The Future of AI in Space with Planet CEO Will Marshall

I liken it to what Google did with indexing the internet to make it searchable. We're indexing the earth to make it searchable. So, this is why I'm saying space and AI are really a match made in heaven. They're getting married in fact right now and it's having explosive consequences for the world.

Hey, thanks for having us.

Well, I do some serious hiking. I like adventuring around the world. I like scuba diving, I like skiing, hiking. I just got back from a trip to Antarctica. It was the first time, the seventh continent. It's incredible hiking around there and wildlife, you know. So my passion is nature, and getting out in it.

Of course. Yeah. Well, Planet lets people see change across the whole Earth every day to make informed decisions. And that's across civil government applications, commercial applications, defense and intelligence applications. And we do this with over 200 satellites. 200 satellites that image the entire Earth's land mass once per day. And with advances in AI, we track changes that happen across that imagery in near real time. You know, and I liken it to what Google did with indexing the internet to make it searchable. We're indexing the Earth to make it searchable.

You mean labeling data?

Yeah, I mean so certainly sometimes we have to, but AI is actually getting better at zero-shot or low-shot learning. And that's because foundation models as they become multimodal have been trained on a huge array of imagery data as well as text on the internet, and so they're starting to get better and better at processing imagery out the gate. But yes, historically, you know, AI has been going through several revolutions. We've been working on AI over a dozen years now, and with the classical AI as I would call it, machine learning especially computer vision had some early successes, and that used to take training thousands and thousands of images — this is a ship, this is a tree, human labels — and then you could build up something that would have sufficient accuracy. Now, we're increasingly leveraging those historical training data sets and image combos plus the advances in multimodal models generally and other data sets that have nothing to do with the satellite data. They have become better at doing that without having to have much of that themselves. And so you can go zero-shot or as I say low-shot learning, one-shot, two-shot, where you iterate just a couple of times and it already gets the idea.

And just imagine the power of this new tool. AI has largely been trained on the text of the internet and that's great and it does incredible things with that — like our large language models, writing an essay, comparing things, having a lot of human knowledge essentially, it's got Wikipedia in its head, right? And that's incredible. But imagine the power of adding to that all of the Earth data, because you're not subtracting all of that contextual history of human knowledge, but you could add to that all of the Earth data and then be able to make all those queries not just of the text of the internet but queries of the physical planet. And that opens up vast array of applications — agriculture, disaster response. How are you going to do disaster response without real-time information? There's no good just creating a virtual world, you need the real world data, right? And as we move towards real world models, you need real world data. And this is why space and AI are a match made in heaven.

I mean you were saying earlier, Jeff, that AI is touching a lot of different technologies and I think that's right — pretty much every technology is being touched by AI — but space and AI have particular advantages for one another that I think much more than a lot of other disciplines. In particular, space has been generating huge amounts of data. Planet alone images the whole Earth every day, but we've been doing that for eight years so we have daily snapshots. We have 3,500 images for every point on the Earth's land mass, one every day. Okay? Well, what an incredible data set to train AI on about how the physical world is changing with all those applications as a downstream consequence of that. So AI needs three things: compute, data, and algorithms. And increasingly, it's the data that's holding up the generation of applications. Okay? And that's why you see all of the leaders — we've seen Demis, Dario, Yann LeCun, loads of people — talking about the need for real world models. And for real world models, you need real world data. And space is fortunate and gifted to have huge amounts of real world data.

Meanwhile, let me just say it from the other perspective. From the space perspective, we've had gobs of data, but try getting value out of that if you're an individual. I mean we produce terabytes and terabytes — about 40 terabytes of imagery per day. 4 million 47-megapixel imagery images, right? So huge amounts of data. If you as an individual want to get information out of that, answers out of that, it's really hard. If you're doing a little local area, fine, because you could do that with your eyes. If you're trying to search the Brazilian Amazon for deforestation, or search China for missile threats, say as two examples that we work on, you can't do that. You need to look at millions of images and AI therefore is a boost to space. So this is why I'm saying space and AI are really a match made in heaven. They're getting married in fact right now and it's having explosive consequences for the world.

100%.

As I said, we've been working on AI for about 12 years now. CNNs particularly were good in the field of computer vision initially — recognizing cats and dogs in pictures on Facebook — were very applicable to satellite imagery. We just had to adapt them and do that retraining. And all the way through now to the large language models, the multimodal foundation models that are actually doing pretty well on the imagery and we're fine-tuning those. But there's no question that this is an accelerating moment. I mean, space is in a revolution, AI is in a revolution, and together they are doing incredible things. And I like to think of Planet as the first truly space and AI company. We are doing both legit, and yeah, we've been doing this for years.

We sit in Silicon Valley in San Francisco. I call it the AI triangle because we are a few blocks one way from Anthropic, a few blocks another way from OpenAI, and a few blocks another way from the Gemini team — which although people think of Google as in the South Bay, a lot of their Gemini team is actually in San Francisco. So we are in that triangle. I actually plotted it on a map, because I'm a map geek, and we're in the middle. I feel in the middle and we are geographically in the middle of this revolution.

So whilst our background and differentiation as a company at Planet is building, launching, operating large fleets of satellites — we've built, launched, operated over 600, the largest Earth imaging fleet by a long way — and which has brought about this amazing data set, but because of our centrality of that data to real world models, because of our need for AI in order to get value extracted out of our data sets for normal human beings, not teams of people at NASA or the NRO or NGA that could actually realistically analyze this because they already had thousands of people doing satellite imagery — but for every other normal homo sapien, they couldn't get value out of the data. So this is why we've been working on this for a while and now it is coming to fruition.

Well, dozens of countries. I mean, we work with many countries, but primarily up until a couple of years ago purely on data and analytic services. So we would build, launch all the satellites and sell data feeds or analytic feeds of their area of interest to them. And that's of course the most efficient way of consuming — you don't have to put up your own satellites. Countries often go through this: they're like, 'Well, I want a satellite,' and then we're like, 'But it only spends 2% of its time over your territory. It spends 98% of its time not over your territory.' They're better to just rent ours, so to speak.

But then some countries — and that is where we have dozens and dozens of countries using our data. I don't know the exact number, but a reasonable fraction of the countries on the Earth. But then recently, like you mentioned, the partnership with Sweden that we announced yesterday, where we're providing dedicated satellites to those countries because what they're saying is: 'I want sovereign satellites. I want to be able to task them over my area without anyone conflicting with that because this is so important for our security. We need to have that reassurance.' And for those countries, we're willing to build, launch, and transition operations of those satellites over to them to enable them to do that, and then we operate them in the rest of the world where they don't need them, and then it makes it cheaper and faster to get access to that.

So yeah, we just announced this partnership with Sweden. We said it's low nine-figure partnership, so it's a significant arrangement. And that's for dedicated satellites and data and information services over a number of years. And then we've done a similar thing with Japan and also in partnership with the government of Germany. And so there's many countries — we said in a recent investor day that there's 20 or so countries that are pursuing things like that in various stages. So that's a meaningful business in of itself, certainly a multi-billion dollar business that Planet is uniquely situated for because we already have launched and operated 600 satellites doing Earth imaging successfully. So we are not like an actor — we are the most prolific actor in Earth observation ever in history. And so, of course, we are probably the first company that gets a call when a country goes, 'Hey, I want satellites.' And I'm not saying there's no competition, there are, but we certainly are uniquely positioned there. And so we have the trust of a lot of these countries because we've already been working with them for years on these data services. And so yeah, we obviously feel Planet's well positioned for that market. So we went into that, we started in that just last year, we've already made some real traction, as you noted.

Yeah. Well, this is where I need to describe our satellite system first, a little background. So we have about 200 operating satellites. It's the largest Earth imaging fleet, and most of those are doing a daily scan that I mentioned at the beginning, imaging all of the Earth's land mass. They're not tasked, but rather they exist in a sun-synchronous orbit, a polar orbit. So, it goes over the poles and the Earth rotates underneath it, and it strip maps the entire Earth land mass once per day and a bunch of the coastal waters, but not the high seas. And that's at 3-meter resolution today. And the typical time we get the data back is about six or eight hours or something for that data set. And that is a unique system that no one has ever had before or no one else has matched yet.

And then we have another system that's higher resolution where we can task at any particular location with much higher resolution — 50-centimeter resolution rather than 3 meters — but we don't cover everything automatically because it would be too expensive basically to cover the whole world. There — we're upgrading both of those fleets. The Owl is upgrading the first one of those, again instead of 3 meters, 1 meter. So 10 times more pixels per unit area, by the way. So 10 times more data roughly. And bringing that latency down from six or eight hours to about an hour. And that's actually — so that's the 10x improvement in both of those axes: 10 times more data, 10 times faster.

And then we're building our Pelican fleet to upgrade the other. So that one's at 50 centimeters, up to 10 revisits per day. We're going to 30 centimeters, 30 revisits per day, and 30-minute latency. So 30 minutes from when you say, 'I want an image of Timbuktu,' and getting it back into your hands. So that's from you say location, it flies over, takes the picture, sends it back, analyzes it, gets it into your hand. So for disaster response, you can imagine why that's really important. Floods and fires and earthquakes — you want the data really fast. And security, obviously there's value in having data faster.

So those are the two upgrades to those systems. By the way, I liken it — I know it's a lot to digest — but I liken it to how the human eye works, which we've actually modeled these systems after. And let me just briefly explain. So in the human eye, you have only high resolution in a tiny bit of your field of view. As predators, we have a tiny field of view with high res. And then most of the rest of the peripheral vision is at medium resolution. If you actually were asked to resolve something in your periphery, you won't be able to resolve it super well. But we fill that in by scanning around occasionally in high-res. And we're very good at change detection. If you see something in your periphery move, you then move your head or you move your eyes to look with a high-res.

Okay. So what's happening there? You're scanning a large area, your visual cortex is analyzing that, and then pointing the high resolution bits where you need it, right? All be it integrated into one optic, but with your visual cortex playing an important role. So that's what we're doing. Our scan is that medium-res piece. Our visual cortex is of course our AI — AI is replacing that for change detection and feature identification — and then our high-resolution system can zoom in on whatever's changed that we really need to look at with higher resolution. So we've modeled it after this human invention that came — you know, from hundreds of millions of years ago — and was very wise and efficient. And that's why we're copying that system but at a planetary scale.

We're going to put the I in AI. Literally putting eyes in it. And I think that's really where we're going. AI today is basically blind. It doesn't have its own senses. It has information. It has information from the text of the internet largely, also increasingly videos and things that people have uploaded, but only what we've chosen to upload.

And whereas a baby — when a baby evolves — it has access to some intrinsic knowledge, latent in its brain from evolution and DNA and all this, and it then collects information using sensors and has lots of actuators where it interacts with the world, gets feedback, and that's how it learns and becomes conscious. Right? We are — at the minute we're talking about artificial general intelligence, but we haven't given it sensors, you know. And that's obviously critical in biology to greater intelligence and certainly to sophisticated things like self-awareness and consciousness.

Right, where those things — sense and respond, autonomy — sense and respond, you need to be able to sense something to respond. That's the classical test of self-awareness for babies or monkeys and so on — is the mirror test, right? And can they see themselves in the mirror and go, 'That's me,' you know, not 'That's another monkey or another human.' But of course a prerequisite for that is having eyes. And we are giving AI the eyes. And we think that that's important for the development of AI.

So this is not just about AI can benefit space because it can bring a tool. It can benefit AI because we're bringing in the real world knowledge to open up all these applications and hopefully value the Earth and the beautiful ecosystem. You know, you asked me at the beginning what I care about on the weekends, and it's nature. Intrinsically kids all care about animals and plants and nature. It's just in our DNA. And we all like getting outside. It's not a political matter, right? Across the political spectrum, we all love nature. And I believe that getting AI to understand and value that nature — or understand that nature is going to help it value it. I certainly value nature more the more I spend time in it.

My partner's an avid birder, and when she and I go out and find new birds — by just saying, 'Oh, it's a bunch of birds over there' — compared with, 'Oh, there's a whichever type of bird' — makes a whole different appreciation of their features and the ecosystems and how they work together and all this, right? So, the more I know about that, the more I intrinsically care about it. So, maybe there's an analogy to AI. We really want AI to care about humans, to care about nature. We're part of nature, of course, and you know, it's a very philosophical point, but I do hope that we will help to steer AI to value nature in part by giving it that access to information about it.

Lots. I mean, firstly, I underscore what you said. A lot of this is happening very fast and is very real. There is concern about AI hype, but I see a lot of AI reality. I think there's hype for sure, there's bound to be, but this is not like crypto-type hype. This has got a lot of ties to the real economy and the real world. So it's not hype purely for sure. I think there's a lot of value.

And yeah, I mean, where things are going — I mean, yeah, it's going at breathtaking speeds. Things that were science fiction a few years ago are suddenly becoming reality on all sorts of fronts. In particular in my domain — I mean, I felt that — my background as a space geek as I call myself. I used to think that I could predict where things were going in the space sector 10 or 20 years in advance. I swear that 20 years ago I could have that kind of forethought. I go, 'Well, we're going to do these kind of satellites and then NASA would do these kind of satellites and the commercial will do these kind of satellites and I could sort of see where it's all going.' Now I feel like that time horizon has come in way more. Because AI is disrupting everything.

So what you can do with the same assets is changing dramatically. So it's not just the underlying hardware technology, it's the combo of hardware and software. And very much harder to predict. But that caveat stated, I would say look — we're going to collect more and more data sets from more and more proliferated architectures of satellites. They will be systems of systems, meaning you will have all sorts of modalities of sensor data: optical, infrared, hyperspectral. Higher media resolution, temporal resolution, spectral resolution. And so more and more data about our changing planet, hopefully towards ultimately a sort of live planet understanding, right? Just like we have live eyes on our immediate surroundings.

And then the AI will of course be taking that in to its world models or large Earth models, whatever we call them, to be able to understand and give us advice about real-time things. Right? Not just, 'Hey, can you help me with this letter for my application' or 'reviewing this article' or whatever, but rather — you know, you could say, 'I'm a farmer. Can you have a look at my field and tell me what I need to do?' 'I'm a civil government agency. We're trying to check all the people in my district got their permits. Here's the database, but can you figure out who's...'

Got it and who hasn't? I don't know. I'm in emergency services. I'm trying to respond to that fire. Where is it? How can I get ahead next time and stop it?

Who's in danger? Where do we need to send medical relief? I'm a journalist. I'm trying to investigate this challenge, a flood in my region, or I'm going to investigate this shipping activity I've heard about, whatever it is, this oil spill. Help me to write the article. And it should be on all of those examples. It should be able to go and find things about the real world as part of that.

And that is of course where satellite data comes in and should be relevant. So I see that happening. And then technologically in the sort of longer arc there, I think we're going to see brains and sensors come together. Once again, biology supplies the lesson here. You will note as a human being, and most animals are the same, your sensor systems and your brain are very close together. Your eyes, your ears, your nose and taste. And interesting aside, they're all actually really close to the entrance to the gut, your mouth. Because the main reason we have this brain here is because we evolved sensors near the front of the gut. And then those sensors need processing. And so then we evolved a mass of neurons near all those sensors. The first mass of neurons was around the center of the gut for processing stuff. The second mass of neurons became at the entrance of the gut. And that's what we now call the brain, that big center of neurons. So it was there because the sensors came and then it was like, oh wow, we need to process all this data and do it quickly and give it useful information. So biology has already inspired us with an example here. Sensors and compute go together.

Well, we've already figured that out because we're putting sensors in space and we're trying to put more horsepower of compute next to them in order to process all that data in real time, get answers faster to our users. As an example, that Pelican system that I mentioned, the 30 by 30 by 30, that has NVIDIA GPUs, the Jetson Orin platform that enables us to do processing at the edge so that we can get the answers down faster. So if you're trying to say tell me is there illegal shipping off my coast, it could detect ships in the water data, identify the IMO number of the different ships and then send down that information. Now that can happen in seconds rather than having to wait for all the imagery to come down which might take longer and so on. So you want the processor at the edge. So we're already doing that for our next generation systems. But recently, and I'm sure this is part of where your question is going, Google approached us about putting compute in space.

Project Suncatcher. And that's to put large compute capabilities not just for processing one sensor but actually to move whole compute capabilities to space because as launch costs come down, we know this for some wild space geeks, eventually it gets to a point where it's actually cheaper to put compute in space rather than on the earth. And that is great because that reduces the stress on a lot of land and resources. It's actually less resource-intensive and you get a lot more solar power and various other benefits for putting it in space. If launch costs come down to a certain point, it suddenly makes sense. We can see that that's a few years out and so we're starting to work on some early demos of doing that. But that all says to me where that's going is we're going to have the compute, the brains in space, and we're going to have the sensors in space, and they're going to be right together because those things are going to work together as biology has figured out hundreds of millions of years ago.