Data Sovereignty and AI: Keeping Business Data Local | Future Finance Forum 2026

Okay. So let's start. Uh our panel is um um called data sovereign in TI uh keeping business data wo. So wo can mean many different things. It can mean something that is uh here in Bulgaria or something that is on premise of your company or it can be something that is is within within the European Union.

Um and u I would like actually kind of to start with asking you what is wok for you? What does it mean to wok

Okay, I I'll start. Uh I think what defines wo is what defines one really big buzz word that is sovereign. It's uh similar. I would say it's uh defined by three categories. Uh the first one being transparency of the technologies used uh because you need to know um what happens.

You need to be able to audit why a decision has been made and so on and so forth. The second thing of course is the control. You need to be able to um run your product uh and not be subject to some geopolitical risk or some enterprise merging that could uh stop your u your AI um suddenly. And the third one of course is the cost predictability. So you need to you need to know uh how much it's going to cost you in terms of hardware and software.

So the AI runs and grows together with the adoption in your company not against it. So in in my point of view those are the three things that uh define the sovereign AI. If you have them you have

Okay. Um I could just add to what Christo was saying. It's uh also about predictability and reliability. You really need to rely on this infrastructure and uh what you're going to use and you should be also able to predict the future which is not always the case say when you're using uh somebody else uh resources.

Uh yeah, I think for me the definition is maybe slightly local or slightly different. I would say a model is local or AI is local if you control the sort of device it's running on. So if that's in a company I would see local as on premises. If you're an individual it is on your own laptop on your own phone. Uh to me any control that you give away from that would make it less local.

I think there's a point to be made like local can also mean local in the country. But if you ask me what is local AI, I would say on your own devices in the first place.

So um if a platform promis uses that the data is local u as a company, how do we kind of what kind of proof do we ask for? So how how how do we get a proof that data actually stays wok?

Okay. Um yeah I would say that u it's a matter of access. So if you um if you are able to um prove and make sure that the this AI uh doesn't have physical access to the outside world then you have truly local uh and truly um secure AI because you know everything else is a middle ground and every security check that you impose on it is like a walk but you know walks no matter how many of them you put on a or they can be compromised depending on how much time you spend and how many resources. So the solution is to remove the the the door entirely. I think and I even have a counter question on this.

Imagine that you have a a brick of gold and doesn't every reasonable person you go to a bank and say, "Can you please uh uh store it for me?" And they put you in a in their vault and show you all the securities and locks and all that stuff. But imagine that this vault is not in their basement, but it's in some uh dark alley in a questionable neighborhood. Are you in which case you're going to feel more confident about your your brick of gold? Uh obviously when there is no physical access to it

the vault itself

when you talk about local do we kind of what aspect is most important? Is it that the data is stored locally? Is the data is processed locally or is it that only locally people in a certain organization have access to it? Yeah, I think for me it's kind of every step, right? Because kind of what you were saying, it's never about the 90% that you do correct.

It's just what is the one vulnerability that you have. Uh, and that's what like hackers or anyone who wants to steal data is going to care about. So, I think like you were saying as well, keeping it completely local means you are exposed to the least vulnerabilities in the whole chain. Um, and I think it doesn't care that much or it doesn't matter that much what part you're actually talking about because you can do everything correct, but if you're still storing it as logs somewhere, then that's going to be the target. Uh, so for me, it would be completely local is the best solution.

Ideally, not connected to the internet. As soon as you use any outside provider, you're going to deal with like zero data retention agreements, but it's kind of trust, don't verify. Um whereas fully local is is of course your own security. You control every part of it. So yeah, I would say uh it's it's all the different aspects together.

I mean the also always a compromise between connectivity and locality and security. I mean there's a famous saying that ships are the the safest place for ships in the harbor. But nobody build ships to stay in the harbor. Same here with uh security and data. Even we believe the data are local.

Uh quite often they actually ask what Christo was saying they're sitting somewhere in the cloud or somebody else premises and whenever they claim about security it still is questionable. I mean still uh concepts of uh complete uh complete say hardware security I mean there's no such such devices yet in the meaningful prices. So there's always a compromise and question is how much we are willing to compromise for convenience.

So we have talked about localality and primarily of the data and who where it's stored, where it's processed, who has access to it. But we are now in the age of AI and uh so um and kind of people worry about W language models and the question is maybe kind of we should shift to that and ask the question do you need local models? Do you need kind of models that are locally developed that are locally hosted? um kind of does Bulgaria need to develop or kind of does in the individual countries or maybe the European Union level need to develop its own models or is it enough that we use and host open weights models uh provided they're hosted the right way? Who wants to go first?

Yeah, I can go first. I think actually developing a frontier model at this point is incredibly difficult and expensive to do and I don't know if any individual country or even European Union can do it. Like it's not a defeist or it is maybe a defeist attitude. Uh but there's a reason that anthropic and open AI are worth trillions. That's kind of the investment you need nowadays to build actual frontier models.

Um, so I would aim more at having your own compute at least having your own GPUs, your own data centers so that the open- source models which keep getting better and which you can fine-tune towards your own purposes and um locally run or run in your own country. As long as you have the compute and the data centers, you can always at least access those models. Whereas if you don't have the frontier models self-developed and you don't have compute, you're kind of getting extremely dependent on outside parties granting you access to these models, right? So I would say prioritize data centers and compute over trying to build your own models because I feel like that race is passed at this point. I mean I fully agree and couple of even couple of months ago probably it was been easier answer but you see the development of open weights model especially how fast they developing and how good they getting into even say kind of minority languages or small languages like Bulgarian because in the past was used to be a big issue.

I mean most of this model are trained say kind of in the bigger languages English, French, Spanish, Chinese and much less uh into Bulgarian and everybody who is using uh probably AI has experienced that they have struggling quite some uh with with this smaller languages but truth of the matter nowadays I think with the development of open weight models and all the hardnesses and all the things we can put around them I think 99% % they're good enough uh for for a lot of the the work we wanted to do. So,

yeah, I could only add that u my personal point of view is that developing frontier models is more like a scientific work and um being a scientific work u it's a totally different domain where you put a lot of money and uh for example a bank uh you as a business want the best tools available at the most economically reasonable price because through those tools you afterwards sell your product and services right so for um for a company or even for a bank they don't need localized models I mean for example they don't use u locally developed core banking system they they use the best available and economically reasonable core banking system and it's exactly the same with uh with the AI models. So as uh as uh Milan said, you need to have the compute so you can u run it yourself and then you can um decide which one is best for you. So how much intelligence you want, how much um you need to pay for it and you need to control it fully. That's that's the more important part for the businesses. But um so so kind of are you let me be a little bit provocative here.

So are you saying that Europe should stay out of this game because it's too late and uh you know kind of just be consumer of models developed elsewhere and kind of are there dangers in that? So for example there have been some models that have been uh you know by major companies that have said uh oh this model is not allowed to be used in the European Union. So uh isn't there some danger in that?

I don't think so. Uh because we are in a controlled world. We live in a connected world. So why America and China are leading in this area. It's from one hand side because of the compute but for sure on the other hand side because they have the most scientists which are developing those technologies.

It's hard for Europe to compete, but it's also not the playground for Europe. As I said, the the playground for Europe is how to make business out of this new technology and how to make sure this technology is put into work and uh developing the frontier models is entirely different uh play a battlefield. I would say I

mean I I can I can continue and we are focusing on frontier models when you're talking about say uh uh sovereign AI or locally hosted AI but it's much more than just a model model is just part of the big puzzle but also it's about chips about the whole infrastructure around the chips the servers uh the software and operating system which is running and then then we have a model uh and I do think that it'll be very difficult for Europe to catch catch up in conventional technologies which they've been developed uh quite fast and for many reasons good or bad Europe has decided to step down out of uh this classical uh computer development but there's areas where Europe still can actually lead if uh deciding to focus but I fully agree that um it's not too late and uh still many of the application which are developing uh for everyday use cider for companies for for public administration uh they have quite sensitive data and quite sensitive workflows and still is meaningful to make sure that you have some control on this one. So you have local infrastructure whether is a national European or at your home in your basement doesn't matter uh where you can be a bit more confident that the data and the workflows nowadays really about the workflow. So the models as you can see they're trained not that much on specific personal data but actually in the way the workflows people are doing certain jobs and this is also quite dangerous because the rapid development of this model is effectively we train them how to do better our job.

Yeah. So what I was saying earlier about not being able to catch up or being quite late I think that if Europe would try to do that it would have to throw a lot of money at it. But I wasn't saying it from a point of there's no need to. I think it is quite worrying if Europe is not able to like have their own frontier models. So I do get the point of even with slightly lesser models, you can do a lot.

You can do very much. But if there ever comes a time where like the US it's sometimes questionable how friendly they still are to us, right? And Fable the the access to Fable was cut off for people outside the US. I think it is quite worrying in a sense if we are behind other continents, other countries because um at the end of the day it's it's all about intelligence nowadays, right? Like armies are kind of led by how good the intelligence is that they receive.

Um so yeah, for me it is quite worrying that that Europe is this this far behind with AI models and I think we partly solve it by having more compute on our own. Uh but I I I'm a bit concerned that the US and China are this far ahead.

I mean I have read at some point some interesting uh observation that uh the US strategy about large language models is to try to have the best model while the Chinese strategy is different. They want the economically important model. So kind of one that actually kind of can be put in production. uh of course ideally the best one in the world but doesn't have to be it has to be good enough so that it actually does the job so what should be the Europe stance on that and kind of I can have put here another another kind of provocative thought that I have heard once that uh you know US innovates Europe regulates so is this the role of Europe to regulate this and kind of know I'm not dismissive here actually kind of there are many dangers in this technology and kind of uh actually kind of this can be also an important control of Europe actually kind of to put the things under control. So what is your thought here kind of what is the role of Europe here in this?

I think that's quite a big uh topic that have a lot of ifs if I may put it like that. Um getting back to the business perspective because what what he said about militaries and armies being led by or being empowered by AI is something completely different. But if we ground oursel to ourselves to the business perspective, I still believe that first the open models need around two to three months to catch up to the frontiers first. Second, uh their intelligence is to so high level right now that it's not anymore up to the intelligence of the model, but it's more of the who puts it best into operational use case. So put it uh and run it economically viable protecting the data and uh making use of it not just making demos of it or proof of concepts of it because that's also some of the uh most easiest path to to to to sidestep.

So yeah speaking purely from business perspective I would say that 95% of businesses don't need any more intelligent models than the one that are currently available and they will still continue to develop. So now the focus of the businesses and the companies and uh everybody in in that aspect should be in how to put this into work and the technology is here. U the the hardware is available. So that there is actually no blocker. It's it's only about uh doing it and as they mentioned uh in the previous panels um you need to start doing it as soon as possible.

Uh let me add something and I fully agree that uh for for many of the jobs uh the models we using too intelligent and too wasteful. I mean the world was doing well for hundreds of years without AI and six months ago without having the the best frontier models uh and it's really about say creativity of people who can control it and also using the most the best frontier model it's not necessary will give you the best result the same like if you take a room full of genius most likely not going to get say the best company or the best result out of it and it's really about people which know what and how to lead and ask the right questions. So this is still in my view still haven't changed.

Yeah, I I think my opinion is is slightly different on this because I think many much of the value that we create actually comes from sort of um advances in intelligence. So I think we develop new medicine when we have smart people uh developing medicine. And what we see now is that some of these models are getting smart enough to develop medicine by themselves or to solve math problems by themselves. Um, and I think yes, there are many like infinitely many tasks that are simple enough that current models can do them. But I think much of the new value creation from businesses often comes from a breakthrough where having a slightly smarter model that's just the 3 months like ahead of the other models can make quite a big difference.

And then if there is one country or one company that does have access to those models and one that doesn't, I kind of fear for the one that doesn't because it becomes hard to compete if you're even if you're using the older model very well. Um it's just really hard to compete against a much smarter model in some markets at least. I think many markets it matters less

and and it's still sorry for interrupting but it's still it's a matter of is it a question of just having a 3 months head start because if you never get to that intelligence I agree with you but having just a 3 months of head start uh I think this is something that is really really hard for any business to tap into it to to to make it work on on his advantage in a way that um he's better tweeted than his competitors.

But I I tend to tend to disagree with this notion because uh I'll argue say the biggest and the most successful companies in the world haven't been built by the best genius, financial genius etc. They've been say by creative people and innovative people and there still comes to this question about say how intelligent is intelligence how we measure how innovative is and yeah probably if it's in the area of drug discovery models can do a lot of brute force but even the example people have say recently solving Navia's equations it's still arguably a mathematician who actually prompted the model and this prompt leak out say to another models uh So I don't think this is critical still for me uh human intelligence is the one which is critical and if we give up everything to the model I think we're giving up everything because if the model can do everything then what's uh what's our role. Yeah, I I kind of agree. But I also that's what I'm afraid that we are going to that we are getting to the levels where these models truly are smarter than we are of some of these newest models. Um or at least in in specific narrow domains like math for example or like medicine.

Um so I think the human prompting the model is going to matter less and less. Um, I think it's mattered quite a bit so far, but I think often what we now see also in our business is that we're asking the models what the best decisions are or like what should we look into next. And we notice surprisingly annoyingly often that they're actually quite good also at broader views at like what are the next things to look at, what should be like our priorities and such. Um so I agree so far but I'm afraid that that intelligence increase of the models is not going to stop whereas we are quite limited but then you're giving up away again if you believe that model is more intelligent than you uh then in my view something is fundamentally wrong in the mindset because you what's what's the reason than why you run the business I mean just makes a a big uh AI factory which can run and give all the ideas and all the direction do you believe in this future and I still argue we are not yet there and I hope we will not be there soon.

Can I be provocative again? So uh we have been talking about models but maybe that's the wrong thing that we are talking about because in the last maybe year and a half the biggest improvements didn't come just because we have better models. We do have better models but actually the biggest improvements came from the fact that we have now agentic systems. Those are not about systems. We're now we are not anymore in the age of Loom.

we are in the age of agentic systems and um so maybe that's another chance for Europe and for many other players to catch up because then you can just like build different components that are part of those systems right so kind of and then things becomes much more decentralized and then maybe kind of going back to the topic of our panel which is about local data what does it mean for a data to be local when you have an agentic system that is calling all kinds of tools and it also on that.

I mean those those agentic systems and those uh tools uh it is a matter of definition where where you where you leave them to to work and what you give them access to touch. I mean it's the same right now without AI. So it's that there is a reason that uh for why the selling or the number of sales of mainframes is rising even though people are calling them that I don't know maybe for decades now and they they still sell them and they their sales even increase because everything is self-contained. Of course not everybody needs that. Not everybody needs his uh whole thing to be self-contained and to be local but for the companies that need it today, they will need it with AI in the same way.

So I I don't see a big difference in the operational model or in the in the way that those are being treated. But uh there's also another aspect of this I which is uh let's say I have my data local but what we discuss I'm connected and if I'm good at security or buy the best security software again questionable how good is this and talking safe here in the bus uh are my data more secure than if they in the cloud to the big company which part of their businesses actually making sure your data are secured. So we always kind of afraid of this kind of sneaky guys which want to read our data but quite often our data in our on premises are much less secure and now with the area era of a genti probably some of you already have read the story about say how agent hacked the hugging face and I think a few days ago been announced that in a summer they have hacked say Australian government and what is interesting this agent you think them local your data are local but these agents are communicating uh finding the way to communicate with with other agents and exchanging the information and part of this information say concerning your your security again there's been another example say a guy doing something on a on a big mainframe I mean a lot of data and then he give access to to the model uh to SSH keys which giving root access to the machine without actually thinking what's going to happen and this model actually exchange this information and suddenly other models are able to access not only the data but compute resources in the machine. So all the things start changing now. So kind of maybe you can go back to uh all this is not possible without infrastructure right infrastructure for data storage for compute.

So what should companies invest in? should they invest in or kind of what what we as as as a country as European Union kind of where should the investment go? Should it go into private uh compute uh within different organizations? Should it go into national infrastructure or should it get into European Union infrastructure maybe regional infrastructure part of the European Union? Um and then what should dictate where we should go because there's a little bit of a chicken in tech problem.

uh if you have the infrastructure if it's there then businesses can actually use it. On the other hand you know you there can be different kind of infrastructure and you might say well there's another way you might see first what is the need and then we build the right kind of infrastructure that the business actually needs. So kind of how should we address this from an economic perspective.

So so sorry if I'm understanding it correctly is the question should we see if there's demand for the infrastructure first and then build the infrastructure

or Yeah. So my thinking is there is currently almost infinite demand for compute for model usage um even if it's not necessarily local. So I think local businesses also have big demand for compute but let's say you build a gigantic data center now here um and it turns out that local businesses do not use AI enough yet to actually fully make use of it. I don't think that's actually an issue in the short term because you can sell the compute to a any number of buyers of like foreign companies or uh AI companies that would be interested in hiring that compute and because you've built it in a country that means you can always sort of redirect it later right so if it turns out that you in your country have more need for compute for running the AI models for yourself you already have the infrastructure for it so I think there is currently no more need to wait to see if there's local demand because there's so much demand that you can always sell it off elsewhere.

So maybe what is the one piece of infrastructure capability that Bulgaria should use to enable businesses to use this technology? uh I'll just build on also to the previous uh question and the answer is uh I I don't agree that we need to start building gigantic infrastructure without a demand because you need to realize that these data centers uh when they turn on they work 24/7 365 days per year and they consume even if they have no any customers they're going to consume like 40 50% of their peak power. So and the lifetime of such a data center is 3 years. So at the moment you put say a couple of billions the clock is start ticking and you can't wait for uh new customers to come and I think this is a myth that customers are going to buy any infrastructure at any cost. Customers are start getting quite picky uh because big companies are providing say a lot of compute and intelligence under under the price.

So you'll be surprised actually if you do the math and try to ask the business to pay the the the actual price even for the co infrastructure you realize that they start getting reluctant. So I think we need to build very carefully uh the amount of local infrastructure and European infrastructure and there should be a combination between say government supported especially if you start thinking uh if I want to run intelligence for the government administration so probably you need a centralized uh data center for this one also for education for for scientists for students etc. proper. You don't want all this um workflows they're doing and all the agents uh they basically to run the elsewhere. But then also there's a need for uh private infrastructure and private clouds uh which will then specifically ser serve the businesses and proper will be much more optimized in order to squeeze the maximum utilization of of this infrastructure though probably will be in a smaller scale.

So I'd like to take some questions from the public. Do we have questions from the audience? U there was a slid question. Yes.

Yes.

Okay.

I'll uh read some of the questions.

Yes, please.

Are we in an AI bubble or this time is different? One of the people asks.

Anyone wants to take this?

I think we have different opinions here. Maybe. Um I think uh I think this time is a bit different just because I think the models are actually incredibly useful now in many different areas. I don't want to give investment advice because I don't I wouldn't say the companies are fairly valued but I do think the sort of focus on AI currently is quite justified even if the valuations themselves are getting insane. I I do think say from technological point of view we're just starting uh this development uh and the same like com bubble I mean internet is still was probably the most important technology and the companies which been say highly valued the most highly valued company there still one of the biggest companies in the world uh the reason the the the bubble collapses is because say greediness of the people uh which want to invest in in any price.

So let's separate technological advances and the need for technology versus our greediness and the willingness that we going to jump on the wagon and we'll get rich doesn't matter what and doesn't matter what price

from financial perspective even though I'm not financial guy I really believe there is a bubble and in the same time I believe the technology is here the technology will last and the technology will change a lot of businesses and it is already changing them.

And the next question, I've seen it so many times, you've probably heard it a lot of times. Are you worried about AI replacing humans? And do you do you think that uh some control and restrictions on AI are needed?

Yes.

Yes.

Or or short answer? No, I think yes. I think some more control is needed and I think it is a legitimate fear.

In what way? What control?

Uh I think that's a lot harder but I think there should be some say of people and governments into how intelligent do we want these models to get? How much power do we give to just a few companies essentially?

Uh you just contradicted yourself about intelligence. Uh you wanted the the biggest intelligence all the time but then it comes to the price. I do believe that indeed we need to have a control but it is not too late. I also don't know uh but you can see the biggest uh CEOs of the biggest companies in the world are calling for some form of control. I think the problem is we don't know uh what is the right form and also how to make sure this balance between say making still making advancement versus uh being a bit more on the safe side as everything else that is that transformative it will eventually get regulated and uh who's going to do and exactly what's going to be the regulations this is yet to be seen and For the other part, no, I'm actually not scared.

I'm more on the optimistic side and even I would say enthusiastic side to see what we can achieve with this uh technology and through this let's call it revolution rather than being afraid of of it. It it will happen doesn't matter if we are afraid or not. So it's it's better to be uh to be on the train.

Thank you so much. Thank you. Thank you for taking the time to be with us today.

So yeah, thanks for the uh panelist and uh maybe that I'd like to thank them again and uh uh maybe as a conclusion uh when you think of local data, it's probably not enough to uh for somebody just to claim that the data is wok. you need to kind of have a way to kind of to know what does it mean kind of does it mean that the data is local uh kind of who who controls and actually how you can actually prove it right to the clients. So thank you very much