- 59 minutes ago
For years, the AI industry was shaped by a simple belief: the best ideas could come from anywhere. Today, building state-of-the-art AI increasingly requires access to vast computing power, specialized infrastructure, and billions of dollars in investment. Yet just as power appears to be concentrating, open models are accelerating, new infrastructure players are emerging, and governments are investing heavily to avoid dependence on a handful of platforms. Is AI becoming an industry only a few companies can afford to build? Or are we entering a new phase where open ecosystems can compete at scale?
Category
🤖
TechTranscript
00:00Please prepare to welcome Danila Stahn, CTO of Nibius, Thomas Wolff, co-founder and chief
00:08science officer of Hugging Face, and Vladislav Tankov, head of technology at JetBrains.
00:16Moderating this session is Amy Thompson from Bloomberg, so please welcome them to the stage.
00:43Hello. Okay, so we've got a really cool one today. So the prompt is, open versus closed,
00:51who gets to build AI? And we're doing this right in the middle of this enormous concentration
00:56of wealth and power around largely, I'd argue, the closed ecosystem. You're all representatives
01:04of different parts of the tech community, from developers to builders to the compute infrastructure.
01:11Looking at what's going on right now with these IPOs, with the biggest companies in the world
01:16right now, how do you see this impacting companies that want to build open-weight models? Does it
01:22change the economics? I'll start with you, Danila.
01:29I think it depends on what exactly the economics is, because obviously, having the gap in AI capabilities
01:37right now cremence and then, like, justifies this enormous valuations and business booming. Although,
01:47sustainability is still an open question. Like, these companies are immensely rich and large,
01:54but their business model is based on the fact that they are growing, not the fact that they are
02:01sustainable as a business entity. So I think we're still in a very early stage of exploration of what this
02:09actually means as a business and as a web community. And because of that, open source is probably even more
02:18relevant than it was before, because this is part of that exploration and part, like, we are not sure yet
02:25whether the closed models will be the source of everything, right? And it will drive this immense
02:32valuations and the whole change in how everything is done. But we can't discount the open source models,
02:38which are very capable, very flexible, provide a lot of not over-dependence on some entities, right?
02:49So I think they're definitely still in the game. Thomas, do you agree with that? Yeah, look,
02:57here is... I think this week was very revealing about that. And here is how I see this basically
03:04unfolding. We have an enormous concentration, like you're saying, like huge, right? And what happened
03:09is these companies start to have some hubris, which is they think they can actually decide what
03:14their customers use their model for, right? So yesterday I was talking with a biotech startup CEO,
03:19and he was like, okay, we used to be able to do, you know, all of these biotech DNA type
03:24of thing,
03:24where they run like foundational model DNA, they predict some things on cloud model.
03:29They cannot do that anymore, because they directly fall in this restriction, right? And so what
03:34happened there is you're a CEO of a biotech company who won't stop working, because
03:41now you're deciding that you're not allowed to use their model. So what they decide is basically they
03:45take their own future in their own hand, and they say, we're going to move to open source model.
03:49We take the new GLM 542 released two days ago, basically competitively open 4.8.
03:54We fine tune this one on our type of data, and we decide for ourselves our own faith.
04:01So I think this concentration is basically making open source even more relevant for many of the CEO,
04:07because they're really scared about that. Yes. Okay. Vlad, what do you think?
04:11Yeah, I would like largely agree with everything. So I think with open source models nowadays,
04:17it's even possible to disrupt some markets, especially if we're talking about some specific
04:23use cases. We've seen, for example, a lot of products playing with private models, like since
04:30I'm from JetBrains, it's mainly about coding, like next study suggestion, similar things, code completion.
04:34And it was possible to disrupt the market of existing private closed models with new open source. And
04:41in general, we do see that open source models and open waste models are becoming a lot better,
04:45basically in terms of how you can fine tune them and get a very good results on specific tasks that
04:51you need. And in general, so 30 billion models nowadays are completely different things than
04:56it was like two years ago or something. So yeah, definitely open weight models are very important
05:03nowadays. They are becoming even more and more important for the industry. A lot more folks in the
05:08industry starting relying on them. They just have to decide models from whom they want to use,
05:15basically. And yeah, with those models, you can even disrupt new markets and existing markets.
05:23Okay. I was saving this for a little bit later, but we're already talking about it. So let's talk
05:27about it. We can't have a discussion about AI in Europe this week without talking about Mythos and
05:34Anthropic and the ban that we all found out about last Friday night. I'm not the first moderator to
05:40mention this, but within our framework, I think it's really interesting. How does this bring the
05:47open versus closed debate home? Thomas, you already started talking about this a little bit, but can I ask
05:52each of you, what were your reactions when you saw this last week? I'll start with you.
05:58My reaction was it tracks. In a sense, everyone was worried that this is going to happen one way or
06:07another. And it is very consistent with what has been happening in the market, because
06:13given what you've described, that biotech company that essentially had to reinvent what they're doing
06:19with open source. But we've had the same push before that. Like, for example, Cursor is a very
06:26well-known example that was relying on Anthropic big time, like up to a point where rumor has it
06:33they've spent, I don't know, like 90% of their revenue to foot their Anthropic bill, which is not
06:38sustainable for a company and for a business. They pushed hard on open source. They took an open source,
06:44open weight model. They post-train it with their data that comes from their product, which actually
06:50is the mode. And here you are. Composer now is one of the most capable coding models in the world.
06:58Maybe it lacks the capabilities of Mythos as of today, but they've done it in an extremely short time
07:07span with extremely limited resources, given they're not a small company, but still like compared to the
07:13behemoths of Anthropic. Oh, relative. Yeah, yeah, yeah.
07:15They're a tiny scrappy startup, right? And they can do that and they can compete with that. So
07:21I think in a sense, this push for the whole Mythos push is, it's misjudged because in a sense, models
07:33are less relevant today. I don't want to say that in a way that you shouldn't use those. You should,
07:40you should always use the absolute best that fits your use case. But in a sense, and we see that
07:46with our customers, for example, like before everyone talked about the largest of the most
07:51capable model. Right now, we see that models are being like becoming rest-level because they become
07:57a building block for a use case. Some of our clients who actually employ agents in production
08:02at scale bring us a new use case every other week to run inference for them. And we see that
08:09the
08:09models there are different. They are always fine-tuned to their use case. They are always
08:13like custom tailored for the whole workflow orchestration that they have. But the model
08:19itself, the starting point is different because they are choosing which one will respond well to
08:24the post-training, to the alignment. How will that behave? What is the track like? What is the user
08:29engagement, et cetera, et cetera, et cetera. So the model itself becomes just a building block in a broader
08:34ecosystem. So despite me being personally offended by the whole mythos thing, I think first, this is
08:45something we've been going towards and it's hardly surprising, but I wouldn't be overly concerned because
08:51despite like this, this is definitely an expression of some form of control over some limited resource.
08:59I think open source is closing the gap and for the actual practical large-scale use cases, frontier models are
09:07less relevant than the
09:09custom tailored and fine-tuned and the lined ones.
09:13Thomas, were you surprised or were you not surprised?
09:17I was very surprised. Not, and to be honest, I fully agree on the risk and I think
09:24Dario is speaking the truth in the way they do that, even though the truth is, I think the way
09:32they
09:32formulate is always extremely in the doom way, very scary, which I think is not good in general for
09:41yet. But I think he really believes in the danger of cyber and bio weapon and I truly agree with
09:46that.
09:47What I really didn't like was the deceptiveness. So we had Fable and we were using it and
09:53it was actually giving us weird response. And now our strongest hint is it was what I had at the
09:59beginning,
09:59which was downgrading the quality of the response without telling you, which I think is a very bad
10:05precedent, which is basically someone is going to manipulate here the output of the model based on
10:11what they think you are doing or what they think you need to hear. I think that's an extremely bad
10:17precedent in general. I'm very, I mean, by nature and bias, I'm very pro-transparency. I want to
10:23understand which model I'm using, if possible, even what he was trained on, but at least I want to
10:28understand which weights are being used and they're constant and I can control that. So I really hated
10:33this part. I think it was very bad. And in general, I think that's just a highlight that you don't
10:40want
10:40to, for something that become core to your business or company, you want more and more to not depend on
10:47a single provider and you want to not be tied to their decision there. But I think to the point
10:55of
10:55Daniel, I think what is getting more and more important now is you have this basic level of
11:00commoditized intelligence, which is basically open source or like type of frontier model. But what's
11:07getting more and more important now is the data that you're using in your business and how you adapt
11:11this in your own business, being for physical AI, being for biology, being for material science, being
11:16for like a customer service agent, like all of that. That's where every business is as a key and you
11:23can
11:23decide as you use this data yourself to fine tune your own model and to take your own destiny or
11:29you
11:29give it to someone else to basically resell it back to you in the form of a better model, right?
11:35Which sounds
11:36really stupid. Like you wouldn't want to do that, right? Ultimately, you're kind of paying for your
11:40own tragic fate, I think. So I think it's really time for people to start training their own models.
11:46And the nice thing is in 2026, it's getting really easy to train your own model. Like there's a lot
11:52of
11:52GPU, there are great cloud providers, there's a lot of like people providing open source model to start
11:57from, the recipes are in the air. And what we've seen as well this year is that the model themselves
12:02can
12:02train good models. So you can ask a GLM 5.2, you can ask a cloud to just prepare a
12:07model that will be
12:09useful for yourself internally. Okay. Now Vlad, coming back to your point, Tom, on Fable and the
12:16performance, what are you hearing from developers about this whole situation, about Fable, about,
12:22you know, what is the developer community saying about this right now? Well, I think developer community
12:27indeed was also offended in a very big sense. Fable was taken out of the out of the models list
12:33very fast. So I don't think actually, there were a lot of folks who did act depending the production
12:38systems, hopefully on Fable 5, like in a day or two. But in general, folks were offended. And I think
12:45it was a great wake up call. So talking about even surprise, we have not been surprised at your brains
12:51by the fact that Fable was taken out. Because we've seen that in general, across all of the
12:57providers, there were, for the last half of the year at least, there were a lot more,
13:03they were tightening of the security. So we got a lot of new guardrails, we got a lot of new
13:09security protocols from the EA providers to be implemented. And in general, it was very visible
13:15that everyone is concerned about the security. I was very surprised that the decision was rushed so fast,
13:21because basically, when you see for a few months that you have to implement nowadays, like,
13:2620 pages of new security controls for the new announced or unannounced model, you would expect
13:32that folks would, like, think about a bit before the release. But once it was released, developer
13:38community, at least as far as I've seen, was very split in terms of whether it's actually such a big
13:44breakthrough or not, because, like, someone is developing frontend, someone is cyber security,
13:48they see completely different results, one of them are losing all of their quota and not get
13:53big results in the frontend, another one seeing, like, a breakthrough in cyber security. But the
13:59fact that it was taken so fast was a wake-up call, because a lot of folks started talking about,
14:04indeed, open source models, the fact that even small models that you can nowadays write on the
14:08unified memory are very capable, and they will stay with you forever until you have this desktop,
14:14or they will stay with you until you have a connection to some cloud provider where you've
14:18deployed those weights. So for a lot of developers, I think it become very understandable that for the
14:26production systems nowadays, you definitely cannot depend on the new model, because you don't know
14:31what will happen with that. I think the same questions would be about GPT 5.6, or whatever it would
14:36be
14:36called, or Gemini, which one, I don't know, 3.5 Pro. The same questions will arise, and folks would not
14:44be rushing that fast nowadays to support in the production. But for the development, in general,
14:52yeah, I think we will see also the move towards open source and towards open weight models and models
14:58fine-tuned for the use cases. Because at the same time, it's a lot cheaper, and it's a lot more
15:04controllable. So if you have GPUs in one of the cloud providers, or even under your desk,
15:09and they are running the weights that you have, you are absolutely sure that they will stay there
15:13until you will decide to move them. So was it this moment where everybody lost access to mythos,
15:21or the threat of it became real, that developers started thinking this way, or had the fear been around?
15:28I think fear was raising, because for example, zero data retention was restrictions were higher and
15:36higher for the new models, and folks who were working like directly with the models already knew
15:42that it's becoming harder actually to use new models, while having the level of security of your
15:48set of your data that you want to have. But in general, the moment when it was abruptly removed,
15:54it was a wake-up call. Because even folks from Claude were not able to update the applications,
16:00a lot of folks were getting just like Mises was disabled as an answer for Mises itself from Fable
16:05model. So yeah, I think it was just a bit of a harsh way to tell you that folks, not
16:13all of the models
16:14will be supported. And the fact that it was based on, at least as it is said right now, on
16:21the
16:21nationality decision, like whether you are in the United States or not, it was also a bit harsh.
16:27Thomas, how do you see this continuing to develop when you're thinking about how governments are
16:31thinking about sovereignty and national security when it comes to AI models? How does this, what's the
16:38logical end point of this? Is open source a national security threat going forward? Help me think
16:48through kind of where this is going. I mean, Hugging Face is a global company, so we're operating
16:56everywhere in the world. But there is certainly one narrative which says this is all part of a very,
17:02very long-term scheme of basically regulatory capture, basically scaring politics enough that they
17:11decide to forbid every models but a couple of companies, right? It happened a couple of years ago,
17:16there was this discussion about you need a license to train a model. And the idea was of course that
17:21this was pushed by the two companies that would get the two licenses, right? We think this would
17:26ultimately be very bad in general. And I mean, it's maybe a bit strong, but very bad for humanity,
17:32I think, to have just one or two corporations that basically own the whole intelligence layer,
17:37right? This is, this is a future where everyone, for instance, in Europe, where we, we have Miss 12,
17:42well, but we have, we have no really like frontier closest model at the moment. And it's flexible.
17:47Everyone in Europe has to go and to send, and not just the data, but that's, that's, that's the
17:53technology that's going to replace part of the workforce. Basically, everything would go through
17:57some data center from control by two company. And if some of these companies, like they start to do,
18:03decide what you can use or not, based on the profile they have with you, maybe in the future,
18:10the collaboration, let's say that they work with two biotech companies, there's two big biotech
18:14companies, but they have a contract with one company. Why not decide that basically this model
18:18is only available to this company, right? So, so if there is a future where, in addition to have this
18:23kind of, you know, like, like uncontrolled decision process on how you can use the model for,
18:30you are also very limited in a choice of model that you have, which is you only have two models.
18:35So basically you're trapped, right? This is, this is a future we, we, we really don't want to see
18:40happening at Hugging Face. We think this will basically kill innovation. This was like basically
18:45kill all the inventiveness in all the entrepreneurs who, who take this idea that we have like very
18:51powerful AI and we can invent new use case. We can create new world models. We can create all of
18:56that.
18:56This will die. I think that's, that's a real call, call to action. And there is of course,
19:02like lobbying around this. We, the, the problem of the open source committee as always is we are
19:09less powerful in terms of capital that we have. And we're also less united because we're less
19:13concentrated. It's kind of, it's kind of the cathedral on the bazaar. It's like David against Goliath.
19:18So we're doing what we can. I think it's a lot of education. It's basically trying to,
19:22to see these events for what they are and not, not trying to pretend anything else than what we
19:28see happening. Now, when it comes to Nebius, you've obviously got more demand than there is supply.
19:36How do you think through who gets that supply in this context? Is there a risk to having,
19:42you know, just two models, I guess, that dominate the industry and that take up all of the GPU and
19:48take
19:48up all of the compute? Um, uh, is the company thinking strategically about, you know, leaving
19:54space for open source? This is a sudden move. Uh, no, well, first of all, we are, uh, definitely less
20:06impactful in the whole market than the OpenAI and, uh, and Tropic are. But, uh, in a sense,
20:12this is the problem that we have to figure out every day. So like we've, uh, just had a conversation,
20:19like we have regular meetings when sales team gathers around, lays out all the potential
20:25partnerships and sounds like, okay, who gets the GPUs? Like, of course there's like, uh, pricing,
20:33et cetera, et cetera. But in the end there is still the demand is higher. So for us, it's important
20:39not
20:39to just get the highest price, but to make sure that ecosystems thrives because Nibius at its core
20:47is, uh, like the independent player, right? So what we are trying to build is like this independent
20:54layer of infrastructure that is not affiliated with a huge company or the company that can like,
21:00so we need to enable the broader ecosystem. So sometimes you would like postpone the contract
21:06with a big player in order to enable a smaller startup or a smaller lab simply because we want to
21:13grow with them and we want to see them strive because if the ecosystem has more variety within it,
21:22independent players like ice have a chance. If the variety dies, uh, we are reduced to just being a
21:31GPU reseller from Jensen to Dario and I'm pretty sure these guys are capable of talking to each other
21:39without us. So like our whole existence is based on the fact that it's like ecosystem has variety,
21:47open source tribes, like there, there, there's everything like let all the flowers bloom.
21:53Um, and in Europe, uh, coming back to mythos, we saw an immediate backlash. Um, uh, a lot of people
22:01in government saying, you know, this just shows, you know, we can't rely on big American models.
22:06There's been a lot of rallying around generally. We've mentioned Mistral a few times, um, uh, which is
22:12an open way company. Um, uh, is there a risk though of, uh, this, this emphasis on sovereignty and
22:21locally grown tech, um, gets overdone and then we're, we're constrained in another way. Um,
22:28especially as you're seeing things like the commission coming out with, um, frameworks
22:32for sovereign cloud and frameworks for, you know, AI and, uh, how they're going to regulate its use.
22:39I'm coming to you first.
22:42Uh, I have a very radical point of view on this.
22:46Oh, great.
22:46Like, uh, first coming to your point about, remember those trade restrictions for training
22:52models, et cetera, et cetera, et cetera. We all know who those restrictions were aimed for, right?
22:59Those were China and the likes. Do you know who is the leading nation after us in closing the gap?
23:07It's like a state of the arts, uh, AI China. So not only those restrictions did not diminish their
23:16capabilities, but it's actually empowered them to build around those, to get creative. The whole
23:23deep stick moment, the whole, like the restrictions never get to anything good because they are very
23:30short term and people being, people will find ways around people being, people being creative,
23:36being like resourceful. They will manage, uh, to achieve their goals anyway. Like it's, it's, it's a
23:43hard, uh, I don't understand the reasoning to, to deny something if in principle this can be
23:51re-implemented. So, and coming back to Europe, I think the whole sovereign play is a little bit off
23:59with the, with the goal, right? The current narrative is more about, we need our independent
24:06infrastructure. Let's build data centers. Let's put GPUs in place. We need our foundational models,
24:14right? Let's find some smart people who will do that. But what we actually need in Europe is,
24:21again, a broader ecosystem, right? We need, uh, clients of those models and startups. And if there
24:29is demand that is, uh, nurtured, facilitated, then you will see people getting up, responding,
24:36capital being attracted, infrastructure being built, but building infrastructure in a green
24:42field where there is no demand, no one is actually like lining up to procure those GPUs. Like we've had
24:49that before. And, uh, and the good example is Middle East. At some point in time, a couple of
24:56years ago, they invested heavily into build out the infrastructure. Now there is a real struggle
25:01because those data centers are just sitting in the desert without being consumed all the,
25:05because there is no layer of growth of smaller companies that will continue to consume those
25:12resources, to use those models that are running there, et cetera, et cetera, et cetera. So I think
25:16the sovereignty play, uh, while of course, like building roads is important, right? This is a
25:24natural function of the government, uh, and what have you actually like making sure that there are
25:30enough cars driving those roads is in my opinion, way more important. And I am like very disappointed.
25:38I don't see that narrative in this current conversation about the sovereignty. I, I don't see this
25:43narrative. We need to enable a broader ecosystem. No, it's usually talking about the simple problems
25:49of building roads. Thomas, you look like you agree. Yeah, very much. Um,
25:56I think, I think it's hard to, it's hard to separate the discussion of future of AI from the discussion
26:02of
26:03both entrepreneurship and reconversion of company. And basically being able to create new AI native
26:08company, being able to convert, you know, your former large company or smaller, but like who has been
26:15around for some years in an AI infused company. I think all of this goes together. And so often my
26:21recommendation for AI is kind of to point is, uh, we need, we need also to help entrepreneur because they'll,
26:27they'll use this in Europe. They'll want to actually use this AI product. So you, you, you need both sides.
26:33Like sovereignty has this, um, infrastructure side, which is like, we need data center here for sure.
26:41But also this usage, we need adoption as well here. And this part, the like regulation part,
26:47like for instance, if you get the AI models, yeah, three months later in Europe, because they are not
26:52allowed to be used, that's, that's very bad. So it can be very damaging because people just want to use
26:57the latest model. So they, or they want to explore something because they're scary. They might fall in some
27:02buckets of regulation and they want to try that. So I think this is super important. Um,
27:08in terms of reading materials for the audience last year, there was one last week, there was one,
27:12one blog called Europe 2031, which was a vision of what could happen in the future of Europe.
27:19If we don't do anything, basically, if we just go quietly, uh, in, in the idea that everything will solve
27:25it magically, I think what it's interesting really is slightly, uh, slightly scary, of course,
27:31but I think it's good to read. And, and some of the key points there are, um, first that we
27:37need to
27:38really invest in infrastructure and each range from data center and compute and energy, but also to some
27:44of the very strong strength of Europe, which are around, um, like, like manufacturing, engineering,
27:51from chips. Of course we have ASML and everyone knows about it and that's a great thing, but also
27:55we have all these, like, car manufacturer, we have all these, like, things that we know how to do.
28:00There's a lot of knowledge and we should try to lean in that as maybe the next step of AI.
28:06I think,
28:06and I take in face, for instance, we've been working a lot on robotics and we think physical AI,
28:11something that's going to become also very important. So I think there is a lot of option for Europe to
28:15stay
28:15relevant. Um, but it's very important to, to start acting now. Like in the second value,
28:22people talk about this. We know that's closing for action. I think it's quite true. Like,
28:26if we just wait five years and we don't do anything, um, I think it's just going to be too
28:31late.
28:32Okay. We're down to our last minute. Um, uh, quick lightning round, coming back to the prompt.
28:38Uh, if you were going to advise an entrepreneur in the audience who's just starting out, they want to
28:43build the next big AI company quickly. Uh, where should they focus? What would your advice be?
28:50Start with Danila. Product. Don't care about the model. Take the best one. Distill or think about
28:56open source later. You need the product, not the model. Thomas. That is a good advice. Um,
29:06I would say specifically for your thing globally, like tackle a world market, uh, directly. And,
29:13and of course, thinks about what is an AI native company in what you want to build. How can you
29:19be
29:19a native small team, a lot of agents, like a lot of things like use, use that for sure.
29:26Yeah. I would say you should first build indeed very good product with whatever you have.
29:31And then only think about like performance, cost effectiveness and so on. So we see that the
29:36AI race is really fast and very likely once you have a good product, you will be able to find
29:42the way,
29:42how to make it work for the customers and yourself. So good product with what best what you have.
29:48And then only think about cost effectiveness and performance. Great. Great. Thank you very much.
29:53So good advice Stephens. Thank you.
29:54Thank you. I appreciate it. Thank you.
29:56Thank you very much, goodälke. Thank you. Thank you.
Comments