Sovereignty by Design: Rethinking Data & AI, Cloud, and Control – with Dr. Maja Himmer, T-Systems

00:00:00: If you look at the AI models that you can host sovereignly in Europe, this is much cheaper than accessing a GPT-Five point five or Fable.

00:00:09: Or those proprietary state of the art models.

00:00:12: they are so expensive now and we see this pattern Now That A Lot Of Companies Are Interested In Reducing Their Costs And At The Same Time Increasing Their Independence.

00:00:21: So This Is Kindof A Sweet Spot.

00:00:32: Hello and welcome to the Data Culture podcast.

00:00:35: I'm Karsten Bange, founder and CEO of BAG And my guest today is Maya Himmer.

00:00:40: She's AI and sovereignty strategy lead at T-Systems One of the providers of sovereign infrastructure in Europe.

00:00:50: I wanted to hear from her first of all, her take on the sovereignty discussion.

00:00:54: So is it really that companies are now starting to invest in or they still only talking about?

00:01:02: And then discuss with her how a sovereignty actually being implemented as, for example infrastructure or platforms as a service today and where all this is leading with special focus on AI.

00:01:19: So we are looking at AI models and were to run them in that specific case the cost considerations and big headaches of many companies.

00:01:32: it's very well aligned considerations they have, because especially in the airfield cost is now one of major drivers for change.

00:01:42: And we discuss what she sees there and what companies are actually doing?

00:01:47: We end with some practical questions to ask if you want address sovereignty better or how can approach that interesting episode.

00:01:55: enjoy it!

00:01:56: Hi Maya welcome to The Data Culture podcast.

00:02:00: hi thanks nice be here.

00:02:02: absolutely Maya, our topic is obviously data sovereignty and everything that's basically related to this topic because you are one of the big experts in this topic.

00:02:15: Let me start with a general question.

00:02:17: what your perception was?

00:02:19: Did latest news like the Fable incident or pulling back from this model... What did they do for their whole discussion?

00:02:32: It did a lot.

00:02:33: So the whole topic of sovereignty, we're now talking to companies that consider themselves global companies and not affected by this whole sovereignty discussion.

00:02:42: And now they are suddenly interested.

00:02:44: so Fable also the GitHub co-pilot price model change was very huge trigger.

00:02:54: I mean We just see how tokens are being consumed.

00:02:59: token consumption is going up exponentially for large companies.

00:03:04: They now consider it a key topic, so this whole topic like where can I get access to the best model and what does it cost me?

00:03:17: And how reliable is

00:03:18: it?".

00:03:19: It's very strong discussion topic everywhere And it's, as you said, highly related to the sovereignty topic.

00:03:27: Absolutely I can confirm that just.

00:03:30: we had so many discussions at The Data Festival recently That two topics are really tightly bundled.

00:03:39: in a way because cost often drives very similar discussions In terms of what do i need now?

00:03:48: What do i use actually driven very much by AI because that's where the new experiences are to be made.

00:03:57: You also mentioned it, tokens for general AI systems other big topics but sovereignty is related to that.

00:04:07: so what then your overall point of view here?

00:04:13: So you mentioned there had been a change.

00:04:18: How strategic is sovereignty now for organizations or companies?

00:04:22: I think you can decide between different clusters of companies, right.

00:04:26: They are the ones in a highly regulated industries health public sector also finance.

00:04:31: they have been considering sovereignty as an important core concept and their businesses anyhow four years already.

00:04:39: And then there other ones that maybe produce intellectual property or are handling some kind of sensitive data, but they aren't really object to strong regulations so that can define their own desired level of sovereignty.

00:04:58: I mean within the let's say GDPR compliance.

00:05:02: and there it is where we have already seen a bit of change in past years when our sovereignty was more about resilience for the businesses that are, for example in production and they're heavily reliant on non-European services.

00:05:23: In Europe.

00:05:23: there was a Stragi study which said that eighty percent of digital services in Europe come from non European vendors or providers.

00:05:34: so this is very strong finding.

00:05:36: already few years ago not much has changed.

00:05:40: now I would say Yes, sovereignty is now more about resilience and about independence.

00:05:48: About having the freedom of choice in case there are any

00:05:52: incidents.".

00:05:53: And then I think that it's this third cluster of companies which also was not really affected by discussion on operational resilience but looks at AI model space.

00:06:09: Yeah, recent incidents that you mentioned earlier where suddenly the newest technologies are not coming to Europe anymore.

00:06:18: And I mean Fable was not the first one.

00:06:20: I think Metta did at beginning a few years or maybe one and half year back with the newest Lama models.

00:06:29: they didn't ship to Europe any more because of the EU AI Act like big hurdle on regulatory issues there had.

00:06:37: so this is now really new.

00:06:38: but The better those models are becoming and the more we build our core businesses, in our daily operational work around it.

00:06:48: The more we are reliant on non-European AI model providers of course cloud providers as well.

00:06:58: so yeah I think that's the evolution.

00:07:02: how sovereignty has been seen increasing interest?

00:07:06: Absolutely If I look at our recent study, or survey on that one of the findings was that more-or-less everyone thinks it's super important.

00:07:16: It is either even of rising importance.

00:07:20: but then if we asked... Or when we ask yeah okay what do you actually about?

00:07:25: Do your invest and have budget for sovereignty initiatives

00:07:30: etc.,

00:07:31: Then obviously enthusiasm wasn't as high.

00:07:36: I think there's a difference between finding the topic super important and then actually doing something about it.

00:07:43: Question to you is, do we see a change?

00:07:45: So did all these latest developments now lead that decision makers act?

00:07:52: Yes!

00:07:54: It's good question because as you mentioned most leaders ship their topics back through IT departments.

00:08:04: they say sovereignty, security.

00:08:07: That's all one thing.

00:08:08: you handle

00:08:08: that.".

00:08:10: On the other hand in the IT departments they are on pressure to save costs and not spend more money on like more sovereign solutions.

00:08:20: And there is no real this point where sovereignty is a strategic topic with budget.

00:08:28: The systems actually I think first company assigned role of chief sovereignty officer.

00:08:35: So, yeah.

00:08:36: I think for systems also not new to talk about the sovereign topic we do that from many years.

00:08:41: so fast.

00:08:42: it's kind of natural to have someone who is dedicated in pushing their topic further.

00:08:47: but this definitely similar to this chief data officer role which had kinda one department and whole organization work with him or her on bringing a data strategy at place in the company.

00:09:03: I think it's very similar with sovereignty because you don't look at sovereignty from single point of view, It is not only about cloud but also where your data sits and only about models.

00:09:17: that touches

00:09:19: everything.".

00:09:23: Your question was whether we actually see companies act upon this now slowly?

00:09:30: Yes.

00:09:30: Okay, but yes so there is change to the better in a way or two to the topic?

00:09:37: But that's I think it's good here.

00:09:39: That's really good

00:09:40: and

00:09:41: you see patterns how companies approach the topics.

00:09:45: So are they certain things like address first of what's biggest importance for them?

00:09:52: So when we split into let say cloud discussion and the AI discussion of it.

00:10:01: Then certainly in the cloud area, this is where when you talk about having your data storage with a European provider or processed within an European sovereign Cloud This usually comes at a higher price.

00:10:16: Not always maybe but In many cases obviously the hyperscalars have other kinds of economics because they're I don't know how many orders of magnitude larger.

00:10:29: So it's maybe even easier now to start with the AI side, because if you look at the AI models that you can host sovereignly in Europe these open models.

00:10:43: this is much cheaper than accessing a GPT-Five point five or Fable or those proprietary state-of-the art models.

00:10:51: they are so expensive.

00:10:52: Now when you look That it's easier and we see this pattern now that a lot of companies are interested in reducing their costs.

00:11:02: And at the same time increasing there independence, so this is kind of sweet spot right?

00:11:07: Yeah I can completely agree!

00:11:12: especially cost pressure that we have here in many European companies, especially in central Europe is driving them to the topic.

00:11:20: A mighty cloud approach seems to be a pretty common way to tackle sovereignty and companies.

00:11:25: do you see this thing?

00:11:26: Absolutely it's one relatively straightforward way to get less dependent right?

00:11:33: so take Deutsche Dillekomm as an example.

00:11:35: of course part of our applications running on the hyperscalers, I think sixty percent or so.

00:11:45: And we also have ten per cent of the stuff still running on legacy on-premise systems and maybe that's like never gonna be fully cloudified.

00:11:54: but we off course use our own sovereign infrastructure.

00:11:57: i think about thirty percent.

00:11:59: now We have on tCloud public or T Cloud related infrastructure Our own dedicated infrastructure whatever And we always push to increase that, because of course it makes a lot sense to utilize what you have also for your own.

00:12:17: For cloud but of course as well for AI.

00:12:21: and how much on-prem?

00:12:23: So you've public cloud or hyperscale... Yeah

00:12:25: ten percent about.

00:12:26: I think so.

00:12:28: Ten percent say we still have on premise and i think mainly legacy systems are not easy too to move out.

00:12:37: Okay, so ten percent on-prem thirty percent in your sovereign own infrastructure and sixty percent in the hyperscalers?

00:12:46: That's about constitutionally.

00:12:47: that is interesting.

00:12:49: let us dive a little bit deeper into because On one side UST systems you have an offering.

00:12:58: On other hand You have big cooperation going with Google.

00:13:01: So how does it work?

00:13:05: How do you do that, doing both in time?

00:13:09: I mean we have corporations and or let's say partnerships with all of the hyperscalers because we're traditionally an IT systems integrator.

00:13:19: We also offer managed cloud services for all U.S.-based hyper scalers.

00:13:24: That is what customers demand.

00:13:25: And then when you mention Google i think What your referring to this?

00:13:33: Google sovereign Cloud?

00:13:34: Yeah, systems has been trying to bring into the market what like two years ago.

00:13:40: It still exists but it didn't really work the way it was supposed to be working.

00:13:46: I would say so...I don't think we have a lot of customers interested in that and i mean..i understand it ..it was not having this same level functionality as normal Google Cloud.

00:14:01: We now moved our strategy entirely to having our own cloud infrastructure.

00:14:06: So I mean, we already have the T-Cloud public for a long time.

00:14:10: that was usually in the past.

00:14:12: it was called Open Telecom Cloud over ten years?

00:14:16: Maybe even fifteen years old...I don't know exact number and It makes absolute sense to promote this as the sovereign cloud offering more than a US based hyperscaler sovereign cloud.

00:14:30: Yeah, so that's why I wouldn't like put any emphasis on this Google thing when we talk about sovereign Cloud.

00:14:36: We talked T-Cloud public our private cloud offerings and now recently also the industrial AI cloud.

00:14:45: So yeah, AI compute, AI factory that we built in Munich and that were operating out.

00:14:52: Interesting okay!

00:14:54: And what is your ambition there?

00:14:56: So do you want to become like a German, European hyperscaler?

00:15:02: Where does that lead or what's the strategy behind it.

00:15:05: I mean we kind of already are.

00:15:09: when you look at recent analyst ratings and so on We're already number one European hyperschaler alternative although you wouldn't probably classify it as a hyperscaler, but more like a superscalar maybe?

00:15:23: Yeah because I said that's bit smaller.

00:15:25: But from the European sovereign cloud landscape It is most mature public-cloud solution and when compared with hyperscale functionality You see an eighty percent feature parity in core features And on the agenda, we have to increase that to a hundred percent of core features until end this year.

00:15:52: So yeah most feature is people are going use maybe not super special serverless non SQL database.

00:16:01: but I mean this was an example, i don't even know if we have with us to be honest.

00:16:06: But yeah so but the things that are used the most we will have and or over here having.

00:16:12: So I wouldn't say there's only a new strategy.

00:16:17: of course it is the strategy position test as leading public quality in Europe.

00:16:22: It has...I mean its Germany for sure?

00:16:24: We also have data centers in Switzerland And Netherlands!

00:16:28: So certainly European offering not just for the German market.

00:16:32: Okay, very good and you talked about core services.

00:16:37: I'd like to know more about that from a specific viewpoint of data in AI.

00:16:42: so obviously there's infrastructure as a service.

00:16:44: but then if we move i'm thinking about these let's say pretty old layers in terms of uh service above that is platform as a surface which i think still works because you need databases you need data integration technology.

00:16:59: What are you offering there?

00:17:01: I cannot tell in twenty minutes, because the hundreds of services that we offer and different types of databases are certainly part of that container registries.

00:17:10: The typical things like virtualization layers even data pipelining frameworks.

00:17:18: it's called Data Arts or things similar to the SageMaker model building platform.

00:17:26: It is called Model Arts.

00:17:29: There is a huge package of services.

00:17:33: You can check out on the website, I think there's also an overview what you can book notes called and i think we even have tables that compare these services to the equivalents off the hyperscaler.

00:17:46: so for anyone interested in moving it's kind of easy to define a map, a migration map.

00:17:53: hows that service called?

00:17:55: In our world And I mean, we also offer migration services to help that transition.

00:18:01: Of course Sam and i think what's special with us is this industrial AI cloud which was really not just... We have put a number of GPUs into our existing data centers but we built real AI factory based on reference gold standard architecture as Nvidia suggests such that you can really bring the most performance out of GPUs.

00:18:30: We put two GPU types together in combination covering a wide variety different workload type and this is our AI engine, which we use for heavy AI workloads.

00:18:47: How do companies get access to that?

00:18:49: What's your motto there?

00:18:51: I mean if you want to go into that specific, so this would be a dedicated access point that we set up.

00:18:59: So it's not reachable via the public internet but you will have a dedicated IPsec tunnel.

00:19:09: since We are already now touching on AI and AI workloads let's dive deeper in.

00:19:16: What's your take on sovereignty and AI models?

00:19:22: So what are companies now concerned about, what they're doing.

00:19:26: Yeah I think the discussion... No we already have it that you think cloud.

00:19:32: then maybe you think data.

00:19:34: Then use those dated train models to ingest them into models so that you can run models And build applications Workflows, agentic systems And along this whole stack, you would want to consider sovereignty and with AI.

00:19:50: You also have those different layers yeah?

00:19:52: You have the AI models.

00:19:55: um maybe let me start with a definition of how we classify these in terms of sovereignty.

00:20:04: so you'll have the proprietary models that are not sovereign at all, because they're like a black box.

00:20:10: You get an API you send something to get something back.

00:20:13: that's no sovereignty involved.

00:20:15: then you have the data sovereignty level where you say I Have control over my data?

00:20:23: I know when my data sits.

00:20:25: i Know Where are my data is going.

00:20:29: Yes At least do you ever feeling you will have some control for your data and this Is what we're in the AI model Layer, I would say the open weight models live.

00:20:40: So you have The weights those parameters that you can download.

00:20:46: You need to architecture of a model and then you can deploy your model on your own hardware.

00:20:52: so That means you have If if you do that on let's say hyperscalars it still goes through the hyperscales but you have control over the Model.

00:21:03: you can you can host yourself and then if we go into the next level of sovereignty, that would be operational sovereignty.

00:21:11: So in a cloud perspective this will be case where you say I have myself the kill switch nobody can switch off my services because i operate the Cloud.

00:21:26: for AI model maybe not exactly mappable to these sovereignty levels.

00:21:32: It's because open weight could also be considered as operational sovereignty.

00:21:39: But the next level when we consider models would be Open Source Models, so there you don't only have access to weights and architecture that can run model but actually look at source code.

00:21:52: how was a model trained?

00:21:55: And this gives more transparency into whether aligned with your values.

00:22:03: For example, if you have insight into the training data You could see If there are any data biases within the training dataset that for example you train Image recognition model to take a very simple example and it has only been trained on white people.

00:22:20: so It will never recognize a black person as a person but maybe some form of animal.

00:22:26: So this happened in the past like ten years ago or so.

00:22:30: It's a very easy to understand example, how bias works.

00:22:35: And of course it is good way.

00:22:37: make sure that the model you're using can be compliant with your local regulations or values and so on.

00:22:46: Then there are various levels.

00:22:48: some providers give you training code but not data.

00:22:53: Some gives insight into.

00:22:55: they chose their training data.

00:22:57: But this whole like training Cycle is not really published.

00:23:02: So it's sometimes a bit there like gray areas.

00:23:05: You cannot really draw strict lines than they are.

00:23:07: certain licenses also Like some are published under a commercially available license.

00:23:12: Some are under a restricted license.

00:23:15: Um, so lots of things to be considered.

00:23:19: but what we consider as the highest level of sovereignty?

00:23:24: The technological sovereignty where you really have control over technology, and I mean when you think of hardware this is really difficult in Europe because we don't have GPUs.

00:23:39: And honestly i don't believe that it's going to change if we have high-performance GPUs for the next few years.

00:23:47: but if you look at models there are some initiatives in Europe that are then fully compliant with the EU AI Act, they're fully transparent and also cover all European languages for example.

00:24:05: That makes sure this model is aligned to our European

00:24:10: values.".

00:24:11: And so most recent examples of that as for example Sophie from the German AI Association like lots research institutes and universities that came together to say, let us train a large language model in Germany from scratch.

00:24:28: That fulfills all these criteria.

00:24:30: so basically we not maybe as personal or company but state or political region can have technological sovereignty over this model.

00:24:43: the SOFI model is actually being trained on our industrial AI cloud.

00:24:47: some are very proud that we already, Ori can serve as such a nice sovereign AI initiative with this data center.

00:24:56: Interesting.

00:24:57: so there are lots of interesting things in there.

00:24:59: just I found it interesting but you're differentiating between open weight and open source models because off their men yet they used synonymously.

00:25:08: So maybe could give us few examples.

00:25:11: where does gets clearer what you mean.

00:25:13: I mean, when you refer to proprietary models obviously You're talking about JetGPT or on the end topic models and so on.

00:25:22: What do you see maybe most popular or often used open-weight and open source model?

00:25:30: To be fair, most open models that we see today are open-weight models.

00:25:34: Open source models are very rare.

00:25:37: there have been some in the past bloom and so on but they haven't really made it into this area of large language models that were talking about today.

00:25:48: actually not sure I think There are some.

00:25:53: So sometimes Deepseek, I believe is considered open source because there was this one model type where they have published it in an open-source kind of fashion.

00:26:04: But yeah as i said so sometimes... It's a bit of a gray area!

00:26:07: So when do you actually consider it opensource or how much of their training pipeline do they have to publish that?

00:26:15: You say okay this is really open-sourced but um so this Open Source that goes into the European open source area, I think that's where we have the most popular options on these Sophie models.

00:26:32: And there are also a few others like Apertus from Switzerland or the previous generation of Sophia that was called Toykin but it wasn't rather small model too.

00:26:44: those i believe our proper open-source models

00:26:47: interesting.

00:26:48: so Again, experience from our client interactions.

00:26:54: I was surprised that there seemed to be quite a big openness towards Chinese models like Quen.

00:27:01: or you mentioned already Deepseek.

00:27:02: so it sounded for me that many companies is not an issue to use them if they now do the step.

00:27:11: but what you also mentioned yeah?

00:27:13: They start looking into alternatives often trying to run our models on-prem, especially to avoid open costs.

00:27:21: So it's really often more a cost issue and somebody comes as the benefit that we already discussed.

00:27:27: so what is your take?

00:27:28: Do you also see there an openness towards Chinese models?

00:27:31: or maybe

00:27:34: No, absolutely.

00:27:35: I mean if you look at the most recent evolvements in the open model space You see how?

00:27:40: The gap is really closing.

00:27:41: to the proprietary ones.

00:27:43: There was GLM five point one released and two weeks later glm-five point two.

00:27:47: so Chinese Open weight models but that are so powerful That they play on a same level like Gemini A flash of the most recently Gemini models or also GPT models our Claude opus four point eight or those models.

00:28:05: So there's really this, you can look at the website artificialanalysis.ai.

00:28:11: that is a project where they aggregate different evaluation metrics so like nine different evaluation matrix and combine it into one single number.

00:28:23: And their Claude Fable for example has an index of sixty.

00:28:28: GLM five point two has fifty-one.

00:28:34: GPT-Five.five has maybe fifty five or so, the gap is really closing really quickly and some people believe that next year there will literally be no difference anymore between what's open and what proprietary models can achieve.

00:28:52: I don't know...I think we'll only know next year but what would definitely have us?

00:28:59: that OpenModels next year would be much more powerful than proprietary models this

00:29:05: year.

00:29:06: Which leads us, I think to an important topic that companies need to basically classify and make up their mind what level of sovereignty they actually meet for these modules.

00:29:17: we see it.

00:29:20: if is true as you just said or really comes through with the case next year then I think the business models of OpenAI and Entrarchic are in pretty big trouble, but let's see.

00:29:37: That is why they push so hard on vendor lock-in this year!

00:29:42: They give you free tokens for seven weeks now with Fable or at the beginning the prices were so incredibly low that they must have lost a lot of money just to give people that... You know that kind of iPhone moment.

00:29:59: Ah, AI is equal to GPD.

00:30:02: so this was I think strategic too.

00:30:04: now be able to rise the price and some

00:30:07: yeah.

00:30:07: So basically people or companies have to understand what level do they need?

00:30:12: What performance?

00:30:14: And then They can decide also which for example Which model should use.

00:30:20: when we talk about sovereignty you Have those different levels but Also not every workload that you have or every use case needs the same level of sovereignty and therefore it's good to also Have a choice.

00:30:33: You can't use for example, For non-sensitive coding you can use the proprietary ones And still fall back two open models for more sensitive workloads.

00:30:44: For example what we have in systems We call this AI Foundation Services.

00:30:48: It is exactly kind of an API gateway That gives access to Models no matter whether they are proprietary and hosted on the hyperscalers or sovereign models, open models that are hosted sovereignly onto systems infrastructure.

00:31:03: I think having such kind of access point to easily switch models depending on what becomes available newly and what fits your needs.

00:31:13: this is... And with that we're already in a question how should implement software energy every day life?

00:31:22: No absolutely!

00:31:23: That's exactly what I was leading to with there.

00:31:26: So how do we do that now in practical life?

00:31:29: Maybe let me ask you a simple question, What questions should companies asked themselves or do they need to

00:31:34: ask?".

00:31:35: Yeah!

00:31:36: You can ask a lot of questions and an easy way to get started is thinking about the... In Germany i always say the five Ws but English one as the How.

00:31:49: so it's Where Who how, what and when.

00:31:58: And then on those five question words.

00:32:01: you can phrase questions across all levels of your technological stack.

00:32:07: so for cloud for example who operates my cloud infrastructure?

00:32:13: For data you can ask where is my data located?

00:32:16: or for the AI level What has been the purpose for which model that was trained, or what is defined architecture behind that model?

00:32:32: Or how would a certain use case change the behavior of the model.

00:32:36: And if you go to applications level so basically the use cases... You can ask when should I expect something changes in the pricing model application in having access to certain types of technology.

00:32:53: So I think those five questions just help, thinking about the different levels in such a deep way that you have an initial assessment where you stand and then you can always say okay so i don't know full control but that's fine for me because...of

00:33:11: reasons?

00:33:11: No!

00:33:11: I think it is useful at times.

00:33:13: yeah there are good starting points And

00:33:18: if you need a more, let's say professional assessment to systems has of course also assessments ready for that.

00:33:26: So hyperscalar risk assessment goes into that direction where we assess the cloud landscape off your company and how they hyperscolars are distributed across their various departments.

00:33:37: Our time is already up, but let me end this conversation with the latest and greatest which are obviously AI agents.

00:33:43: So what's your take?

00:33:45: How does building AI agents now change the sovereignty discussion?

00:33:50: when you have an agent do you have of course models behind that?

00:33:55: so I could say one model one agent in multi-agent landscapes yet to pick.

00:33:59: they also different models for different tasks on them all as not everything unique memory unique data source is connected to your agent, you need tools and all of that builds the agent harness.

00:34:18: That's how we call it.

00:34:19: yeah so there's a model and there's the harness for every single thing.

00:34:36: autonomy to do things on their own is super risky and a lot of companies including us are now building enterprise platforms for that, so in including governance into this.

00:34:47: So for example look at different agendic platforms.

00:34:51: there's these proprietary agent frameworks that help.

00:34:55: they're closely tied too For example an proprietary AI model And it worked very well with it.

00:35:01: but once you want to move away from the model you have to also move to another agent framework because it doesn't support swapping out models.

00:35:09: So just by defining your initial agents stack, You already define how easy or difficult there will be later on for you to switch vendors and the same is true of protocols that are being used For the Agents To Communicate.

00:35:24: There Are More Proprietary Frameworks Like I Think It's Called A Genetic Commerce Protocol which is also from one of the US hyperscalars.

00:35:36: And then there are more open protocols like, I mean most famous being MCP for sure and having just the awareness about all these different things that are tied together.

00:35:50: to bring agents to life not only where they're running helps too of course, by accident.

00:36:02: systems also builds an agentic AI hub right now to do exactly that in open ecosystem and vendor agnostic fashion.

00:36:12: I would call that sovereignty-by design...

00:36:15: Yeah!

00:36:15: Exactly.

00:36:16: And you build up your strategy and the architecture.

00:36:21: how you want to build agents.

00:36:22: then you should really think about sovereignty from the ground up at very beginning.

00:36:27: Then you will save a lot headaches or problems later.

00:36:30: I mean, it's also what people do in software engineering.

00:36:34: So back in the days you built monolithic applications.

00:36:38: nowadays You build modular microservice based softwares to exactly that To be able to swap out features and functionalities.

00:36:46: Yeah so we basically should reuse What we have learned them and applied here.

00:36:52: Maya!

00:36:52: That is a great conversation.

00:36:53: Thanks for much sharing your insights Thanks

00:36:56: for inviting me Of

00:36:57: course, and all I can say hope to see you soon again.

00:37:00: Bye-bye!

00:37:01: Thank You bye.