By 2040, we expect general-purpose robots to be part of everyday life. Their influence will reach us through the products we buy and the services we use, long before everyone has access to one at home. The route there is already taking shape on the factory floor.
The race to artificial general intelligence may be creating a second wave with an even larger physical footprint: physical general intelligence, or PGI. At Fellows Forum 2026, Boris Power, OpenAI’s Head of Applied Research, put it plainly:
“I would really suspect there to be a huge boom in physical AGI and we'll probably have a shortage of robots parts.”
Power compared the potential bottleneck with the graphics-card supply chain. Frontier models could help train smaller, faster models to control robots. Robotics would then inherit capabilities developed through the wider AGI race. The constraint could shift towards manufacturing enough machines to put that intelligence to work.
For factory work, our definition of a humanoid is practical: a robot with two coordinated arms that can reproduce human manipulation at human workstations. It can be stationary, move on wheels or have legs if they make sense. One gripper holds a part while the other fits or fastens it. That familiar arrangement makes human demonstrations a natural starting point for teaching tasks.
For this generation of our factories, we are choosing wheeled, dual-arm robots. A wheeled base provides greater stability, room for larger batteries and more onboard compute. A stable base also makes it easier to stop safely at almost any point in a task, without balancing on legs. Those advantages matter for repeatable assembly and recovery.
This useful form factor will become increasingly commoditised. For example, Unitree shipped over 7,000 humanoids and AGIBOT around 9,700 in H1 2026. At two arms per robot, that implies over 14,000 and 19,400 arms respectively, or over 200,000 actuators from just two companies in six months. Over the last 12 months, hardware prices have fallen sharply. A mobile dual-arm system can now be bought for under €15,000. Compare that with a single stationary ABB-style industrial arm with a 5 kg payload, typically costing over €25,000.
Intelligence moves faster than adoption
As we write in our Physical General Intelligence by 2028 article, we believe broadly capable physical intelligence is close. Widespread adoption follows a different clock.
Goldman Sachs’ 2024 forecast put annual humanoid shipments at 1.4 million by 2035, four times its previous estimate. Morgan Stanley has also raised its projections. Elon Musk has called Optimus potentially “the biggest product of all time by far.” We share the conviction that this market will be enormous. But its growth must be built physically and it’s a good reminder that it took Tesla 12 years to manufacture its first 1m cars.
For home robotics, our adoption analogy is fibre broadband reaching European households. ChatGPT arrived through devices and networks people already had. Fibre needed capital, crews and cables in the ground. Robots need components, assembly, delivery, servicing and safe integration into unfamiliar spaces.
A breakthrough can therefore arrive quickly while deployment takes years. Software can improve across an installed fleet overnight; expanding that fleet takes factories. We expect industry to finance and refine the capabilities that eventually make home robots practical.
The early adopters know automation
Manufacturers have an immediate reason to act: labour shortages, cost pressure and the need to add capacity. Humanoids can earn a place by taking over particular tasks before they can perform an entire person’s job.
Two great such stories from the last Humanoid Robotic Summit that our CEO Martin moderated in Stuttgart: France’s Renault is testing Wandercraft’s Calvin at Douai, handling two tyres weighing about 30 kg together. Its roadmap calls for 350 robots across French and Spanish factories deployed in 2027. South Korea’s Hyundai plans to deploy Boston Dynamics’ Atlas in its Georgia operations in the next two years. Trials and production plans are different stages, but both bring humanoids into companies that can measure their usefulness in real industrial roll outs.
That is the advantage of starting in a factory: known cycle times, experienced operators and recurring jobs. A failed pick can become an engineering adjustment and another attempt. Manufacturers provide both a paying application and a place to improve.
The data-centre expansion adds a powerful customer to this story: AI itself. NVIDIA and Foxconn are developing factory simulation and robotics while expanding production of AI infrastructure. Foxconn has demonstrated a wheeled humanoid performing tasks including force-controlled screw fastening. The commercial incentive is direct: the companies expanding intelligence also need more physical work done.
We see a feedback loop beginning. AI helps develop robots; robots could help manufacture the infrastructure that expands AI. Industrial demand gives the next wave somewhere to start.
What useful work looks like in 2027
Start with those arms. Unitree’s G1 uses five actuators per arm, or seven with the additional wrist joints, before counting the fingers. Each is a compact motor-and-gear assembly. Like many electric-car motors, its motor uses current in copper windings interacting with permanent magnets to produce rotation. Miniaturisation makes it light enough to carry; gearing trades motor speed for the turning force needed to extend an arm and lift a part.
Munro’s Unitree G1 teardown shows the packaging challenge: a planetary gear stage sits inside a hollow motor, while a sensor after the gearbox measures the joint’s actual position. That feedback matters when a hand meets a part.
Repetition makes heat decisive. Every lift sends current through resistive copper windings; rapid cycles can build heat faster than it escapes, forcing the robot to slow down or cool off. Doubling current quadruples this heating at unchanged resistance. Unitree reports 30% less heat at the same torque from G1+’s upgraded arm motors. That is progress towards sustained assembly work, beyond what a single impressive lift reveals.
For repetitive assembly, our planning loads per arm are roughly 1 kg for R1, 2 kg for G1, 3 kg for G1+ and 5 kg for H2. These are 120ft working assumptions, subject to reach and cycle-time validation, rather than continuous-duty manufacturer ratings. They explain our current focus on products weighing 5 kg or less.
The next gains will come from longer productive hours, fast charging, safe operation without oversized exclusion zones, and coordination between machines. Who takes over when one robot charges? Can a failed attempt be retried without calling an operator? Those answers determine output.
Our expectation for 2027 is more humanoids doing useful work in deliberately designed workflows. A slower robot can still add capacity if buffers separate its task from the line’s pace. Each productive deployment builds the case for the next. That is how the AGI race becomes an industrial wave and how robots eventually enter everyday life with capabilities that have been proven in use.