DePIN’s Biggest New Deal: Valeo x NATIX | Alireza Ghods

Between May and the end of the year, we collected over 100,000 hours of driving data, which translates to over 600,000 hours of video data. The data size was over 1 pabyte.

We just closed a deal with Valeo, one of the largest automotive OEM suppliers, to build a world foundational model. We're very proud to work with them and build arguably one of the largest open-source multi-camera world foundational models.

A world foundational model can basically generate a simulated version of our reality with the physics of it. The simulated reality is like, if we took a couple of pictures of someone walking here, we can have the world foundational model have them continue to walk. So then it's like everyone that's a static image becomes a moving character, but it's simulated.

Welcome back to Proof of Coverage. Glad we get to catch up here maybe every 6 months or so. I think the last time was in July in Africa at the DePIN summit, right?

So maybe give us a couple of the quick high-level updates for what's new with you and Natix since July.

July was right after the launch of our VX360, which is the Tesla U dongle that enables you to collect multi-camera data using your Tesla. We ramped that up quite well.

Between May and the end of the year, we collected over 100,000 hours of driving data, so that's over 600,000 hours of video data. The size of the data was over one pabyte, which is insane.

Last quarter, a lot of our efforts have compounded, especially with respect to data sales, so we actually signed multiple contracts. We also did a very cool integration with a platform that Graph provided to us. We actually called it World Seek.

It's not public yet, but we did a demo to our community. It's a VLM based technology, so vision language model, and it enables you or anybody to search our entire imagery database with human-readable queries.

I think we did an announcement on it, I believe it was like mid-December. You can just write in any human-readable queries, even complex ones, and the visual language model is going to find exactly where it was.

You can even select the part of the map that you want the data in, and you can, for example, search for potholes or manholes or tilted traffic signs, or whatever that you want to.

I'm extremely bullish on VLMs in general, as the whole robotics world and autonomous driving world is bullish on because it's different from the traditional detection models.

Back in the time, 5 years ago, when you wanted to detect a traffic sign or a manhole, you had to have a lot of training data, train a computer vision model to detect pothole or traffic sign, and then run it through your data. And it would only do that job, nothing else.

When a customer comes in and says, "Hey, can you detect different traffic sign," you had to go through the process of data training, all of that again. And this was quite difficult and expensive.

With the VLMs, they are very generalistic models. With a human-readable query or a prompt, they can basically search the entire database in a quite efficient way and find whatever you're looking for, whatever a human eye would have seen, even complex scene analysis.

For example, find me, even in terms of a video, find me when a car slows down because of a bicycle in front of it. It's a very complex explanation.

You run that as a VLM agent on your video data and it actually finds you all of those cases.

I love when there's a tool that retailer average folks can demo or at least see use and I can see how big corporate paying customers would be interested in that too.

So that footage is being generated from all the apps, all the NX apps, the VX360 and the Grab kind of like delivery driver mapping devices as well. It's like meshed together.

For now, we're not actually using the data of the graph because most of our use cases are outside Southeast Asia and Grab is inside Southeast Asia. But we do have a data partnership actually.

There are a couple of cases where customer wanted Southeast Asia data and we have been chatting with them and providing to the customer, but we are outside Southeast Asia predominantly, so Europe, US, Japan, Middle East.

Grab is inside Southeast Asia, Philippine, Singapore, Indonesia, so in ways we're complimentary and we provide data that Grab doesn't have outside Southeast Asia to them.

Generally speaking, we have been combining or getting a lot of support from their tech side both in the hardware and software and that's why we've been building much faster.

The focus has been last year on data monetization, use of the data data monetization, and it compounded over the last quarter.

I think today I'm here for one of them, which is arguably one of the largest news Natix have had in the history of our company and project, but we have multiple other ones that we're going to gradually announce and we're working with them.

Also, a lot of leads that I haven't closed yet, but they're in the process. Strong pipeline.

Well, yeah, let's hear about that first deal then. It sounds like a new customer, almost like a new product as well. So curious to hear the download from you.

You guys are the first media that we wanted to announce this with. So, we just closed a deal with Valeo, one of the largest automotive OEM suppliers to build a world foundational model with our data and their knowledge around building world models.

I will explain again what world model is and why the heck Natix is building one and why we're so bullish on a world model. Valeo is one of the largest automotive players actually out there.

They have over 100,000 employees, over $20 billion of revenue in 2024. I think the numbers are Fortune 500 company and they're supplying all kinds of technologies including ADAS and autonomous driving technology to pretty much all of the car manufacturers.

European ones like Mercedes, BMW, you can go to Japanese ones like Toyota or you can even go to the Chinese ones or Koreans or even American car manufacturers.

They're very well known for building automotive technology and they have an arm called Valeo AI which is only focusing on the AI side of the story and these are top-notch scientists and researchers that I have came across.

We're very proud to actually work with them and build a world foundational model arguably one of the largest open-source multi-camera world foundational models.

This sounds like an abstract statement so we need to dig a little bit deeper into what the hell world foundational model is but yeah happy to do that for sure.

Yeah maybe if you could tell us about what the model is and then how the Natix data is used to create it basically and and why they chose you guys specifically as the partner on this.

Word foundational model you can think of it as an LLM but for the real world. LLM understands words it can generate language and ideas and it's trained on internet data so to say.

A world model can generate four-dimensional reality three-dimensional space fourth dimension being time. Why do we need a world model and that comes hand in hand with physical AI.

Robots, humanoids, and autonomous cars, which is a new hot topic. I think Nvidia is not touching anything other than physical AI these days. That's where they're saying the next aha moment of AI would be.

Humanoids are extremely hot. Autonomous driving has never been closer to reality than today. But a world foundational model can basically generate a simulated version of our reality with the physics of it.

Right now, why do we need a world foundational model? Why do we need a simulated version of our reality?

Because you can train a robot or an autonomous car purely inside this simulated version of reality to do the jobs that you want it to do, without it having ever been in the real world.

I give you an example in the automotive case. If Valeo wants to expand their jurisdiction to London or Germany or Netherland usually they used to go and collect data from those jurisdictions because it's different traffic behavior.

London is even on the other side of the road you have more cyclists than you have in San Francisco it's different traffic behavior. Netherland I mean Amsterdam is insane with the cyclist that you have so being able to.

Exactly so you had to go collect data in those places and then come train your model to be able to drive in those scenarios. With a world model, you can train your AI completely inside a simulated version of London or simulated version of Amsterdam.

It actually can drive at the end of the day and it just increases the speed of production and global adoption by 10 100x if not if not even more. That's why every physical AI company in one shape or form is trying to create a world foundational model.

Now here's a caveat. Now world foundational model is replacing real world data. It's a synthetic version of the data. But to build a world foundational model, you need a lot of data.

The same way that for building an LLM, you needed a lot of data, you still need a lot of data. Companies like OpenAI have been investing a ton into getting that data, scraping the data, getting that data.

Same thing goes with the with the world model. You need a lot of real world data and you need diversity. You cannot build a world model that can generate London if it has never seen any data from London.

Same thing goes with the rest of the world. So that's kind of where Natix comes to the game because we have been doing an amazing job in collecting data, a very diverse set of data.

We have over, as I said, 600,000 hours of video data. This is in Japan, this is in US, this is in UK, this is in Europe. So it's pretty diverse pretty much all of the places that the automotive industry would care about right now as a highest priority and that's why you know building a world model that is using a diverse set of data and a lot of it is is extremely important and that's why they chose us for for for having you know for building the world model.

Yeah, very exciting. I think something that helped me conceptualize maybe what that would look like was when you said like it's 4D. So obviously like Google Street View is 3D, right? like they they took static images and kind of stitched them together and you can like walk down a street in Google Street View, but it sounds like the world foundational model is 4D in the sense that like if you capture enough high quality video footage from a street, you know, you it's 3D in the sense you can walk down it, but then the 4D part is like the simulated reality of like, you know, if if we took a couple pictures of someone walking here, we can have the World Foundational model have them continue to walk. And so then it's like everyone that's a static image becomes a moving character but it's exactly that that's interesting.

As a matter of fact what what you can also do with world models. So these are other use cases of world model right. Let's say you have some clips of data from London or even some frames right of of roads in London.

What you can ask a world model to do is, hey, generate me a minute clip and add a certain edge case or a longtail scenario into it. So a car coming for example from the front and and and and losing the sight and and wanting to hit me, right?

Or change the weather condition, make this into a harsh rain, make this into a fog kind of very foggy kind of. And why are these important? Because suddenly you can create all these different variations of your data.

I'm not sure if you followed, but when Nvidia recently published an open-source end to end model for driving, they put it on a Mercedes-Benz CLA, and everybody was basically, you know, shorting Elon. Never do that, by the way.

The history has shown that whoever shorts Elon will lose the game. Elon was saying, "Yes, I mean, it's good. the the the think and drive, but the detail is in the longtail scenarios. And whoever is going to win this autonomous driving game is the one that has trained their model on the longest set of longtail scenarios, right?

Edge cases, situations that you know like dangerous overtaking, harsh weather conditions, all these different variations of situations that are not normal. It's not a boring drive in a highway and just keep the lane, right? And no crazy stuff will happen.

I think that's why Tesla is a few years ahead of the game right and again that's why world model in any shape or form will help us and any company to reach there much much faster because it can generate all of these longtail scenarios and edge cases synthetically and and and make your model even more resistant towards all of these special kind of cases.

I did send you a link by the way if you want to show something to your viewers there's a company called wave they also have a world model out there or there is another project open source ones we can also show to the audience so you kind of get an idea of what it means to also generate a multi-camera kind of version which is completely synthetic right these are not real world data.

You can see like there are one of the one of the famous world models that are out there and then there are multiple of these right in the in the so this is closed source by the way so they are very good actually in world models and and this is not open source but then there are also open source ones that are out there.

Our goal is that the version that we're going to publish will be open source which I think would be a great contribution also to the to the industry it's not just you know and then we're going to take that and and try to make it even better but it was important for us also for this to be open source for sure.

And it sounds like Wave could potentially be a customer as well, right? If if I don't know how they're sourcing their their road footage that they're using to create synthetic footage today, but I'm sure eventually you guys will surpass if if you haven't already their their ability to capture real world data.

I think we're going extremely fast and when I look at the numbers again what we have achieved and we're amplifying that as well in the coming months is but I think we're going to have arguably one of the largest data sets out there in terms of real world multi-camera data that can be used for autonomous driving applications like world model or end to end model.

The larger the data set and the more longtail scenarios you have in these data sets the better the world model will get which means that the better the the the final model that you're going to train based on a world model will get right.

Whoever has the longest set of long tails will lead the space in my opinion and that's that's not just my opinion actually that's that's what what Elon has been saying as well and I think you know Tesla did this the best way right they are a data engine and that and their system is getting better and better over time.

It's extremely important for having a large quantity of data and crowdsourcing that actually is better than just collecting it in a centralized way because it it's more natural. There are more long tail scenarios and edge cases.

Because you have natural driving, you know, people that are not just driving to get to, you know, because these data collection vehicles, they're even trained drivers. So, it's a little bit different. You don't you you maybe get a little bit filtered version.

When you have completely crowdsource approach, you actually get the the true taste of the real wild world.

Maybe just to round out the Valeo deal news. So, how they're they're an OEM, like you said, a vehicle OEM. Like, how are they going to how are they planning to use the world foundational model that you're building for them? Are they going to go try to sell access to that to autonomous driving companies? Are they like you know simulating the performance of their parts on the road? What what is their use case?

Both of us are actually working on use cases on top of that. World model has two use cases. One of them you can generate synthetic data for training purposes. So you can train models right agents navigation models that can actually operate the car.

The second one is for testing and that is like once you have created your agent now you want to test it in the simulation environment and that is again you know inside the simulation environment you can use you have to generate scenarios right to test your car or like a simple ones are right turn left turn you know lane change the more challenging ones are the edge cases right dangerous overtaking would how would your car behave in this situation right.

For testing purposes you use a lot of synthetic data which is generated by the world model but you can also that's again why Natix is very important because you cannot only rely on synthetic data for both training and testing and validation because AI will hallucinate so we will have let's say we would need 10 to 20% real world scenarios and edge cases to test and that is also very important so it's going to be both for training and for for testing inside simulation environment.

I was curious like what's the latest with the the Bit Tensor Subnet like part of the business or or idea. I remember seeing something about that a while ago, but I'm curious like where where that's all gone because I know robotics has obviously continue to become a big, you know, deal and concept in in crypto.

I should come another time on your podcast because we're going to announce this soon. I don't want to jinx it but we are going to repurpose the subnet and making it a little bit more generic but also for physical AI for robotics and for autonomous driving.

What that would look like let me you know let us talk about that once the announcement is out but I think it's going to be huge and and we we want it actually not to be even just Natix related. We realize certain needs in the physical AI space.

There are so many synergies between whatever is happening in autonomous driving as well as whatever is happening in the humanoid and robotics side of the story right world models is the same end to end models is the same there they're differences but the direction that the industry is going is pretty much the same real world data versus synthetic data it's it's it's very very similar that's why Tesla is a very good bet for building humanoid because they have a very good you know team for for for autonomous driving And a lot of these things are the same.

We realize gaps in the physical AI by doing it ourselves. We want to repurpose this subnet a little bit and making it not just useful for Natix having physical AI application but also to robotics companies and and and autonomous driving companies having physical AI application working with larger scale data.

Having all that high quality real world footage is a lot of use cases, you know, including outside of autonomous driving like you mentioned that that that data could be used for. So that's exciting.

What else would you want to cover today slash like what could we look forward to from Natex in 2026?

We will keep on selling the data and closing more deals. For us many of these deals are extremely strategic as well. On the other side we are going to build world models.

This is a big statement. Why are we building world models? because synthetic data will take market share from real world data. You like it or not, that's what the entire industry from robotics and autonomous drive from physical AI is is is working on and this will happen very very soon.

We want to keep market share and I think we have something that many companies do not have. So we want to be not just a data seller, we want to be part of the world model game especially for for autonomous driving and that's that's what we have been working on a lot.

You're going to hear even more stuff that we have been working on along this direction. 2025 was a lot of work around data and monetization. 2026 is going to be even heavier and productization.

World model for me is a productization of our data which increases our defensibility and as I said I said this quite some time ago we went away more and more from mapping and we're right now heavily focused on autonomous driving physical AI right different requirements different league different game but this is where our focus is we're still selling data to the map makers just the front-facing camera but you know we're we're having different different requirements for the for the entire data and you know it's it's Again it's very very different.

It sounds a bit similar because yes both are collecting images from from a car moving car but you know from the type of the data even what you do on top of the data how you filter the data how you collected you know size of the data everything is different.

We keep that path right for me I have never been more bullish on physical AI. I think whoever has been talking with me, I had a conversation with Salaron one year ago and and I was shilling him so hard this physical AI and he was like okay I'm going to go and and read about it you know more I mean he knew about it but he was like okay I'm going to go like like do a deeper dive. So that has not changed and we're going to go all in on this direction.

Well that's exciting. Yeah, I think a lot of our our audience is probably pretty familiar with you guys, but is there any kind of call to action or way to get involved that you have for the listeners today?

So, twofold, if you want to contribute with the data, especially if you own a Tesla, get one of our devices, you know, join our network. If you are a company or a researcher that is having anything to do with the world models, end to end models, you know, reach out. we're happy to chat and and support.

That's what I said. It's it's our priority. World foundational models end to end models for 2026 is our priority and I think we're going to do a lot of stuff. We already did a lot of stuff compounded in the last quarter but you know we're going to keep on announcing and also closing more stuff around that.

If you can work with our data alongside world model simulation, if you hear this from automotive industry, you know, feel free to reach out. We we're happy to support.

Thanks for coming on to break the news and congrats on the deal. We'll talk soon.

Thank you, Connor. Thank you for having me.