Nebius Group N.V. Class A Ordinary Shares 0 Earnings Call

Let's get started, I will go now to the console because we're not going to do slides today. I will just go to the console and go over the new features and show you how they are in the new console and how we can use them. The first one is Nebius Echo. I'm not sure if you heard of this before or if you used it before. If yes, please put some pluses in the comments. Nebius Echo is our AI agent. If you've been using AI coding assistants before, you've probably gotten used to describing what you want and then asking for it. This is not normal chatbot. It actually answers your questions, but it does something more than that.

For example, if I want to create a VM, I will tell it, "Create a VM on a single H100," and then I will delegate creating the VM. Let's say I'm new to the Nebius console, and I don't really know where the creating the VM workflow happens. I will just need to ask Echo here and then wait few minutes. Not few minutes, few seconds for the answer, and let's see what it's going to give back. It's taking some time. I hope it's not a few minutes. Let's try again. Works hard.

Can you tell us in the comments if this is the first time you've seen Nebius Echo, or anyone has worked with this before? If it doesn't work, we need to refresh. We'll keep this in the background and then, Narek, maybe you can tell us about Nebius Echo until it works. I will not remove this.

Yeah. Actually, I use it in a little bit different mode, to be honest. A console is one of the ways to use it. Another way, actually, you can use Nebius Echo from CLI. You can write Nebius Echo or Nebius Ask, it knows everything about your tenants, knows everything about your infrastructure. I actually created a skill for my cluster in Cloud to go and refer to Nebius Echo as a professional expert who knows about Nebius Cloud and Nebius Token Factory, like Echo knows about the Token Factory as well. Yeah. Oh, that's very cool.

Yes. For example, I recently was doing a good experiment. I was just asking, "Cloud code, can you just run this workload in Nebius?" He went to Echo. He found some capacity, like Nebius Echo can also query available capacity. Echo queries where he can find the GPU, which fits to the amount of the workload that I wanted to run. Found this GPU, and run a serverless job on top of it. Actually, this is a vision of why we're building this Echo. Basically, it's not only an interface for human, it's also interface for other agents, like to use our platform. You can use it by console, like how Marwan is showing it to you.

Yeah. Yes, please. Sorry to interrupt.

I am just trying to get it to work very quickly, I just want to say that it does not do anything before requiring the approval. Here it gives you back all the details about what you need to create, I will just click on Accept and see if it actually creates it in the console. Narek, please continue. Yes. Actually, this is its first implementation, how to use it.

Our vision under Echo is, in general, that this is an AI that basically works as a very good trained customer architect or customer support that knows everything about you, like your tenant. It knows everything about our documentation, our platform, our code. It basically represents our global vision. Our global vision that we work with Frontier AI customers, create the best platform for them, we generalize knowledge for everybody in the world. This is basically continuation of this vision. Our Frontier AI customers are very lucky. They have solution architects working with them who knows everything around platform. Like smaller ones, not everybody has this privilege, we want to democratize this privilege.

A lot of knowledge of value of CSA is that it is people who knows everything about customer context, about internal documentation, about external documentation. We are collecting this into Echo, we will continue to make it more reliable to build blueprints. We also understand that AI is super creative, I would say. That is why it is not reliable sometimes. We want to make reliability of the answers that it provide much better. What I mean is that, for example, you wrote, "Just create me a virtual machine." It is very easy to reproduce it. If you go and ask, "Please create me a cluster for the training," it can be very creative in understanding what does it mean a cluster for training.

We are actually building, it will be next releases, it is already in the block of Echo, but we will build pipelines and blueprints which will be reused by AI multiple times. We will basically build best practices. These best practices are baked in our solution architects repository. We will bake in them, we will give them to you will not only have professional helping you, we will also have blueprints, which basically will be reused by multiple customers, it will allow us to provide a better user experience and for you to get better results. This is basically what we are definitely building. Next stage will be, we want to allow you to build, not only to build by host end-to-end applications, not only for AI, but if you build something with AI.

Basically, I believe future will be inside solo founders. Solo founders will build AI-enabled companies. AI-enabled companies, they will build software for them, not only buy software for them. We are preparing our platform to be the place for that. It is definitely future. This is important building block for that. Very reliable agent who knows everything about Nebius and who knows best practices how to build on top of Nebius.

Yeah, interesting. By the way, the VM, Nebius Echo says that the VM has been created successfully. We will just check if it is. Here is our single VM that we can check. If we want to do anything with it, or create something on top of it, we could also go back and ask Echo to do that. One curious question from my side, Narek. Do you think that in the next few months or even in the next few weeks, that this could even become this standard type of communication that users and developers could communicate with cloud environments and especially with Nebius Cloud?

Good question. To be honest, I am not sure about short-term. I think it will depend on the coverage of the scenarios and value. Definitely in short-term, you as users, you will be able to debug problems of AI. This is what we are enabling, AI debugging, because a lot of things, there are sometimes problem in platform, of course, when you have problems. A lot of times there are problems which are related to the configuration, and usually the problem for people is to, okay, my workload is not working fast or reliable. I don't understand why. You have to go to the monitoring, observability, see the metrics, do decomposition, all of this. We will offload this part. I believe UI interfaces will be more used to do this one. We will enable building something complex. Then more and more capabilities to operate- to day two.

Right now, I believe the best thing that you can do with it, you can definitely get reliable answers about documentation, your quotas, everything around your tenant. You can also start to use it with debug of the problems. It will also help. You can also query capacity that is available. For future use, we will cover more and more use cases. Then we will just do them as fast as possible. In reality, I recommend you, if you have some questions about any question, just go to Echo, ask this question, give us feedback. With more and more usage, we will also prioritize more and more use cases, which makes sense.

Yeah. That's true, that's your invitation, everyone, to go try Echo and come back to us with the feedback.

Yes. Again, you can also use it not only by console. If you have your Claude.

With CLI. You can just add CLI, just install our CLI tool and ask Claude to ask our CLI if we had anything related to Nebius, it will magically go and forth.

Yes, thank you, Narek. One thing that we're going to shift towards now is another thing that you can actually create with Echo, which is a SkyPilot server. For anyone who hasn't used SkyPilot before, just a reminder, SkyPilot is an open source project that makes it much easier to run AI workloads on the cloud. The concept is instead of manually provisioning infrastructure, you just describe the resources you need, and SkyPilot handles that for you. Until now, until before this release, using SkyPilot on Nebius meant deploying your own SkyPilot API server, which was not particularly difficult, but it's still yet another service that you needed to manage yourself.

Now with this release, how it looks is that let's go to AI orchestration, and in SkyPilot, Nebius now mounts the control plane for you, and you can create a managed SkyPilot instance in just a few clicks. You can choose a name, and choose the platform you want your SkyPilot instance to run onto, and the preset, the CPU preset, and when you click on deploy the application, it's just deployed in one click. Narek, anything you want to add on this managed SkyPilot? I know that there was some asks for this. Can you tell us a bit more about it?

Yeah. I'm very happy that we finally created fully managed version of the SkyPilot server. It's also put in using AI orchestration part of our navigation. What does it mean? It means that actually our customers, you, they run AI workloads in our platform in multiple ways, to be honest.

Some of them run in Separator, some of them run in SkyPilot, some of them run in Anyscale, some of them run in Serverless. This is actually part of our vision. We want customers to build AI the way they want to build. We don't want to say, "Customers, this is the way how you want to build it." We want to say, "Build it the way you want. We will provide you best primitives, best infrastructure, best low-level capabilities that needs to be done." We invest a lot to make the best training and inference in our platform in lower levels, and it benefits all the platforms. This is first thing that's important.

SkyPilot. We want to provide more and more native integration with tools that are getting more and more demand and traction. We see that SkyPilot is one of the next big schedulers to run AI, to run AI training workloads. It's not, of course, as popular today as Slurm or Kubernetes, but I believe guys created very good products to run multi-cloud AI workloads. They combined two important capabilities. First is aggregation of scarce infrastructure. SkyPilot actually can be used to work in multiple clouds. In our cloud, it can be used also to work with multiple data centers. It's the same problem, to be honest. Orchestration problems are the same, and SkyPilot solves them very well. Second thing that SkyPilot solves very well is AI-focused primitives that you can use, because you can aggregate infrastructure very easy with multi-data center, multi-cloud Kubernetes.

It's again, Kubernetes. It's very good interface for DevOps. DevOps love it. It's not as friendly for AI researchers, so researchers want something simple. SkyPilot combines points of them both world. That's why we believe in this product. That's why it's getting traction in our platform. Customers use SkyPilot as all Customers who use only Nebius use SkyPilot. Customers who use Nebius and something else use SkyPilot. I think good example is public document of Shopify. They use SkyPilot to run data preparation GCP and then run training in Nebius. This is how you should use it, to be honest.

We provide you a good way to host control plane on the SkyPilot. We don't charge for this control plane. We charge only for the compute that you use in our platform. Actually, you can use Nebius as your main hosting platform for SkyPilot and operate in multiple clouds, multiple data centers using kits, and yes, running your workloads. They recently announced reinforcement learning primitives. They announced, I think in roadmap, sandboxes primitives. I believe it's one of the ways to do it better than in Separator, for example. There is adoption problems. We know that Slurm is known to every AI researcher of the world. SkyPilot is still something new. We will continue supporting everything that is getting adoption in the market. This is our main vision and strategy. Yes. Good times ahead. We're going to leave this our SkyPilot instance provisioning, and we're going to go and shift towards a bit of a different bracket, which is security and governance.

One thing that we added and included in this release was a feature called Budget. As the name tells, it's about budgets. Please ignore these numbers, these are not real. The budget concept, as the name says, it's one of those really simple features, but once you start working with them, you see how important they are. At least in my case, it was always a problem if you forget some machine, and then you check the cost, and then that's not what you want to see. The budget concept is that you set a budget that you want to spend.

You set a spending limit either for your team, or for a project, or for a product, and you set a time for it. Let's suppose I have my budget, one that runs monthly, and it's either for all the usage of my Nebius account, or I have conditions for it. For example, let's say I'm using a new product, like a GB200 or a B200, and I want to set a budget only for this product monthly. I would say I want to spend not more than $1,000 on it. What happens is that when I set my budget and I add an alert, if I put here 50%, when this threshold will be exceeded, I will get a notification to the email I will add here, obviously.

I will get a notification to my email saying that the consumption has exceeded the budget that you set for the month, or the threshold that you set for the month. It's worth noting here that Nebius and the console will not stop your resources, but it will just alert you and send you an email so that you know if you decide to stop or keep it running. Narek, what can you say more about the budget feature?

I know that it was demanded for some amount of customers. I believe it's just something that we must have built before. We just didn't have enough time to build it before, to be honest.

Yeah. In general, we're building stack for enterprise readiness.

What I mean by enterprise readiness, we are enabling our platform to be used by large companies where there is multiple teams operating simultaneously within infrastructure, and they need to work together in some secured governance environment. We're doing a lot to create features for that, like budget. We create special features that allow to integrate with current environments. To be honest, if you have your FinOps tool to operate budgets, you can export all the billing data to your FinOps tool. We have this feature for the export and- From Nebius your systems.

Yes. This is basically something that is needed for any mature org which has multiple customers. I believe it is already in. It is good that we finally shipped it. We will ship more and more features, which I believe we have to be shipped before. As you know, we are a software company. We have 10X backlog of the capabilities. More and more features will come. Please come to our support, to our managers, to your sales managers, ask for some feature requests. It also helps us to prioritize something that we definitely need to build before.

Yes. Yeah, definitely. We will stay under security, and we will go to the second feature, which is Key Management Service. Everyone knows that by default, your data is encrypted in the cloud, and especially on Nebius, but many organizations need more control than that. They want to manage their own encryption keys, decide who can use them, rotate them regularly, or even destroy them if needed. This is why we introduced Key Management Service. With Key Management Service, you can create your own customer managed encryption keys, CMEK, and use them to protect your workloads. There are 2 types of keys that you can create here, either symmetric key or asymmetric.

Symmetric is a single secured key to encrypt and decrypt data, and asymmetric gives you a pair of keys, public and private, and it makes it useful for more encryption for things like authentication, sign-in, or secure communication. You can just define your rotation period. It is the period that you want the key to rotate, and then create the keys and use them afterwards. Narek, what I want to ask about this feature in particular is that, is it mostly coming from enterprise customers, or is it something that you are seeing across the board of the product team?

Good question. To be honest, explicit request definitely is coming from enterprise ones, but when we released it, we got some positive feedback from AI natives. I believe it's also necessary feature for us to be trusted provider for secure data. It's KMS exist in every hyperscaler.

Yes. KMS is integrated to a lot of security tools in open source and commercial ones.

We will go and do these integrations as next steps. What important is that, again, it's, I believe, first step. Our mission is to allow using customer encryption keys. Right now, we use hardware security modules to create the keys, but these hardware security modules are Nebius managed, and it will allow you to gain personal encryption. It definitely gives another level of security in encrypting the data in our platform. There is more levels to achieve, and we are definitely going there. We will definitely, I believe by the end of the year, maybe a little bit later, we will enable customers to bring their own keys. In general, it will allow customers to store the data and be sure that nobody except customers will see this data.

What's important is that it will use the same service. You can start using this service, encrypt this data. When we're going to have bring your own keys, you will just change settings in your keys.

You will get more secure level, definitely.

Interesting. Okay, before we move to the next feature, I just want to highlight that if you guys are interested or curious about our security features in general and our compliance guidelines and the key security features, you can go to nebius.com/trust-center and there is every information about our compliance guidelines and security features. Our next feature is actually about storage. I feel like in every product release we talk about storage. It's something that everyone is needing. Here, what we introduced is intelligent object storage. For those of you who are familiar with using storage, we used to have here standard and enhanced, but now we have intelligent.

The concept of intelligent is pretty simple and straightforward, is that let's say you have a large data set, for example, and you're planning to use it next week, but it doesn't happen, then you forget about it, or there are other priorities, and it gets buried somewhere in other workloads. What happens is that intelligent storage works in tiers. It's like cold and warm tiers. Your storage first goes to warm tiers, and if it's unused for 30 days, then it goes to the cold tier and you actually pay less for it. If your storage is there for more than 30 days unused, you start paying less for it until you use it, which could cost you less money eventually.

Narek, what I want to ask exactly about this, because definitely and very simply, you can see the value in the cost efficiency of this feature. My question is more about is there any added value more than the cost efficiency of intelligent storage?

No, actually main value is, there is cost efficiency that you don't need to configure by yourself.

What's important? Important that if you train models, if you run simulations, particularly when you work with synthetic generated data, industries like physical AI, or customers who train multimodal models or create, run multimodal inference, or physical AI inference with both models they get, they're going to another level of the scale from the storage capacity. Scale is tens of petabytes of data that need to be hosted. This scale, ideally, everyone wants to get unlimited fast storage.

Yeah. This is ideal world.

The problem of this ideal world that it costs a lot. This scale of tens of petabytes is usually when cost of the storage started to hit cost of the cluster. It started to become visible. In smaller scale, cost of the storage is negligible. It's hundreds of terabytes, single digits of petabytes. If you run inference, usual inference, you just host model. You just need 1,000 terabytes. If you do training without simulation, without reinforcement, without synthetic data generation, it's also single digits of petabytes. It's not negligible if you compare to the price of 1,000 GPUs that you buy. If you go to tens of petabytes, it's becoming serious, a serious part of the cost, serious part of the TCO of the cluster.

The only way to cut the cost here is to have multiple tiers of the storage. To have caching tier, to be honest, that's why we also introduced local disks to have a high performance level of the storage, which you use to current data. To have capacity layer of the storage for the data that you don't need to access frequently. With this, you can get your cost back to not negligible. We working a lot to make this happen on the scale. Intelligent tier is definitely, again, one of the steps that we are making here. If you use storage in serious scale, when you see that your storage cost is hitting you need to think about tiering. This is way to enable your tiering is very easy way.

This pattern 30 days is something that we found out from working with Frontier AI Labs. It's some working pattern, I believe. It's not going to be the only option, but it's for you to save much more money if you go and run scenarios that require synthetic data generation. You can use this expertise. You don't need to be Frontier AI Lab to use features that Frontier AI Labs use.

That's true. For the features, that would be the main features that we will cover today. You will have a poll pop up in front of your eyes to ask you about whether any of these features are useful for you and for your future workloads. We're not going yet to go to the Q&A because I have another thing that we have launched within this release that our dev rel team is pretty proud of, which is the Nebius Builder Program. I have seen a question before about how to get credits. I couldn't think of a better way to get credits for Nebius platforms rather than the Nebius Builder Program. In simple words, it's a program in which we are collecting Nebius Builders and making you part of our journey and of our community.

It's free registration, by the way. By registering to it, you have a lot of benefits, including credits, including office hours with Nebius engineers that we're going to launch in the next weeks, and also access to our Nebius community that we're growing. There's also another thing that we launched within this release, which is the Nebius certification, which through the Nebius Builder Program, you're going to get for free. Nevertheless, if you want to go and get certified, I see a lot of people are getting certified even though it was launched less than two weeks ago. We already have two certifications live. Go ahead. If it resonates with you and makes sense, get Nebius certified and brag about it in social medias and tag us. We are watching all these certified Nebius Builders over there.

There is something I want to discuss with you, Narek. There has been some ups and downs about the credits of the Builders Program. Do you have any information that you can tell us around that? I know a lot of people in Discord are asking about that. I would really appreciate some clarity on this.

That's a good question. Our mission is being, continue to be, to enable everyone to run AI, not only Frontier AI Labs, but small customers. This program is about it. We are doing some efforts to sponsor you with credits, sometimes with compute, to help people build some new AI models, applications, anything. What's important that it's a challenge. It's a challenge to build such program. We will continue to build features to allow to work in a scale. Sometimes we will disable these features because there is people who have good intentions to use free compute in mind. There are people who don't have good intentions to use free credit and compute in mind. Actually, we are continuously investing to allow people with good intentions and stop people who doesn't have good intention to use our platform. We will continue doing that.

It's a journey for us, but I believe it's our mission. What does it mean? It means that you will have some updates about you will have more credit to use. Sometimes it will be less credit to you. Sometimes you will have free tier enabled, sometimes disabled. It's work in journey for us. We are committed to be in the journey, so be with us here, and just follow our programs, follow our webinars, workshops. These are places where you can get more and more data from us, more and more news for us to build.

Yeah. Please, guys, reach out to us for the Builder Program or for credits if you have any issues. I see in the chat some issues. Reach out on Discord, and we can figure something out together. Before we go to the last Q&A section, I just want to ask you, Narek, about what's ahead. What are the plans for the next release or even for the next half of the year?

Plans. We have so many plans, to be honest.

Give us the best ones.

A lot of plans. Yes, a lot of plans. First big thing that we will be one of the first in the market to provide GB300 from the platform. It's a big thing. GB300 in general is very interesting platform. It's only right now available to Frontier AI Labs, hyperscalers in the world. It's not available for smaller use cases. I mean, smaller is order of magnitude of customers who can pay tens of millions of dollars. It's funny to say that. It's only available for people who can pay $ billions today. We want to democratize it. GB300 will be available in smaller scale in our platform. We will be, I believe, first in the market who provided such democratized access.

We are also definitely providing GB300 to bigger ones like Microsoft and Meta. This is what actually the whole market is doing right now. We want to change this market. Second thing is, big thing is that we will continue to invest to pay-as-you-go model. Our platform right now is great if you can come and buy a reserve for some amount of money. Even most of the customers are pressing us to buy long-term commitments today. We understand that this is good short-term strategy for us. Long-term, we're investing in multiple ways to run compute, to combine resources and on-demand. Like today in our platform, we can combine pay-as-you-go with preemptible and reserve. We want to go further. We will launch ability to get preemptible machine with auction model.

You will be able to bid. Based on your amount of bid, you will get some compute guarantees. You can combine multiple types of compute here and not use static price. This is first thing. We will continue to evolve this model to provide more and more spot-like capabilities to our platform. Another important big thing that will come is that we will make our platform a multi-data-center by nature. Right now, it's like your workloads are bound to specific data center. We want to create services. You can build multi-data-center environments very easily. You can do it today right now as well. We want to give all the capabilities, features, and infrastructure to build multi-data-center workloads, like multi-data inference and production output. This is also something that is coming from the big things.

I think also interesting thing that I believe we will build is we know that all the mature customers. Some customers start with inference, some customers start from AI builders. You start either from inference or for the training. From inference, it's usually if you build AI product with state-of-the-art model. With the training, if you start to build your product with model. It's very easy. When you mature, when you scale, you combine both, to be honest. Our vision is that customers will combine training and inference within one cluster.

We are building capabilities for customers to get one cluster for us and combine it with training and inference, and for inference use Token Factory, for training or unmanaged inference use the cloud. This is also important capability that we are building to support multi-scenario clusters in our platform. A lot of enterprise features, something that our enterprise customers, the most demanding features of our enterprises today is cloud interconnect. Capability to connect your data center with our platform. It is also something that we are already doing with first preview customer. Soon we are going to make it fully generally available to allow any customer to connect their data center to our platform. I think from the big things, this is it, and there is 100 important low-hanging fruit that we are building to create the best compute in the market, best user experience in enterprise readiness systems.

build- Yeah a lot of fundamental features.

Yeah. This also brings us to one of the questions from our builders today, and it is asking about Serverless endpoints. The question says: Can we expect any improvement in Serverless endpoints deployments, like easy deployment for open source LLMs or simply like Hugging Face endpoints without creating a container registry?

Yes. Good question. I'll put it to two things, how we think about it. First, in our cloud platform, we recently released jobs and endpoints. Current version of endpoints is mostly to double check that job that you fine-tune it is working right. Basically, it's a developer endpoint, it's not production endpoint. Definitely, we will make these endpoints production ready with auto-scaling capabilities, with multi-data center capabilities. It's definitely part of our roadmap in our serverless tech platform. Important to tell that Token Factory also evolved. Our serverless tech will be around you get your container, and we orchestrate your model container very efficiently with understanding of AI. What does it mean to orchestrate AI native workload? It's like efficient container orchestration for AI native inference. Token Factory is also part of this vision. It goes a little bit up to the stack.

Token Factory is about bring your own weights, and we will orchestrate your weights for it. You will not control the containers. Containers orchestrated, controlled by Token Factory. If you use LLM models today, you can also bring your own weights to Token Factory. It's maybe not available using self-service interface, but you definitely can contact them through support, and they can make this feature available for you, definitely.

Exactly. This is definitely something that we haven't.

Inference is something that we need to cover fully. That's definitely manage inference part. Yes, definitely. I also think this answers another question because Alan is asking, will there ever be serverless hosting of custom AI models that you're charged with inferences actually running for your model, and that's what we offer in Token Factory.

Yes, solid answer to this question is that if your model is LLM model that uses vLLM or SGLang, you can just bring your models today already to Token Factory. Please contact them. If it's something else, like protein folding, multimodal thing, please wait a little bit and we will allow it to happen in Serverless stack. Yeah. Yeah. One question about KMS and cryptography, how safe is the cryptography's key, especially against quantum?

Are we planning anything when quantum arrives because it could break our security codes and keys?

We have this conversation about quantum a lot inside our company. First is that, today these keys are created by hardware security modules. I don't remember exact vendors, but it's something that super secure from the perspective that it's not software. It has all the HSM guarantees. It has all the hardware guarantees that keys are rotated. There is hardware enforced guarantees that keys are encrypted, rotated, everything is stored in secured manner, everything around it. Second thing about quantum, I'll be honest, we don't have post-quantum encryption enforced everywhere yet. We're definitely doing it. We prepare, I think, during this year or next year, we will fully move to post-quantum encryption. In reality, we're not going here faster because we understand that, yes, there is a risk, but we also understand that quantum computing is still in emerging phase.

If somebody will use quantum computer to do it will be some very serious government, quacks, agencies.

Yeah All of this. If they need to do something with-- These agencies will not only use quantum computers to get what they want.

They will use a lot of social engineering, a lot of other interesting tricks to do real red teaming-like attacks. In this case, we are preparing for this. Yes. Definitely it's going to be covered. Risk exists, but we believe that probability today is still very small for the whole industry.

All right. Yeah. One other question about, could you explain to us the vision solution provided by Nebius? I think it's about VLMs that we offer in Token Factory, if I understood correctly.

We will allow to host, I think Token Factory, at least I think for the next year, it will be very focused about language models. Main value of Token Factory today is that, okay, you have your language model, open source, custom, and most of the models that are hosted and used in Token Factory and people pay for Token Factory is mostly some custom model, which is based on some open source. Language model and main value of Token Factory is that you can go to Token Factory guys, give them some amount of money and, with managed inference capabilities that Token Factory provides, you will get much more value, much more tokens, that you will host it by yourself. There is optimized runtime, optimized orchestration, a lot of low-level things that in general give total cost of ownership value to the customers.

Either you save money or you can get more value from your budget, basically. This is the main business metric. Today this is about large language models. We're gonna evolve these capabilities, but we understand there is a lot of other models like vision models, protein folding, small language models, a lot of them. We will allow customers to host them in general purpose orchestration in our serverless. This is something that we, I believe, will bring end of this year or beginning of the next year. Any model that you can put to the container that it's not LLM, that it's not very specific to the runtime, will be able to run with us. It will not be as efficient on the low level but it will allow you to not to think about production, operating day two production use cases.

All the auto-scaling capabilities, high availability capabilities will be in it. For you, it will be just you don't need DevOps to hold this model. We'll provide you this value, definitely. For any type of the model.

Yeah. Well, thank you, Narek. It's always a pleasure to share this with you, and thank you everyone who joined us today. Please try to join our Nebius Builder Program and feel free to reach out to us on Discord, on email, and see you in the next one. Bye. Bye bye. Thank you, Focus.

What interesting. Keep in touch.