The hand is the bottleneck

The clearest read on where humanoid robotics is actually stuck came, in 2025, from the company with the most to prove. Through the second half of the year, multiple outlets — reporting first surfaced by The Information and carried by TrendForce, Tom's Hardware, Digitimes, and Electrek — described Tesla scaling back Optimus production plans, from a target near 5,000 units toward roughly 2,000, and building only a few hundred, with partial robot bodies stockpiled and left incomplete. The part they could not finish was the same part every account named as hardest: the hands and forearms. The reported failures were mundane and physical — motors overheating, grip strength too weak, joints failing too often to run a line.

A humanoid that can walk, balance, and carry a tote but cannot reliably close its fingers on a part is not a worker. It is a torso on legs. Tesla could build the torso and the legs at volume in 2025. It could not build the hands. That asymmetry — everything above the wrist essentially solved, everything past it not — is the single most useful fact about the state of the field.

The hand is hard because it concentrates almost every unsolved requirement into the smallest, most crowded volume on the machine. A capable hand needs high articulation — Sanctuary AI's Phoenix hand claims 21 active degrees of freedom for in-hand manipulation, against a human hand's roughly two dozen — and every one of those axes needs an actuator, a transmission, and its own control, packed into something the size of a fist and expected to survive an industrial duty cycle. Tendons and linkages crowd the palm; the motors that drive them throw heat with nowhere to go. Sanctuary chose miniaturized hydraulics for exactly this reason, citing flow resolution and sensitivity, and reports its valve actuators passed over two billion cycles without leaking. Those are company figures, not independent benchmarks — but the engineering choice tells you how tight the constraint is.

Then there is touch, which vision cannot replace. As Sanctuary's Jeremy Fishel put it, with video alone you do not know you have touched something until well after the collision has already moved the object. Fine force sensing closes that gap: Figure says the fingertip sensors on its Figure 03 robot, unveiled in October 2025, can register forces as small as three grams — the weight of a paperclip — enough to tell a secure grip from an impending slip before the part falls. Sanctuary quotes sensitivity near five millinewtons, approaching the human threshold of about three. But the sensors themselves remain immature. A 2025 peer-reviewed review in Science China Materials names the open problems plainly: insufficient durability, limited coverage, and poor conformability to the hand's curved, jointed surfaces. The signal you most need is the one the hardware is least ready to deliver reliably.

There is no settled answer for how to build the thing, which is itself a sign of how unsolved it is. One school packs in more motors and accepts the heat, cost, and fragility. Another underactuates hard: researchers at the University of Bristol built the Tactile SoftHand-A, which reaches 15 degrees of freedom from just two degrees of actuation through an antagonistic tendon mechanism, with a 3D-printed vision-based touch sensor formed directly into the fingertips. Fewer motors, lower cost, more compliance — at the price of direct control over every joint. The field is still arguing the tradeoff in public.

Even Tesla is visibly mid-argument. Patents surfacing in April 2026 described an Optimus hand with four degrees of freedom per finger and two at the wrist, roughly 25 actuators moved back into the forearm and each finger pulled by thin tendon cables routed through the wrist — a design meant to get the heat and the bulk out of the hand itself. Yet Musk has since said publicly that the patented rolling-contact mechanism did not work in the real world despite working in simulation, and separately claimed the hand problems were being overcome. What Tesla actually ships is, honestly, uncertain right now. A patent is an idea under test, not a solved subsystem — and the most-watched hand program in the world is still cycling through architectures.

Capital has noticed that the hand is both the bottleneck and a business. Goldman Sachs, which raised its humanoid total addressable market to about USD 38 billion by 2035 — more than sixfold above its earlier USD 6 billion estimate — still names robot manipulation, the plain act of grasping objects, as one of the significant remaining bottlenecks in the way of that number. The gap between the size of the prize and the state of the hand is where the investment is flowing. In China, the dexterous-hand segment alone drew more than 20 financing rounds and over 3 billion yuan — roughly USD 415 million — from 2024 onward, per tallies from the Sci-Tech Innovation Board Daily carried by 36Kr, with degrees of freedom and tactile range cited as the competitive dividing lines. That figure is a floor; later rounds have run larger.

What this describes is a distinct market layer sitting beneath the robot: the end-of-arm tooling — hands, grippers, fingertips, tactile skins, and the fixtures that hold and present the work. It is engineered, sold, and funded separately from the humanoid it attaches to, because it is a separately hard problem. The robot is the general platform. The tooling is the specific, physical interface to the task — and it is where the last unsolved engineering, and a meaningful share of the spend, actually lives.

Moduloa's thesis holds that the durable moat in physical AI is the framework — the standards, data, certification, tooling, and routing — not the robot, which will commoditize like every general platform before it. The state of the hand is direct evidence for that shape. The one subsystem that is not converging, that every serious team is still redesigning, that is drawing hundreds of millions on its own, is precisely the physical interface between a general machine and a specific job. A commodity humanoid still cannot do useful work without a purpose-built, reliable, certified way to touch the part — and that layer is exactly where a framework compounds and a robot does not.

The honest limits are worth stating plainly. General-purpose dexterous hands good enough for a factory floor do not yet exist at reliability and cost; much of the best public data on grip forces, cycle life, and sensitivity is company self-reported rather than independently benchmarked; and Tesla's current architecture is genuinely unsettled as this note publishes. What the year establishes is not that the hand is solved, but that it is the thing left to solve — and that betting on the layer beneath the robot, rather than the robot itself, is the more defensible bet while the hardest inches remain open. This note scores nothing on the register; it is evidence about the terrain those predictions cross, not a test of any dated claim. See the register →

This is a synthesis of company announcements, industry-trade and business reporting, peer-reviewed research, and analyst commentary. Non-obvious claims are cross-checked across at least two independent sources; company performance specs (grip forces, cycle life, sensitivity, degrees of freedom) are self-reported vendor figures echoed by trade press, not independently benchmarked, and are attributed as such. Key references:

TrendForce — Tesla scales back Optimus production as hand-design issues stall assembly · Tom's Hardware — Optimus production plans halted · Electrek — Optimus program head exits, production delayed · Teslarati — Optimus V3 hand and arm details revealed in new patents · Figure — Introducing Figure 03 · Sanctuary AI — In-hand manipulation and hand dexterity · The Robot Report — Sanctuary AI integrates tactile sensors into Phoenix · International Journal of Robotics Research — The Tactile SoftHand-A · Science China Materials — Tactile sensing for dexterous robotic hands (review) · Goldman Sachs — The global market for humanoid robots could reach $38 billion by 2035 · 36Kr — Dexterous-hand financing (Sci-Tech Innovation Board Daily tally)

Limits of this note: the Tesla production figures originate largely from one initial report (The Information) corroborated by trade outlets, and describe mid-to-late 2025; Musk has since given conflicting public statements on whether the hand problems are overcome, so the current shipping design should be treated as unsettled. The Goldman TAM figures were first published in an earlier report and later reaffirmed. The Chinese financing tally is a directional count from trade coverage, not audited disclosure. Corrections are welcome: shk@moduloa.com.