Every conversation about AI eventually runs into the boring question underneath it: where does this stuff actually run? On this episode of the Cloud Build Podcast I sat down with Erwan from Crusoe’s product management team to get an answer. Crusoe is a NeoCloud, meaning a cloud built for AI and nothing else. It started as a Bitcoin mining company. Two high school friends wanted to monetize the gas being flared off on oil fields, built modular data centers to do it, noticed what GPUs could do, and in 2025 sold off the mining business entirely. Today Crusoe has five gigawatts under management and a pipeline of 40. Erwab is a telecom engineer who spent years building AI services on hyperscalers without once thinking about where the electricity came from, which turns out to be a useful place to be coming from.
Five things worth pulling out
Why do we need a new kind of cloud? Hyperscalers are already good at this.
Power density broke the old assumptions. A hyperscaler rack from under a decade ago drew about 15 kilowatts. An Nvidia GB300 rack draws 130. Everything downstream of that number changes: the building, the floor, the cooling, the networking. Then it changes again in software. If storage goes down mid-run on a thousand-GPU job, weeks of work is gone, so the components everyone stopped thinking about are suddenly load-bearing. Erwan’s phrasing: “storage is cool again.”
What’s actually holding things up?
Energy. There are more projects underway than there is power to run them. Crusoe’s answer is to own every layer and to staff the leadership team with energy people, data center people, and software people, who all think on wildly different timescales. Energy people think in years. Software people ship tomorrow. Sites being signed now deliver in late 2027 or early 2028.
Who’s using this, and did that turn out how you expected?
No, and Erwan said so. He joined assuming AI natives would want raw hardware and nothing else. Give me the GPUs, I know what to do with them. Managed Kubernetes now runs 70% of the fleet, because once your fleet is big enough there are failures constantly, your researchers need the thing available all the time, and even an infrastructure engineer will take the help.
Is anyone really moving off frontier models onto open weights?
At real scale, yes. Erwan described a coding assistant company that cancelled a frontier lab contract worth triple-digit millions to move that workload onto an open model served by Crusoe. Cost was part of it, but not all of it. They also wanted to decide when the model gets retired, and to modify it.
And data centers in space?
Crusoe is participating in the assessment and Erwan finds the energy case genuinely interesting. He is also the person watching GPU failure rates and writing software to catch problems before they kill a training run. His line: “we need to perfect the basics of that infrastructure on Earth, because I’m not going to be able to ship somebody to repair the GPU.” He wouldn’t put a number on the failure rate, only that it’s “still pretty frequent.” His counterweight is Starlink. He’s a telecom engineer and a sailor who spent years on bad satellite service, and fifteen years ago he’d have told you it was never going to work.
Watch full episode on YouTube
Full transcript
Lightly edited for readability.
Priyanka: Welcome back to the Cloud Build Podcast. Today we’re talking about a very key, important part of AI, which is: where do we run all of the AI models and inference? For that, I’ve invited the product management team from Crusoe. Erwan, how are you?
Erwan: I’m very well. It’s fantastic to join you today.
Priyanka: Tell me a little bit more. What is Crusoe, and what do you do?
Erwan: Sure. Crusoe is a so-called neocloud, meaning a cloud purpose-built for AI. We do AI, AI only. We help you train and inference models and build agents. That’s what we’re focused on. The name comes from a very, very early focus on energy, like Robinson Crusoe on his island. That’s where the name of the company comes from.
Priyanka: So take me to the history. Where did Crusoe start from? Because it’s a very non-traditional path to AI, and I’m excited to dive into where it came from and what it became today.
Erwan: It is non-traditional. I think the moment allows for it. You have so much disruption in the tech stack and the infrastructure and the demand that there is space for nonlinear trajectories for new players.
So, Crusoe was born in Bitcoin mining. Two founders, high school friends, had early careers separate from each other, but they met and decided to go after Bitcoin mining using specific sources of energy that were completely wasted, like flaring gas on oil fields. And so they started building modular data centers. Then they uncovered GPUs, and how helpful they could be. And eventually in 2025, Crusoe completely divested its Bitcoin mining business to go full steam on a business that started around 2022, which is AI infrastructure, an AI-dedicated cloud.
Priyanka: That is very cool history. Now, describe “neocloud” and what AI infrastructure means, and why was this not here before? The need for the new clouds.
Erwan: I think the simplest way to put it: it’s a combination of very profound disruptions in the technology that you need to mobilize, and the vast amount of demand.
Clouds are amazing, hyperscalers are amazing. They’re general purpose. They can do a ton of different workloads. They can take care of your ERP and your website. But here’s the thing. When you do AI, you’re using hardware that has a power density we’ve never seen before. Take a data center from a hyperscaler less than ten years ago. A single rack, one of those big fridge-style boxes where you stuff all the compute, was typically requiring 15 kilowatts per rack. You take a GB300 rack from Nvidia, which we’re shipping every day, the shiny GPU of the moment, and you’re at 130 kilowatts per rack.
The consequences of that simple fact are plenty. My data center is different, the floor is different, the way I cool is different. And then of course the density of compute has a lot of implications on networking and so on. And so far I’ve touched only physical stuff.
Now let’s go logical. You’re training a model. You mobilize thousands of GPUs to run jobs for multiple days, sometimes multiple weeks. If storage is down, your run is ruined. So even within the components of cloud infrastructure, the dependencies are changing. Storage is cool again. I’m a former storage guy, but storage is cool again, because utmost availability is mission-critical to extract value from very expensive GPUs.
And when you go even further up the stack, you have major disruptions as well. Take a topic like inference. There are many different ways to do it. But now that people are in production at such a large scale and really care about the cost per token in software, we find a lot of ways to improve that. Whatever is true at day n, you seriously want to challenge at n plus six months. And that pace has not slowed down since the GPT moment. So: disruption, innovation, every layer of the stack, you need to reinvent it all. And it’s kind of cool to start from a blank page.
Priyanka: And why can’t the hyperscalers or the other data center providers do this? They already have the infrastructure, the locations, some sort of cooling mechanisms, the physical things you talked about. Why can’t they just spin up more of this for AI workloads?
Erwan: Actually, they do. Crusoe is an AI purpose-built cloud, but we have another line of business where we build data centers only, at very, very large scale. We were the developers of OpenAI’s first target data center in Texas. We’re doing similar engagements with Microsoft and others. Think of that as gigawatt-scale campuses.
And so even these hyperscale players found value in engaging with a new dedicated company who would rethink that entire data center design A to Z. And we’re happy to help them there. I think they have a challenge we don’t have, in repurposing old data centers. I’m really glad I don’t have to get involved with that, I’m sorry to say. But I think they’re formidable contributors to the industry, and the fact that they are themselves attempting a self-disruption is pretty telling about the moment.
Priyanka: We are living in a very interesting time, where everything from the infrastructure to the storage layers to how we connect to the rest of the world, and the apps that already exist, legacy and new, how do we bring all of that together?
Which brings me to this. You talked a lot about the GPU aspects, but agents require a lot of CPU and compute, basically taking a request, queuing it up, and then sending it to the GPU when you really need to get the inference from the AI model. Walk me through that flow for an application that runs on Crusoe.
Erwan: Actually, we’re still very GPU-centric in the way we talk about the business, because training has been the historical workload. Now inference represents more traffic, but is also using a lot of GPUs, and GPUs are a very expensive asset. But you’re right, the amount of agentic functions and data pipelines that are happening in the very neighborhood of that GPU cluster is growing.
If I look, for example, at a metric like how many terabytes of storage we put per GPU at a full data center level, we used to go one GPU, four terabytes. We’re way north of ten terabytes now. Why? If you take a physical AI workload as an example, you’re going to inference, you’re going to train the models, then you get those models to run on the robot, and the robot comes back with a lot of data that you want to clean up in a data pipeline and inject into a reinforcement learning loop. A lot of those tasks around the data pipeline are CPU-hungry and storage-hungry, as opposed to GPU-hungry.
And then you take an agent. Everything you put around inference itself is typically running on CPU as well. So the infrastructure is getting much more diversified than GPU clusters, as we take AI into production and build capabilities beyond the model.
Priyanka: And in those types of use cases today, does the user have to interface with wherever the rest of their infrastructure is, and come to Crusoe for the inference and the training and the fine-tuning aspects?
Erwan: We welcome two kinds of users, because we spend 95% of our time or so with AI natives.
Sometimes we find people who are incredibly opinionated, want a low level of control and access, turning a lot of knobs. These are very much infrastructure engineers. They have a team of researchers or developers who tell them, “can you optimize this or that?”
But the second persona that has emerged very strongly in the last year is more like the agent developer. These expect services that are more abstracted, like managed services. They don’t need to be involved in even selecting the GPU, but they want to be sure that they rely on a state-of-the-art stack. So our inference, for example. GLM is kind of one of the most popular models right now. We extract 430 tokens per second, any time of the day, and there are leaderboards for that. This is reassuring for a developer who depends on that model performance, both throughput and latency, and most importantly cost, to power the agent. So having an API endpoint from us where the model of their choice, even their modified version of GLM, can be served at the right performance and cost, that saves them time. They can spend more time on the harness and the business process they’re trying to help.
So that second persona is why we’ve developed more managed services, where the knowledge is served not as a configurable set of knobs but more like a prepackaged API endpoint with an SLA, and you can go focus on the upper layers of the stack.
Priyanka: And we see more and more of this emerging, with people wanting to use open-weight models, do customization on top, and just have that deployed and hosted themselves. I don’t think we keep talking about the fact that “oh, I’m going to deploy my own model in my own infrastructure,” which to me is always going to be one of the clouds. And if this cloud is the Crusoe cloud, optimized for AI workloads, then you want to go use the cloud that’s optimized for AI workloads, rather than having a rack in your own room and desk, because that requires maintenance.
Erwan: Yes, right.
Priyanka: So I see a world where, and I think today about 25 or so companies signed this letter to America, Nvidia is one of them, encouraging more open-weight models. This is the kind of push I see the future having, where people want the right amount of control, which is what you’re talking about. But they also want to offload a lot of the running of the open-weight models.
Erwan: I mean, we all remember Llama 70B, end of 2023. Then we were like, oh wait a minute. Open models can do something very, very valuable as compared to a frontier model. Then we got the DeepSeek moment, early 2025. I would argue that GLM 5.2 and Kimi K3, that’s kind of like the next moment.
And so now, I was chatting with a coding assistant company who’s using all sorts of models depending on the prompts. But they cancelled a frontier lab contract, triple-digit million dollar, to move that particular part of their workloads onto GLM 5.2, to be inferenced by Crusoe, because they have control. They decide when to retire the model. They can modify the model. And the cost structure is much more competitive to their product, the coding assistant.
Priyanka: Exactly. We’ll keep seeing more and more of that.
Now, I want to contextualize this discussion with Erwan by sharing with you that training models on your own data sounds easy, but it’s actually a big headache. Now you’ve got to deal with GPUs, manage infrastructure, and babysit clusters. Crusoe’s serverless fine-tuning takes care of all of that for you. So all you have to do is upload your data and the model tunes itself. Here’s what that actually looks like inside the Crusoe platform.
Fine-tuning lives right in the Develop section. First step is to pick your base model. I’m going with Qwen 3.5 here, but there’s a range of open models to choose from. Next, you load your training data. Just browse and select your files. And then, just as important, is adding your validation dataset, so the platform can check how the model’s actually performing as it trains. And from there you’ve got a few tuning parameters like the batch size, the learning rate, the number of epochs. Dial those in based on how aggressive you want the model to run, and hit start.
Then Crusoe spins up the infrastructure behind the scenes. There’s no GPU provisioning that you need to worry about, no cluster management. It’s just training on its own. And once the training wraps up, you will want to evaluate the results before doing anything else. And when you’re happy with them, that’s your cue to move to self-serve deployment and spinning your fine-tuned model up as a dedicated inference endpoint, which is ready for real world use in your AI apps. You select the parameters, the number of replicas, and then you deploy. I highly encourage you to try this yourself. The link is in the description.
Going back to the data center aspect of it, what is, in your experience building Crusoe, the biggest bottleneck right now at the data center level?
Erwan: Supply chain constraints at every level. I think the ultimate bottleneck is energy. Very clearly there are more projects underway than available energy.
And I’m a software person. I spent all the recent years building AI services in clouds, frankly taking GPUs or TPUs for granted, and certainly never thinking about the energy. Today I work at a company where, around the leadership room, I have people who spent decades in energy sourcing and production. People who spent decades in acquiring land and permitting land to build something. People who spent decades building data centers. And of course then all the cloud people show up.
I don’t know how you do that well if these relationships are all supplier-customer type stuff, with everybody trying to carve out their piece of margin and having their own agenda. Because to land a site, and right now we sign papers to deliver sites late 2027, early 2028, you need to synchronize body clocks that are so vastly different. Energy people, they think in years. Data center people used to think in years, they think in quarters now, at least the Crusoe ones. And then of course software people were like, I can ship another version tomorrow. And so synchronizing those clocks is a true leadership, human, entrepreneurship challenge that I think is key if you want to deliver on time. Because in an energy-constrained environment, your ability to deliver that compute capacity on time is all that matters.
Priyanka: Timing. And how are you managing that with the vendors and suppliers, especially on the energy side, because it’s years?
Erwan: So right now we have five gigawatts under management, deployed or under active construction, and we have a pipeline of 40 gigawatts of energy.
Priyanka: Wow.
Erwan: And that’s the way you do it. Then you consider options of being behind the meter, so you produce energy on site to augment what the grid can give you. We’re also very good at hunting, I would say, the most climate friendly sources. So we have a data center in Iceland because we have fantastic geothermal energy. We’re in Norway because you have hydro energy. In Nevada, I should invite you there, we have modular data centers powered by recycled EV batteries, themselves coupled to a solar farm, and we run a lot of managed inference from there.
So we integrate every layer, like energy, data center with our own modular data centers, GPUs, and the whole software stack to inference your favorite open model. And I think that vertical integration will pay off even more in the future to deliver capacity on time.
Priyanka: Is that going to be a competitive differentiator for companies?
Erwan: It feels like it. I mean, it’s a very humbling market. I think we’re all learning every day. But these modular data centers, we call them Spark. Now that I see them being deployed, and predictable timelines between that container exiting our own manufacturing facility and then ending up in the field, powered, GPUs powered, software stack on it, tokens coming out of it, and we have line of sight on making this predictable and being able to turn on sites at an industrial scale. I think that’s a terrific advantage.
Priyanka: That’s amazing. And how do you decide where you put these Sparks?
Erwan: Geographically it’s obviously a combination of demand and optimizing constraints. There is, I think, a growing case for inference at the edge. So the gigawatt factories are amazing for training and global inference. But there is a case for breaking down the inference distribution into smaller data centers. For example because you want the last mile latency to be optimized. And from a building point of view, at let’s say 20 megawatts, 40 megawatts, the acquisition, site preparation, permitting, energy sourcing, all these problems are much more manageable than at a gigawatt scale.
Funny enough, just using my product hat, the per megawatt economics are highly comparable. So obviously 20 megawatts is much smaller than a gigawatt, so you perform different jobs. But on a per megawatt basis, if you believe like me that energy is the ultimate unit, we’re talking about monetizing megawatts. And so if I can give you the same economics in terms of cost of the whole stack, energy, data center, GPU, software stack, if I can give you the same economics at 20 megawatts with modular data centers as at a gigawatt factory, that’s pretty compelling. Because then cost is not exactly the factor to decide to go small or big, and you can really optimize site acquisition, preparation, permitting, energy sourcing.
Priyanka: And multi-performance. Exactly. Awesome.
If you go back to, because I was reading a paper that you just wrote on electrons to tokens, if you can walk me through it, because I don’t think most people understand what that means. Walk me through electrons to tokens, that thought process, that philosophy.
Erwan: Okay. So the currency of the AI economy, for you, me, and anybody using or any builder building an agent, is a token. So we all care about dollar per token, and then latency per token. Everything is per token.
But as we quickly touched on earlier, the ultimate bottleneck, the most critical piece in the supply chain, is the energy. And so if you connect these two dots better than your competitor, you bring more value to your client. That’s why we force ourselves to think of the stack from electrons, whether I get them by plugging into the grid, or I build a wind farm, or I get creative to be among the first to use nuclear energy in the years to come. All this research and development on energy is what allows you to optimize the whole chain.
And so we really want to be a company who can think from electrons to tokens. We have the same chips as the other cloud providers. We’re not building our own chips. But if we optimize the energy, if we optimize the data center, if we optimize the software stack, eventually with a simpler margin structure, because it’s not like six companies piling on top of each other, we’re going to deliver more reliable and cost-effective tokens.
Priyanka: That is very insightful. And then if I extend this idea to the concept that I keep hearing about from Elon Musk as well. He’s talked a few times about putting these data centers in space, and it sounds compelling because of the harnessing of the energy. But curious on your take on what that could look like, what that means, putting a data center in space.
Erwan: I think actually that topic is really good to illustrate the challenge of building these clouds.
I am, and we are as a company, really interested in the energy aspect of this. I am also an operator who’s watching the failure rates in the GPUs, and building all sorts of software to detect these problems before they block a training job and proactively suggest to our customers to swap, and so on. Having the promise is amazing, and we need to perfect the basics of that infrastructure on Earth, because I’m not going to be able to ship somebody to repair the GPU. I don’t want to give you a number, but it’s still pretty frequent. Let’s just put it this way.
But we decided to participate in that effort of assessing the opportunity. Because if I would have told you 15 years ago that Starlink would have existed, you probably would not have believed it. I grew up as a telecom engineer, right? And I’m a sailor. I go on oceans, far away from land, relying on very bad satellite service, up to Starlink. And as a telecom engineer, I never thought this would get any better, for a bunch of technical reasons that we don’t need to cover now.
So I think we need to also recognize that we are at an amazing moment of innovation. AI accelerates innovation. So I think as a company we’re recognizing that long innovation cycle paying off, and so we are invested, we are participating, we are exploring. And at the same time we’re very focused on getting the GPUs on Earth to not hurt our customers’ jobs as much as they fail.
Priyanka: So the idea is to perfect the art here, and then once we are at a point that we have less failure rates, and also automation on those failures so that we can self-correct them, that’s when it makes sense to do that, right?
Great. Okay, so coming back on Earth. The open source models, the open models, fine-tuning and serving your own models, is really where most people are using a lot of the GPU and the compute. What part of AI workloads is Crusoe helping companies accelerate?
Erwan: At the core, it’s scale and availability at scale. If your job is big, I promise you, you don’t want to manage those GPUs individually yourself.
Now from a developer point of view, it’s also the opportunity to do easily the typical task of optimizing the model to your business. So there is a given open model, GLM 5.2, Kimi K2.7, whatever. You like it because you like the benchmarks. But then your next step is to apply it to your use case. Then you uncover that the model is maybe better, or less good, than you thought. Then you can take your data and start having the model operate within your context. Then you can flank the model with a corpus of information that serves as context. Then you can distill the model to specialize it on a given task.
There are many things you can do from the open model. It’s a substrate from which you can derive your model. And if you’re good at this, this is your model, not the one of your competitor. You have really taken that substrate everybody has access to and derived a differentiator for your business.
We want to help you do all these operations of model lifecycle, model optimization, in a cloud where you don’t need to worry about the GPUs being available and running well, and your job not breaking. We want you focused on that model optimization and then the agent development. So we have an Intelligence Foundry, which is kind of a user surface where either you bring your model or you pick from the library of popular open models, and then you start serverless fine-tuning it. You can evaluate the model. Later this year we’re bringing you reinforcement learning capabilities. And our ambition is to help you build a library of models for your business, starting from those open models as a baseline, as a substrate.
Priyanka: And it is your IP, and you own the context, you own the data, it’s all in one place.
Erwan: Yes.
Priyanka: When we, you just touched on physical AI and robots. If we could drill into that a little bit further, how does managed intelligence change when the workload is physical?
Erwan: So there are a couple of challenges that are, I think, very specific.
A lot of the models are on the robot. That model, which is kind of the brain of the robot, competes with the actions the robot is taking, from a compute and power management point of view onboard. So it’s fascinating to see, we have the privilege of helping companies like Figure and 1X, it’s amazing to see every research they’re putting into perfecting the models that are onboard and perfecting their energy consumption.
The next thing that is unique is that these robots bring a lot of feedback. Like, you have a lot of video recording of everything they do, and that is precious data. You can derive synthetic data out of it. You have one situation, you can create ten situations that are derivatives of it, if you have the right data pipeline. And then you need to have a way to clean the data and do targeted post-training runs that are intentionally trying to improve the model behavior in this or that situation. So this whole feedback loop is absolutely amazing.
And then I think the next frontier that is still to be written. But I think if we do a good job as clouds, we can not only help them train and take the feedback and improve the model, but maybe bring inference of some of these models close enough that they can alleviate the power consumption on the robot, and use the cloud to do part of the job. And I think this is fascinating. There is so much innovation there. And I just touched on kind of a robot-type physical use case. Physical use cases are so much more.
Priyanka: And what you were describing, I kept going back to, you need a data center like Crusoe’s, and a cloud like Crusoe, to be able to do so much of that feedback and back and forth, and then the loop of taking that data, cleaning it up, putting it back into a retraining of the model. You need that type of compute and GPUs.
Erwan: Yes. And the volumes involved, combined with the speed of innovation these companies aspire to, really deserve a cloud partner that is kind of dedicated. That’s the only thing we do. We’re not distracted by running an ERP or a website or whatever database.
Priyanka: I think that’s part of the co-innovation.
So if I’m a CTO today picking a neocloud, what should I be looking for? What are the questions I should ask?
Erwan: In terms of the services you’re going to consume, I think it’s important to have a panel of services from a low level of control to managed services. So, can you help me train or inference, and give me different levels of help?
The reason I’m going there is that when I joined more than a year ago, given that we work almost exclusively with disruptive AI natives, my assumption was that they would not be interested in anything but the hardware. Give me the hardware, I know better what to do. Today, for example, our managed Kubernetes offering is powering 70% of the fleet. Why? Because we’ve built some self-healing capabilities, detecting issues and taking proactive actions on the GPUs. And even AI infrastructure engineers, and I’m not talking about a model researcher, I’m talking about an AI infrastructure engineer, the fleet is bigger, there’s failures left and right, the researchers need this thing available all the time. So I’ll take help.
And so I think when you select the neocloud, of course look at what GPUs they have, but look also at what value-add they can bring, and whether their customers are finding value in their managed services.
And I think the second criterion is to think a few quarters ahead. Am I partnering with people who build their data centers and know how to find energy? And do they deliver that capacity on time? Because it’s very, very tempting today to say, “oh, I’ll get you 500 megawatts in a year and a half.” But there’s quite a few stars in the sky you need to align to get that done. It’s true knowledge, it’s true know-how. So you probably want to check if the people you’re partnering with are going to be relevant two years down the road as we continue scaling this industry.
Priyanka: Very insightful. Anything else you’d like to add as last thoughts?
Erwan: I think the fascinating part of this is the demand. I’m a telecom engineer by education. I saw the internet bubble. I am incredibly humbled by the demand, and also by the constant innovation. This is amazing. Every month there is something, and it’s been like this since the ChatGPT moment, and there are no signs that it’s going to slow down.
We could have talked about inference-specialized silicon, which is the next thing that’s going to take the token economy to the next improvement. We could have talked about nuclear energy, that is abundantly available in some countries, not so much in others, but it’s certainly extremely promising, and there’s a ton of innovation going on there. So to me, that massive innovation wave is both very energizing and very humbling.
Priyanka: And that’s very cool. The demand and where things are going. We had the SaaS moment, and now it seems like there’ll be this AI apps, AI agent apps moment, and everything in between that is all of this infrastructure that’ll make all of that possible.
So let’s do a rapid fire. I know you’re not ready for this.
Erwan: I love it. I’m ready for it.
Priyanka: Training or inference in five years?
Erwan: Inference.
Priyanka: Most overhyped claim right now in AI?
Erwan: All this debate about demand. It’s the least of my worries. There’s so much demand.
Priyanka: Besides Nvidia chips, the input that matters most in three years?
Erwan: Energy.
Priyanka: Vertical integration, real moat or expensive distraction?
Erwan: No, real moat. If you want to deliver on time with quality, you’d better understand everything from electrons to tokens.
Priyanka: Most underrated part of the AI stack?
Erwan: Energy.
Priyanka: Managed services, margin play or land grab?
Erwan: Lifesaver for the developer.
Priyanka: Great. Oh my god, you did amazing. Very, very exciting. Thank you so much for joining me today.
Erwan: Thanks for having me. Thanks for having Crusoe.
Priyanka: Thank you for joining us. And if you have any questions about Crusoe, neoclouds, anything else about AI infrastructure, drop them in the comments below. I will get back to you.
