S2E8 | How do we draw agentic borders? — The Shift Podcast by Microsoft Azure

0:00 We're building all these product capabilities and, you know,

0:04 controls for our customers, but how should they think about sovereignty?

0:08 Oh, that's like the million dollar question, right?

0:11 $10 million?

0:13 I thought we discussed it was 10 billion.

0:14 Yeah, $10 billion.

0:15 That's right.

0:15 (upbeat music)- [Evelyn] "The Shift," agentic edition,

0:18 connects the people building AI tools with those creating agents,

0:23 so we can all level up together.

0:25 Today we're taking a beat with the Microsoft Azure team, Evelyn,

0:29 Edouard, Meena, and Karim, about an open question we heard from you.

0:35 How do we draw agentic borders?

0:38 (air humming) (keys tapping)- I actually think

0:42 it's really interesting 'cause from a Microsoft standpoint,

0:45 we define digital sovereignty in a way where it's talking

0:49 about like the capability to participate in the digital economy securely,

0:52 independently, and with self-determined controls.

0:56 And I'm like, what does that mean?

0:58 (Evelyn laughing) Yeah,

0:59 and I think we try to kind of think about positioning it more so around,

1:05 it's not isolation, it's not about autarky, right?

1:09 So participating in the digital economy really means that, you know,

1:14 we operate globally.

1:15 We have to work with partners all over the world.

1:17 We have our operations all over the world, so.

1:21 I just came back from Mobile World Congress this week,

1:26 and what I really experienced is that sovereignty

1:28 is not like a predefined one thing.

1:33 It meant different things for different people that we

1:36 spoke with, and it almost felt like a spectrum.

1:39 On one end, one of the partners that I met with was kind of trying

1:44 to define this concept of sovereign enough for them

1:49 to be able to meet the regulatory needs,

1:53 the compliance, and build an economically viable

1:57 solution for, you know, their end users.

2:01 And then fast forward in that spectrum on the other end,

2:04 it was all about, you know, in-country, their infrastructure,

2:10 their hardware, building, their own AI model, you know, confidential compute,

2:14 like everybody's talking about sovereignty in their own way.

2:18 So obviously the sovereignty for a national security customer

2:22 is different from how a retailer looks about sovereignty.

2:26 And it's usually once you start

2:28 that conversation that you discover actually that, okay,

2:31 what is really the trade-off that they're ready to take?

2:35 What do they really need in their workload?

2:38 And that type of conversation comes if you

2:40 spend enough time with the customer and spend enough

2:43 time to try to understand and even help them

2:47 discover what is their real need in that process.

2:51 But it's never an easy conversation.

2:54 Yeah, I think most of the time when I meet customers,

2:58 their conversation often steers towards like, show me where the cliff is.

3:02 Like, when will your solution stop working?

3:05 Or when will I not have capabilities versus my-- Which is the worst scenario.

3:10 Job is to make sure there is actually no cliff.

3:12 You can actually have a successful journey

3:15 without ever having to hit that cliff, rather than jumping to the solution,

3:19 let's understand what the problem is and how we can jointly solve that problem.

3:25 I think that's a great point because I think what we've

3:27 been seeing is that digital sovereignty is a very emotive topic, right?

3:33 It's very emotional.

3:34 And so I think as humans,

3:36 our first thing to do is go to worst case scenario, as you just said, Edouard.

3:40 What's the worst thing that could happen

3:42 and how do I prepare that for myself versus,

3:44 you know, let's be a little bit more practical and kind of figure out,

3:48 you know, when we're thinking about sovereignty,

3:51 we should be thinking about what amount of, you know, which workloads, right?

3:55 The emotional part is interesting.

3:57 You know, the thing that I would say is

4:00 with all the innovation that's going on, you know,

4:02 everyone talking about AI and wanting to get involved,

4:05 people really feel compelled to use more technology

4:07 than they ever have in new novel ways.

4:10 But our world is now as or more uncertain than it's ever been.

4:14 And so when you combine those two things,

4:16 the discussion of sovereignty or digital sovereignty is what we need

4:19 to do to build confidence to operate in that digital economy.

4:24 Yeah, and I also think that there's some kind of confusion as well,

4:28 which is around sovereignty versus security,

4:32 versus there's aspects of sovereignty that are focused on data residency.

4:36 And I think sometimes that all gets conflated into one bucket of sovereignty.

4:42 Yes, to the point we discussed before that everyone has their own definition.

4:47 The most obvious would be data residency,

4:49 and that's the definition of sovereignty for most.

4:53 But actually when you dig, it's not always the case.

4:55 and it started that way because of GDPR,

4:58 because of the European data boundaries and so on.

5:00 We see requests coming from Asia and Latin America

5:05 and Central America and so on and so forth.

5:07 And in those regions, it might not be always the data residency,

5:12 sovereignty will mean different things depending on where you are

5:14 and what type of business and what type of workload you're operating.

5:17 Yeah, I just wanna build on top of what Edouard said,

5:21 like in addition to like data residency,

5:23 the most concern we hear from customers are related to business continuity.

5:27 Like no matter what happens,

5:29 they want to ensure their critical infrastructure in the line of business

5:33 that they serve continues to operate without

5:36 impacting their end users and their business.

5:39 Majority of customers define sovereignty as business continuity.

5:44 The end goal for them is to ensure that critical infrastructure keeps running.

5:49 I know we were talking about this earlier, is like,

5:52 what is the insurance that are willing to pay

5:54 to ensure that their critical infrastructure is running?

5:57 So that like resonates super strongly.

6:00 Like when we built up the Microsoft cloud,

6:03 the things that are like Office, Microsoft 365 and Azure,

6:06 the apps and services that we build that go

6:09 and span all these regions of the world, we look at failure modes.

6:12 We think, you know, as engineers,

6:14 what are all the combinations of things that could go wrong?

6:17 And we gradually have been fixing and addressing all those concerns.

6:20 And, you know, every once in a while, there'll be something we miss and we say,

6:23 okay, great, we need to go and address some concern.

6:25 But, you know, more and more, you know, losing one or even many pieces of all

6:30 these interconnected systems doesn't break our customers,

6:33 doesn't cause outages to the global fabric.

6:36 So the failures stay isolated.

6:38 And obviously some of the requirements that are

6:41 now bubbling up due to sovereignty and all

6:44 the geopolitical aspects and the fact that it's

6:48 central to most of the conversations today,

6:51 some of those requirements are legit actually.

6:56 Bring the industry forward, it helps us to get better and, you know,

7:00 making sure that we provide the right level

7:02 of controls to customers and the right level of audit.

7:06 And so it's actually a pretty positive trend on that front.

7:11 Whether it would've happened outside of that geopolitical context, probably.

7:16 As you were talking about the changing governance and regulations,

7:20 almost like every day we are hearing about like a new law,

7:23 new country coming up with some requirements.

7:26 I can't keep re-architecting my workload to meet act A and act B and act C.

7:32 Like I need to make sure that I build

7:34 something that is more sustainable and scalable over time.

7:38 And I need to have those controls at every level

7:41 of my technology stack so that I'm

7:45 well-protected against all these changing requirements.

7:48 A, they they want to be able to take advantage

7:51 of all this capabilities to innovate and participate in the digital economy,

7:57 like shielding themselves to make pace

8:00 with that innovation while kind of, you know, remaining protected.

8:04 As we talk about all these national policies, now we're starting to hear a lot

8:11 of countries mandate using their nationally developed generative AI,

8:17 large language models, and then they want to govern who gets access to it.

8:23 It's like how do they ensure

8:24 that that data remains within the country's jurisdiction,

8:29 especially with capacity constraints in, you know,

8:32 global regions and kind of partnering with new clouds for capacity.

8:36 They're concerned that they might not be able to protect their data,

8:41 either at rest or during training or during transit.

8:45 How are we, from an engineering perspective,

8:47 kind of building out different features or like keeping sovereignty in mind?

8:51 Yeah, I mean, I could speak to a couple

8:54 of the layers and the ways we think about it.

8:56 Confidential compute is a piece of it, that's,

8:58 you know, data in use being protected through confidential compute.

9:01 A big part of it as well is, you know, the data at rest,

9:04 like what happens for that data if it exists in another place and how you do it.

9:09 So encryption at rest and then management of who accesses, you know,

9:14 the keys or how those keys exist to protect and encrypt

9:17 that data is like a secondary way to think about it.

9:21 And the way I would say it is, you know,

9:23 every time you use layers of encryption or layers of protection,

9:26 you make the data, you make the processing opaque

9:29 to the operator or you make it opaque to someone else.

9:31 So if you control the keys to your data,

9:34 that data is yours and it's effectively in your virtual national boundary.

9:40 And so, you know, when you're training on large data sets,

9:42 if you're gonna process that data outside of boundary,

9:44 then confidential compute comes in from a processing in use place.

9:48 But if the data exists outside the boundary at rest,

9:51 then you're gonna encrypt it and you know,

9:54 you'll definitely want to have someone, you know,

9:56 intermediating the keys or an operator's access to the keys of that data.

10:00 And then that gives you a much higher degree of control of that content.

10:05 You can define all these sort of layers of keys and protection,

10:07 but even then there's more isolation

10:10 warranted or needed by those classification models.

10:13 That's where, you know, the local staff comes in to go and protect that.

10:16 And I guess a question I have then is,

10:18 you know, are people using local for training?

10:21 Are they doing local training and data?

10:23 And how do you think about that with things

10:25 like foundry local and other AI models being brought locally?

10:29 The majority of the use cases we see

10:31 with Azure Local is often with inferencing because that's more

10:36 of a versatile and affordable for a lot of customers--

10:38 You don't have to buy-- You able to quickly, yeah.

10:41 Hundreds of millions of euros of GPUs.

10:43 Exactly, you just lead enterprise-grade service

10:44 and you can stand up with existing, you know, data center infrastructure.

10:48 You don't need like fancy power requirements and water

10:52 cooled infinity band racks to be rolled in.

10:55 So the cost to get started is much faster in the inferencing side.

11:02 We're starting to see customers looking into fine tuning models because

11:07 they wanna be able to cater to their industry or their clientele.

11:14 And so tuning the models is becoming more of a pattern that we

11:18 see that requires slightly more powerful GPUs-- But doing that locally as well.

11:23 But doing it locally.

11:24 Yeah.

11:25 Training models at the edge is a very expensive proposition

11:29 that I don't think anybody has solved the mathematical equation to say

11:34 that this is going to be an economically viable solution

11:38 where I can amortize the CapEx of putting these very high-powered,

11:45 expensive infrastructure and be able to reap

11:48 the CapEx of that infrastructure at the edge, yeah.

11:51 For all the data.

11:52 Build models on top of the data with massive amounts of compute,

11:55 and then you do that at some interval.

11:57 And so between those intervals,

12:00 that hardware needs to get used for something or it will sit idle.

12:04 Yeah.

12:04 And to our point earlier on the trade offs, actually,

12:07 we do have alternate solution that allows for training to happen

12:11 in a given region or where you have some level of control of where

12:16 your training happens that is in our data center with Azure Local

12:19 or is it acceptable to have that level of training happening in the region?

12:23 And usually-- Yeah, there are some national

12:27 partner clouds that, not Microsoft but partners,

12:30 the customers of other countries are trying

12:32 to figure out how they can do one themselves,

12:34 but the way they think about, you know, amortizing the investment is to be able

12:39 to bring on more startups and bring more models

12:44 into the ecosystem so that they're trying

12:47 to keep that inventory fresh and running and utilized, rather than like, okay,

12:52 I build out like two models a year

12:55 and then what am I doing with all that infrastructure?

12:59 That's amazing.

13:00 I think the thing that we know and we certainly

13:02 have observed the last couple of years is demand far

13:07 exceeds supply for the compute infrastructure for GPUs or GPU

13:12 racks and in fact for power for all of those things.

13:14 So, you know, we're in this place where I think everyone

13:18 is very motivated to maximize what they're getting out of their investments.

13:23 The question is, how many places can make the scale of investments

13:26 or have the resiliency of investments

13:28 needed to deal with those sovereign concerns?

13:31 You are in deep thoughts, Evelyn.

13:33 I know, I was like, mm, yes.

13:36 I mean, I don't know if this is kind of forking a little bit,

13:40 but you just touched on national partner clouds,

13:42 but when is it that folks are really leaning into private cloud?

13:46 Is it more just for business continuity as you mentioned earlier?

13:50 Yeah.

13:50 Or is it a safeguard?

13:51 Is it that insurance or is it,

13:53 what are the different scenarios that you're seeing?

13:55 So often we see the same customer kind of picking a spectrum

14:01 of solutions for them depending

14:03 on their workload needs and the workload classification.

14:06 It's not that one customer picks one solution like

14:09 the sovereign public cloud and that is it for them.

14:12 So we see some workloads that can sustain an extended

14:16 period of disconnectivity and their function doesn't suffer as much,

14:20 or the intellectual property is not as highly classified.

14:25 And so they're very much within the realms

14:28 of their compliance to take advantage of the sovereign public cloud.

14:32 Can you give an example of like a disconnected or something disconnected,

14:34 like what's a use case?

14:36 So if they do a good business continuity backup solution in region where

14:43 they have multiple regions for backup

14:45 and if they lose connectivity to one region,

14:47 they're able to fall back to another region.

14:50 And so their workload might have some level

14:54 of slowness depending on the traffic all being

14:58 routed to one location and how they have

15:00 done load balancing and scalability in the region.

15:04 You know, how it, there are options-- I

15:05 thought you were gonna say the other way,

15:06 which are solutions designed to run in a disconnected way for extended periods,

15:11 yeah, that's the converse-- So this is like for the public cloud.

15:13 But then there are, you know,

15:16 workloads that cannot sustain any disconnectivity and they

15:19 have to run at the edge-- That word.

15:22 What?

15:23 Disconnectivity?

15:23 Yeah.

15:24 Is that a word?

15:25 I don't know.

15:25 (panelists laughing) The wrong person to ask.

15:27 We're making up a word.

15:28 But I love it.

15:29 You may have said it first.

15:30 Okay.

15:30 I'm gonna trademark it.

15:32 Yeah.

15:34 There are workloads that can sustain a certain amount

15:38 of disconnectivity that are running on-premises and you know,

15:42 what happens is, customers can still access the infrastructure,

15:46 the virtual machines and containers that are

15:48 running on the infrastructure keeps running,

15:51 but you just don't have access to the control plane

15:54 APIs to make any kind of mutating changes on them.

15:58 And so, you know, in a matter of, you know,

16:01 a few days you can restore connectivity and you know,

16:03 things will just keep running as though nothing happened.

16:05 But then there are highly-sensitive workloads

16:09 that just cannot sustain any disconnectivity.

16:14 Yeah, sure.

16:15 But why is that?

16:16 Like I've worked on SaaS solutions in our cloud

16:19 and the hyperscale things and you know,

16:22 there's always some threshold where these big

16:24 applications we've built need to be

16:26 connected to some other part of those systems that make up what they are.

16:30 [Meena] Think about like the submarine scenario.

16:33 Perfect.

16:34 Yeah.

16:34 Or-- I was hoping you'd go for something like that.

16:36 Yeah, yeah, yeah.

16:37 So often like, you know, once every six months, they come offshore and you know,

16:41 catch air and then maybe can take an update

16:44 and then after that-- They're offline again for some period, yeah.

16:48 They're just offline for like extended period of time.

16:49 So it's not a matter of like days, like you need to be able to access

16:52 that infrastructure over an extended period of time.

16:55 And so running a completely disconnected operation

16:59 becomes paramount for those type of scenarios.

17:01 and the way we've built the technology is that you're

17:05 not coming up for air to give a heartbeat.

17:09 Someone's giving you the bits from the cloud so you can just

17:12 go back down underwater-- Download as fast as you can and move on.

17:15 Yeah, so this, I mean, our Azure Local disconnect operations were

17:18 built for the submarine scenarios where you

17:22 have an extended period of time where there's just no connectivity at all.

17:27 But then as we have progressed

17:29 into these sovereign conversations in the past few months,

17:33 we see customers wanting to have the optionality

17:38 to have a local control plane as a failover mechanism.

17:42 They want to ensure that if I lose access to the public cloud,

17:46 I want to make sure that my data and my workload can fail over and run

17:53 from my local cloud without having to take

17:56 a dependency with the public cloud or connectivity concerns.

18:00 And so the Sovereign Private Cloud with disconnected

18:04 operation is becoming more and more of a meaningful

18:08 alternative for customers to ensure that critical

18:11 infrastructure that they need is available for them.

18:15 Makes sense.

18:16 Are there security trade-offs when you have a disconnected operation?

18:22 Well, every technology we build at Microsoft,

18:24 we build based on security as the primary pillar.

18:29 Number one thing.

18:29 Yeah.

18:30 So we build security at every layer of our infrastructure.

18:34 You know, it comes with secure boot, it has trusted VM with guest attestation.

18:40 We're investing in bringing confidential compute as well, both for VMs and GPUs.

18:46 All of the Azure's security and governance policies are

18:50 brought in as part of our Azure Local solution,

18:53 even in the disconnected operations.

18:55 But it's a limited set of Azure services that we are bringing

18:58 to the edge and be able to run with the local control plane.

19:02 But the goodness here is the consistency of those services, the Azure portal,

19:08 the same APIs and PowerShell commands,

19:11 so customers don't have to learn a new technology.

19:15 And Meena, this is the point that comes across in the customer conversations.

19:17 Like you mentioned before,

19:19 that most of the customer doesn't go for public or for private,

19:22 and usually it's kind of a mix of both.

19:24 And so being able to manage their workloads depending on their requirements,

19:29 consistently with the same technology, across public and private,

19:35 is what lead them usually to pick that architecture, so.

19:39 Yeah, I mean this is a common question that we get asked by customers,

19:43 how is the lifecycle management within Azure Local Disconnected?

19:49 And the first surprising response that they hear is, it's always disconnected.

19:54 It never has to send a heartbeat back to the public cloud.

19:58 It's like the diametric opposite of the cloud that I work on.

20:01 (panelists laughing) Which is, you know,

20:04 how many hundreds of thousands of updates every single day,

20:07 but we actually count on our ability to go as fast as possible

20:12 with safety to get updates and features

20:15 and innovation and experiments to that cloud.

20:18 And so while the resiliency concern is absolutely top of mind for us,

20:22 I'm like, I need to go really, really fast.

20:24 You know, I need features deployed or fixes in, you know, out-- In minutes.

20:28 In minutes worldwide.

20:29 And so the amount of work and the kind of work that you

20:33 have to to have a solution like that be both secure to your point,

20:36 but also, you know, stable and resilient,

20:38 like you may not get another chance for six months to fix those issues.

20:42 So with Azure Local, whether connect or disconnect,

20:46 in both cases, we actually push out updates every month.

20:50 So every month customers have an opportunity to get access

20:53 to those-- Come up for air and then download the bits, yeah.

20:56 Yeah, the security patches, updates and features, you know,

20:58 depending on, you know, the cadence in which we push updates.

21:02 The submarine analogy is amazing,

21:04 super helpful to like think through resilience concerns, operational concerns.

21:09 But I like, I'm curious what the equivalent or what a different,

21:13 like what metaphor would you use to talk about our public cloud

21:16 and how customers think-- I don't think it would be as perfect.

21:18 But basically if you look at the concern,

21:21 a lot of our customers are in regulated industries, in the public cloud.

21:24 And so if you look at a bank,

21:26 they have sense of constraint, they want to run their workload,

21:29 they usually have a lot of regulatory constraints that they have to comply with.

21:33 Like what?

21:34 What's an example?

21:35 Local regulation and making sure that they comply

21:38 with the standard financial regulation in their country.

21:42 And so that-- One example of that would be PII information.

21:46 Like your citizen, like your clientele personal identifiable information.

21:53 Purchase patterns, or-- Stored in a way or stored in place.

21:56 Who has access to it, yeah.

21:58 Got it.

21:58 So here you want to make sure that basically your data stays in the country,

22:03 that they're encrypted if necessary.

22:05 And so that's where confidential computing is useful to them.

22:09 But now more and more, you can throw actually AI in the mix and extrapolate some

22:15 of your results and manage some of your risk through AI.

22:20 And the way you control where your data is processed,

22:23 where your model is trained.

22:25 So it's always deployed usually in a very

22:28 secure environment and sometimes with Azure Local as well.

22:31 So even though it's not a submarine, they also are interested into having

22:36 that disconnected operation capabilities for continuity purposes.

22:39 You would've the local AI, or control your AI piece.

22:42 Some of the isolation pieces, yeah.

22:45 And you would've also the new operator control or new operator access.

22:49 So you basically control who has access to your data,

22:52 when, and you're fully in control of your data environment.

22:58 That's all.

22:59 When you talk to customers about their concerns.

23:01 And obviously part of what you do is you're

23:03 trying to elicit a set of questions from them,

23:06 you know, what are like the last few conversations that you had, like?

23:10 The risk scenario is a very good one.

23:13 So it's how can I leverage AI without going into crazy investments

23:20 because I don't have the bandwidth or the-- Five megawatts of power, yeah.

23:26 Yeah.

23:26 So I still want to benefit from the power of the public cloud,

23:31 but at the same time I want to keep that control over my IP.

23:35 And so I want to protect that.

23:37 And maybe I'll kind of ask a slightly different question.

23:41 I mean, we're building all these product capabilities and you know,

23:45 controls for our customers, but how should they think about sovereignty?

23:50 Oh, that's like the million dollar question, right?

23:53 $10 million?

23:54 I thought we discussed it was 10 billion.

23:56 Yeah, $10 billion.

23:57 That's right.

23:57 Even more than that I think.

23:58 I mean there are definitely optionalities that we're investing

24:01 in, but how does customer decide like when to choose one,

24:05 especially when our customers are global customers.

24:08 And so how does our solutions scale by region, by country?

24:14 Right now we offer the sovereign public cloud.

24:18 Anyone can kind of leverage the sovereign controls and features that we

24:22 offer in the public cloud in which they already operate, right?

24:25 So, you know, we have 70 plus data

24:28 center regions operating across, what is it now?

24:30 400 plus data centers.

24:33 Then you kind of layer on the Sovereign Private Cloud,

24:36 which gives them kind of that flexibility

24:39 to not just have the sovereign public cloud,

24:41 but they can kind of operate hybrid,

24:44 or disconnected as we just talked about, in terms of disconnected operations.

24:49 I think right now, a lot of our focus has been around Europe, right?

24:54 So we've got the European Union and EFTA,

24:57 which is European Free Trade Association.

24:59 It's been really easy for us to kind

25:01 of think about how do we tackle a large market

25:05 that has harmonized regulations and it's much easier for us

25:09 to kind of address sovereignty in a big area there.

25:13 Our next challenge as we think about it is,

25:14 you know, how do we bring that out globally?

25:17 Already we've made announcements in India, we've made announcements in Canada,

25:22 then you go into Asia and it's, you have to kind of pick things out

25:26 in different countries based on whatever their regulations are,

25:28 what their requirements are, what exactly they're needing.

25:32 And so, you know, we're kind of partnering

25:35 around the contractual and legal assurances as well, right?

25:40 I think now we're just trying to expand that out as we go around the globe.

25:45 So, you know, we have a number of public cloud locations

25:49 where customers can take advantage of sovereign controls we're investing in.

25:53 But then we're aware that we don't have

25:57 region in every country or every location jurisdiction.

26:00 And for that we have Sovereign Private Cloud

26:03 where customers can deploy on-premises and then choose

26:06 to connect to a public cloud for the control

26:09 plane if that is within their jurisdiction boundary.

26:12 Or can have a locally deployed control plane

26:14 with Azure Local Disconnected and take advantage of, you know,

26:18 the Azure goodness, Azure consistency while running

26:22 on-premises and meeting all the regulatory needs.

26:24 Like this is a pretty broad range of customer cohorts that we talk with.

26:28 [Evelyn] Absolutely.

26:29 There are many, many capabilities that we can continue to build

26:32 by talking to customers in particular industries and for particular

26:36 concerns that turn out to be amazing ways to go

26:39 and solve for the next set of problems that come up.

26:42 And so like, I'm bullish that even if

26:46 we don't know exactly what the world looks like,

26:49 we're building towards whatever might be coming our direction,

26:53 it might be coming the way of our, you know, vast customer base.

26:57 Karim, I would just add to that.

26:59 Most of everything that was built in the first

27:01 place was built for public sector and regulated industries.

27:05 And now basically every customer, no matter their industry,

27:09 cares about sovereignty and so all those feature that were usually

27:12 limited to very specific industries are now kind of a mainstream.

27:17 One of the investments we're making with Sovereign Private

27:20 Cloud is bringing Sovereignty AI to the edge as well.

27:25 And what that means is we're bringing some of the goodness of our Foundry,

27:31 Microsoft Foundry capabilities to the edge.

27:34 Satya talked about Foundry Local on Azure Local,

27:37 where we are bringing access to the thousands of like thousands

27:42 and thousands of open source models that can be deployed at the edge.

27:45 So customers don't have to, you know, build their own models but can take

27:50 advantage of global models that are available.

27:53 With Foundry Local, we're also extending to bring

27:56 more agentic AI capabilities at the edge with all

28:00 the IQ workflows that now is available in public

28:03 cloud will soon become available in private cloud as well.

28:07 Both in connected as well as disconnected.

28:09 So this really helps customers get a jumpstart in their AI innovation,

28:14 particularly as we have pivoted to investing in more

28:18 of the agentic capabilities that customers want to build.

28:21 (light upbeat music)- We're still offering

28:23 that continuity between public and prime.

28:24 Yeah, exactly.

28:25 (light upbeat music continues)- [Evelyn] Thanks for hanging out with us.

28:29 Check out the link below to get our latest whitepaper.

28:32 See you next time.