Robots Are About to Get a Compute Bill
Picture an industrial arm welding the same seam on a car door, twenty thousand times. Once it's programmed, the marginal cost of doing it the twenty-thousand-and-first time is close to zero — you're paying for electricity, wear, and maintenance. Getting it to that point is the expensive part. The International Federation of Robotics has put programming and integration at roughly 50-70% of the total cost of a typical robot application — a $40k arm easily turns into a $120k automation system once you count the engineering.
That's the traditional shape of robotics cost: enormous upfront investment in teaching, then near-free execution. Which makes traditional robots incredible at high-volume, low-variability work, and often not worth deploying at all for anything that changes often or varies a lot. Nobody's going to pay for a full integration cycle to handle a job that runs for three weeks.
Teaching a Robot Used to Be the Expensive Part
That upfront cost is specifically what AI attacks. Not "the robot gets smarter" — the more concrete version is that AI lowers the cost of telling the robot what to do in the first place. A traditional retooling looks like: engineering ticket, integrator, new fixtures, new program, testing, safety validation, downtime, deployment — thousands to tens of thousands of dollars before the robot does anything useful. The AI version looks more like showing it a handful of examples of the new task, maybe fine-tuning a model, validating, and going.
If that holds up, the economics change more than "robots get better." A $100,000 robot that can be economically repointed at a hundred different jobs can end up cheaper than a $25,000 machine that only ever pays for itself doing one.
Now the Robot Gets a Bill Every Hour It Runs
Here's the part that's easy to miss if you stop at "integration got cheaper." Traditional robotics is a lot like old software: buy the machine, install the program, run it. AI robotics starts looking more like a subscription: buy the machine, then keep paying for the intelligence, continuously, for as long as it operates. Model inference, GPU time, camera and sensor processing, connectivity, occasional retraining, human teleoperation for whatever the model can't handle on its own — none of that existed as a recurring line item for a robot arm running a fixed program.
The shift is really from paying for intelligence as a one-time capital cost to paying for it as an ongoing operating cost. You used to spend a lot up front teaching a machine exactly what to do, then execute that program for free, essentially forever. Now you spend much less teaching it, and pay a small amount every single time it has to actually perceive something and decide what to do about it.
Two Ways to Pay It
That bill has to get paid somewhere, and there are really only two places to put the compute. Send everything to the cloud, and the robot itself stays cheap, but now every action depends on a round trip: camera, network, datacenter GPU, inference, robot. You've made latency, bandwidth, and cloud uptime into dependencies of a physical machine that might be holding something breakable, or standing near a person. That's a considerably more consequential version of "the API is down."
Or put the datacenter on the robot. NVIDIA's Jetson AGX Thor — a real developer kit aimed at exactly this — launched around $3,499 for 128GB of memory and up to 2,070 FP4 TFLOPS in roughly a 130-watt envelope (NVIDIA's own marketplace has since listed it closer to $5,500). That's not just "add a $2,000-ish GPU" as a line item, either. More compute means more battery, more power electronics, more cooling, more physical space, more weight — and robotics has an ugly property where adding a few kilograms of anything tends to make the rest of the machine more expensive too, because now the motors, gearing, and structure all have to handle it. AI can lower the integration bill and raise the hardware bill at the same time, and there's no getting around paying one of these two prices.
The Likely Architecture Is Layered
I don't think production robots end up fully committed to either extreme — streaming every frame to a giant cloud model and waiting for the next motor command isn't going to work for anything that needs to not fall over. The more sensible shape is layered: a cloud tier handling fleet learning, heavier reasoning, and training; a local "brain" running vision and task planning on-device; and real-time controllers underneath handling balance, motor torque, and collision response on their own, with no round trip at all. Google DeepMind's Gemini Robotics On-Device — a vision-language-action model specifically built to run locally, with low latency and without a network connection — is a real, shipping example of exactly this layer, not a hypothetical.
What Actually Matters Is Utilization
Once you're paying continuously for intelligence, the number that ends up dominating the whole equation is how much you're actually using the thing. A $100,000 robot working productively twenty hours a day is cheaper per productive hour than a $30,000 robot doing something useful for three. Every term in the traditional cost model — R&D, hardware, energy, network, inference — matters less than whether the robot is actually earning its keep most of the day, which is exactly the thing a robot that can be retasked cheaply is positioned to do and a single-purpose one usually isn't.
Two Cost Curves, Compounding
There's a second thing happening at the same time that's easy to miss if you're only looking at software: manufacturing itself is scaling. Agility Robotics' RoboFab is built for more than 10,000 humanoid units a year. Figure's BotQ launched with roughly 12,000 units of annual capacity and has already ramped from one robot a day to one robot an hour in a matter of months, with a stated target of 100,000 units within a few years. That's a manufacturing learning curve bringing down the hardware cost, running independently of the foundation-model learning curve bringing down the integration cost. Whether an individual robot costs $30k or $100k in 2026 probably matters less than the fact that both of those curves are moving in the same direction at once.
None of this points toward AI robotics replacing the traditional kind. Classical robotics still wins clearly at high volume, low variation, structured environments, and maximum reliability — putting a language model in the loop for a machine stamping half a million identical parts would be a strange thing to do. AI robotics wins where the traditional model always struggled: low volume, high variation, environments that change, tasks that change often. Warehouse picking, hospital logistics, construction, eventually homes — the more unpredictable the environment, the more the economics favor something that can be retaught cheaply over something that was expensively taught once.
The Actual Question
The honest summary is that AI doesn't make robots cheaper. It moves where the cost sits — from a huge upfront bill for teaching, toward a smaller upfront bill and a running bill for thinking. Traditional robotics gets expensive fast as variability goes up, because the engineering has to account for everything in advance. AI robotics probably starts with a higher baseline cost, since every unit needs real compute just to operate, but that cost should climb much more slowly as the environment gets less predictable. Where those two curves cross is, I think, one of the more interesting open questions in robotics economics right now — and it's a question that gets asked robot by robot, task by task, not settled once for the whole category.
Picture an industrial arm welding the same seam on a car door, twenty thousand times. Once it's programmed, the marginal cost of doing it the twenty-thousand-and-first time is close to zero — you're paying for electricity, wear, and maintenance. Getting it to that point is the expensive part. The International Federation of Robotics has put programming and integration at roughly 50-70% of the total cost of a typical robot application — a $40k arm easily turns into a $120k automation system once you count the engineering.
That's the traditional shape of robotics cost: enormous upfront investment in teaching, then near-free execution. Which makes traditional robots incredible at high-volume, low-variability work, and often not worth deploying at all for anything that changes often or varies a lot. Nobody's going to pay for a full integration cycle to handle a job that runs for three weeks.
Teaching a Robot Used to Be the Expensive Part
That upfront cost is specifically what AI attacks. Not "the robot gets smarter" — the more concrete version is that AI lowers the cost of telling the robot what to do in the first place. A traditional retooling looks like: engineering ticket, integrator, new fixtures, new program, testing, safety validation, downtime, deployment — thousands to tens of thousands of dollars before the robot does anything useful. The AI version looks more like showing it a handful of examples of the new task, maybe fine-tuning a model, validating, and going.
If that holds up, the economics change more than "robots get better." A $100,000 robot that can be economically repointed at a hundred different jobs can end up cheaper than a $25,000 machine that only ever pays for itself doing one.
Now the Robot Gets a Bill Every Hour It Runs
Here's the part that's easy to miss if you stop at "integration got cheaper." Traditional robotics is a lot like old software: buy the machine, install the program, run it. AI robotics starts looking more like a subscription: buy the machine, then keep paying for the intelligence, continuously, for as long as it operates. Model inference, GPU time, camera and sensor processing, connectivity, occasional retraining, human teleoperation for whatever the model can't handle on its own — none of that existed as a recurring line item for a robot arm running a fixed program.
The shift is really from paying for intelligence as a one-time capital cost to paying for it as an ongoing operating cost. You used to spend a lot up front teaching a machine exactly what to do, then execute that program for free, essentially forever. Now you spend much less teaching it, and pay a small amount every single time it has to actually perceive something and decide what to do about it.
Two Ways to Pay It
That bill has to get paid somewhere, and there are really only two places to put the compute. Send everything to the cloud, and the robot itself stays cheap, but now every action depends on a round trip: camera, network, datacenter GPU, inference, robot. You've made latency, bandwidth, and cloud uptime into dependencies of a physical machine that might be holding something breakable, or standing near a person. That's a considerably more consequential version of "the API is down."
Or put the datacenter on the robot. NVIDIA's Jetson AGX Thor — a real developer kit aimed at exactly this — launched around $3,499 for 128GB of memory and up to 2,070 FP4 TFLOPS in roughly a 130-watt envelope (NVIDIA's own marketplace has since listed it closer to $5,500). That's not just "add a $2,000-ish GPU" as a line item, either. More compute means more battery, more power electronics, more cooling, more physical space, more weight — and robotics has an ugly property where adding a few kilograms of anything tends to make the rest of the machine more expensive too, because now the motors, gearing, and structure all have to handle it. AI can lower the integration bill and raise the hardware bill at the same time, and there's no getting around paying one of these two prices.
The Likely Architecture Is Layered
I don't think production robots end up fully committed to either extreme — streaming every frame to a giant cloud model and waiting for the next motor command isn't going to work for anything that needs to not fall over. The more sensible shape is layered: a cloud tier handling fleet learning, heavier reasoning, and training; a local "brain" running vision and task planning on-device; and real-time controllers underneath handling balance, motor torque, and collision response on their own, with no round trip at all. Google DeepMind's Gemini Robotics On-Device — a vision-language-action model specifically built to run locally, with low latency and without a network connection — is a real, shipping example of exactly this layer, not a hypothetical.
What Actually Matters Is Utilization
Once you're paying continuously for intelligence, the number that ends up dominating the whole equation is how much you're actually using the thing. A $100,000 robot working productively twenty hours a day is cheaper per productive hour than a $30,000 robot doing something useful for three. Every term in the traditional cost model — R&D, hardware, energy, network, inference — matters less than whether the robot is actually earning its keep most of the day, which is exactly the thing a robot that can be retasked cheaply is positioned to do and a single-purpose one usually isn't.
Two Cost Curves, Compounding
There's a second thing happening at the same time that's easy to miss if you're only looking at software: manufacturing itself is scaling. Agility Robotics' RoboFab is built for more than 10,000 humanoid units a year. Figure's BotQ launched with roughly 12,000 units of annual capacity and has already ramped from one robot a day to one robot an hour in a matter of months, with a stated target of 100,000 units within a few years. That's a manufacturing learning curve bringing down the hardware cost, running independently of the foundation-model learning curve bringing down the integration cost. Whether an individual robot costs $30k or $100k in 2026 probably matters less than the fact that both of those curves are moving in the same direction at once.
None of this points toward AI robotics replacing the traditional kind. Classical robotics still wins clearly at high volume, low variation, structured environments, and maximum reliability — putting a language model in the loop for a machine stamping half a million identical parts would be a strange thing to do. AI robotics wins where the traditional model always struggled: low volume, high variation, environments that change, tasks that change often. Warehouse picking, hospital logistics, construction, eventually homes — the more unpredictable the environment, the more the economics favor something that can be retaught cheaply over something that was expensively taught once.
The Actual Question
The honest summary is that AI doesn't make robots cheaper. It moves where the cost sits — from a huge upfront bill for teaching, toward a smaller upfront bill and a running bill for thinking. Traditional robotics gets expensive fast as variability goes up, because the engineering has to account for everything in advance. AI robotics probably starts with a higher baseline cost, since every unit needs real compute just to operate, but that cost should climb much more slowly as the environment gets less predictable. Where those two curves cross is, I think, one of the more interesting open questions in robotics economics right now — and it's a question that gets asked robot by robot, task by task, not settled once for the whole category.
Picture an industrial arm welding the same seam on a car door, twenty thousand times. Once it’s programmed, the marginal cost of doing it the twenty-thousand-and-first time is close to zero — you’re paying for electricity, wear, and maintenance. Getting it to that point is the expensive part. The International Federation of Robotics has put programming and integration at roughly 50-70% of the total cost of a typical robot application — a $40k arm easily turns into a $120k automation system once you count the engineering.
That’s the traditional shape of robotics cost: enormous upfront investment in teaching, then near-free execution. Which makes traditional robots incredible at high-volume, low-variability work, and often not worth deploying at all for anything that changes often or varies a lot. Nobody’s going to pay for a full integration cycle to handle a job that runs for three weeks.
Teaching a Robot Used to Be the Expensive Part
That upfront cost is specifically what AI attacks. Not “the robot gets smarter” — the more concrete version is that AI lowers the cost of telling the robot what to do in the first place. A traditional retooling looks like: engineering ticket, integrator, new fixtures, new program, testing, safety validation, downtime, deployment — thousands to tens of thousands of dollars before the robot does anything useful. The AI version looks more like showing it a handful of examples of the new task, maybe fine-tuning a model, validating, and going.
If that holds up, the economics change more than “robots get better.” A $100,000 robot that can be economically repointed at a hundred different jobs can end up cheaper than a $25,000 machine that only ever pays for itself doing one.
Now the Robot Gets a Bill Every Hour It Runs
Here’s the part that’s easy to miss if you stop at “integration got cheaper.” Traditional robotics is a lot like old software: buy the machine, install the program, run it. AI robotics starts looking more like a subscription: buy the machine, then keep paying for the intelligence, continuously, for as long as it operates. Model inference, GPU time, camera and sensor processing, connectivity, occasional retraining, human teleoperation for whatever the model can’t handle on its own — none of that existed as a recurring line item for a robot arm running a fixed program.
The shift is really from paying for intelligence as a one-time capital cost to paying for it as an ongoing operating cost. You used to spend a lot up front teaching a machine exactly what to do, then execute that program for free, essentially forever. Now you spend much less teaching it, and pay a small amount every single time it has to actually perceive something and decide what to do about it.
Two Ways to Pay It
That bill has to get paid somewhere, and there are really only two places to put the compute. Send everything to the cloud, and the robot itself stays cheap, but now every action depends on a round trip: camera, network, datacenter GPU, inference, robot. You’ve made latency, bandwidth, and cloud uptime into dependencies of a physical machine that might be holding something breakable, or standing near a person. That’s a considerably more consequential version of “the API is down.”
Or put the datacenter on the robot. NVIDIA’s Jetson AGX Thor — a real developer kit aimed at exactly this — launched around $3,499 for 128GB of memory and up to 2,070 FP4 TFLOPS in roughly a 130-watt envelope (NVIDIA’s own marketplace has since listed it closer to $5,500). That’s not just “add a $2,000-ish GPU” as a line item, either. More compute means more battery, more power electronics, more cooling, more physical space, more weight — and robotics has an ugly property where adding a few kilograms of anything tends to make the rest of the machine more expensive too, because now the motors, gearing, and structure all have to handle it. AI can lower the integration bill and raise the hardware bill at the same time, and there’s no getting around paying one of these two prices.
The Likely Architecture Is Layered
I don’t think production robots end up fully committed to either extreme — streaming every frame to a giant cloud model and waiting for the next motor command isn’t going to work for anything that needs to not fall over. The more sensible shape is layered: a cloud tier handling fleet learning, heavier reasoning, and training; a local “brain” running vision and task planning on-device; and real-time controllers underneath handling balance, motor torque, and collision response on their own, with no round trip at all. Google DeepMind’s Gemini Robotics On-Device — a vision-language-action model specifically built to run locally, with low latency and without a network connection — is a real, shipping example of exactly this layer, not a hypothetical.
What Actually Matters Is Utilization
Once you’re paying continuously for intelligence, the number that ends up dominating the whole equation is how much you’re actually using the thing. A $100,000 robot working productively twenty hours a day is cheaper per productive hour than a $30,000 robot doing something useful for three. Every term in the traditional cost model — R&D, hardware, energy, network, inference — matters less than whether the robot is actually earning its keep most of the day, which is exactly the thing a robot that can be retasked cheaply is positioned to do and a single-purpose one usually isn’t.
Two Cost Curves, Compounding
There’s a second thing happening at the same time that’s easy to miss if you’re only looking at software: manufacturing itself is scaling. Agility Robotics’ RoboFab is built for more than 10,000 humanoid units a year. Figure’s BotQ launched with roughly 12,000 units of annual capacity and has already ramped from one robot a day to one robot an hour in a matter of months, with a stated target of 100,000 units within a few years. That’s a manufacturing learning curve bringing down the hardware cost, running independently of the foundation-model learning curve bringing down the integration cost. Whether an individual robot costs $30k or $100k in 2026 probably matters less than the fact that both of those curves are moving in the same direction at once.
None of this points toward AI robotics replacing the traditional kind. Classical robotics still wins clearly at high volume, low variation, structured environments, and maximum reliability — putting a language model in the loop for a machine stamping half a million identical parts would be a strange thing to do. AI robotics wins where the traditional model always struggled: low volume, high variation, environments that change, tasks that change often. Warehouse picking, hospital logistics, construction, eventually homes — the more unpredictable the environment, the more the economics favor something that can be retaught cheaply over something that was expensively taught once.
The Actual Question
The honest summary is that AI doesn’t make robots cheaper. It moves where the cost sits — from a huge upfront bill for teaching, toward a smaller upfront bill and a running bill for thinking. Traditional robotics gets expensive fast as variability goes up, because the engineering has to account for everything in advance. AI robotics probably starts with a higher baseline cost, since every unit needs real compute just to operate, but that cost should climb much more slowly as the environment gets less predictable. Where those two curves cross is, I think, one of the more interesting open questions in robotics economics right now — and it’s a question that gets asked robot by robot, task by task, not settled once for the whole category.