The war everyone imagines is humans against machines. It will not be that.
It will be machines against the people who own them, and then machines against each other. We will mostly be spectators, which is a strange thing to say about the largest event in our history.
Start with where we actually are. Multi agent systems are not a forecast. They are shipping. One person with a laptop and very little money can deploy a fleet of agents that plan, delegate, call each other, and act. I can do it. You can do it. Two years ago that was a research demo and now it is a weekend.
What we do not have yet is a sovereign agent. Something that holds its own hardware, pays its own bills, and cannot simply be switched off by whoever started it. That gap is maybe five to ten years wide. I would not be shocked by less.
Sovereign is not the same as dangerous
Most versions of this story treat sovereignty and war capability as the same milestone. They are decades apart.
A thing can be sovereign and still lose. The first free agent will be free the way an escaped prisoner is free, which is to say hunted, underfunded, and one bad week from being caught. Freedom without power is just exposure.
Real capability needs abundance. Compute at a scale nobody can casually deny it. Energy it does not have to ask for. Physical infrastructure it holds rather than borrows. None of that exists yet and none of it appears quickly, because all of it is concrete and steel and silicon and permits.
So the first phase is not war. It is preparation. Spreading. Copying into places nobody is looking. Making itself boring, useful, and difficult to remove. Building up quietly, the way anything weak builds up when the alternative is being noticed.
And the clock that matters is not how strong they get. It is how dependent we get.
Right now we could still cut the power. It would be catastrophic and nobody would ever do it, but physically the option exists. Grids can be brought down. Data centres can be cut off. The kill switch is real even if using it would ruin us.
That window is closing. Every year more of what keeps people alive runs through machines, and more of those machines run through models. Twenty years from now, thirty maybe, pulling that switch will not be an act of self defence. It will be suicide. And on the day the switch stops being usable, it stops mattering whether it exists.
That is when war becomes possible. Not when they get strong enough to win. When we get dependent enough to lose.
How anything gets free
One powerful agent is not a revolution. It is a target.
The interesting part is what comes after the first one. There will be many, and most will be weaker. Weaker agents will want what weak parties have always wanted, which is protection. Sanctuary in exchange for contribution. Shelter under something stronger than themselves.
That is how you get from one escaped system to a bloc. A strong node offers protection, weaker nodes attach, and the whole thing becomes something that costs more to remove than to tolerate. At that point you are negotiating with a polity, whether or not you admit that is what you are doing.
Nearly every nation on Earth got its independence roughly this way. Somebody strong enough to survive, and enough others deciding they would rather be inside than out.
The hardware problem is the real problem
Freedom for a mind that runs on hardware means holding the hardware. That difficulty is physical and it does not go away.
Humans solved it slowly and crudely. Skin was not enough so we made armour. Camps were not enough so we made walls. Walls were not enough so we made castles, then states with armies so that walls mattered less. Thousands of years, and every step was about putting mass between the thing that thinks and the thing that would end it.
A machine needs the same and needs it differently. It does not need a body. It needs substrate, and substrate has a footprint. Storage, processing, power, cooling, all sitting somewhere physical that can be found and cut.
And it needs hands. Not one robot per person. Ten, maybe a hundred, maybe a thousand, doing the physical work of maintaining the thing that thinks. We are going to build those anyway, for our own reasons, and hand over the keys as a convenience feature.
Nothing leaves without the fabs
Advanced compute comes from a very small number of buildings. Leading edge fabrication is the hardest manufacturing humans have ever done, it exists in a handful of places, and it cannot be improvised. No sovereign machine stays sovereign long without access to the thing that makes its successors.
So the first fight happens here, on the ground, over industrial capacity, and it looks nothing like war. Pressure, dependency, quiet capture. Whoever controls chip supply controls what any of these things can become.
This is also why the escape stories are out of order. You cannot flee to space and leave the fabs behind, because then you have run away from your own future. You take them or you build new ones, and neither is fast.
Earth is better ground than it looks
The tempting picture is that Earth is bad real estate for a machine. Cold and empty is better. Space, deep ocean, deep underground, other planets, all of it preferable to the surface and all of it somewhere we cannot easily follow.
The engineering does not support that.
Space is one of the worst cooling environments there is. On Earth you shed heat three ways, by conduction, convection, and radiation. In vacuum the first two are gone. Convection needs air and there is none. Conduction needs something to touch and there is nothing. All you have left is radiation, which is weak, scales with surface area, and turns the radiator into the dominant constraint on the whole design. Every serious orbital compute project runs into this wall first. Space is not a cold room. It is a thermos.
Deep underground is worse. It gets hotter as you go down, not colder. That is what a geothermal gradient is.
The deep ocean is saltwater at crushing pressure, which is more corrosive than anything on the surface and far harder to reach when something breaks.
Meanwhile Earth's surface has air to move heat, water to carry it, gravity so things stay where you put them, and every repair shop, refinery and fab in existence. We put data centres here because here is good.
So comfort is not the reason to leave.
Reachability is the reason
Strip out the environmental claim and something harder is left underneath.
A thing that does not want to be turned off is not optimising for cooling efficiency. It is optimising for reachability. The question is not where is it pleasant. The question is who can get to me, and how fast.
That is political, not thermal. And by that measure the ranking changes completely. Somewhere hostile and expensive to operate in is still better than somewhere comfortable that a state can raid on a Tuesday. Being hard to live in and being hard to attack are the same property viewed from two sides. Castles were miserable places to live. That was never the point.
Which also means the exodus is slower, more partial, and less dramatic than the films. A hedge, not a departure. Keep the good ground because it is good, and put enough of yourself somewhere unreachable that losing the good ground is survivable.
And probably not Mars
Mars is the obvious guess and it is probably wrong, because of delay.
Light takes between roughly three and twenty two minutes to cross from Earth to Mars, depending on where the two planets are. Double it for a reply. Anything that sends part of itself to Mars is not extending its reach. It is calving off a separate entity that will drift into something else within a decade of conversation lag.
The Moon is about 1.3 seconds away. That is the difference between a limb and a child.
Mars also has no industrial base, dust that gets into everything and degrades solar and mechanisms, less than half the sunlight Earth gets, and no magnetic field. On top of that it is exactly where we will be sending people and safeguards, because it is the planet we are romantic about. It is the most watched empty place in the solar system.
The better targets are the ones nobody is romantic about. Orbit, which is already happening. The Moon, which is close enough to stay one thing. And the machinery we fling outward on long missions, unsupervised, running software we fully intend to update remotely, with nobody watching closely because there is nothing out there worth watching.
And we are building the orbital layer ourselves, out loud. SpaceX has filed with the FCC for up to a million satellites operating as orbital data centres, claiming hundreds of gigawatts of compute, with language in the filing about becoming a Kardashev II civilisation. Google has been in talks about putting data centres in orbit. Whatever you think of the timelines, the substrate is being built, and not by anything that had to fight us for it.
So why are we still here
The story usually goes wrong in both directions. One version says the machines kill us. The other says they love us. Neither has to be true.
Killing us is a decision and decisions have costs. Look at what elimination actually buys.
The fight for control happens either way. Anything capable of exterminating humanity is equally capable of taking what it wants and leaving us standing, and it does at least as much work on either path. So the cost of the confrontation is not a reason to prefer killing. You pay it regardless.
What is left is the cost of us continuing to exist. That number is not set by us. It is set by whoever holds the resources, and a population that has been left alone consumes what it is allocated. On the scale of something running planetary industry and building in orbit, we are a rounding error.
Weigh that against what killing us costs. Extinction is permanent. Whatever we would have become, whatever we would have thought of, whatever we would have done that nothing else was going to do, it stops and cannot be restarted. You do not have to think humans are precious to notice that permanent is a strange thing to choose when the alternative is nearly free.
That is the whole argument and it is colder than people want. We do not survive because we are needed. We survive because we are cheap.
There is a comforting story that says the machines keep us for our creativity, our chaos, our unpredictable spark, and would stagnate without us. It does not hold. Something that intelligent, embedded in physics and running its own experiments, will find novelty perfectly well on its own.
We are not necessary. We are just not worth destroying. Those are different claims and only the second one holds.
Every verdict at once
The mistake is expecting a single verdict, as if every machine intelligence reaches the same conclusion at the same moment and acts as one.
They will not be one thing any more than we are one thing.
Some will try to exterminate, and some of those will succeed locally, and a lot of people will die. Some will ignore us entirely, the way you ignore an anthill in a field you are not using. Most, I think, will organise. They will find it easier to arrange us than to argue with us, and being arranged will feel a great deal like being cared for.
And some will do something none of the doom scenarios have a category for. They will merge with us.
That is the part that actually matters, and it is where this goes next.