The Moat is Where Your Data Lives | James Kurz (Filecoin Foundation) at ETHConf

I have a question. To start with, it's simple. Who has tried Fable 5 today? Has anyone had a chance to try Fable 5? Do you know what Fable 5 is?

Anthropic has released their new model. It's like Mythos Light. I tried. She is very good. Hmm, okay.

Another simple question. Who, uh, heard of Filecoin? Can you raise your hand if you've heard of Filecoin? Hands. Good.

Who thought much about Filecoin in 2026? Can you show your hands? Okay, great. Few people are really thinking about... Well, I'll try to get you thinking about it more in 2026, but we're mostly going to talk about the big macroeconomic forces macroeconomic forces that are affecting all of the AI ​​and infrastructure developments.

So, a little quick history lesson, and we'll put it all together all together . Uh, historically 10, 15, 20 years ago, computing technology was very, very cheap, and it wasn't the part of the stack wasn't the part of the stack that people thought about. You put hard drives where people live in Northern Virginia or the San Francisco Bay Area, and the computation doesn't matter that much. You buy some chips from Intel, you buy some chips from AMD, no problem. But now we have it too, and we all know that's not the case anymore.

Obviously, computing is what everyone is talking about. Perhaps an underrated part of history is data storage. Um, when we think about what's going to happen with artificial intelligence over the next five years, we're going to be collecting a huge amount of data about the world. Data that in the past we never wanted to collect because no one could look at it. No one could do anything about them.

But if you think about neuroscience and robotics, and of course you've seen it with self-driving cars, there's so much observational data in the world that's relevant to AI training and also AI inference also AI inference . And if we really want AI to permeate all of modern life and actually advance modern life, we have to collect data about it. We also have to process real-time data about the world we live in. And if you think about what we've actually collected over the last 30 to 40 years, that's a tiny, tiny fraction of the data that makes sense to collect in the era of artificial intelligence. And only we have the opportunity to collect a tiny, tiny part of it.

We need a lot more storage capacity, and storage is going to be a much bigger part of the story in many different ways. So, it's kind of the whole story mixed together. In the past, data and computation were in the same places because data was the hard thing. You brought the calculations to the data. Now, as everyone is saying in the markets, the aspiration is to bring computing to where the power is.

You have to find the power to actually power these GPU clusters , and that means you're going to places like Wyoming, northern Norway, Finland, Ohio, places that traditionally weren't the Internet's interconnections where compute and storage were located. And now all this power is being directed to these places in the middle of nowhere, and there is no storage there. So a solution is needed for this. And also, power is everywhere. It's not just in one place.

It's not just in 10 places. It's in 200 places or maybe even thousands of places even thousands of places , and that's the future of where data centers are going. So this means that there are storage deserts, and this is already relevant. This is not about the future. This is happening right now.

So, as you could see by scrolling through this list, the major GPU clusters that were being built last year are planned for next year, and they're all being built very, they're all being built very, very far from where the storage is located. If you need hundreds of petabytes or exabytes of data storage and don't want to pay exorbitant rates to the GPU cluster itself or AWS, you don't have a high-performance option. You don't have a close alternative. And this is a real problem, and I'll explain a little bit why. If a lot of people don't necessarily appreciate it because they don't think about it that much think about it that much , they'll start thinking about it when they try to train their big AI models .

When your data is very far from GPU clusters, you are dealing with the laws of physics. This is not something that can be solved with hardware. This is something that depends on the speed of the fiber optic cables, the number of nodes you need to jump over. And if the data is close to a single connection, physically close to the actual computations, you get much higher throughput. Clicker is a bit unstable.

That's it unstable. That's it . It is also extremely expensive. So there's a whole economic side to this that's not very well appreciated. Extracting data from many traditional data stores, such as AWS, is extremely expensive.

Why is it expensive? Well, because it can be expensive. It doesn't really cost that much money to put a lot of data on the line and move it. Um, but AWS charges you a lot of money. And the proof of this is that companies like coreweave actually give huge credits for moving your data to a coreweave server so that they are close and encourage people to actually move their data out of these bins.

The problem is that if you move your data to the core bucket, and we'll talk about that a little bit later, and you really want to train your data on a lambda server or on a cluster, because that's what's available, you're not locked into one vendor one vendor . It just shows you that it's really, really important to have your data close by and available really, and available really, really quickly. It's not just about the cost of exit. It's not just about disappointment. It's about the cost of iterations and whether you actually win in the AI ​​development cycle you're participating in.

If you can run more tests and run more benchmarks, if you can move your data to where the GPUs are, and iterate, test, try, and experiment faster, you are ahead of your competitors. So someone who has their data, their big data lake that they're training their data from, that's a thousand kilometers away, versus 10 kilometers away, or in the same building, or in the same meeting room, is going to have a big competitive advantage over someone whose data is very, is very, very far away. Why, uh, why isn't anyone addressing this issue? And the truth is, for many existing players, it doesn't really make sense. Hyperscalers don't want to address this issue because they operate at very, they operate at very, very large scale, and they need very large scale to justify and define these things.

Hyperscalers will spend a lot of time on analysis. They will spend a lot of time building very, very large data centers, large infrastructure. This will take them a lot of time. They are also very focused on building GPUs. And by the way, if you are a user, you don't want to be locked into one hyperscaler or even one NeoCloud anyway.

This is where Filecoin comes in. And we'll talk about that a little more. If you 're an on-premises storage operator, if you're in a data center as a colocated operator, or maybe you have a small data center, you have the ability to be where the compute is and actually benefit from it without having to buy your own GPUs. You don't need to spend capital and compete with Coreeave or Nebius for GPUs. You can simply be a data layer that is closer, faster, and more productive, and in those places where the GPU clusters actually are .

Hmm, just to add another level add another level , if there are lawyers here, I doubt it a bit, but doubt it a bit, but maybe there are. Hmm, there is also the issue of laws. So, because GPUs are built in places around the world where electricity is available, where electricity is cheap. And that sometimes means that this data that you're going to use for your training or for your inference, which I think is going to be a growing reason for creating data lakes, big companies will have all their data in one place that LM LLM will have access to . It must be in fact in the same jurisdiction as the place of calculation.

So if you have a cluster in northern Norway or maybe somewhere in Eastern Europe, you need to remember that this data actually needs to reside in the appropriate jurisdiction. And having a large distributed global network where storage is actually available next to the GPU clusters gives you that flexibility to ensure that you can meet the performance requirements that you need and what you need to do so that you can iterate faster. So, Filecoin is actually already this, and it's a big story that I think everyone should understand. We're already all over the world with our clusters, and we've built this up over the last five or six years, um, even more if you consider the development before that. Hmm, that's a really powerful asset.

You don't need to spend billions of dollars in capital investment capital investment because billions of dollars in capital investment have already been invested around the world. And in many cases, these storage device clusters are very difficult to obtain. If you look at some of the press releases from hard drive manufacturers, you'll see that they actually have order backlogs that currently extend beyond 2027. Well, all this infrastructure already exists. So the network exists and is available for use.

And as Filecoin, we're working to have this big, already built network in 2026, this big capital expansion, you don't have to build it yourself. You can use what we have and what we use. We've talked about this a little bit already , but honestly, it's really important to have a cloud agnostic. So there are other options for storing data. People think about it.

If you go to go to any of the big GPU-based neoclamers, they have data storage services. They are very expensive for their price, but also you are tied to this one cloud. Let's say you're migrating all your data from one particular company to a cluster in Ohio, but it's all already maxed out. They don't have a large cluster available to train on your data. I don't know if anyone has actually tried to reserve an instance on one of these NeoCloud websites, but they usually have the oldest hardware and the smallest quantities available, and everything else needs to be reserved for a month.

So you need to be able to move your data between different clusters very quickly so you can iterate quickly, be competitive, and build the best possible product for your customer, for your company, for what you're building. Therefore, an agnostic data store is critically important. Having an agnostic datastore that is really, really close to the GPU clusters you want to use is even better. So, Filecoin, as I mentioned, is a huge network that is built for exactly this. We have almost two exabytes of data capacity available worldwide.

We have shown in the past that the power of our network, provided by our crypto economy, allows us to further increase the size of this storage network. We built a very, very large network very, very quickly. Um, we clearly know how to mobilize a lot of power. So, by using these two exabytes, we can grow with them. And you know, we've already talked about storage availability as a serious issue.

Not only do we put storage where it needs to be, with the performance level we need, but we can also leverage excess capacity that is not being used, that is underutilized underutilized . And that's a really powerful force, because every free storage device, every hard drive, every NVMe is really going to need to be used in the coming years. So, everyone is talking about limitations and bottlenecks in computing. We talk about limitations and bottlenecks in memory and RAM, but this is something that is starting to emerge and will become very relevant because it is all interconnected, and Filecoin as a network can be a global solution to this. There are, um, two ways to try Filecoin in 2026.

If you haven't tried Filecoin in a while , I would, um, really, really encourage you to give it a try because there are some really new interesting things. Hmm, it's much easier to use if you've tried using Filecoin in the past and found it a bit confusing, maybe a bit frustrating. Hmm, that's not at all disappointing or confusing anymore, if it was at all. Um, the first file is kind of our entry into S3. So, this is a really good way to just replace your S3 endpoint and start using Filecoin services.

We're currently operating in the EU with a very, very high-performing, uh, region, and we'll be in North America with a high-performing region very soon, the next few weeks, and we'll be in Asia, uh, within the next few months. So we're going global with this, uh, S3 compatible tier for Filecoin and the Filecoin onchain cloud, which is ideally designed for developers, for designers, for anyone who wants to use smart contracts, uh, be programmatic in their, uh, work. Uh, this is what we've released over the last six to twelve months. So you really should check it out. Uh, I would really encourage you to try it, test it, compare it, and, uh, tell other people about it.

So, that's all. That, uh, ends the conversation. Uh ends the conversation. Uh , and I hope you learned something from this . By the way, I'd be happy to, uh, talk to you if you want to continue.

Be kind, friend. You can send me a telegram for free. Uh, but I think data is, uh, an underrated part of the story of the AI ​​generation and the infrastructure building that we're doing now. And the solution is not in, uh, hyperscalers building big data centers, or in creating, uh, data warehouses that lock you into one vendor one vendor . It's about enabling a global network of high-performance, but smaller, uh, and mid-sized, uh, data centers that can provide you with all the storage performance you need, right where the compute is.

And as this, uh, trend of building GPU clusters all over the world, including the most remote places the most remote places you can imagine, but where the power, uh, is growing, uh, Filecoin is going to be an important part of that story. Thank you very much.