RAISE Summit '26 | Erwan Menard, Crusoe

Clips
More from RAISE Summit '26

Erwan Menard

SVP Product

Crusoe

play_circle_outline Inference-First Vertical Integration: Crusoe's Gigawatt GPU Campuses, Edge Zones, Managed Kubernetes, and Sovereign Open-Model Services

play_circle_outline Model lifecycle management: quick retuning, sharding, API endpoint switching

play_circle_outline Supply chain constraints affect capacity planning: SSDs, memory, GPUs availability

play_circle_outline Outcome-based pricing and product-money fit: aligning inference cost to business value

Info

Erwan Menard, Crusoe

Erwan Menard

SVP Product Crusoe

search

John Furrier

>> CUBE Studios connecting, Silicon Valley and Wall Street.

Erwan Menard

>> I'm John Furrier, the host of the CUBE here with Gabe Olave, my co-host. Welcome to the RAISE Summit coverage brought to you by Celadon with support from Tensordyne, d-Matrix, and Argentum.

John Furrier

>> Welcome back to the CUBE, everyone. I'm John Furrier, your host here in Paris, France for the RAISE Summit. This is where all the action is turning out to be an international global AI infrastructure and AI enablement accelerating applications. Everybody's here across the world talking about the growth of AI, particularly the applications are starting to emerge, the real use cases from physical robotics. Everything's happening. Of course, the AI infrastructure is powering it. Our next guest with Crusoe, they're one of the big reasons why it's all working. Erwan Menard is here, SVP of Product. He's the product czar. At Crusoe. Super busy. Thank you for taking time out of your day. I appreciate it. Great to see you again.

Erwan Menard

>> always a pleasure.

John Furrier

>> head of product is a big role, but you're under a ton of pressure. I just wrote a post 2 days ago around this. We're midpoint in this token economy.

Erwan Menard

>> Yeah,

John Furrier

>> I use you guys as an example because you guys are moving so fast. First, you have the big data center in Abilene, Texas. I kind of use that as a backdrop. But it's not just big capacity. You guys are delivering now services. You have to turn on the product chops, start meeting the demand. Yeah, because you got to— you can only build it as fast as you can build. Well, where are you with the product? What are you guys shipping? What's happening?

Erwan Menard

>> So thank you for mentioning the data center roots. this is a vertically integrated company. So knowing how to build a gigawatt campus is certainly one of our expertise. But to your point, we've been shipping GPU clusters over the last few years, typically managed Kubernetes form factor. But the big news is that we're now playing in inference and helping people modify an open model, make it theirs with their data, and then inferencing it the most effective way. And doing all of this in a managed service fashion that you don't need to get involved with managing the GPUs. So managed inference, managed fine-tuning.

John Furrier

>> Yeah, and I think the demand is forcing that function because we're in a speed game, speed of delivery, which is building. I was just talking to AMD and Cerebras. They got to make their manufacturing plants faster. They got to stand up new ones faster. You got to build faster. You got to deliver services. That's because of the demand curve.

Erwan Menard

>> Yeah.

John Furrier

>> What has changed on the demand curve? Is it more cloud-managed services? Is it more inference tied to specific use cases? Has it changed the target audience? Give us a taste of what's the most impactful change in the past year.

Erwan Menard

>> I think, of course, we still have a lot of training going on, and this has been the genesis of the business over the last few years. But the fact that people are building agents going to production, their number one cost is inference. And if you look at the open models and how quickly they improve, You need to be able to tune the model, quickly retune the new model starting from a better baseline. And I think that's where we need to help. you pick a model, GLM-5.1, Kimi K2.6, whatever the model you like. Quickly you get your data, you derive your own version, and it becomes the brain of your agent. And you need to do that loop very often.

John Furrier

>> I don't like to brag very often, although I do pat myself on the back, but when 3 years ago, Dave Vellante and I published the power law of models. Yeah, you had OpenAI and Anthropic. Google was just coming on the scene with Gemini. And we put out a thesis that said there'll be a power law, there'll be a kind of a head and then a torso and a tail of models. And the thesis was specialty models or small language models. And if you look at Alex Karp's recent rant on CNBC, if people don't know what that is, just search Alex Karp. He's the CEO of Palantir, he basically lifted up the curtain or pulled the curtain back on some of the big proprietary, big frontier models, OpenAI, Anthropic, saying you don't need them only, which validates one, our thesis. But that's been validated. But this is now coming into practice. You're seeing people build their own models. Open source has been a big driver of that. You're seeing a lot more concentration of the value proposition in the models. Decoupled from the hardware. We've been covering that on the CUBE for a year. What's the impact to you? Because now you start to see real engineering going on at cloud scale.

Erwan Menard

>> Yes.

John Furrier

>> You're a cloud company by default, because you have the hyperscale infrastructure.

Erwan Menard

>> The impact is that you need to invest so much more in the software stack to help run these models cost efficiently. The inference engine we've built, I can onboard a new model in a matter of days, optimize to deliver throughput or latency, whatever your preference, and serve your business at the best cost factor. And you need to do that quickly. And what's interesting is that instead of being a static software piece, like everybody would download the same thing, the techniques are evolving almost week by week. So you need researchers, forward deployed engineering to give that inference software magic to the developer in a few clicks.

John Furrier

>> We're at RAISE Summit. This is more of a big industry show. And you started out by saying Kubernetes, which is oh, I got triggered. That's KubeCon, that's Linux Foundation, that's cloud native.

Erwan Menard

>> Yes.

John Furrier

>> Okay. So you're bridging cloud native to large-scale cloud native with AI workload-like capabilities. Explain why that's happening and what that means for a customer. Do they onboard like a cloud? How do they engage? Take us through this. Yeah, this is now next level.

Erwan Menard

>> You log into a console, you want to have a cluster of X many GPUs. And there they show up as a managed Kubernetes space where basically under the hood we're self-healing a number of issues. I'm sorry to say GPUs continue to have issues. It's very fragile hardware. And so the value out of the service provider is to detect those faults ahead of them happening, self-healing them, giving you options to avoid ruining a training run or affecting inference quality.

John Furrier

>> How has the— in the past 5 months, your business grown? What's been the customer feedback? What's been the learnings?

Erwan Menard

>> The business is on the infrastructure as a service side, it's the adoption of this managed Kubernetes. The clusters and the fleets are becoming so big that people need help to manage all this hardware. So that's now 70% of our fleet that is managed Kubernetes. On the inference side, it's the model lifecycle. I pick a model in our Intelligence Foundry, I modify it with my data, serverless fine-tuning, and I inference it with the latency or throughput recipe. And in a few clicks, I've been able to power my agent with my version of Kimi K2.

John Furrier

>> cloud is not easy, as they say in French, it's très difficile.

Erwan Menard

>> You're tapping into my dual citizenship here.

John Furrier

>> It's très difficile for Americans. Europe is hot. Sovereignty is here. How are you guys looking at the sovereign equation? There's a lot of cloud discussions because AI is not just a privacy problem. There's energy and the economics involved. That's money. You start to see people digging in in their countries, but they still got to run globally in a distributed way.

Erwan Menard

>> Yeah.

John Furrier

>> How are you guys looking at the European expansion? What are your thoughts there?

Erwan Menard

>> So we're already present in Iceland because we're energy-minded. So a lot of geothermal energy. We have a big presence in Norway. There is great hydro energy. And we're looking at projects in Western Europe. You're absolutely right. I think it starts with data and compute residency. I'm able to compute and save or store all the prompts and outputs of the model locally. we also have an interesting form factor with modular data centers that I can give you as a company, full exclusive use while you control it with a global control plane. I give you an edge zone, a Crusoe Edge Zone, that's a container mini AI factory. We manufacture ourselves and that gives you that data and compute residency. That's how we're approaching the topic. I think open models are also part of it because open models mean control. I control the version of the model, I control when I retire or launch a new model, and I have full control on the training, post-training, reinforcement learning of the model. And I think that plays a great role in those sovereignty requirements.

John Furrier

>> I had a couple of customers that do large-scale work say to me, hey, why don't you just switch your model? He goes, well, John, it's not that easy, because there's cost involved. Talk about some of the, I won't say tire-kicking kind of deployments, but people kick the tires, check things out, but there is cost to change.

Erwan Menard

>> Yes.

John Furrier

>> Can you share your insights around what you've learned on that? Because people might think, oh, I'm going to do this for a while, and next thing they're over their skis.

Erwan Menard

>> I think you're hitting a very important point. Yesterday, I was catching up with a customer who's a coding assistant company, They're sharding models all the time because new models show up, they improve their own models. And so our ability as a service provider to quickly change the API endpoint, stage a new model is actually mission critical. So the way we try to help with that is that we give you an abstracted service. You have an API endpoint, we take care of sharding the model, reconfiguring the GPUs, and we're able to do that very quickly so that you can take advantage of the new version of whatever model coming out.

John Furrier

>> Has data scale come up a lot with some of your customers, especially in physical AI?

Erwan Menard

>> Because you're going to have a lot of data preparation, you got so much feedback coming back from those devices that you want to bring into a reinforcement learning loop. So you have data preparation pipeline. We're seeing a lot of CPU, general purpose CPU usage side by side with the GPUs as well as a lot of storage for the sake of preparing data and feeding reinforcement learning.

John Furrier

>> Not to get in the weeds, but I have to ask, has data pipelining changed? the old days it felt like you dig a ditch, you put the pipes down, you put the grass on it. Yeah, static. Now we're in a generative model mode where you got all this inbound data coming in, especially on inference.

Erwan Menard

>> Yes.

John Furrier

>> What's been the data pipelining direction that you're seeing good use cases and good practices around?

Erwan Menard

>> First of all, I think you need to be very thoughtful about sizing. We used to deploy let's say 4 or 5 terabytes per GPU. I'm talking storage. We're now way in excess of 10 terabytes. So first, because of shortage in supply chain, plan, have proper CPU and storage capacity beside the GPU. Second, I think we have a lot of real-time pipelines that are in the works. We work with physical AI companies Figure is a customer as an example. I think we have a lot of value we can build. Over the next few months for these customers as they start getting a tsunami of data back from the robots that are in production. And I think the whole data pipeline needs to be reinvented. At least we're doing a lot of research work there to be able to quickly, at the right cost, find the right data and bring it in and generate synthetic data to augment the dataset and again feed the reinforcement learning.

John Furrier

>> You brought up physical AI, I want to just I'm going to double click on that for a second because I think physical AI is a telltale sign for what's coming fast. And it's not just robotics, I love robotics, but you got cars, you have other things. It shows the real-time nature, and physical AI will move to everything, retail, telecom, anything physical. And those edge devices are going to throw off a lot of new data.

Erwan Menard

>> Correct.

John Furrier

>> It might need to be trained, reinforced, processed.

Erwan Menard

>> And actually, at first filtered.

John Furrier

>> Yes.

Erwan Menard

>> What do I want to keep? Because it's a very noisy data flow. And then you need to generate synthetic data out of it. you get a rare situation that you want to create derivatives on. So the whole data pipeline has to be faster, has to filter almost on the fly, has to generate synthetic data. these are new challenges.

John Furrier

>> All right, vertical integration is a big thing. You guys are vertically integrated. Are customers going down that route? How are they thinking about their stack right now? As these architects are looking at building their next generation, this is generational. This is not hey, let's bolt onto a data center. Yeah, there's a little bit of that.

Erwan Menard

>> Sure.

John Furrier

>> But the real work is, I need to architect my business future globally and/or regionally or wherever. These are big decisions.

Erwan Menard

>> Big decisions. I think the first bifurcation is, Do I need infrastructure for training? Do I want to control lower in the stack, or do I want to rely on a managed service abstracting the infrastructure management? The second thing is, who am I working with for a time to market that's going to hold? Because the supply chain has stress everywhere— SSDs, memory, GPUs. You want to work with people who can give you a date by which the cluster is available and you can rely on it. And then the next aspect is the software stack. This inference engine conversation, it's amazing. We have so many customers who are testing us against all the other options and it's really fine-grained optimization with forward deployed engineering that gets you to the right balance of throughput and latency for the token price.

John Furrier

>> On my post 2 days ago, which I featured you guys in and others about the infrastructure enablement, I made a statement kind of buried in the post. It's kind of like an Easter egg and I was testing out the concept. But I said tokens are the new bandwidth. And I got some reaction from some senior engineering people. It was a dog whistle kind of a phrase. But my thesis was networking becomes super important because you got to move packets. Sure, that's networking in the classic sense. But tokens are now money.

Erwan Menard

>> Yes.

John Furrier

>> Money moves. So moving tokens is going to be a thing, I think. What are your reaction to that? Because that brings up the technical concept of networking, which we know what's going on with KVCache and memory and moving data into the right areas at the right time. Moving tokens and the role of networking, a new kind of networking. How would you describe this phenomenon of the new kind of network? This isn't your classic connect the campus. And move a packet from point A to point B. It's a lot of touch points, omnidirectional, a lot of threads.

Erwan Menard

>> I think you have that aspect. I think you said the word cache, which I think is the first order of optimization we need to deliver now. Because if you keep the cache big enough and fast enough, you can save a lot of recomputing. And so you save, you mask the networking challenge. Then comes the storage challenge, and I think networking is next. And if you think about building a cloud, we were very focused on GPUs, managing all their failures. We have that more or less under control right now. new SKUs come all the time, and you need to be humbled and do proper new product introduction. But then it's the cache. we've invested so much in KVCache, bundling the memory, giving you a very powerful cache. I think networking is the next topic once we have that cache thing under control.

John Furrier

>> All right, I have to ask you, since you're the product guy, I love the phrase— I've loved, actually I should say past tense, I love the phrase product market fit. My new phrase is product money fit, economics, because tokens are money.

Erwan Menard

>> They are.

John Furrier

>> So value is whatever we want is getting out of AI. So the value coming out of AI is better productivity, but that translates to dollars.

Erwan Menard

>> Correct.

John Furrier

>> Okay, so how do you guys view that? Because at the end of the day, these workloads are built for a purpose.

Erwan Menard

>> Correct.

John Furrier

>> To drive a value.

Erwan Menard

>> Yes.

John Furrier

>> That produces an income.

Erwan Menard

>> It's a really good point. In this managed inference business, we currently go into agreements with customers which are about a certain amount of spend because they can correlate it with their revenue tied to the agent solving a real-life problem. And then underneath, we're going to tune the model, we're going to change the configuration to take advantage of technology improvement, but we're not talking about Can you give me this model at this throughput? From a price point of view, we look at, let's create this value for your agent managing the biggest cost item in the agent cost structure. And then we'll tune the technology to serve that business goal. So this outcome-based pricing, I think, is very sound. Because we're not experimenting. We're shipping agents that are now helping millions of users.

John Furrier

>> Time to value is time to money.

Erwan Menard

>> Absolutely.

John Furrier

>> Great to have you on. Congratulations again. Love the AI infrastructure. Again, you guys got to go faster. You're pedaling as fast as you can. Congratulations. Great to see

Erwan Menard

>> you.Don't we know it.

John Furrier

>> You're turning into a mega cloud.

Erwan Menard

>> Well, it's a great adventure, and it confirms that if you want to do AI well, you need to focus on it. So I think being an AI purpose-built cloud at this moment is nothing but a

John Furrier

>> privilege.And the real-world opportunities are hitting too this year. And next year, physical is going to explode. So you're ahead of that. Congratulations. Good to see you again.

Erwan Menard

>> Good to see you, John.

John Furrier

>> I'm John Furrier, your host of the CUBE here at the RAISE Summit in Paris, France. We'll be right back after this short break.

Erwan Menard, Crusoe

search

John Furrier

>> CUBE Studios connecting, Silicon Valley and Wall Street.

Erwan Menard

>> I'm John Furrier, the host of the CUBE here with Gabe Olave, my co-host. Welcome to the RAISE Summit coverage brought to you by Celadon with support from Tensordyne, d-Matrix, and Argentum.

John Furrier

>> Welcome back to the CUBE, everyone. I'm John Furrier, your host here in Paris, France for the RAISE Summit. This is where all the action is turning out to be an international global AI infrastructure and AI enablement accelerating applications. Everybody's here across the world talking about the growth of AI, particularly the applications are starting to emerge, the real use cases from physical robotics. Everything's happening. Of course, the AI infrastructure is powering it. Our next guest with Crusoe, they're one of the big reasons why it's all working. Erwan Menard is here, SVP of Product. He's the product czar. At Crusoe. Super busy. Thank you for taking time out of your day. I appreciate it. Great to see you again.

Erwan Menard

>> always a pleasure.

John Furrier

>> head of product is a big role, but you're under a ton of pressure. I just wrote a post 2 days ago around this. We're midpoint in this token economy.

Erwan Menard

>> Yeah,

John Furrier

>> I use you guys as an example because you guys are moving so fast. First, you have the big data center in Abilene, Texas. I kind of use that as a backdrop. But it's not just big capacity. You guys are delivering now services. You have to turn on the product chops, start meeting the demand. Yeah, because you got to— you can only build it as fast as you can build. Well, where are you with the product? What are you guys shipping? What's happening?

Erwan Menard

>> So thank you for mentioning the data center roots. this is a vertically integrated company. So knowing how to build a gigawatt campus is certainly one of our expertise. But to your point, we've been shipping GPU clusters over the last few years, typically managed Kubernetes form factor. But the big news is that we're now playing in inference and helping people modify an open model, make it theirs with their data, and then inferencing it the most effective way. And doing all of this in a managed service fashion that you don't need to get involved with managing the GPUs. So managed inference, managed fine-tuning.

John Furrier

>> Yeah, and I think the demand is forcing that function because we're in a speed game, speed of delivery, which is building. I was just talking to AMD and Cerebras. They got to make their manufacturing plants faster. They got to stand up new ones faster. You got to build faster. You got to deliver services. That's because of the demand curve.

Erwan Menard

>> Yeah.

John Furrier

>> What has changed on the demand curve? Is it more cloud-managed services? Is it more inference tied to specific use cases? Has it changed the target audience? Give us a taste of what's the most impactful change in the past year.

Erwan Menard

>> I think, of course, we still have a lot of training going on, and this has been the genesis of the business over the last few years. But the fact that people are building agents going to production, their number one cost is inference. And if you look at the open models and how quickly they improve, You need to be able to tune the model, quickly retune the new model starting from a better baseline. And I think that's where we need to help. you pick a model, GLM-5.1, Kimi K2.6, whatever the model you like. Quickly you get your data, you derive your own version, and it becomes the brain of your agent. And you need to do that loop very often.

John Furrier

>> I don't like to brag very often, although I do pat myself on the back, but when 3 years ago, Dave Vellante and I published the power law of models. Yeah, you had OpenAI and Anthropic. Google was just coming on the scene with Gemini. And we put out a thesis that said there'll be a power law, there'll be a kind of a head and then a torso and a tail of models. And the thesis was specialty models or small language models. And if you look at Alex Karp's recent rant on CNBC, if people don't know what that is, just search Alex Karp. He's the CEO of Palantir, he basically lifted up the curtain or pulled the curtain back on some of the big proprietary, big frontier models, OpenAI, Anthropic, saying you don't need them only, which validates one, our thesis. But that's been validated. But this is now coming into practice. You're seeing people build their own models. Open source has been a big driver of that. You're seeing a lot more concentration of the value proposition in the models. Decoupled from the hardware. We've been covering that on the CUBE for a year. What's the impact to you? Because now you start to see real engineering going on at cloud scale.

Erwan Menard

>> Yes.

John Furrier

>> You're a cloud company by default, because you have the hyperscale infrastructure.

Erwan Menard

>> The impact is that you need to invest so much more in the software stack to help run these models cost efficiently. The inference engine we've built, I can onboard a new model in a matter of days, optimize to deliver throughput or latency, whatever your preference, and serve your business at the best cost factor. And you need to do that quickly. And what's interesting is that instead of being a static software piece, like everybody would download the same thing, the techniques are evolving almost week by week. So you need researchers, forward deployed engineering to give that inference software magic to the developer in a few clicks.

John Furrier

>> We're at RAISE Summit. This is more of a big industry show. And you started out by saying Kubernetes, which is oh, I got triggered. That's KubeCon, that's Linux Foundation, that's cloud native.

Erwan Menard

>> Yes.

John Furrier

>> Okay. So you're bridging cloud native to large-scale cloud native with AI workload-like capabilities. Explain why that's happening and what that means for a customer. Do they onboard like a cloud? How do they engage? Take us through this. Yeah, this is now next level.

Erwan Menard

>> You log into a console, you want to have a cluster of X many GPUs. And there they show up as a managed Kubernetes space where basically under the hood we're self-healing a number of issues. I'm sorry to say GPUs continue to have issues. It's very fragile hardware. And so the value out of the service provider is to detect those faults ahead of them happening, self-healing them, giving you options to avoid ruining a training run or affecting inference quality.

John Furrier

>> How has the— in the past 5 months, your business grown? What's been the customer feedback? What's been the learnings?

Erwan Menard

>> The business is on the infrastructure as a service side, it's the adoption of this managed Kubernetes. The clusters and the fleets are becoming so big that people need help to manage all this hardware. So that's now 70% of our fleet that is managed Kubernetes. On the inference side, it's the model lifecycle. I pick a model in our Intelligence Foundry, I modify it with my data, serverless fine-tuning, and I inference it with the latency or throughput recipe. And in a few clicks, I've been able to power my agent with my version of Kimi K2.

John Furrier

>> cloud is not easy, as they say in French, it's très difficile.

Erwan Menard

>> You're tapping into my dual citizenship here.

John Furrier

>> It's très difficile for Americans. Europe is hot. Sovereignty is here. How are you guys looking at the sovereign equation? There's a lot of cloud discussions because AI is not just a privacy problem. There's energy and the economics involved. That's money. You start to see people digging in in their countries, but they still got to run globally in a distributed way.

Erwan Menard

>> Yeah.

John Furrier

>> How are you guys looking at the European expansion? What are your thoughts there?

Erwan Menard

>> So we're already present in Iceland because we're energy-minded. So a lot of geothermal energy. We have a big presence in Norway. There is great hydro energy. And we're looking at projects in Western Europe. You're absolutely right. I think it starts with data and compute residency. I'm able to compute and save or store all the prompts and outputs of the model locally. we also have an interesting form factor with modular data centers that I can give you as a company, full exclusive use while you control it with a global control plane. I give you an edge zone, a Crusoe Edge Zone, that's a container mini AI factory. We manufacture ourselves and that gives you that data and compute residency. That's how we're approaching the topic. I think open models are also part of it because open models mean control. I control the version of the model, I control when I retire or launch a new model, and I have full control on the training, post-training, reinforcement learning of the model. And I think that plays a great role in those sovereignty requirements.

John Furrier

>> I had a couple of customers that do large-scale work say to me, hey, why don't you just switch your model? He goes, well, John, it's not that easy, because there's cost involved. Talk about some of the, I won't say tire-kicking kind of deployments, but people kick the tires, check things out, but there is cost to change.

Erwan Menard

>> Yes.

John Furrier

>> Can you share your insights around what you've learned on that? Because people might think, oh, I'm going to do this for a while, and next thing they're over their skis.

Erwan Menard

>> I think you're hitting a very important point. Yesterday, I was catching up with a customer who's a coding assistant company, They're sharding models all the time because new models show up, they improve their own models. And so our ability as a service provider to quickly change the API endpoint, stage a new model is actually mission critical. So the way we try to help with that is that we give you an abstracted service. You have an API endpoint, we take care of sharding the model, reconfiguring the GPUs, and we're able to do that very quickly so that you can take advantage of the new version of whatever model coming out.

John Furrier

>> Has data scale come up a lot with some of your customers, especially in physical AI?

Erwan Menard

>> Because you're going to have a lot of data preparation, you got so much feedback coming back from those devices that you want to bring into a reinforcement learning loop. So you have data preparation pipeline. We're seeing a lot of CPU, general purpose CPU usage side by side with the GPUs as well as a lot of storage for the sake of preparing data and feeding reinforcement learning.

John Furrier

>> Not to get in the weeds, but I have to ask, has data pipelining changed? the old days it felt like you dig a ditch, you put the pipes down, you put the grass on it. Yeah, static. Now we're in a generative model mode where you got all this inbound data coming in, especially on inference.

Erwan Menard

>> Yes.

John Furrier

>> What's been the data pipelining direction that you're seeing good use cases and good practices around?

Erwan Menard

>> First of all, I think you need to be very thoughtful about sizing. We used to deploy let's say 4 or 5 terabytes per GPU. I'm talking storage. We're now way in excess of 10 terabytes. So first, because of shortage in supply chain, plan, have proper CPU and storage capacity beside the GPU. Second, I think we have a lot of real-time pipelines that are in the works. We work with physical AI companies Figure is a customer as an example. I think we have a lot of value we can build. Over the next few months for these customers as they start getting a tsunami of data back from the robots that are in production. And I think the whole data pipeline needs to be reinvented. At least we're doing a lot of research work there to be able to quickly, at the right cost, find the right data and bring it in and generate synthetic data to augment the dataset and again feed the reinforcement learning.

John Furrier

>> You brought up physical AI, I want to just I'm going to double click on that for a second because I think physical AI is a telltale sign for what's coming fast. And it's not just robotics, I love robotics, but you got cars, you have other things. It shows the real-time nature, and physical AI will move to everything, retail, telecom, anything physical. And those edge devices are going to throw off a lot of new data.

Erwan Menard

>> Correct.

John Furrier

>> It might need to be trained, reinforced, processed.

Erwan Menard

>> And actually, at first filtered.

John Furrier

>> Yes.

Erwan Menard

>> What do I want to keep? Because it's a very noisy data flow. And then you need to generate synthetic data out of it. you get a rare situation that you want to create derivatives on. So the whole data pipeline has to be faster, has to filter almost on the fly, has to generate synthetic data. these are new challenges.

John Furrier

>> All right, vertical integration is a big thing. You guys are vertically integrated. Are customers going down that route? How are they thinking about their stack right now? As these architects are looking at building their next generation, this is generational. This is not hey, let's bolt onto a data center. Yeah, there's a little bit of that.

Erwan Menard

>> Sure.

John Furrier

>> But the real work is, I need to architect my business future globally and/or regionally or wherever. These are big decisions.

Erwan Menard

>> Big decisions. I think the first bifurcation is, Do I need infrastructure for training? Do I want to control lower in the stack, or do I want to rely on a managed service abstracting the infrastructure management? The second thing is, who am I working with for a time to market that's going to hold? Because the supply chain has stress everywhere— SSDs, memory, GPUs. You want to work with people who can give you a date by which the cluster is available and you can rely on it. And then the next aspect is the software stack. This inference engine conversation, it's amazing. We have so many customers who are testing us against all the other options and it's really fine-grained optimization with forward deployed engineering that gets you to the right balance of throughput and latency for the token price.

John Furrier

>> On my post 2 days ago, which I featured you guys in and others about the infrastructure enablement, I made a statement kind of buried in the post. It's kind of like an Easter egg and I was testing out the concept. But I said tokens are the new bandwidth. And I got some reaction from some senior engineering people. It was a dog whistle kind of a phrase. But my thesis was networking becomes super important because you got to move packets. Sure, that's networking in the classic sense. But tokens are now money.

Erwan Menard

>> Yes.

John Furrier

>> Money moves. So moving tokens is going to be a thing, I think. What are your reaction to that? Because that brings up the technical concept of networking, which we know what's going on with KVCache and memory and moving data into the right areas at the right time. Moving tokens and the role of networking, a new kind of networking. How would you describe this phenomenon of the new kind of network? This isn't your classic connect the campus. And move a packet from point A to point B. It's a lot of touch points, omnidirectional, a lot of threads.

Erwan Menard

>> I think you have that aspect. I think you said the word cache, which I think is the first order of optimization we need to deliver now. Because if you keep the cache big enough and fast enough, you can save a lot of recomputing. And so you save, you mask the networking challenge. Then comes the storage challenge, and I think networking is next. And if you think about building a cloud, we were very focused on GPUs, managing all their failures. We have that more or less under control right now. new SKUs come all the time, and you need to be humbled and do proper new product introduction. But then it's the cache. we've invested so much in KVCache, bundling the memory, giving you a very powerful cache. I think networking is the next topic once we have that cache thing under control.

John Furrier

>> All right, I have to ask you, since you're the product guy, I love the phrase— I've loved, actually I should say past tense, I love the phrase product market fit. My new phrase is product money fit, economics, because tokens are money.

Erwan Menard

>> They are.

John Furrier

>> So value is whatever we want is getting out of AI. So the value coming out of AI is better productivity, but that translates to dollars.

Erwan Menard

>> Correct.

John Furrier

>> Okay, so how do you guys view that? Because at the end of the day, these workloads are built for a purpose.

Erwan Menard

>> Correct.

John Furrier

>> To drive a value.

Erwan Menard

>> Yes.

John Furrier

>> That produces an income.

Erwan Menard

>> It's a really good point. In this managed inference business, we currently go into agreements with customers which are about a certain amount of spend because they can correlate it with their revenue tied to the agent solving a real-life problem. And then underneath, we're going to tune the model, we're going to change the configuration to take advantage of technology improvement, but we're not talking about Can you give me this model at this throughput? From a price point of view, we look at, let's create this value for your agent managing the biggest cost item in the agent cost structure. And then we'll tune the technology to serve that business goal. So this outcome-based pricing, I think, is very sound. Because we're not experimenting. We're shipping agents that are now helping millions of users.

John Furrier

>> Time to value is time to money.

Erwan Menard

>> Absolutely.

John Furrier

>> Great to have you on. Congratulations again. Love the AI infrastructure. Again, you guys got to go faster. You're pedaling as fast as you can. Congratulations. Great to see

Erwan Menard

>> you.Don't we know it.

John Furrier

>> You're turning into a mega cloud.

Erwan Menard

>> Well, it's a great adventure, and it confirms that if you want to do AI well, you need to focus on it. So I think being an AI purpose-built cloud at this moment is nothing but a

John Furrier

>> privilege.And the real-world opportunities are hitting too this year. And next year, physical is going to explode. So you're ahead of that. Congratulations. Good to see you again.

Erwan Menard

>> Good to see you, John.

John Furrier

>> I'm John Furrier, your host of the CUBE here at the RAISE Summit in Paris, France. We'll be right back after this short break.