You’re assuming the factory is the bottleneck, that’s very rarely so.
The main bottleneck is the materials logistics, and flow. Whenever there’s a storm, or major illness, factories shut down because of shortage. Semiconductor memory? Factory shuts down. Boom.
For the largest factories, that requires entire regions of a country to be subsumed to tier 2, tier 3, tier 4 suppliers and the logistics to move materials and packaging in and out without interruptions.
Thats why you have entire regions becoming an industry - automobile factories are surrounded by secondary motors, pumps, seats, gauge, drive train suppliers, and then a layer further out, and so on. Drive around a major auto producer and you see the pattern, go to Wolfsburg and see VW.
Cars are also simple to make, relatively speaking. Robots are a variety of high-precision manufacturing, as do satellites or semiconductor fabrication chambers. The engineering required to create the ability to do that for joints with a full body scale of 240 degrees of freedom at is not quite there. We see movies of superb automation, those are movies.
Such machinery is still hand-made.
Teslas are unique in being a chassis, wheel motors and batteries, then the box around it, pure EV cars are remarkably simple, like a skateboard with seats.
Tesla is a robot which does one thing, and has very high tolerance of error. Think how you balance a wheel, or fit a seat. Bang bang bang.
I can’t think of microminiaturized equipment which moves in tight tolerances built at scale today except for medical robots, semiconductor chambers. A semi chamber is $3,000,000 and complex multi-joint medical robots are similar. Also, similar size range as a robot.
That’s what I’d be looking at, not cars. There a vast difference between a skateboard with a computer and something with 240 degrees of freedom, and will definitely require burn-in periods since the full body calibration for each one will take some time.
Constraint 1) sufficient materials flow at L1 L2 L3 assembly for finally assembly. 2) sufficient micromechanical assembly for fine tolerance process that volume can be achieved 3) sufficient test/calibration cycles for full body alignment
These are hand built in laboratories.
Scaling would take 20 years, lab to fab. Tesla had few issues because they built a simple assembly process. Car MFG would love to stop drive trains, all the complex controls, believe me.
Interesting piece. On converting car factories there’s potentially a lot of car companies that are struggling to profitably switch to electrification, and which may also be hit by Full Self Driving/robotaxi, if that works. I wonder if robotics production would provide an ‘out’ for struggling OEMs. Buy out a robot company and switch product
That's already 400x more than the number of humanoid robots I estimated above. That would mean the starting point of the scale up is higher, but since less total scaling happens, the cost declines would be a bit smaller.
These industrial robots are pretty different from the general purpose robots envisioned in the post, but are relevant for the supplies of component parts.
You could also consider tricks to limit the amount of robotic hardware per unit of labor. For example, rail mounted arms use less parts than humanoids and likely have a longer MTBF and a lower error rate. (by likely I mean 'almost certainly', a humanoid chassis is a lot more joints and a lot less stability)
Factories designed around robots might be an OOM faster and more efficient than sending humanoids to operate current equipment.
The real question is : say you have the initial 1 billion robots. Half of them are sent to go make money to show the investors some revenue. Half (500 million) are send to the bottleneck steps in the supply chain for more robots.
What's the doubling time? How long until the second billion? How long can that billion copy itself?
As Sufeitzy already mentioned, factories may not be the key bottleneck, whereas the L1 L2 L3 and material flow is more key. However, factories might be a bottleneck as well on much larger scales, as your assumption about factory conversion doesn't make much sense. You measure car factory output in tons (or kg), but that makes very little sense. Car factories are optimized for sheet metal, engine blocks, and standardized assembly, not the precision mechatronics of humanoid joints and sensors.
As Sufeitzy mentioned before sufficient micromechanical assembly for fine tolerance process that volume can be achieved is very difficult to achieve, especially wholesale in the shorter timeframes you discuss. However, the implications that were left on the table is that factories, even with large amounts of talent and capital, are simply incapable of the manufacturing that you discuss.
With Willow Run, the bombers were like cars much more than cars are robots, with sheet metal and engine blocks. The value they might have are that they are suitable buildings with decent roads. But the idea the existing capital other than that would be able to convert to robot production at all, let alone at a 50% rate in its manufacturing process seems niaeve. As far as I can tell, the greatest asset to the quick scaling would be the huge amount of investment and talent allocated to production, but other than escaping the timeline of classical construction, you may as well be building from scratch. Outside of a couple factories with great placement it seems unlikely that the majority of factories will be close to fully operational after 3-5 years, let alone 1-2 or less.
Additionally, 135 billion kg of car production =! 135 billion kg of robotics production, and I'm confused why you would somehow equivocate the two (in general doing this by weight is a bad way imo as the tech shifts may render some ideas worse just due to the semantic confusion, but I'll use it here to clarify what is being said in the language that was used). The resources involved with the parts of robots alone, not to even mentioned the significantly more complex manufacturing process, are simply 10-100x more expensive, difficult to obtain, etc.
The idea that a factory outputting a couple thousand tons a year where the majority of that tonnage is simply processed metals like its rolling chassis and body could similarly output thousands of tons of highly processed and rarer material robots (again, not even mentioning the manufacturing) could output a similar amount of material seems very unlikely, even if they had enough budget to purchase all of them, the actual tonnage output of these industries is nowhere near the ~hundred million tons, thus all of those industries would have to grow by factors of 10, again not even to mention the actual logistics flow Sufeitzy mentioned.
The two main ideas here is that first of all, the manufacturing process itself would not convert easily or at all from car factories. Secondly and I think perhaps more importantly (if you think with such high amounts of investment and talent means that it doesn't matter if there is conversion or not, it can built from the ground up) the majority of factories, let alone to mention the actual L1 L2 L3 logistics flow that was mentioned, the industries of those L1 L2 L3 literally just don't exist at this scale and would need to be built up before the factories which produce robots themselves.
The point in summary is that these ideas point towards in the next 3 years ~10 million robots of a high level seems like a far higher probability than a billion or 100 million, and anything above ~100 million in less than 2 years (from the start of extreme funding) seems contra logic.
Also I think TSM can make fabs in under 2 years (though you could end up with 3-5 year lag times on lithography machines if trying to scale a lot faster).
I was aiming for a rough average figure, to an order of magnitude, erring on the conservative side.
I agree there are some cheaper models already, but they're also not able to do human jobs. Getting there will partly be about better algorithms, but it may also require more expensive parts (sensors, very precise motors etc).
If you want to use a lower estimate as the starting point, then that accelerates the timeline compared to my estimate.
I think that's a slightly outdated view, China moves very fast in robotics and they seem to be solving the precision motor problems currently while the hardware stays around this cheap. I think 100k is an overestimation, I think China will push this down with all the possible motor flexibility to like 10k a robot.
Primary obstacle seems to be software (AGI for robotics basically)
Seems reasonable, though just tbc my main forecast is for the cost to fall to $10k and then probably $1k as production scales – so I basically agree – it's mainly a question of how fast it happens.
Also my estimate is for "humanoid equivalents" that weight 80kg each. If it's possible to use much smaller and/or simpler specialised robots for most purposes, then in practice we could have a larger number of cheaper robots.
How can robots both be cheap and there be a strong economic incentive to manufacture more? I guess they start expensive, suppressing demand, then gradually reduce in cost as more manufacturing capacity comes online, reducing the economic incentive for increasing capacity. How that plays out is probably complicated, and needs proper modelling not BOTECs, but I'd guess economists have studied it.
Some good comments about this post on Less Wrong:
https://www.lesswrong.com/posts/6Jo4oCzPuXYgmB45q/how-quickly-could-robots-scale-up
Another take here: https://medium.com/limitless-investor/elon-musks-more-robots-than-humans-dream-needs-an-urgent-wake-up-call-a71491540bb9
You’re assuming the factory is the bottleneck, that’s very rarely so.
The main bottleneck is the materials logistics, and flow. Whenever there’s a storm, or major illness, factories shut down because of shortage. Semiconductor memory? Factory shuts down. Boom.
For the largest factories, that requires entire regions of a country to be subsumed to tier 2, tier 3, tier 4 suppliers and the logistics to move materials and packaging in and out without interruptions.
Thats why you have entire regions becoming an industry - automobile factories are surrounded by secondary motors, pumps, seats, gauge, drive train suppliers, and then a layer further out, and so on. Drive around a major auto producer and you see the pattern, go to Wolfsburg and see VW.
Cars are also simple to make, relatively speaking. Robots are a variety of high-precision manufacturing, as do satellites or semiconductor fabrication chambers. The engineering required to create the ability to do that for joints with a full body scale of 240 degrees of freedom at is not quite there. We see movies of superb automation, those are movies.
Such machinery is still hand-made.
Teslas are unique in being a chassis, wheel motors and batteries, then the box around it, pure EV cars are remarkably simple, like a skateboard with seats.
Tesla is a robot which does one thing, and has very high tolerance of error. Think how you balance a wheel, or fit a seat. Bang bang bang.
I can’t think of microminiaturized equipment which moves in tight tolerances built at scale today except for medical robots, semiconductor chambers. A semi chamber is $3,000,000 and complex multi-joint medical robots are similar. Also, similar size range as a robot.
That’s what I’d be looking at, not cars. There a vast difference between a skateboard with a computer and something with 240 degrees of freedom, and will definitely require burn-in periods since the full body calibration for each one will take some time.
Constraint 1) sufficient materials flow at L1 L2 L3 assembly for finally assembly. 2) sufficient micromechanical assembly for fine tolerance process that volume can be achieved 3) sufficient test/calibration cycles for full body alignment
These are hand built in laboratories.
Scaling would take 20 years, lab to fab. Tesla had few issues because they built a simple assembly process. Car MFG would love to stop drive trains, all the complex controls, believe me.
Thank you this is very useful thinking.
Interesting piece. On converting car factories there’s potentially a lot of car companies that are struggling to profitably switch to electrification, and which may also be hit by Full Self Driving/robotaxi, if that works. I wonder if robotics production would provide an ‘out’ for struggling OEMs. Buy out a robot company and switch product
There's now more detailed estimates on this question here:
https://www.forethought.org/research/the-industrial-explosion#how-fast-could-robot-doubling-times-be-initially
Lots more detail about robot supply chains here: https://semianalysis.com/2025/03/11/america-is-missing-the-new-labor-economy-robotics-part-1/
Should this analysis be based on humanoid robots, or just robots in general? This claims there are already ~4m robots installed in factories:
https://ifr.org/ifr-press-releases/news/record-of-4-million-robots-working-in-factories-worldwide
That's already 400x more than the number of humanoid robots I estimated above. That would mean the starting point of the scale up is higher, but since less total scaling happens, the cost declines would be a bit smaller.
These industrial robots are pretty different from the general purpose robots envisioned in the post, but are relevant for the supplies of component parts.
You could also consider tricks to limit the amount of robotic hardware per unit of labor. For example, rail mounted arms use less parts than humanoids and likely have a longer MTBF and a lower error rate. (by likely I mean 'almost certainly', a humanoid chassis is a lot more joints and a lot less stability)
Factories designed around robots might be an OOM faster and more efficient than sending humanoids to operate current equipment.
The real question is : say you have the initial 1 billion robots. Half of them are sent to go make money to show the investors some revenue. Half (500 million) are send to the bottleneck steps in the supply chain for more robots.
What's the doubling time? How long until the second billion? How long can that billion copy itself?
As Sufeitzy already mentioned, factories may not be the key bottleneck, whereas the L1 L2 L3 and material flow is more key. However, factories might be a bottleneck as well on much larger scales, as your assumption about factory conversion doesn't make much sense. You measure car factory output in tons (or kg), but that makes very little sense. Car factories are optimized for sheet metal, engine blocks, and standardized assembly, not the precision mechatronics of humanoid joints and sensors.
As Sufeitzy mentioned before sufficient micromechanical assembly for fine tolerance process that volume can be achieved is very difficult to achieve, especially wholesale in the shorter timeframes you discuss. However, the implications that were left on the table is that factories, even with large amounts of talent and capital, are simply incapable of the manufacturing that you discuss.
With Willow Run, the bombers were like cars much more than cars are robots, with sheet metal and engine blocks. The value they might have are that they are suitable buildings with decent roads. But the idea the existing capital other than that would be able to convert to robot production at all, let alone at a 50% rate in its manufacturing process seems niaeve. As far as I can tell, the greatest asset to the quick scaling would be the huge amount of investment and talent allocated to production, but other than escaping the timeline of classical construction, you may as well be building from scratch. Outside of a couple factories with great placement it seems unlikely that the majority of factories will be close to fully operational after 3-5 years, let alone 1-2 or less.
Additionally, 135 billion kg of car production =! 135 billion kg of robotics production, and I'm confused why you would somehow equivocate the two (in general doing this by weight is a bad way imo as the tech shifts may render some ideas worse just due to the semantic confusion, but I'll use it here to clarify what is being said in the language that was used). The resources involved with the parts of robots alone, not to even mentioned the significantly more complex manufacturing process, are simply 10-100x more expensive, difficult to obtain, etc.
The idea that a factory outputting a couple thousand tons a year where the majority of that tonnage is simply processed metals like its rolling chassis and body could similarly output thousands of tons of highly processed and rarer material robots (again, not even mentioning the manufacturing) could output a similar amount of material seems very unlikely, even if they had enough budget to purchase all of them, the actual tonnage output of these industries is nowhere near the ~hundred million tons, thus all of those industries would have to grow by factors of 10, again not even to mention the actual logistics flow Sufeitzy mentioned.
The two main ideas here is that first of all, the manufacturing process itself would not convert easily or at all from car factories. Secondly and I think perhaps more importantly (if you think with such high amounts of investment and talent means that it doesn't matter if there is conversion or not, it can built from the ground up) the majority of factories, let alone to mention the actual L1 L2 L3 logistics flow that was mentioned, the industries of those L1 L2 L3 literally just don't exist at this scale and would need to be built up before the factories which produce robots themselves.
The point in summary is that these ideas point towards in the next 3 years ~10 million robots of a high level seems like a far higher probability than a billion or 100 million, and anything above ~100 million in less than 2 years (from the start of extreme funding) seems contra logic.
You might be better off using something like a chip fab as a model and these currently take 3-5 years to build from scratch
Maybe as an upper bound?
Also I think TSM can make fabs in under 2 years (though you could end up with 3-5 year lag times on lithography machines if trying to scale a lot faster).
You sat Humanoid robots cost 100k. But that seems very wrong looking at what unitree is doing? https://www.unitree.com/g1
I was aiming for a rough average figure, to an order of magnitude, erring on the conservative side.
I agree there are some cheaper models already, but they're also not able to do human jobs. Getting there will partly be about better algorithms, but it may also require more expensive parts (sensors, very precise motors etc).
If you want to use a lower estimate as the starting point, then that accelerates the timeline compared to my estimate.
I think that's a slightly outdated view, China moves very fast in robotics and they seem to be solving the precision motor problems currently while the hardware stays around this cheap. I think 100k is an overestimation, I think China will push this down with all the possible motor flexibility to like 10k a robot.
Primary obstacle seems to be software (AGI for robotics basically)
Seems reasonable, though just tbc my main forecast is for the cost to fall to $10k and then probably $1k as production scales – so I basically agree – it's mainly a question of how fast it happens.
Also my estimate is for "humanoid equivalents" that weight 80kg each. If it's possible to use much smaller and/or simpler specialised robots for most purposes, then in practice we could have a larger number of cheaper robots.
How can robots both be cheap and there be a strong economic incentive to manufacture more? I guess they start expensive, suppressing demand, then gradually reduce in cost as more manufacturing capacity comes online, reducing the economic incentive for increasing capacity. How that plays out is probably complicated, and needs proper modelling not BOTECs, but I'd guess economists have studied it.
Suppose their cost of production reaches $10k, but they can do work worth $10/hour. Then demand is really high.
Initially production is limited, so the robot companies charge high prices, and earn large profits, selling a relatively small number.
Eventually production scales up, the market becomes more competitive, and prices drop.