AWS and Qlik Strategic Collaboration Overview

Perfect. Uh, so incredibly excited to chat with everyone today. Um, Anna, dive into AWS and Qlik, uh, strategic collaboration Agreement and partnership.

Uh, I mentioned my name is Ola Wale Olohan. Wale Olohan. I'm the technical director for AWS supporting our software companies, ISV, and I'm joined by p Hello, everyone.

My name is Posh Bora. I'm Principal Solutions architect, and I also support ISVs, uh, customers, and Qlik is one of the customers that I support. Perfect.

Perfect. Um, so I think a couple things are really exciting for us. Um, one, when you think about AWS, it's, you're very used to like the infrastructure as a service to scale, compute, networking databases, uh, and then you go even higher up the stack things like serverless and containers.

And what's been great with the partnership with Qlik is that this brings together essentially a decade of partnership. Uh, Qlik bringing their data and analytics competencies to our customers, thinking about how we can blend together, not just the incre incredible like infrastructure and scalability, but the data analytics intelligence that you've heard of, uh, from earlier today. And so it's been great that we've had this decade long partnership.

Uh, and I also wanna point out some of our joint customers. Uh, one of them exciting, uh, Vanguard. I was here last year, Vanguard speaking about their data modernization using Qlik, using AWS to essentially rethink how they do kind of financial services and customer experience for their own end customers.

And so when we think about Qlik and AWS talk about a few things, um, a couple things that you've seen on AWS side is we really focus on how do we help our customers think about their modernization and data kind of migration strategy all up. And there's a realization that customers are gonna run their data in multiple different places. Customers are running SAP systems, and when they think about SAP, they have to think, how do I get real time data out of that system?

How do I be able to drive better analytics and insights on that? And that's where Qlik provides a great offering along with AWS in real time, trying to get answers to your questions, whether exist on-prem or exist in the cloud. Uh, similarly we've seen customers think about mainframes.

It's very funny. I started 10 years ago at AWS and they're like, mainframes, they're gonna go away. Now you're never seeing mainframe again in like two years, right?

No. These are core mission critical systems that customers run. They invest and they understand that they need.

And so what we've seen over time is customers saying, how can I tap into that mainframe data information very quickly, again, near real time to start solving my challenges? And this is where Qlik, through their integrations, are able to pull and integrate from these mainframe systems. And then you think about the rest of the AWS services, our data warehousing services, our data analytics services, tying those together to start driving insights and value.

And of course, we can't go too far without talking about ai, uh, and machine learning. So there's some exciting things between Qlik and AWS on one click uses and integrates with AWS AI services, our SageMaker, our bedrocks of the world. This allows you to train models, um, then run your inferencing either through a combination compute containers or more managed services like Bedrock, Amazon, bedrock.

And then the other thing we hear a little bit more about click answers is how you can integrate contextual personalization into what is commonly a rag based architecture. Rag is a way you can pull data in and say, I wanna understand it, I wanna interact with it, get personalized insights from my own data by leveraging these LLMs. So is this, are you saying that click answers rides on AWS?

Yeah, it integrates with, uh, Amazon Bedrock. Uh, so Bedrock is our generator AI service. And what it allows be, what Bedrock allows you to do is basically choose the best models for your workload.

And, and so that's what Qlik answers is using to run. Yeah. Correct.

Okay, Thanks. Now I'll hand it to p She's actually gonna dive deeper into the architecture as well. Yeah, sure.

Thank you all. Alright. Uh, yeah, so teeing up on, uh, what, uh, teeing off from what, where left?

Uh, let's start. I was thinking about what, you know, even before generated ai, um, became a buzzword and we started talking about it a couple of years ago, Qlik AWS partnership goes long back, right? And I was thinking about what workload should we talk about?

And it wasn't a tough de decision to pick click answers, uh, as a, uh, as a reference, uh, talk here. And let's dive a little bit deeper into how AWS and Qlik are working together to revo revolutionize data analytics, uh, through click answers. Um, so click answers, as many of you must be aware, is a plug and play, uh, generative AI powered, uh, knowledge assistant that brings together data from various data sources, provides business users with contextual, um, uh, relevant answers to their questions.

And these answers come with explainability, uh, so that the answers that you get are trustworthy, uh, are, uh, consistent and with, uh, they're with full explainability. So, you know, you'll have data sources that their answers are coming from those information also available. So, uh, you'll always know where things are coming, uh, came from and have access to those data data sources.

Uh, now I would like to go deep into how click answer, how AWS powers click answers. Uh, but before we dive into that, let's go through AWS, uh, service and let me provide you with a, you know, overview of AWS generated AI stack. So, um, at AWS, we have played, uh, you know, a key role in democratizing machine learning services even before generative AI came into picture.

Uh, over the years, uh, the main, uh, you know, goal for AWS was to how do we democratize machine learning services for our end users so that users and, uh, developers and builders at any level should be able to consume machine learning services and they should be available at their fingertips. So the way we like to see, uh, AWS gene tag is in three layers. One, uh, the bottommost layer is the infrastructure layer.

Um, if you're building your own models, AWS is relentlessly focused on providing everything that you need to build your, uh, machine learning models, whether it's traditional machine learning models or foundation models, right? The best chips, the most, um, advanced virtualization, uh, petabytes scale of networking, hyperscale clustering, and all the tools that needs to go along with it to provide you with everything that a developer will need to build machine learning models. Um, and that's why, uh, we have built Amazon SageMaker.

Now, as you mentioned, Amazon SageMaker provides you with the flexibility of whether you are want to build the model from scratch, or you want to fine tune an existing model, whether it's a foundation model or traditional machine learning model, or if you want to go a step, uh, uh, ahead and then, you know, even, uh, pre-train a model, uh, foundation model from scratch. Or if you just want to draw an inference from e existing foundation models, Amazon SageMaker provides you end-to-end support and tooling around developing your machine learning capabilities from, uh, building the model to delivering the model and deploying the modern production at scale. And with AWS grade enterprise grade secured and privacy, uh, context.

And the next layer, the middle layer is where we provide Amazon Bedrock. Amazon Bedrock is a, uh, you know, a hundred percent managed services. Our customers do not need to manage any infrastructure, and it provides foundation models, variety of foundational models as a service.

Um, you just need to invoke API to invoke a foundation model, and you will be able to build application generated via applications on top of, uh, Amazon Bedrock. Um, and the, it provides variety of foundation model that includes anthropic, um, AI, 21 Labs, um, uh, cohere meta stability, and Amazon, uh, proprietary models. Now, we know both Qlik and Amazon's customers, AWS customers, they care deeply about data security and privacy.

And that's number one priority of all the services that we develop. And same with Amazon Bedrock as well. So when you build your generat application, or when our customers builds our, uh, their Generati application on top of Amazon Bedrock, it has the Enterprise Gate great features that make it secure so bad customer can trust their data remains protected.

And that, so if, if, if other customers are using Amazon Bedrock and Foundation models available via Amazon Bedrock, they are assured they are, you know, that to train the underlying foundation and fine tuning the foundation models, the fine tune model will stay private and only they will have access to that models, the data addressed and in transit so that our joint customers, AWS and Qlik customers are assured that the generative applications like click answers provide all the enterprise grade data security, uh, and, uh, accessibility, uh, controls that it requires. Um, we were talking, we touched upon, um, agents as well, and we all know that agents are AI agents are industry buzzword, uh, last year and this year. So with AWS we are using the same approach with agents as well.

Like we were using with machine learning, providing democratizing the way and providing optionality and flexibility to our customers to use and build on agents the way they want. So it spans from, um, using out of box agent with services like Amazon Queue, where you can do workflow automation using agents, or if you want to build agentic ai, uh, product and services, you have Amazon Bedrock that provides, um, managed agent services. Or if you want to bring your own agents or you want to bring build agents using open source framework, you can do that as well, uh, and use AWS infrastructure to build on top of that.

So again, uh, that flexibility and optionality is there, uh, depending on what your use case is, what, uh, or what the skill sets are and how you want to build the agents. Let's come back to the click answers. How click answers is built and how AWS powers the click answers.

Uh, like WI mentioned, right? Bringing together clicks, um, uh, value proposition in data, uh, movement and data governance and data quality areas, uh, data pre-processing areas, and how do we bring it together with AWS infrastructure capabilities and providing some of these services like Amazon Bedrock and how click answers, uh, uses flexibility and opportunity that, uh, AWS provides. For example, when it comes to, uh, when a user asks a question, that question gets embedded, uh, in vectors, right?

And it still gets stored in Vector database. That's where, uh, Qlik uses, uh, Amazon SageMaker and cohere, uh, embedding model to, uh, to vectorize it and, and to embed it. However, when the prompt is prepared, um, using, um, you know, uh, re-ran curve, so that's where Amazon, uh, click uses Amazon, um, AWS, uh, sorry, cohere re-ran model on SageMaker to, uh, bring those results for, uh, to pass it on to the la uh, large language models.

And that's where click uses Amazon backrock, uh, managed services and underlying, uh, foundation models as anthropic, uh, uh, uh, cloud model, uh, to bring the response back to the, uh, to the end user. So that's how, that's where the flexibility and optionality, where what model and what purpose build model needs to be used and what tooling needs to be used for right purposes is what accelerates the innovation for our customers like Qlik. And that's where the value proposition of Qlik and a s coming together to accelerate that innovation for our joint customer comes into play.

Um, I'll stop here. Um, and hopefully it gave you a good insight into how to answer your question, how Amazon services powers click answers. That's good questions.

So, question, you mentioned, um, excuse me, um, for Coherent Bedding, and you mentioned it's being stored inside a database, right? What, what, uh, do you have a choice of database? Do I have to Yeah, we have choice of databases, but I think with, uh, Qlik, uh, we are using Open Search, open Search Vector db.

Oh, I see it. Okay. Yeah.

So walking back to the original premise, I have my traditional business critical systems, whether we're talking about mainframe, SAP, et cetera, where I don't necessarily want to use those systems to develop net new capabilities, not on the systems themselves. I want to use these modern cloud abstractions. The basic capability is Click does the connection between that and gets me two Bedrock and two Q where I can now build my, what we would call back in my day, bolt on applications to these systems to extend my mission critical capabilities and data into real time analysis, et cetera, et cetera.

Is that the overall story? Yeah, that's the story. And, and you know, without going specifics into SAP or Mainframe mm-hmm.

But yeah, that's the story. That how click data connectors and click value propositions and bringing all the data together from different sources, um, and integrating with Amazon AI service AWS AI services to provide a value even with, you know, traditional or mainframe, um, uh, you know, legacy, uh, databases or data sources. So In theory, I could create, in theory, I'd have an action that happens on, I have some batch processing that happens on my mainframe, and that connector would eventually lead me to being able to do something like a, a lambda enabled, uh, agent that runs some Amazon, some queue type stuff.

So I get the ability to have the complete Amazon stack extended from my mainframe environment. Is the mm-hmm. Is the overall value prop between click answers, bedrock q, et cetera, is that's how I get the magic.

Yeah. If, if it helps the, when I talk to customers too, around like the mental model of like AI data, data integration, data modernization, it's being able to make a decision about where you want to store your data, process your data, and then gain intelligence on your data. And so you look at a couple different modes.

You can see a world where in that batch processing mainframe example, I need to move data closer to a certain system. 'cause I need to have it be in the right format. I know the kind of patterns I want to ask it.

And so I'm making the decision about where my data needs to be so I can build an AI application on team. Mm-hmm. There's also customers who say, look, I've got all these disparate data sources.

I need bits and pieces from all of them because I'm trying to build something that new. Um, I have an airline customer as an example. And what they were struggling with is they had legacy, like, uh, like customer data, like legacy flight information.

And like, we're not gonna reform in all this. It's been around for decades and decades and decades, but we want to tie it to our new loyalty system. So how can we take capabilities like a click, like a data modernization migration to start to bridge the gap between what we can do in our legacy system and what we can do in a new system.

So I, I normally say you kind of have like different inflection points. Where do you wanna store your data? Where can you store your data?

What kind of decisions you need to make? Is it real time, is it batch? And then where is the best place for your ai, your machine learning to sit on top of that?

And I think between Qlik and AWS, we're trying to make sure you have as many options available. Um, so you don't just have one, Just a, like a short follow up to that one as well in terms of like the citation part of click, click answers. And you talk about the different sources that sit there and the legacy parts and the more modern parts.

What are some of the techno, 'cause I don't believe you've like referenced the citation parts specifically. So how does that work, especially when those sources are spread in quite a disparate way? I guess You mean through click answers or through AWS Yeah, through through click answers.

I guess the way it iss done for through AWS For click, I don't know if you wanna save it for the, Yeah, so we'll get into it, my section here in a second. But it depends on whether we're sourcing information from a knowledge base, as you saw on the slide that peush actually still up on, on, on the slide here, uh, where we're sourcing the chunk information from the document that has the correct response to the question or the query that the user has asked. Um, it's a little different if it's working in, in operating on structured data, which is another component of what we are now seeing now with, uh, with agents, um, using Qlik applications as a source for structured data where you get the query basically mapped to the specific fields and the specific filters that you, that, you know, we interpreted based on the user query.

So then, you know, what application and what data source it's coming from. So ultimately the citations are different because they're different document types or different information, but we believe strongly, and I think AWS supports this, that explainability is key to making AI a reality for organizations. And so you have to have that transparency.

It was interesting to hear you say there is fully managed, which was in that center kind of Goldilocks area, but you, you still enabled DIY on the right and on the left you of course had, you know, this kind of outcome, you know, so, you know, from q to to bedrock to here's all the primitives you could ever ask for. Um, which ones of those do you believe are the fastest speed to value for an organization? Yeah, so, um, again, speed to value is a related term, right?

So depending on the use case, uh, if there are, depending on the use case, uh, you can, let's say for example, if you wanna build a knowledge base, right? And for your internal, you know, customer base and you just want to upload some documents and get insights from those documents, right? Then click answers, right?

And is is one play, right? But then if you want to, you know, build a model or fine tune a model, uh, a foundation model, that also gives you speed to market if you start using SageMaker, because then SageMaker takes a lot of, uh, you know, undifferentiated, undifferentiated heavy lifting from our builders in terms of, you know, creating those pipelines and many, managing those infrastructure behind the scenes and pointing some additional tools for you to be able to, you know, develop that pipeline for machine learning models and partnering models. So yes, depending on this use case, wherever you're sitting or what your use case is and uh, how customize, uh, you want to, uh, how, how, how much you want to customize the model model or underlying models and how much of a control you want over your application, uh, uh, you will use certain set of services.

And in that category, it'll definitely provide you some speed to, uh, market if you start using those tools and services. Yeah. I'll, I'll add in something to that.

Again, like a couple things I've found helpful for customers, I normally try to meet them where they are really quickly. Um, I see Q for business is one of our things help you pull together like different business applications. You walk into a legal team, well, we haven't done anything with AI or generative ai.

Where do we get started? Oh, do you have documents internally? Let's start with Q for business.

Um, if you're someone who's already training models, you're like, I already do personalization, I do natural language processing, then something like SageMaker, I'm really familiar with the tools. I know how to use PyTorch and all the other capabilities that are already built in for customers who want to just say, look, I don't care what the LLM is. I just need to be able to start engaging with it in a really easy way for me to understand.

Then you start with bedrock, like, we kind of obscure away all the scaling work you need to do and just go pick a model, pick the version you want, what features do you want enabled, and click a few boxes. And then it's, it acts like an API. So I, I normally try to work backwards from a couple of use cases.

Um, and depending on how much experience a customer has, we tend to kind of work down the stack and up almost helps. Yeah. And it's not that one or either or will fit, uh, uh, uh, a customer's need or we have customers in fact is a very good examples.

And it's combination of these, uh, you know, layers. That was my follow Click answer. Percentage of what you've done would you say is managed service versus like extolling the benefits of the, you know, extreme primitives that you, that you have a dedicated team that thinks about nothing but the obsess over the best primitives available.

Yeah. But is managed services what you're seeing more consumed by the Qlik team as an ISV? Yeah.

If I, if I, um, I'll click also speak a little bit more around their architecture choices. I, I see customers using more, more than one. Okay.

It, it's, it's, it's for a few reasons. Um, I mentioned that like each line of business is different. So are you gonna, like, are you gonna buy a Ferrari when all you need is like a scooter?

Like, oh, like the marketing team needs access to these documents very quickly. Um, and then you have other use cases. How do I embed in my product now?

Enterprise grade security is important, reliability is important. Scalability, traceability. Okay, I need to use this, this, uh, service like a bedrock.

Normally I see customers deploying multiple one because it gives 'em the flexibility. Two, because they've already been using AI to some degree. Like SageMaker has been around for years and years and years.

Bedrock's been around for a few years. So yeah, I've already started on SageMaker. I've already gone containers with open source.

So like, it's a little bit of like a a and I would say versus an or. Um, and I think that'll continue. That's Why I was, that's why I was like the relative ratio, you know?

'cause there was the announcements in 2024 announce there's the announcements for 2025. So do you see a, you see an increase or a growth in the use of say a managed service component as compared to the prior, like, we got it working, we shipped it, you know? Yeah.

Now it's like, well, I don't like that particular operational pain. There's a managed service that I can use from AWS natively to do that. Um, and just trying to tie a bow around it.

Is, is that something you see increasing or do you see like, we're just innovating so fast. We're actually, we, we we're set of out outpacing even what the managed service can deliver. Which, which one do you think you're in right now?

Oh, it's interesting. So your question is, um, given how quickly this space is evolving Yes. Where do we see the most growth potential slash usage?

Correct. Um, it's actually interesting. If I were to say that when I think when I put that hat on, I'd probably put it in maybe two buckets.

Um, because this space is evolving, customers look to manage services to simplify the complexity. Um, as you know, uh, anthropic launches more and more models and every couple of months, like how do I keep track of all the infrastructure, the guardrails, the agent orchestration. So like, then they look at managed services.

'cause it allows 'em to at least simplify the foundations. You get security networking, your guardrails all configured for you. So like to keep up with the pace.

I've seen customers lean towards managed service for that reason. Um, then there's always a trade off around flexibility. Okay, I want to do a ton of things in open source like land graph and I can't easily do that.

You can do some of in bedrock, but it's not as easy as doing it just on EC2. So if I do think about pace of innovation, um, the managed services to try to abstract the way the undifferentiated heavy lifting so you can move faster. And then I think what other customers ask is like, okay, I've gotta, if I have investments already on, on, on like open source technologies, then they start gearing towards, okay, I really need to manage myself.

Manage is probably growing faster if you ask me. Yeah. In terms of just like year over year usage and adoption.

And, and that would support the thesis is that businesses are trying to put these things into action. And I think the theme I've heard just so far is like, okay, it's time for some action. Yeah.

And I think as the economies of scale get better, you've seen like price per performance, like throughput, inference get cheaper and cheaper and cheaper and cheaper models going faster and faster and faster. What's the most important part of your business? Your data?

Do you have the right business context, your guardrails? How do you make sure legal, marketing, et cetera, is tied in? That's where to manage services is abstracted that away.

So you can go, I've got security baked in, I've got my guardrails, I've got. So like that's where I think it allows you to move faster. And do you, do you see any of the underlying infrastructure, you know, powering this primitive, do you see a shift from, um, or, or an expansion in use of like the traum, the info, the graviton lines, uh, that are part of your underpinnings?

Do you see that growing as it relates to maybe even click answers? Or is it that's it's, it's over time with other customers and as those other customers see, it becomes part of what click answers is powered by. It's underpinned by things like your own silicon.

Oh. Um, the click answers may be a click answers Okay. Questions.

But I would say, um, training is still, uh, customers when they think about training and train inia, it's again about like choice, cost, cost performance. Yep. And I also think there's a difference between people who are building large foundational models or large domain specific models.

And those building small models. If you're like, I'm in a business of building more general purpose LLMs, larger domain specific LLMs, then you think about I need to have like flexibility in my compute and cost training in, if you get people who are like, I need a small language model, oh, you can probably do that on couple GPUs. Yeah.

Um, so that's probably where our, the differentiate it. And then we look for great partnerships like Qlik to say, what's the right offering for you? What technology makes sense?

And then we obscured away from Qlik customers. That's that's What, yeah. So that was the real question, right?

It's like, where does customer choice get forced into that conversation? They, did they have to think about this? Or is it like, I'm using click answers.

I don't, I don't need to know what's going on necessarily below the covers. Exactly. I appreciate what goes on, but I don't Need to know exactly.

Okay. Yeah. The idea is you, you don't need to know.

Um, or maybe maybe yet. It's just like it's better managed for you to make sure we're providing you the right kind of options and capability appropriate. So, so I'll follow up with a question on that because that turns into me a big question around security if it's managed and I, I I, I, I like where it's coming from, but if it's all managed for you, uh, what kind of assurances are being given that, number one, you're getting the answers that are relevant and the what they should be and what kind of assurances are given?

I, I appreciate the slides you put up first that talked about, um, the data not being used to train the underlying data and you know, of course the, uh, encryption at rest and in flight, which needs to be there. Um, but how are we sure that what's being built, um, if I'm just taking the pieces from a very high level as a click customer, that it's getting the pieces that I need to have for my customers to have a good experience and to have a correct experience. So those assurances available all the way down to the, uh, services that are being provided and managed by all.

Mm-hmm. Right? Yep.

Yep. So yeah, the way, the way to look at, uh, at this is the shared responsibility model, right? Is where like whatever a w services we are providing, when, when, when I was talking about, uh, your data is not being used for training the underlying model, right?

Or encryption, it's all, um, you know, documented and it's, uh, you know, it's a, uh, it's, it's well documented that you can refer to. Uh, so that's an assurance and coming. And when our customers like click the developed click, uh, applications, like click answers, right?

They take ownership also of making sure that because if, if they are processing the data or pre-processing the data that all the data is secured at that layer as well, and then they're providing all the instance, it's kind of shared responsibility model between click and a Ws, where a WS takes care of securing the infrastructure and providing enterprise grid security and compliance. And there are compliances, uh, if, if we talk about Bedrock, for example, right? There are compliances that, uh, needs that bedrock caters to including apa, uh, you know, and others.

Yeah. So just as a follow up real quick, when SA providers first started out to provide services, we are all aware of what the shared, um, the shared model is in SaaS providers, but this is kind of like a new stage of the data being processed without any really insight into what's going on. Mm-hmm.

And when we first started doing things like email and, you know, and a SaaS provider, the SAS providers did not make it very clear, um, how data could be lost. So is there any kind of, I'm just thinking there's probably a correlation because enterprises won't know what they don't know to protect. Mm-hmm.

And I'm wondering if there's any places that are gonna be gotchas for customers as far as, um, correct and, um, privacy of, I think the privacy is pretty well, that seems pretty well walked down, but of things being correct, or as we're seeing now with generative AI that the more you train the model, the kind of dumber and worse advice it gives back mm-hmm. In the same topics, right? And I forget what that word is, but that's happening.

Yeah. So how, how is what, what has to be, what are the gaps that customers need to be mindful of, that they need to put in their own backstops to make sure that their customers or their employees are protected? Yeah, I, I'll answer that and I'll do a broader than drill down to your question.

I think one of the unique parts, especially working in AWS is that we started with security and, and observability and traceability first. So you kind of look at how this space has evolved is like you call an API and it's kind of often an ether. We took a different approach of this needs to sit in your VPC, it needs to be, which is your security, your security boundaries network.

We need to have traceability and inspection. We need you to apply judgment on top of the model. So we have these things called LLM as a judge.

So when we started, we said, how do we start with least privilege, like we would with kind of like AI or data? And then when it comes specifically to things like bedrock, um, to answer your question, you can do traceability inspection and things that are happening. You can say, I wanna understand how this model comes to some kind of net assumption at the very end.

And then you can apply guardrails on top of that as well. You met, we talked about kind of hallucinations, kind of accuracy. How do you make sure you don't have model drift with your prompts?

How can you make sure it adheres to your corporate standards? So like, these are certain things that we never want to re return. So you actually bake this in on top of the model.

That's actually where we spend a lot more of the time. We said, we need to make sure that these models, by default, you have the right level of visibility throughout. And then two, instead of just saying send something and respond, we wanna make sure there's multiple checks essentially sitting on top of that so that as a customer you can kind of tap into each one.

Yeah. The last data point actually was data sovereignty. I'm not sure if you used that term, but it was basically like the, where the API is processed is where the data is processed.

Mm-hmm. And I almost feel like you, you know, that if you had led with that, that's that, um, that is a really important thing. Mm-hmm.

Uh, increasingly so I think we're gonna see in the news, um, when you, when you talk to customers, there's obviously, there's the cloud operations part, there's the security, and then there's the resilience, and there's obviously some resilience in saying like, okay, yes, and it's gonna be contained to this one geographical boundary. And those are just like, that's the rules for this data processing. Um, how do you help them balance that, uh, you know, possibility that maybe getting outside of a region, um, would make sense for them in their architecture or their model?

Oh, um, I'll give 'em where I start with customers. Um, when I start with MP really jump in. Sorry.

Sorry. Um, I normally tell them to start saying like, how do we think about your regional approach first? Yeah.

Uh, most of the time when customers think about regional approach, it's, it's data sovereignty, it's eula, what did we, what are the terms and services we said we're gonna offer to our end customers? We have to start there because we can't lose trust. Right?

We started establishing that the data needs to sit here and process it. Um, then we start looking at like other workloads where they say, well actually, like, this is a new application I'm building and what I really care about is maybe the accuracy or the latency. And so I'm gonna make a different decision.

Um, but I do think the foundations always to start region up. We keep launching more AWS regions for this purpose. We, we launch local zones to get them closer to customers.

So I would say like by yes, outposts as well. I think like by default, giving customers as much choice in region to process. And if they want to do something outside of region, that's where we just say like, well, what problem do you wanna solve?

Right? Like, there's a, there's a give and a take there. So.