Blog
Someone gets lost in twenty square kilometres
Written by Louise Beattie · 5 August 2026
From a Beyond Finance community call, in conversation with Miguel Enrile of Swarm, Subnet 124.
The drone that could find them already exists and costs less than a car. What's missing is a pilot who happens to be free at that exact moment, knows how the terrain has changed since the last map, and can fly more than one aircraft at once. Miguel's case for Subnet 124 is that the search doesn't fail on hardware. It fails on people. And once you see it that way, the same shape turns up in a ski resort waiting on avalanche work, and in a body connected to the Australian government that counts koalas by flying a drone and watching the screen.
Swarm doesn't build drones. It builds the intelligence that flies them, and it trains that intelligence on Bittensor.
The industry's problem, in his framing, is operational cost, and operational cost is people. As he put it on the call, "the expensive part of this industry in general is not the drone, it's the pilot. Because the drone without the pilot is basically something useless." The best of them run to about €30,000, roughly a mid-range car, and most are far cheaper. The person flying it has to be awake, available, close enough to get there, and trained for that particular job. By his estimate more than 99 per cent of missions work exactly this way.
So why has nobody built the autonomy? Because the industry keeps bundling it. He describes a trade fair in Poland where every stand was selling its own drone with its own software welded on. "There's no, you know, Tesla Autopilot for drones. That doesn't exist. That's not a thing."
The missions that need a brain
Not every drone job needs a brain of its own, and he's precise about which do. The line he draws is whether the environment changes while the mission is running.
Inspecting an electrical tower doesn't. Neither does spraying a field. You can survey the site, draw the route, fly it. A pre-programmed flight path does the job, and he says plainly he isn't chasing that work. Mapping is the same, "very easy to do because it's just flying on top of the place."
Now the other side of the line. Go back to the person lost in twenty square kilometres of mountain. You cannot just draw a grid and fly it, because the best way to search the area may not be the obvious one, and the ground has moved since it was last mapped. A drone drifting into an airport's approach, where the target is moving and making decisions. A drone hunting a mosquito, which he offers as the purest version, because no human hand could do it at all. A drone lifeguarding a coast where people who can't swim are going into the water.
Someone on the call asked the sensible follow-up: wouldn't you accumulate a map over time, the way ships gradually build up a picture of the seabed, until you had the terrain solved? He answered that the world is already mapped, more or less. It's the changes that get you, three metres of shift after an earthquake, or snow, or a fallen tree, or wind that puts the drone somewhere other than where it thinks it is. "It doesn't depend on our maps being perfect because the maps are not perfect in real life."
When I raised heat detection over landscape nobody is watching in a high fire-risk week, he took the thermal camera thread somewhere else: an entity connected to the Australian government that wants to count koalas, and thinks automation would do it better than the manual flying they do today. Australia has committed 200 million Australian dollars to koala preservation. Once you stop needing a pilot, the list of things worth flying for gets very long.
What they build, and what they buy
What do they sell into that? Four layers to the stack.
Layer one is the drone and its sensors, which they buy in from DJI or Skydio or whoever makes the right thing. Layer two is the flight controller that keeps it steady in the air, already good and not their problem, though they have a workshop and will modify hardware where a job needs it. Layer three is the autonomy, and the word carrying the weight is horizontal: it is meant to fly anyone's drone, not one manufacturer's, provided the aircraft can be piloted from outside. Layer four is the mission, which is a fleet working out for itself how to do the job it has been given instead of being told where to fly. Those top two layers are Swarm's focus.
Which changes the arithmetic of the search. Today those twenty square kilometres get one drone and one pilot. What Swarm's building is ten drones and one supervisor over the same ground, and eventually ten drones and nobody.
How the models actually get made
That's the problem they've picked. The question is where the intelligence comes from, and Miguel is blunt that the company itself isn't hard to copy. "Anyone could build this startup in reality." Then he puts the obvious question to himself and answers it: "What's our defensibility? That our AI is trained in Bittensor, basically."
A conventional firm has to hire a specialist for every environment it wants to fly in, and if that expertise isn't among the hundred people it employs, it doesn't have it. Swarm builds simulated environments instead and lets miners compete inside them, then takes the winning models and puts them into real drones.
And in his telling it compounds. "As our subnet is performing well, we attract more miners, and more miners means better quality. Better quality means we can create better simulations and better real lab tests." Better lab tests make the products sellable, and sales fund the next round of environments.
Four challenges are running. An autopilot that gets from A to B without crashing, across six terrain maps, among them city, forest, open ground, village and mountain. An interceptor that has to catch another drone actively flying away from it. A search-and-rescue agent that has to find someone lost across those same maps, which is the mountain problem rebuilt as a competition. And swarm versions of the autopilot and the search-and-rescue agent, with anything from two to eight drones at once.
Miners have turned up from Australia, China, Bahrain and Argentina, which he knows because he asks them where they're from when he's been talking to one for long enough. He prefers that to an in-house team for the spread of approaches it buys. "Also, miners explore many approaches, not only one. Even if we were able to do good models internally, we would be lacking you know, other opinions."
If you want to see whether the models are improving rather than take his word for it, he points at Swarm's own site, where he says every challenge carries a line showing the top miner's score, and the lines climb.
What a swarm actually is
The name carries a picture that isn't right, and he's quick to correct it.
Those displays where hundreds of drones make shapes in the night sky are not intelligence. "That's basically code telling each drone where to go and filling the points." Move one thing they weren't expecting and they crash.
Fleet autonomy comes in two architectures. One central brain flies everything, which is much easier and perfectly fine for somewhere between two and ten drones. Past that it stops working, because the communication load scales quadratically and there is only so much you can push through the air. The alternative is each drone carrying its own intelligence and talking only to the drones near it, the way a flock does. That's the one that lets you put a thousand up, because adding a drone adds one drone's worth of thinking rather than squaring the problem. Swarm runs both, and the name tells you where it's going.
Then the drone meets the wind
None of which counts for anything until it flies. The simulation is the easy half, and he says so himself.
"Simulation robotics is basically like PowerPoint finance. You can write whatever you want, you know, in a PowerPoint. The PowerPoint holds everything."
Then you put the model on a real drone and it crashes, because wind exists and because the world isn't quite where the simulation said it was. Closing that gap is why Swarm brought in a robotics CTO, Julia Marsal, who has spent around ten years building robots at startups, including robots for data centres and robots of a kind that hadn't existed before.
When we spoke they had just built the test for exactly that. They'd rebuilt their own office as a simulated map, which is the point of choosing it: it's a room they can also fly in for real. A model that wins in the simulation goes straight into a drone in the same space, and they measure it against what happens. They wanted to know whether the miners' interceptor could beat the one the team had trained internally. The announcement was due the week after the call, so the result isn't in this piece.
Needing somewhere physical to fly things is also why he calls the company a lab rather than a drone company. Swarm is a robotics lab pointed at drones for now, because a drone is one kind of robot and the intelligence they build tells a machine where to go, not how to move. Quadrupeds are already on his mind, humanoids possibly later.
Andorra, and the order he's doing things in
If the hard part is the physical world, then how fast you can get into it matters, and that is what the location is for.
Swarm is based in Andorra, a country of around 80,000 people between Spain and France, and it's a deliberate choice, not an accident of where the founder lives. "The regulations here are much more, much less strict. So this is not the European Union, so we don't have to comply with the European Union if we don't want. However, it allows us to test very quickly whatever we want, whatever we want to build." Test fast under light regulation, then take a working product to the EU and homologate it there.
Being large in a small country compounds the advantage. Whether a drone flies today is one man's decision, and Miguel has his mobile number. With no airport there is no aviation ministry, and so nobody to go through. Swarm works with the national ski resort on the avalanche problem, which right now means taking a helicopter up or walking in and throwing explosives to bring the snow down before the slopes open. He compares working with them to "working with Messi in football because he's the biggest company here". ENLIRA, the country's largest startup accelerator, selected them too.
The same logic sets the order of the roadmap, which starts with the customers he can actually reach. Rescue agents come first. It isn't the biggest market, but the big players are all pointed at defence, so he can walk straight to a ski resort or a mountain rescue team with a problem they already have. Someone goes missing, fifty of your people mobilise, and in winter the person you're looking for freezes overnight. That is the twenty square kilometres, and it is where Swarm is starting.
After that, industrial infrastructure. Hobbyist drones have closed airports across Europe, where the problem used to be birds and the answer used to be hawks, and interception there is ordinary industrial work rather than defence. Warehouse security sits in the same tier. Then disaster response, firefighting and earthquakes, sold mostly to governments. Defence comes last. "It's very sexy, but we're not there yet, but it's on the roadmap."
The projects he takes now are the ones that build the autonomy platform he intends to sell into the larger tiers later.
Who's building it, and how it's paid for
All of it rests on a very small number of people. Miguel is a telecom engineer who was a miner before he was an owner, and by his own account a strong one. That's how he knew how to train models, and knew Bittensor well enough to start alone. The team is fully doxxed and he makes a point of it: "you can come to our office whenever you want. So this is not a crypto token where you don't know who are you buying into."
He hires slowly and on a rule. "I only hire someone which brings to the table things that I cannot do myself", or who does something he could do but does it better. Julia leads the robotics and Miguel the AI. A hardware integration specialist and an AI engineer are due in September, with eight to ten people by the end of the year.
He bought the subnet slot with his own money rather than raising first. "I bought the subnet, I bought the slot, so I didn't go to find investors. I believe in myself." That, in his words, "gave me 100% ownership."
The investors came afterwards, and by Miguel's account they had watched him handle the company's money carefully and consistently. Evergreen&Blue put in around €370,000 and then locked their tokens perpetually, and while an owner locking is routine and easy to do, an investor doing the same is rare.
He gave up only five per cent of the company, and says that was his own choice rather than a limit they set, because he didn't want to hand over more of Swarm at that point. Twenty per cent is the usual figure at that stage.
The subnet belongs to the company (Swarm), not to Miguel. It's registered at the notary. And the company doesn't sell its tokens. In a year he has spent two or three per cent of owner emissions. He takes no OTC deals and pays nobody to shill.
"We don't pay those things because we think it's basically appearing to do a pump and then dump. And for that, I prefer not to do it because I think it's worse. You pump, then you dump, and then you never get the confidence again. I prefer to, to grow steadily and to treat people honestly."
The discipline is personal as well as corporate. He pays himself $4,000 a month, up from $2,000, out of the investment rather than the token. At a conference where other subnet owners took the €2,500 VIP pass, he took the €600 one. He has severely underpaid himself, in his words. "I still haven't made basically 1 TAO from the subnet until this month."
What's done, and what isn't
Worth being exact, because the commercial story is mostly ahead of them.
They've won one project: the avalanche drone, worth around €36,000, though the money doesn't arrive until the regulator approves it, which he expects to happen.
A grant application goes in during September, targeting around €200,000 from a €2 million Andorran startup pool, and what it would build is that drone: an autonomous machine that goes and finds the person on the mountain. He describes the Australian koala work as a prospect that came to them, and says they're selling it now.
And he still has to stop being only the technical lead and start selling. He's been waiting on the team to make that possible.
Going outside Bittensor is his own standard, not one anyone else set for him. "Now we will need to go outside of Bittensor to prove value to Bittensor." From September the plan is a list of every country in the world and who in each of them wants drone projects. Asked separately whether there was money in the map data the drones collect, he said he'd sell that too: "I will sell everything to everyone."
He wants a lab inside a year that can take on research projects with budgets in the millions.
"This is gonna be a top lab in Europe at least. So Switzerland is not the only one which can do robotics."