and to infuse that into smaller models. And that resulting smaller model allows you to have lower latency or lower cost. Very useful. Then the third thing that they added that I think is exciting is they added something called prompt caching. And so if you think about it, let's say you're a law firm and you have lawyers who are asking the same questions, by and large, over and over again, whenever somebody asks a question in a prompt, the prompt then takes certain actions, it generates different tokens and that consumes a bunch of compute and time. And so what prompt caching allows you to do is it allows you to cache those questions and those prompts so that the next time they get asked, all that processing is done already and ready for you to use again, saves you a lot of time and money. And then they also added something called Intelligent Prompt Routeing. And so this is a great feature for if you're gonna actually ask questions and you don't really know some of your questions you ask a model, need the most sophisticated, largest version of that model. But the problem is when you have to go to that most sophisticated model, the latency is higher and the cost is higher. So whenever you can have a smaller version of that model be able to answer it, you get better latency and cost. So this feature in Bedrock, this Intelligent Prompt Routing, you set what your thresholds are around price and around accuracy, and then we will automatically route to the right model to optimize for you on that combination. And then the last feature I'll mention that Bedrock added is something called Automated Reasoning. And so we've been using Automated Reasoning for a long time inside AWS and it really uses effectively logic and math to be able to verify that an answer is right. So we do this a lot of times for identity access control. And so Bedrock now allows you to list a bunch of policies around your data and your own knowledge bases. And then Automated Reasoning makes sure that when we're giving you the inference and the answer that it actually is correct and matches what's accurate and in your policy on your knowledge base. That's a pretty significant set of new features for Bedrock, which is already the most performant inference machine. Now, I think what's gonna happen moving forward is that just like in every application you have compute and you have storage, and you almost always have database and usually use analytics. And now increasingly AI, what you're gonna see is that every application is gonna have generative AI or inference in it. And what you're gonna want is you're gonna want inference service that not only has the broadest array of models you can use, but the broadest array of features that allow you to be accurate, to be fast, and to be cost-effective and to take work away from you that you would've otherwise had to do. And if you look at the rolling thunder momentum of what's being added to Bedrock almost every month, it is very compelling and that's why I think people are using it so much for inference.
Third area I'm gonna mention is around Q. So Amazon Q is our AI-powered assistant that is the most capable assistant for software development and for leveraging your own data. And so one of the capabilities in Q that I think people are finding very compelling is what we call Q transformation. And so if you're a developer, you know that one of the most tedious things that you have to do is you often have to move from one version of something to another. Oftentimes for security, sometimes for cost. And nobody really likes to do that work. It's very tedious. And so an example inside Amazon is we had to move from one version of Java to the next version of the JDK for security reasons and to be updated. And we had tens of thousands of those packages that normally would take us many, many months to do. And instead, we use Q to do the transformation automatically for all those upgrades. And we did tens of thousands of those upgrades and saved 4,500 developer years and over $260 million using Q transformation. That's very compelling, that's very practical use of AI. And our customers were very excited about this idea and they said, 'Could you please build transformations of other things that we know we have to transform but we don't wanna spend the time doing?' And so we just launched three of them. The first is we have a lot of companies that are interested in moving from Windows to Linux for cost reasons, for security reasons, for flexibility reasons. And so we've built a Q transformation that makes it easy to move from .net to Linux. The second one that we just announced is for people that are trying to move from VMware into native EC2 or the cloud. And this is something that companies have been steadily marching toward for many years now. And then with the change in ownership of VMware where the cost of VMware has gone off by so much, it's been a significant catalyst for people to move more quickly into the cloud. And so we built a Q transformation that lets you move from VMware to EC2 instances. That's very compelling. And then the third Q transformation that we just launched, I think handles may be the very most difficult transformation that companies need to make, which is trying to move from a old mainframe into a cloud environment. And with mainframes, you know, people don't even know where the data is or how it's laid out anymore. The people that created that and know a lot about that mainframe often, you know, don't work at the company anymore. And that process of moving from a mainframe, which we've helped a lot of customers do historically takes several years. And so we built a Q transformation that basically dissects and pulls apart that mainframe into the different pieces and the documentation surrounding it. So it's much easier to be able to move from a mainframe and now you'll be able to move from a mainframe to the cloud instead of it taking 3, 4, 5 years, you can count it in inside of a year in terms of months. So those are three very powerful transformations. They're gonna help enterprises move from legacy implementations into more modernized implementations and into the cloud more easily. By the way, if you wanna do generative AI, try doing generative AI in conjunction with a mainframe, it's nearly impossible to do so. This will help enterprises take advantage of this new technology. So those are some areas I'm excited about, and I'll see you tomorrow.
