#749: re:Invent 2025 - Swami Sivasubramanian Keynote

0:00This is episode 749 of the AWS podcast, released on December 4th, 2025. five.

0:09Hello, everyone, and welcome back to the Adibus podcast. I'm Les here with you. Great to have you back. And of course, joined by Julian Ford. G'day, Julian. Day two, day two. Well, it's never day two. It's always day one. Oh, well, you all know the reference. We know what we mean. We know what we mean. Day one. Reinvent 2025, day two. Here's what's going on. There's been lots happening. And in this episode, we're going to cover two things. We're going to cover some of the announcements from the keynote. that Swami did, but also we'll be going through the pre-invent stuff. So all the things that happened before re-invent that we didn't get to cover, there's about 60 of them.

0:47So it's going to be a fair bit, but let's get into what's the latest and greatest. So we're happy to introduce AWS AI factories. These provide rapidly deployable, high-performance AWS AI infrastructure in your own data centers. So by combining the latest Tranium Accelerators and NVIDIA GPUs, specialized low latency networking, high performance storage and AWS AI services, this accelerates your AI build outs by months or years compared to building independently. So you get immediate access to leading foundation models without having to negotiate separate contracts with individual model providers.

1:24You get to use Bedrock, you get to use SageMaker and these AI factories operate as dedicated environments it's built exclusively for you or your dedicated trusted community, which means you have complete separation and operating independence while integrating the broader set of services. So this is really useful for a variety of use cases across industries or just places where you just couldn't get the cloud in there. Now you get a factory in there. Now, speaking of hardware, we're announcing Amazon EC2 Tranium 3 Ultra Servers for faster, lower cost generative AI training. This is our fourth generation chip called the Tranium 3 and it's our first three nanometer AWS AI chip that's purpose-built to deliver the best token economics for all the work that we're doing these days.

2:10So each AWS Tranium 3 chip provides 2.52 petaflops of FP8 compute. It increases the memory capacity by one and a half times and the bandwidth by 1.7 times over Tranium 2 and it gets to 144 gig of HBM3e memory and 4.9 terabytes per second of memory bandwidth. So Tranium 3 is designed for both dense and expert parallel workloads with advanced data types. So you've got things like MXFP8 and MXFP4 and improved memory to compute balance for real-time multimodal and reasoning tasks. Now the Tranium 3 Ultra servers can scale up to 133 Tranium 3 chips so that's 362 FP8 petaflops in total and you can get them in ultra clusters that scale to hundreds of thousands of chips.

3:01So a fully configured Tranium 3 Ultra server will give you up to 20.7 terabytes of HBM3E and 706 terabytes per second of aggregate memory bandwidth. Now this next generation Trainier 3 Ultra server also has the Neuron Switch V1 which is an all-to-all fabric that doubles interchip connectivity bandwidth over the previous generation. So lots of big stuff basically 4.4 times higher performance, 3.9 times higher memory bandwidth and four times better of performance per watt compared to the Trinium 2. So lots to look at there, Jillian. Speaking of more things to look at, this one's like a two-for-one kind of.

3:45So we've announced within Amazon Bedrock and within Amazon SageMaker AI, capabilities that make advanced model customization accessible to developers at any organization. Reinforcement fine-tuning in Amazon Bedrock and serverless model customization in Amazon SageMaker AI with reinforcement learning. This is going to make it really simple to be able to create these efficient models that you can be able to fine-tune. It's going to be a lot faster, more cost-effective, and more accurate compared to just using base models. So these techniques, now that they're available in Bedrock and SageMaker just makes it a lot faster to be able to get started.

4:35I'm personally excited about this episode because I got to interview one of the folks that has been working behind the scenes on this, and we've got an episode coming out on that. Yeah, it'll be on fine-tuning. Should you fine-tune? What are the different techniques? And of course, this is going going to be mentioned as well. So stay tuned for that. And we've also announced two new AI model training features within Amazon SageMaker HyperPod. Checkpointless training, which is an approach that mitigates the need for traditional checkpoint-based recovery by enabling peer-to-peer state recovery and elastic training.

5:16And another one is elastic training. Elastic training maximizes cluster utilization as training workloads automatically expand to use idle capacity as it becomes available. This is going to save you a lot of time if you're been used to spending a lot of time having a stop, stop, restart, loading, configure. It takes like an hour or two depending on the size of the cluster. So this is now minutes, which is cool. Really cool. something else that's cool is a new model amazon nova 2 omni which is an all-in-one model for multimodal reasoning and image generation now this is the industry's first reasoning model that supports text images video and speech inputs while generating both text and image outputs so you get multimodal understanding image generation editing using natural language and speech transcription so you don't have to stitch lots of things together you can just get up and running The model supports a 1 million token context window, 200 plus languages for text processing and 10 languages for speech input as well.

6:22So lots of cool capabilities there. And speaking of cool capabilities, something that Jillian and I have been discussing together for a while now, because we knew we had to get this done for all of our amazing listeners is all the announcements that happened just before reInvent. And it's kind of a weird time because it's Thanksgiving, people are away, stuff's going on, lots of lead up to reinvent sort of focus, but lots of cool stuff came out. So we're going to take you through it. So let's start with the AWS Marketplace. So the AWS Marketplace now has agent mode and an AI enhanced search to let you discover solutions better.

6:57There are 30 ,000 listings more than, so having an agent to help you figure out how to do things properly and quickly will make your life a lot easier. There's also an AWS Marketplace MCP server as well. So you can sit in your interface of choice and get access to what you want. AWS Marketplace has also introduced express private offers for fast, personalized pricing. So rather than having to talk to a sales team, et cetera, you can get it done automatically using, guess what, AI. And there is now a new automated integration for CrowdStrike Falcon DexGen SIEM in AWS Marketplace as well. So this unifies your threat detection, et cetera, and it correlates lots of stuff.

7:37and it means you don't have to do any manual setup when you're creating it as well. We're also now happy to introduce multi-product solutions in the marketplace. So this is a combination of products and services from one or more Adabuse partners tailored to meet a specific use case or industry component. Each component has its distinct pricing and terms but it means you can get it all in one sort of one up solution. And the Adabuse marketplace now also supports variable payments for professional services. So this is a new billing option that lets professional services sellers bill customers as work is delivered.

8:12Next up, we've got analytics. AWS announces the Apache Spark Upgrade Agent. This is a new capability that accelerates Apache Spark version upgrades for Amazon EMR on EC2 and EMR serverless. The agent converts complex upgrade processes that typically take months into projects spanning weeks through automated code analysis and transformation. 80Biz Glue now supports materialized views, a new capability that makes it easier for data teams to transform data and accelerate query performance. Amazon EMR Serverless now offers serverless storage that eliminates local storage provisioning for Apache Spark workloads.

8:55This is going to reduce data processing costs by up to 20 % and prevent job failures from disk capacity constraints. Amazon Kinesis Video Streams now supports a new cost-effective warm storage tier. The warm storage tier enables developers of home security and enterprise video monitoring solutions to cost-effectively stream data from devices, cameras, and mobile phones, while maintaining extended retention periods for video analytics and regulatory compliance. Now, let's move on to artificial intelligence. Yes, more updates in that because we're deep in that wonderful virtuous cycle where customers get their hands on new services and capabilities, and they say, if only it did X, Y, and Z, and the team works towards it.

9:45That's why there's lots of updates. and spoiler alert, wait till we get to the connect section. Oh my goodness, that seems to be busy. But let's talk artificial intelligence first. Amazon Bedrock Agent Core Runtime now supports bi-directional streaming. So now you can have real-time conversations where agents listen and respond simultaneously while handling interruptions and context-changing mid-conversation. So this gets through that friction of the sort of, you know, like talking on a handheld radio. It's much more natural. Amazon SageMaker Catalog now exports asset metadata as a queryable data set, so you can get it through an Apache Iceberg table through Amazon S3 tables.

10:22And Amazon SageMaker HyperPod now supports programmatic node reboot and replacement, so you can do it through an API. And we're also announcing a preview of the AWS MCP server. So this helps AI agents and AI native IDEs perform real-world multi-step tasks across one or more AWS services. And so this is consolidating the capabilities of the existing AWS API MCP and the knowledge servers into one unified server. So you can get access to documentation, generate calls to over 15 ,000 APIs, and also use pre-built workflows called agent standard operating procedures or SOPs that can guide you through normal tasks.

11:02We're also happy to announce the AWS AI League 2026 Championship. So this expands the flagship AI tournament with new challenges and doubles the prize pool to$50 ,000 for builders to compete and innovate. So this allows you to have a quick orientation and then focuses on tournaments with two challenge tracks. The model customization challenge using SageMaker AI and the Argentic AI challenge using Amazon Bedrock Core to build intelligent agents. So you can participate at AWS summits and I think a lot of folks will have a lot of fun with this. Another quick update, multimodal retrieval for bedrock knowledge bases is now generally available.

11:44Previously, you could only search through text documents and images. Now you can get all different formats through one interface. So terabytes of meeting recordings, training videos, visual documentation, a whole bunch of stuff. This can handle it all now. SageMaker HyperPod now supports managed tiered KVKH and intelligent routing for large LLM inferences. So this means you can optimize your inference performance for long context prompts and multi-turn conversations. Managed tiered KV cache addresses the challenge by intelligently caching and reusing computed values, whilst intelligent routing directs requests to optimal instances.

12:20With this, we're getting up to 40 % latency reduction, 25 % throughput improvement, and 25 % cost savings compared to baseline configurations. Amazon SageMaker Catalog now provides automatic data classification using AI agents. And Amazon SageMaker HyperPod now supports custom Kubernetes labels and taints. We're also announcing a major expansion of the AI competency, formerly the generative AI competency, in the largest specialization launch to date, which includes 60 validated partners across three new agentic AI categories. Agentic AI tools, agentic AI applications, and agentic AI consulting services.

13:03This lets you identify and work with the AWS partners who specialize in developing and implementing autonomous AI systems that can perceive, reason, and act with minimal human oversight. Now onto business applications, which I think at this point, we should just call it the Connect show. Amazon Connect now allows you to test and simulate contact center experiences in just a few clicks, making it easy to validate workflows, self-service voice interactions, and their outcomes. Amazon Connect now gives business users greater control over daily contact center operations without requiring technical resources.

13:38With new capabilities to create customer UIs that adjust cues, routing behavior, and customer experience settings in real time, business users can respond to changing conditions immediately while maintaining enterprise-grade governance and security. Amazon Connect launches real-time AI agent assistance and contact summarization for Salesforce Contact Center with Amazon Connect. Amazon Connect now allows you to bring your own Amazon Bedrock knowledge bases and supports multiple knowledge bases per AI agent, giving you greater flexibility in how you organize and access knowledge content for your AI agents.

14:17Amazon Connect now supports third-party speech providers for end-customer self-service, giving you greater flexibility in how you deliver voice experiences. Amazon Connect is introducing agentic self-service capabilities that enable AI agents to understand, reason, and take actions across voice and messaging channels to automate routine and complex customer service tasks. Amazon Connect launches MCP support, enabling AI agents for end customer self-service and employee assistance to use standardized tools for retrieving information and completing actions. You can now use flow modules as MCP tools to reuse the same business logic across both deterministic and generative AI workflows.

15:01And of course, you can integrate this with Amazon Bedrock Agent Core Gateway for all of the fun MCP-ification of things. Amazon Connect now provides analytics and monitoring capabilities for AI agents across self-service and agent assistance experiences. Amazon Connect now provides improved analytics and monitoring for AI agents. Amazon Connect now supports creation of custom metrics for use in dashboards and APIs. Amazon Connect now provides businesses with the ability to automatically evaluate the quality of self-service interactions and get aggregated insights to improve customer experience.

15:41Amazon Connect now allows you to automate email responses and agent routing logic using keyword and phrase conditions, helping organizations increase self-service, reduce manual handle time, and improve routing accuracy. Amazon Connect is launching an AI-powered predictive insights that transforms how businesses understand and serve their customers. So an example is maybe you'll have like a recommendation algorithm, such as recommended for you or similar items. So this is currently in preview, but with this Amazon Connect customer profiles, you'll be able to get these different types of individual user recommendations.

16:24Amazon Connect now streams messages for AI-powered interactions. This new capability shows connect AI agent responses as they're being generated, which reduces perceived wait times and improves the customer experience. Amazon Connect now makes it easier to link related contacts such as email replies, call transfers, persistent chats, and queued callbacks to the same case so agents can view the complete customer journey and resolve issues faster. Amazon Connect now provides AI-powered case summaries. Amazon Connect now provides granular access controls for performance evaluations. Amazon Connect provides managers with new criteria while setting up automated evaluations, making it easier to identify relevant contacts for evaluation and providing additional insights to automatically populate evaluation forms.

17:18Amazon Connect Chat now supports in-flight data redaction and message processing. Amazon Connect Chat now supports agent-initiated workflows. Amazon Connect Outbound Campaigns now supports multi-step, multi-channel customer engagement journey builder. Amazon Connect now allows you to customize and visualize the appearance of the agent workspace. Amazon Connect enhances its agent assistance capabilities. These AI agents analyze conversation context and customer sentiment in real time, actively completing tasks such as preparing documentation and handling routine processes. That might have been the most Amazon Connect updates that we have done.

18:01That's crazy. That was amazing. They're really, really busy. They are. Now let's talk about compute. We have new instances. we have the new compute optimized amazon ec2 c8a instances these are powered by fifth generation epic processes these are the turin processes and they deliver up to 30 percent high performance and 19 percent better price performance than c7a instances as an example they're up to 57 faster for groovy jvm allowing better response times for java-based applications so that's kind of an example where you can just do an instance update and get half off your performance which is pretty amazing.

18:42We're also announcing the Amazon EC2 general purpose M8AZN instances in preview. These are also fifth generation EPIC processors. These have twice the compute performance than the previous generation M5ZN instances as well. We're also announcing Amazon EC2 M4 Max Mac instances in preview. So this is the latest Mac Studio hardware if you're building for Apple, you can get access to those. And we also have got the new EC2 P6E GB300 Ultra servers, which are accelerated by NVIDIA GB300 NVL72. These are now generally available. These are honking great powerful servers that you can get access to straight away.

19:27We're also announcing the Amazon EC2 memory optimized X8i instances. These are powered by custom Intel Xeon 6 processors and these give you one and a half times more capacity up to six terabytes and 3.4 times more memory bandwidth than the previous generation X2i instances as well. These will be SAP certified and give you 46 % higher SAPs compared to the X2i instances. So if you're looking for lots of SAPs, that's the place to get it from. And Adibus previews the EC2C8INE instances. These are the sixth generation Intel scalable processors. These are the Granite Rapids ones and the latest Adibus Nitro version 6 cards.

20:06So you get up to two and a half times higher packet performance per vCPU than the previous generation C6IN instances. You get twice the higher network bandwidth through internet gateways and up to three times more elastic network interface compared to the existing C6IN. So if you've got packet processing workloads, telco type stuff, virtual security appliances, DDoS protection, they're the ones you want to have a look at in preview.

20:38One update in databases. Amazon Aurora now supports BoseGross 17.6, 16.10, 15.14, 14.19, and 13.22. Next up, management and governance. Amazon CloudWatch now offers configuring deletion protection on your CloudWatch log groups, helping customers safeguard their critical logging data from accidental or unintended deletion. Very important. Amazon CloudWatch now enables automated quality assessment of AI agents through agent core evaluations. This new capability helps developers continuously monitor and improve agent performance based on real-world interactions. Amazon CloudWatch launched incident report generation capabilities with an AI-powered root cause workflow that guides customers through the five whys analysis technique.

21:35The feature is modeled on the correction of errors process used by both teams within Amazon and our customers to improve their operations. The FiveWise process is an excellent dive deep process. It really helps you understand the root cause of problems by really helping you ask the right questions in the right area. So definitely a big plus one on using that. Now let's talk networking. We have probably one of the biggest announcements we've had for a long time in networking. and we've had some big ones, we are announcing a preview of AWS Interconnect Multi-Cloud. This provides simple, resilient, high-speed private connections to other cloud service providers.

22:15And it's starting in preview with Google Cloud as the first launch partner, and then with Microsoft Azure later in 2026. So customers have been adopting multi-cloud strategies while migrating more applications to the cloud. And they do have many reasons for interoperability requirements. They may want choice. They may want speed. They may want ease. They may be integrating previously existing systems, et cetera. AWS Interconnect Multicloud is the first purpose-built product of its kind and a new way of how clouds connect and talk to each other. It enables customers to quickly establish a private, secure, high-speed network connections with dedicated bandwidth and built-in resiliency between your Amazon VPCs and other cloud environments.

22:59So you can get up and running really quickly instead of weeks or months. Now, Interconnect Multicloud is available in preview in five AWS regions, and you can enable this capability in the console. And CSPs can also adopt via a published OpenAPI package on GitHub. So lots of information on this one. Another version of Interconnect is now a gated preview of AWS Interconnect Last Mile. This is a fully managed connectivity offering that allows customers to connect their branch offices, data centers, and remote locations to AWS with just a few clicks. So again, no friction of discovering partners and complexity, et cetera.

23:37This is a collaboration between AWS and Lumen. And this really takes advantage of Lumen's extensive network footprint. So it's currently a gated preview. If you're interested in this, you need to reach out to your account team. And this is for our customers in the US. And one more networking update is the Amazon API Gateway now supports MCP proxy, which allows you to transform your existing REST APIs into MCP compatible endpoints. Now that's interesting. So this new capability enables organizations to make their APIs accessible to AI agents and MCP clients. And through integration with Amazon Bedrock's agent core gateway service, you can securely convert your REST APIs into agent compatible tools whilst enabling intelligent tool discovery through semantic search.

24:22So this gives you three key things. It lets REST APIs communicate with AI agents and MCP clients through a protocol translation, so you don't have to modify your application. Secondly, it gives you comprehensive security because you've got dual authentication. You're verifying agent identities for inbound requests, while managing secure requests to REST APIs for outbound calls. And it also enables your AI agents to search and select the most relevant REST APIs to best match the prompt context. And now to finish off, let's talk a little bit about storage. And there's a good one in here that we'll get to in a moment because it changes the answer to a lot of things.

25:00But firstly, Amazon S3 batch operations introduces performance improvements. In fact, it now completes jobs up to 10 times faster at a scale of up to 20 billion objects in a job. That's a lot of objects. That's a big job. And it now pre-processes objects, execute jobs, and generates completion reports up to 10 times faster with no additional configuration or cost. So if you're a batch operations user, this is great. Again, you're old, hey boss, I just sped everything up by 10 times. If batch wasn't doing it for you from a performance perspective, you now can. Speaking of improvements, this is a big one.

25:37Amazon S3 has increased the maximum object size to 50 terabytes. So that's a 10 times increase from the previous five terabyte limit. Now, Now, this is fun because one of the questions we'd often ask SAS and perspective SAS about S3 when we're interviewing or getting trained up was, hey, you can store as many objects as you want in an S3 bucket, but what's the maximum object size? And the answer, of course, was five terabytes. Well, now that is not the answer. The answer is now 50 terabytes. So you can store up to 50 terabyte objects in all S3 storage classes and use them with all S3 features.

26:16Now you want to optimize upload and download performance for those really large objects by using the latest AWS Common Runtime and S3 Transfer Manager in the AWS SDK. And so this is really important to make sure you're using the latest SDK to get access to this, but 50 terabytes, that's a big object. That's a lot. Speaking of a lot, there was a lot today. So we've had re-invent stuff, we've had pre-invent stuff. We've kind of got folks up to date and tomorrow's episode is going to be a little more infrastructure focused and we'll catch you up on any other announcements that have been happening.

26:50But, Killian, it's keeping us busy. I think this is keeping a lot of people busy, for sure. There's lots to absorb. Thank you, everyone, for listening. We love to get your feedback. AWS podcast at Amazon.com is the place to do it. And until next time, keep on building.