Why Physical AI Is Stuck on Actuators, Not Models

Overview

The debate over whether AI’s next bottleneck is intelligence or mechanics is being won by physical constraints, not algorithmic ones. While industry figures like Boris Power have predicted a general robot parts shortage, the reality is far more specific. The industry is not facing a broad component deficit but rather a tight chokepoint at two precise nodes in the supply chain: high-precision actuators and rare earth magnets.

These bottlenecks are governed by the rigid laws of manufacturing physics, which expand on timelines measured in years, not the exponential curves of model capability that expand in months. As foundation models like GPT-6 Astra demonstrate near-perfect semantic generalization, they are colliding with a hard ceiling in micro-force control—a domain where software alone cannot solve the problem.

The Capability Split: Macro vs. Micro

Recent benchmarks reveal a stark divergence in what current large language models can achieve when translated into physical action. In September 2026, Robocurve tested GPT-6 Astra against Claude Fable 5.1 on real robotic arm tasks, producing results that split sharply along the axis of precision.

On a block-into-bowl task, Astra achieved a 95% success rate, completing 19 of 20 attempts compared to Fable 5.1’s 8 of 20. However, on a puzzle piece insertion task requiring millimeter-level alignment, Astra’s success rate plummeted to 10%, matching Fable 5.1’s ceiling exactly.

The same model, the same arm, an 85-point gap between two tasks. The difference is not in world understanding. Block placement is macro-spatial manipulation where semantic generalization works. Puzzle insertion requires micro-force control.

This performance gap highlights that the limitation is not a lack of world understanding. Cameras cannot capture the tactile resistance changes necessary for fine assembly, and language models cannot reason about millisecond-scale force adjustments in microscopic spaces. Only tactile sensors can provide this feedback, yet they remain absent from most generalist robotic fleets.

Supply Chain Realities

The structural constraint lies in the bill of materials. According to McKinsey’s 2026 supply chain analysis, actuators account for 40% to 60% of a humanoid robot’s cost. Within the actuator itself, the gearbox represents 30% to 50% of the expense.

The market for harmonic and strain-wave reducers is historically concentrated, with Harmonic Drive and Nabtesco controlling 70% to 80% of global share, although China’s Green Harmonic has recently expanded to over 35%. Building new capacity for these components is not a matter of capital expenditure; certification cycles and manufacturing physics prevent any compression of timelines.

The deeper constraint resides inside the motor: neodymium-iron-boron (NdFeB) permanent magnets. Each Tesla Optimus unit requires approximately 3.5 to 4 kilograms of these magnets, exceeding the demand of a standard electric vehicle drive motor.

China currently controls approximately 91% of global rare earth refining and 94% of NdFeB magnet production. While mining concentration is lower, the binding constraint is clearly in refining and fabrication. CRU Group forecasts that by 2040, humanoid robot demand could reach 363 kilotons of NdFeB—roughly 157% of current global demand. Under even a low-case scenario, projected 2030 capacity outside China amounts to less than 60,000 tons, compared to China’s 290,000 tons.

Rare earth magnet processing capacity is not GPU fabrication. It does not follow Moore's Law. It does not accelerate because foundation models get smarter.

Vertical Integration and Data

In response to these hardware limitations, OpenAI confirmed in September 2026 that it will independently develop humanoid robots. The company’s robotics team has expanded from 11 to 19 positions, with core roles focused on end-to-end custom electromechanical actuator design.

This pivot follows a strategic lesson from 2021, when OpenAI terminated an earlier robotics program due to insufficient real-world data. Four years later, while cognitive capabilities have transferred across embodiments, the data pipeline for the perception-control loop remains the critical gap.

To move from 10% to 90% precision success, models need data on what the robot tried, how much force was applied, and what physically moved. This data can only be generated by running robots at scale. Owning the hardware fleet is therefore a strategy to control the throughput of this essential training signal.

The Tactile Sensing Signal

GPT-6 Astra’s 10% success rate on precision assembly tasks serves as a direct pricing signal for the tactile sensing sector. The ceiling for precision manipulation without touch is mathematically fixed at this low level. Bridging the gap to 90% requires advances in sensor precision, data collection scale, and force-control algorithms.

Capital markets are already responding to this technological reality. Since July 2026, at least eight financing events in the embodied AI tactile sector have raised over 2.5 billion yuan. Notably, two companies—Pasini and Yimu Technology—have each secured approximately 1 billion yuan in recent rounds, reaching unicorn status.

The financing pace in this specific sub-sector is emerging as the leading indicator for physical AGI’s transition from laboratory demonstration to commercial deployment. As long as the actuator and sensor layers remain constrained by physics rather than code, the hierarchy of bottlenecks will favor hardware manufacturers over pure software firms.