← 返回

2026.09.22 | TB | Cloud | Neocloud: Nebius & Mistral

原标题：Neocloud Market – Nebius & Mistral AI – Former Senior Executive, Strategic Partnerships at Nebius Group NV

#cloud #datacenter

Third Bridge · 2026-09-22

Client-led • 9 Sep 2026

Companies: NBIS US, ASML NL, NVDA US, PLTR US, IREN US, CRWV US

Expert Former Senior Executive, Strategic Partnerships at Nebius Group NV

Agenda

Questions

[00:06:09] What are your views on the smaller players entering the neocloud market? Are Nscale and Volta both credible players, despite Nscale being larger? What is the revenue potential of these companies?

Expert [00:07:21]: There are two important developments. First of all, Nscale, they have the large contract from Anthropic. This is something that was expected. Nscale would have eventually a large contract from the closed-source modules. This has basically been done in order to speed up the valuation of Nscale for the IPO in that regard and to provide additional capital to the company. Now that the Anthropic contract has been announced, Nscale can proceed with the IPOs with more solid fundamentals. But the question still remains, is Nscale, well-funded? The contract to Anthropic, while it is definitely a very good one in that regard, we also need to take in consideration that it's going to strengthen the position of Nscale quite significantly.

At the moment, the contract more or less is going to be about USD 45bn on a six-year agreement. At the same time, it's going to scale up to probably like 1GW, or 2GW, actually. That is going to be more or less the scope. At the moment, the contract itself is quite solid. It's going to be in the Monarch, in this case, the site of Nscale. How much is going to be unlocked from this contract? First of all, we need to get into that because, at the moment, of the contract, Nscale can definitely move forward with the USD 3bn IPO raised. It is going to be used for expanding the current business model, but we did also take into account that the legitimisation from Anthropic at the moment, is coming at a very good time for Nscale because, like I mentioned, there were many doubts. By clearing out the doubts at the moment, Nscale consolidate itself as a credible player in the market, able to raise additional funding and to scale further. This is my take on Nscale.

Client 1 [00:09:45]: You're bullish, I guess.

Expert [00:09:46]: Bullish more from a technical perspective. I see that the company at the moment has done what it needed to do in order to pass to the next level. That was the most important, because in order to progress, they needed to achieve another large contracts in order to be able to unlock the highest IPO valuation and also additional funding. Bullish, like I mentioned, even though at the moment, they have been definitely legitimised by the market, they can definitely provide large bare-metal contracts. The thing is about their expansion. The USD 3bn, while it's a very good amount, is not enough at the moment to expand the cluster expansion further. They need much more capital than that, probably between USD 10bn to USD 20bn that they need to raise. This one, they could come from our, let's say, contracts or, let's say, for other different financial instruments, because you need to take into account that when the company is going to be public, they start to have their disposal convertible notes that, for example, they could use for the shares, because having the shares now increases the amount of financial interest from institutional investors and also different investors.

[00:11:16] Can Nscale become the new Nebius if it accesses the capital?

Expert [00:11:23]: Not the new Nebius at the moment. It's not possible to reach the scale of Nebius, but they can definitely compete against Iren, for example, in that regard. Nscale has more pipeline capacity than Iren at the moment. They are targeting similar segments, the enterprise and also the large contracts from hyperscalers, as well as closed-source models. They can definitely become a very good competitor of Iren. How do I position them? At the moment, definitely lower than a Nebius, for example, also because Nebius now, with the Palantir contract, has a far stronger AI sovereign value proposition than Nscale can offer at the moment.

Client 1 [00:12:14]: Is the market so big that there's space for almost everyone?

Expert [00:12:24]: Exactly, yes. At the moment, definitely there is space. The important is that opens up the financial instruments that Nscale needs in order to scale further. Because Nscale needs a massive amount of capital that needs to become independent from contract from hyperscalers and closed-source modules, because these contracts, while they are very good for cash flow, they're normally used for CRM and sales. You have some, let's say, funding that is left from these contracts, but normally used for the company operations, while you need more profitable GPU-as-a-service or enterprise contracts, or AI natives as well.

[00:13:03] What's your view on Volta?

Expert [00:13:16]: Volta, I have a very strong view as well because even if it is a new player on the market, the team is very skilled. First of all, they just started with a USD 10bn contract with, in this case, Anthropic. The site is going to be about 1GW in that regard, so it's quite a large site, and they have access directly with the Vera. Congratulations. Just to come into the market, having access to Vera and the massive amount of funding that they have, USD 5bn just to start with, and then they plan to actually collect even more. I'm aware that the Volta team is also very well-connected with Wall Street, and also because the two founders are coming from Brookfield Asset Management. They have a very well-developed network, not only from the infrastructure side, for the actually the financial instruments, as well as end customers that are interested, in this case, to lease from them. They have quite a large portfolio at the moment.

Also for them, the main bottleneck is capital and, at the same time, the sites, the power in this case, in order to get to the sites on time. For the rest, I see them as quite solid. Do we see them, let's say, becoming an end scale? At the moment, they do not have the capital or the capabilities, because Volta currently, they are just developing their own team. It means they are able to execute these projects, but they are not able to structure the business in a more complex way. They do not have resources and the time in order to structure themselves to provide also enterprises at the moment. They are focusing more on getting very large contracts and executions.

Client 1 [00:15:32]: Is it more of a VC play, a long-term, high-risk bet?

Expert [00:15:37]: Yes. I would not say high risk because they are well-capitalised. Yes, they are quite well capitalised. They are very professional. They're very professional in what they're doing. The execution is actually the question because, of course, the team has experience in executing the projects, but to not such a large scale. We need to see. I would not see them actually as a speculative or high risk. I see them as quite able to deliver and, at the same time, to proceed with very complex projects, because mastering a 1GW of Vera, it's impressive. It's really, really a complex undertaking. Not everyone, at the moment, is able to do so. It's only the most skilful and skilled team that can actually deploy such a large deployment. Then also how they are operating, it means that they have quite substantial experience in the field and regarding also the pipeline, the contract they have. If you get an Anthropic contract, it means that you are so well connected and, definitely, you're going to have other large customers, like the hyperscalers, for example, is also possible.

Client 1 [00:17:01]: Volta has good investors backing it, which is further validation. Azora, which is one of the asset managers providing financing, has a family connection with one of the founders, so that probably helps.

Expert [00:17:24]: Yes, definitely. Also, this is the first, let's say, credit facility from Azora. Then I'm expecting also more others in this case. Of course, not to over-leverage themselves because we're talking about USD 5bn then going to probably another tranche of USD 5bn, but definitely, they can leverage substantial financial resources.

[00:17:47] Is Volta potentially a USD 10bn company? Is that the kind of scale we're looking at, if it executes well in the medium term?

Expert [00:18:05]: Yes, definitely. Like I said, the quarter overall is going to be about USD 20bn, what I've seen on the conditions. It is just like one contract itself. Think about it, the other ones that they have in the pipeline, because the pipeline is at the moment is full. We need to also take into account that one. They just need the capital and the right sites to basically deliver to what their customer portfolio needs at the moment.

[00:18:33] Mistral recently announced the biggest European fundraising. Does this give the company some credibility or capital to pursue its ambitious plan? What is the potential sovereignty and durability of the technology in the long term? How will Mistral perform in terms of compute, capital and being a fully vertically integrated player?

Expert [00:19:38]: The thing that I've seen is definitely, at the moment, one of the main participants to the Series D is, in this case, ASML. Mistral have taken and doubled down on the sovereign AI in Europe in that regards for the expansion, definitely. At the moment, the thing is that for them, it's going to come easier to actually compete in Europe against the minor players that they are facing. Nebius at the moment is a threat that is becoming important, definitely, against Mistral. Of course, Mistral cannot compete against Nebius on the capital, but Nebius cannot enter the French market as Mistral can operate them. It is one of the major strengths at the moment, and the value proposition of Mistral is actually stronger for some companies than, for example, it is for Nebius. The other part of the market that actually Mistral can definitely compete and outcompete is the smaller neoclouds at the European level. We're talking about Sesterce, for example, or also, let's say, other minor neoclouds like Nebul or Lyceum. All these very small, emerging players. Sesterce is from France. Then you also have a couple of others in France, very small ones. Nebul is from the Netherlands as well, Lyceum that is from Germany. These very small ones, they can actually be out-competed by Mistral without any kind of issues.

Mistral started also to have a more sovereign AI per country, in this case based compute, because while Mistral is deploying in the Nordics they have a main cluster also in France. We had the contract from Microsoft as well for the Vera. This is what I see from Mistral. USD 3bn is not going to get you very far in the US, but in Europe, it can definitely get you very far at the moment. We need to take into account that if Mistral will start with a more aggressive expansion in Europe, then definitely, there are neoclouds that will start to feel squeezed in that regard. Nebius, like I said, it can actually have a position there in France, but because Mistral is French, it has, of course, the favouritism from French companies, as well from the French government. It's very difficult to compete. The question is going to be, how successful it's going to be, Mistral, to expand outside, let's say, France? Also, instead of serving, in general, sovereign AI in Europe, it's also able to deploy in every single country? That's actually the other question, because USD 3bn are good, but you cannot deploy in all Europe at once. You need to target specific markets.

The first one that I have in mind is going to be, of course, the UK. This is the most profitable one and at the same time, they have very good enterprise customers. They can try with Germany as well. It's another market. Of course, there are minor markets like, for example, Italy or Spain also, but they are far less profitable, for example. This is what I see Mistral moving forward. Can they become full stack? They are definitely moving towards becoming a very important player in the inference-as-a-service and especially in the sovereign inference-as-a-service, what we've seen of it. Their performance are quite high. The problem is that there are not enough, let's say, companies, European companies that are providing open-source modules but are also very good in performing. It means that the open-source module strategy is relying on providing the sovereign AI infrastructure, but then to use the Chinese models mainly because they are the most competitive at the moment. This is how I see Mistral.

Can the team succeed? They can succeed, like I said, because they are focusing in Europe. Still, they're going to suffer from brain drain from, for example, the US because I've also seen that some, let's say, researchers, of course, are moving to the US instead because of higher salaries. Of course, the other thing is that, with the acquisition of Hugging Face by, in this case, Nvidia, and also basically the almost de facto acquisition of Poolside, let's say that the number of possible customers from Mistral is actually diminishing, and in that regard, the importance of Nvidia is increasing. Nvidia, of course, is also presenting its own, let's say, sovereign AI version with Nemotron, as well as, for example, now the Poolside commandos, in this case, open-source models. Of course, as the competition is heating up, Mistral can still have access to the enterprise customers because it is a European company. This is how I see Mistral moving forward.

[00:25:03] Can Mistral succeed outside of France and meet the demands of large industrial aerospace and defence enterprises that value sovereignty? How will Mistral's political closeness with the Macron presidency evolve? Can Mistral stay relevant and expand outside of France?

Expert [00:25:50]: At the moment, they are getting quite aggressive. As you can see, they started to hire actually in Germany as well as in the UK, as well in the Netherlands. They are expanding quite aggressively. The question is going to be, how these governments, like you said, the political questions, are they going to accept Mistral? That's also the other part. If you're going to set Mistral as the sovereign AI player in Europe, then you can probably think when Mistral is a good option in that case. At the same time, you have countries like the Netherlands, where they play the middle ground. Also, they have an emerging, let's say, neocloud themselves, called VOLT. They are quite large at the moment in capacity deployments and they also want to become like another neocloud in Europe. This is going to be, let's say, a difficult part for Mistral to play in the Netherlands because in that case, the deployment of VOLT is going to be the Netherlands and then the rest of Europe. Mistral at the moment has other options as, for example, the UK, where there is high competition. Enterprise is still quite good. They have HSBC as a customer.

Then, of course, the German in this case, enterprise. As well, there is Italy, another market where Mistral is also trying to catch, but there is also local competition, Domyn, that is entrenched there. So it's difficult for them to, let's say, expand. Especially because they are competing with entities that do not have substantial funding, but, because they are entrenched in those specific markets and also because they are originated from those specific markets, it's difficult for them to compete directly because they have the favour of the government. It was the same for Nebius. When you are trying to expand like in France, then you have Mistral there, you're never going to be the first player. In the UK, you actually have no competitors because, technically, Nscale is present, but is very limited in deployments there. That's why it's quite open to others. The Italian market, it's very not super profitable in that regard. The German market is difficult to enter. That's some of the different considerations.

[00:28:29] How does the neocloud business model work? For example, Nebius own roughly 70%-plus of its capacity and operates at much higher margins compared to other neoclouds, such as CoreWeave. Why does Nebius trade at a much higher multiple relative to CoreWeave, even though CoreWeave is roughly 4-5x larger in revenue? What are the two different models and the dynamics?

Expert [00:29:26]: Of course. First of all, we need to understand what investors are pricing. CoreWeave, while it's actually an established neocloud and have their own cloud, is no match at the moment for the software integration of Nebius. The full stack of Nebius is far more advanced at the moment than CoreWeave. We see it as well on the case regarding, for example, integration with Palantir. Palantir is a very important strategic partnerships for Nebius, and basically Palantir can be deployed in sovereign clusters that Nebius is deploying at the moment in every country of interest, where they're going to be completely separate from the Nebius cloud. Nebius is going to provide the bare metal, and then, in that regard, Palantir is going to run on the bare metal and to provide these dedicated clusters to customers. At the moment, CoreWeave's focus is more to serve very large contracts, bare-metal contracts rather than going in that direction. That's why the level of sophistication is higher.

Second, also the team. You need to take into account that Nebius at the moment has a very strong team, starting from Arkady, but also Roman and as well the US team, for example, with Marc Boroditsky. At the moment, this team, just like to name a few, this team is very strong, very well positioned. Also have a clear vision, especially voice from Arkady, that they want to go enterprise and also they want to become, like, providing more full-stack services. This is a direct vision, but at the moment, Nebius is consolidating to investors, of which Goldman Sachs at the moment is one of the major ones. More than 10% of Nebius is owned by Goldman Sachs, and this is actually where they can also get the majority of the funding. The third point is that because Nebius, at the moment, is providing quite sophisticated bare metal to Microsoft [and] Meta, as well as professional services on top of that, it has a stronger value proposition than CoreWeave.

The fourth is because Nebius, at the moment, is diversifying heavily not to rely on a few customers, but to try to get as much a diversified portfolio as possible and expanding in multiple regions at the same time. Something that CoreWeave is doing, but is more targeting larger deals rather than to have a very broad, in this case, deployment. The other point that you need also to consider is execution. At the moment, Nebius is very well ahead in the execution, actually surpassing the CoreWeave in execution capabilities moving forward. This is because they are actually, the engineering team for Nebius, coming from Yandex, and, of course, the new members that have been added to the infrastructure team who have experience in the execution of large-scale projects. Also, that Nebius at the moment is hiring many Amazon and former Google employees that, at the moment, are going to be tasked to scale the company further. Investors, they see a vision, they see a direction, they see a value proposition that is basically, at the moment, not the same at CoreWeave.

CoreWeave, like I mentioned before, is more focusing on the larger customers rather than to have a more, let's say, broader vision to the enterprise. CoreWeave is also now focusing on sovereign AI. CoreWeave, at the moment, is following, like I mentioned, demand. The 300MW deployment that they have in Southeast Asia can be considered a sovereign AI deployment, yes, but it's not pursuing a real sovereign AI in this case, in practice, like Nebius in that regard. These are all things that investors are noticing. At the same time, Nebius builds out and CoreWeave builds out. The majority of the build out from CoreWeave are located in the US. While the majority of the build out for Nebius, which is also located in the US, we need to account that Europe is also very well covered with the new deployments. Asia and the Middle East at the moment is another story. Nebius is planning to expand in both regions. While CoreWeave at the moment is expanding aggressively to Asia, the Middle East is in assessment still. That's why we see that Nebius is getting ahead on CoreWeave in those markets at the moment.

[00:34:28] Where is it cheapest to run compute globally?

Expert [00:34:30]: Where is the cheapest? It depends where you are, because the cheapest compute doesn't mean that, the location, you can actually use it. Because, for example, if you have a strict, let's say, sovereign AI requirements, it is usually now always the case [that] you need to be in the region. You cannot be outside. In the past, for example, you can maybe run very cheap data centres if you were in Europe, in Morocco, for example, as well as Egypt. Since then sovereign AI has been implemented, this has basically lost any kind of relevance. That's why, at the moment, Nebius is actually deploying in the target countries or regions where, of course, it can actually be more expensive, but it is a requirement. Where is the cheapest? The US, America is still one of the cheapest markets in the developed world. If I take into account Europe, Europe at the moment, the Nordics or Spain are very cheap to run energy, to run, in this case, data centre power there, very intensive workloads as well.

Client 2 [00:35:45]: What about France? It has an abundance of power.

Expert [00:35:50]: Yes, but it's expensive to run France in that case. While, let's say, it's affordable in that regard, in comparison to other countries, deploying that is still quite expensive to run energy in comparison to Spain. Spain has a surplus of energy most of the time. It means that you can actually have very favourable contracts that are going to be lower than, for example, France itself. This is because, at the moment, the demand for energy in Spain is very low in proportion to how much we produce. France has a lot of energy, yes, but normally that is quite a long-term contract. Also, you normally deploy in France just because you need and you have to deploy in France. Otherwise, I will be happy to deploy in the Nordics always.

In this case, it is normally expensive to deploy in France. French deployments, in comparison to Spain, for example, can sometimes be even double the price. It depends, of course, but during summer, normally, the price of energy is fixed on the French nuclear reactors. Also, you need to take account that, in summer, they go offline quite often. While instead in Spain, normally, you have a huge amount of energy that can actually be supplied directly. While in the Nordics, usually, at the moment, you still have quite substantial amount of energies between nuclear as well as, for example, hydropower and the minor, like wind, for example, and they are very small countries. That's also thing. They have surpluses there, but normally, they are localised in different zones that they are divided. You cannot really export energy from, let's say, the remote regions of, let's say, Finland to Estonia in that regard. It doesn't make a lot of sense. You can actually have very good deployments over there. France, instead, has more or less standardised power pricing because, basically, running mostly on nuclear power. Let's say, the renewables, at the moment, are quite, let's say, between 25%, if I'm not mistaken, of the mix. The nuclear is providing the majority so it means the power is not going to be cheap.

Power is going to be affordable, moderate, but not cheap because you do not have an oversupply. Technically, you have oversupply of power, the nuclear power, but normally this power has been contracted already to be exported or, for example, can be contracted to be supplied. You're producing more power than you actually consume, and then because the prices are going towards negative, you can actually agree on certain prices. The nuclear power is always the same pricing. It's not possible to actually to negotiate a lower pricing. Sometimes it's also higher price because, like I mentioned, because of maintenance or because of problems related to climate change in that regard. That's why the price remains attractive, because you do not have problems of power supply in France. That's the most important. I don't have a problem to actually find sites to have a data centre. If it is France proper, then, for me, it makes sense to have a data centre there because I cannot deploy anywhere else. If I can supply just for everyone else in Europe, I can just have the Nordics or Spain. For me, it makes a difference in the pricing power.

Client 2 [00:39:41]: Does 1GW of compute sell for roughly USD 35bn-50bn in the spot market from a unit economics perspective?

Expert [00:39:51]: We need to take into account which kind of cluster we are talking about. Are we talking about H100, H200? This is the most important question.

[00:40:11] The figure in most news outlets is that 1GW of compute is roughly USD 50bn. How much does it cost the compute provider to procure 1GW? How much does that provider sell 1GW for?

Expert [00:40:40]: In this case, let's start with 1GW of compute, we need to take the largest ones. In this case, for example, we can take the Vera and GB300 because these are usually the ones that are going to be deployed in the future because, at the moment, 1GW of compute of one single cluster, it does not exist. At the moment, there are multiple clusters in that regard. How much does it cost? The price we have, depending on the model, is usually between USD 30bn to USD 80bn for 1GW of compute at the moment. These are actually the costs. If you are taking into account how much is going to be the revenues [for] 1GW of compute, it does not exist at the moment in practice.

At the moment, I have USD 38bn for the GB300 deployments, while instead, for the Rubin, it's going to be USD 47bn. Of course, these are very high-efficiency deployments because 1GW, in this case, at the current pricing of September 2026, but they can definitely scale further. The USD 80bn figures that I mentioned was, for example, after the Rubin. That's why it's going to be far more, in this case, expensive to deploy. When I said the annual recurring revenues that, at the moment, I see for the GB300, for such a deployment, is going to be approximately USD 16bn. Between USD 16bn to USD 21bn for the GB300. While instead, for the Vera, the one that I see at the moment is going to be approximately USD 31bn, but I can also imagine that it can actually go even to USD 35bn, probably, because I'm also looking at other contracts that have been signed at the moment. Definitely, I'm expecting something between USD 31bn to USD 35bn or USD 40bn because we need to take account also Vera.

Client 2 [00:43:50]: Is this net of cost? It sounds like the provider sells it lower than what it buys it for, unless you're quoting net-of-cost numbers.

Expert [00:44:00]: This is actually the annual recurring revenue. You need to take into account that it is just for one year.

Client 1 [00:44:05]: Payback is effectively one year?

Expert [00:44:10]: Yes.

Client 2 [00:44:12]: Payback would be 1.5 years, or something like that?

Expert [00:44:16]: Yes. Exactly. Normally, Nebius, at the moment, is saying 22 months, the payback period is. The industry is about 18-22 months, so it is normally the average.

[00:44:34] Mistral had a USD 3bn raise and USD 900m in debt, which are small numbers compared to US peers. Will this limit Mistral's capacity?

Expert [00:44:49]: Yes. I have a negative view of Mistral because it's very small. Still, let's say that, for example, with USD 3bn, what we're going to do? Definitely, they can have a cluster that is going to be a large cluster for European standards, of course, if you exclude in this case, actually, Nebius, because we need to take into account that, at the moment, Mistral is also trying to consolidate its own deployments instead of relying just on CoreWeave, because CoreWeave costs about USD 1bn, from what I know, in contract that they have. They have tried to consolidate their financials. With USD 3bn, yes, they cannot buy a huge amount. Let's say, if they are going to do probably…

Client 2 [00:45:41]: 200MW max. Something like that?

Expert [00:45:48]: Even lower, because we need to take into account, in this case, that the smaller is the deployment, the higher is going to be the cost.

Client 1 [00:46:16]: I assume Mistral will probably raise more money going forward.

Expert [00:46:21]: Yes, in Mistral, of course.

Client 2 [00:46:26]: It is the largest European race, so it should give Mistral a lot of credibility in the European landscape. But in the larger context, this is nothing compared to what is happening in the US.

Client 1 [00:46:44]: Is the existing 200MW around USD 8bn annual recurring revenue?

Expert [00:46:57]: If you take into account the GB300 in that regard, and we said USD 3bn in deployments. The pricing is going to be higher, so we cannot use that one completely because the deployment is going to be smaller.

Client 1 [00:47:40]: They trade at 20x.

Expert [00:47:47]: The thing is that, normally, you cannot use the figures from the 1GW and then go down. The reason why is because costs are actually the…

Client 1 [00:48:23]: Economies of scale.

Expert [00:48:24]: Exactly. The CAPEX are actually higher. In that regard, you need to adjust the CAPEX first. When I have adjusted CAPEX, I can give you the right results, because otherwise it's not correct to use the 200MW because it doesn't make sense at the end. At the moment, with the GB300, let's say, for the pure IT hardware perspective, the GB300 is possible to acquire about 63,000 GPUs. From the Vera perspective, we are talking about 47,000 GPUs.

Client 1 [00:49:14]: Is that enough? Is that 1% of what Mistral needs?

Expert [00:49:31]: Ballpark, what they need, they need far more than that. Definitely, this is something that is not enough for them, but definitely it's good to become more independent in that regard. This is something that is possible.

[00:49:49] What would be 63,000 GB300 out of Mistral's total need? Would it be 10%, 50% or 1%? How much would Mistral need?

Expert [00:50:00]: Mistral, at the current level, they probably need 3GW of deployments. It's basically nothing. To be honest, it's because the plan that they have is about 1GW. These ones, about 200MW, is going to be a deployment that they can serve for Europe overall, but it's nothing in comparison what they need. They need far more than that. If you do the scale, they need at least 1GW in that regard. If they want to expand and to be the dominant player in the European market, they need 3GW. If we take into account, like the cost of the 3GW, we are talking about USD 120bn. It is more or less what we need if they want to have the monopoly. As you can see, there is quite discrepancy. The minimum that I'm aware of is that they want to deploy at least 1GW in Europe. Just that 1GW is going to cost between USD 40bn to USD 47bn.

That's the theme. They can only deploy these very small deployments of about 200MW, more or less. This is how they are going to move forward. It's sustainable. They can do it, but, of course, Nebius is going to deploy 300MW just in Finland next year, and then it has another 240MW that is going online, basically, in 18 months in France. It has basically passed Mistral deployment already, without counting the other giga-factory that is coming online in Estonia.

[00:52:02] The underlying assumption of these calculations is that GPU prices do not come down. Could GPU prices decrease as technology develops and new players start developing their own chips?

Expert [00:52:31]: At the moment, like Arkady mentioned in a conference the other day, basically Nebius is selling GPUs at 15-20% higher than the pricing on the market because they are one of the few players that have GPUs at the moment. I'm expecting, definitely, the price of GPUs to continue to rise. Also because, at the moment, say, Meta, they're not releasing large numbers of GPUs because they are not ready for commercial deployment, just bare-metal deployment. It means that the GPU prices are going to be constant for a long time. It means that the strategic reasoning of Mistral is that they are going to deploy fewer GPUs in that case, but they're going to be probably GB300 as well as Vera 200. Let's say, that if you do like a USD 3bn investment just in Vera, even if it is a small investment in that regard, it can generate you about USD 2bn in annual recurring revenues. Mistral at the moment is running about, if I'm not mistaken, USD 1bn contract with CoreWeave, and then they have about USD 1bn in annual recurring revenue. This funding is going to help them to actually double the capacity that they have available. While, of course, it's very small on what we really need, it's still good for them to become more independent and also to have more capacity for their customers.

Client 1 [00:54:09]: If you consider the deployment engineer model, can this be a business that has USD 4bn-5bn in ARR in the medium term?

Expert [00:54:26]: It's a high number because it's more a professional service rather than, for example, hardware. You could have something that is like this if it is a hardware and mixed deployment with engineers in a large scale. At the moment, this starts as a professional service.

Client 1 [00:54:45]: Both, not just the services business?

Expert [00:54:48]: They do not have the current cloud computing capacity to reach USD 5bn.

Client 1 [00:54:56]: In five years?

Expert [00:54:59]: They need more capital. That's the thing that they need. At the moment, five years, they need to raise at least another USD 9bn, probably, more or less, to expand. I will say at least, conservative, it's going to be another USD 3bn, conservative, because if we had the USD 3bn that they had it, in this case, it's going to be USD 2bn in ARR. We consider the USD 1bn that they have deployed at the moment, so we're going to be about USD 3bn when they deploy another Vera, so it's going to be USD 5bn in that regard. It means that they need USD 6bn, but we need also to take into account that they're also running a business. They are not going to survive just with this little money. They need also additional capital. Probably, they need at least another USD 3bn for hardware and USD 2bn-3bn for operations in order to get to the USD 5bn.

Client 2 [00:56:13]: Can Mistral do this through raising debt, which is what it's been complementing the equity raises with?

Expert [00:56:19]: Yes.

[00:56:21] Samsung led this round. What strategic benefits does Samsung have? Why would it want to lead a round on Mistral?

Expert [00:56:38]: Definitely because the experience with ASML on Mistral. They are interested in the kind of solutions that Mistral is also implementing at ASML. These are tight, basically dedicated, let's say, artificial-intelligence-as-a-service, I call it, where basically they are serving directly, in this case, the company. You are not relying on the other third parties, but you are relying on actually, like, a professional service that is going to be completely dependent on you in that regard. It means that it sees it as an on-premise deployment. In this case, like the unit also, it's very important because the partnership was not led, let's say, just from Samsung, but it was led by the Samsung Device Solutions. Like I mentioned, it is more a strategic partnership because we've seen the experience with ASML. It is for the on-premise industrial AI, process optimisation as well, hardware-software synergy, and regarding the possibility that, like I said before, they can have a directly in-house capabilities and they are not going to be dependent from external models. This is the main value proposition. Basically, from the ASML project, Samsung got interested, but we've seen it a more larger and broader application.

[00:58:13] Who is the best player in the inference-as-a-service space?

Expert [00:58:31]: Together AI at the moment is one of the best players. Definitely, they have also the most solid value proposition. Fireworks is, at the moment, with the higher valuation proposition, but on the technological stack how they are positioned they are inferior to Together AI. Their solutions, they are not at the same level, but the investors are actually pricing that Fireworks is also offering training. That's why in that regard, the investor base see that, potentially, Fireworks has a larger market potential than Together AI. Together AI is more specialised on just the inference-as-a service, the agentic AI, the orchestration, as well as massive token output and very complex engine for very complex engineering deployments of open-source models. While, for example, Fireworks is not at the same level in engineering, but because it has a broader audience, it has a higher potential, in this case, of annual recurring revenues than Together AI. Baseten, at the moment, is the weakest one between the three because, first of all, it's the newer player, the newest player between the three and also because it's targeting mainly enterprises that do not have, let's say, enough capacity to expand, not only on GPUs, but also regarding financing. Even though they have raised substantial funding lately, but still they are not at the same level.

These three players are going to dominate the market. It's not going to be one dominant player over the other ones, because, like I said, they have different value proposition and different complexities, and they have also different target audiences. That's why they have targeted different markets. Sometimes they overlap between each other, but most of the time they are competing for the same customers. Of course, when you're talking about sophistication in engineering or availability of GPUs. For example, Together AI or Fireworks, they definitely beat, in this case, Baseten. That's the main thing, Together AI, like I said, because they had the limitation of the inference mostly, so their clusters, they do not provide, like Fireworks, the training at the same scale. That's why they have a different value proposition. They cannot target the same identical audience than, for example, a Fireworks, but they can target the most sophisticated.

Client 1 [01:01:35]: Does the real value sit in the training rather than the serving?

Expert [01:01:42]: It's not the real value, the training. It actually depends how do you see it because let's say that the inference is the stickiness, the training, you can also change the provider. The thing that you cannot change is the size of the provider. A provider that can provide you a multi-gigabyte deployment, it's unlikely that you can find it everywhere. The inference that Together AI, you cannot find it. The solution of Together AI at that level of sophistication is not possible to replicate on other platforms. You can replicate similar solutions, but not at the same level. This one is actually the value proposition. To summarise, it all depends about the software stack. That's the thing. The training is not the main stickiness, it's the ability to own the cluster, actually. That is going to be the differentiator. If you own the cluster, then you're going to have the highest training service. If you don't own the cluster, like Fireworks, they're going to own the cluster in the future, but right now, it is not sticky at all. The software, instead, is the value proposition, the stickiness that all of three, they have.