+First time on the top traders desk?
The desk runs its own numbers and publishes its calls next to the account that trades them. The short route:
by 1X just shipped the hands. "An API to the physical world."
1X unveiled a new generation of hands for its NEO humanoid: 25 degrees of freedom, tendon-driven, force-transparent, with tactile skin that feels slip before it happens. Nothing like it has existed on the market. And they can build 10,000 of them this year.
- 1X unveiled a 25-degree-of-freedom hand for its NEO humanoid (22 joints in the fingers and palm, 3 at the wrist), tendon-driven and force-transparent, with tactile skin that detects a slip before the object falls.
- The hands run quasi-direct-drive tendons at a ~5:1 to 15:1 gear ratio versus the field-standard 100:1-200:1, so all 25 joints are backdrivable and natively force-controlled; every joint doubles as a force sensor.
- Specs include ±0.2mm positioning accuracy, 45N distal flexion force, IP68 food-safe sealing, and wrist joints validated past 2 million cycles under load.
- 1X says it can build 10,000 hands in 2026 off a dedicated in-house line covering tendon materials, motors, polymers, skin and the tactile stack.
- CEO Bernt Børnich frames the hand as the humanoid's API and argues the hardware ceiling is now gone, so the remaining barrier is data, not dexterity.
- The TT desk call: don't bet on the humanoid, bet on what every humanoid must buy: precision actuators, low-ratio transmissions, tendon materials, tactile sensor stacks, IP68 polymers and high-yield manufacturing lines.
- Watch whether any rival ships comparable dexterity at comparable volume within 12 months; if not, 1X's vertically integrated data flywheel compounds alone.
Watch it first
Before any description, watch what these hands do. Twenty-five seconds of grasps, in-hand rotation, tool use, fine pinch.
1X: "Twenty-five ways to ask a question."
Not sped up. Real-time speed, dexterity and precision.
"A humanoid is a computer whose API is its hands"
That's 1X's own framing, and it's the right one. If the robot's brain is the computer, the hands are the interface: they set what it can know about the world and what it can do in it. A humanoid with a two-finger gripper exposes three verbs to developers (pick, place, push), and every application ever written on that platform is a composition of those three, executed blind. The ceiling was never in the software. It was at the end of the arm.
Most robot hands are write-only devices. You command a position; the hand goes there; nothing meaningful comes back.1X — NEO's Hands
What's actually new in the engineering
22 actuated in fingers and palm, 3 at the wrist. Distributed anatomically, not evenly: the thumb carries roughly as much as the rest of the fingers combined, because a thumb that genuinely opposes is what makes a grasp.
Quasi-direct-drive tendons. The field standard is 100:1-200:1, where friction swallows contact force before it reaches the motor. Low ratios make every joint backdrivable and force-transparent.
Small-object regime: screws, coins, USB-C, LEGO. Which is where most human labor actually happens.
3.5 Nm peak thumb CMC torque, 2.6 Nm at finger MCP, 17.75 Nm at the wrist. Motors sit in the forearm and pull tendons through the wrist, like yours. Light hand, low inertia, high force.
NEO washes its own hands. Wrist joints validated past 2M cycles under load; full finger assemblies through millions more.
Hundreds already off a dedicated in-house line: tendon materials, motors, polymers, skin, tactile stack. The strategic number, not a manufacturing footnote.
The part that matters most: it reads back
Every one of the 25 joints is natively force-controlled and doubles as a force sensor. Push a finger and it yields, and reports exactly how hard you pushed. 1X calls this force transparency: force flows out and information flows back through the same physical path. Layered on top, the fingertips carry high-resolution tactile sensing that measures normal force, contact location and shear, so the hand detects a slip beginning and re-grips before the object falls.
The consequence is that the hand always knows its own pose and effort without looking, the same way you can touch your fingertips together with your eyes shut. Motion becomes smooth by construction and fine motor control stops being a demo. Every grasp becomes a labeled experiment, which is exactly the substrate learning-based manipulation has been starved of.
Clips: 1X. Also demonstrated: screws and coins out of a wallet, light-bulb install, screwdriver use, tea pouring, catching a ball, wine glass, sign language.
The founder's claim
Our goal was never a hand that just looks impressive on paper. With these hands, NEO crosses a critical threshold. The robot can now do the things humans do with their hands, every day.Bernt Børnich, CEO & founder, 1X
Why this is an investment question, not a demo
1X is right about the hierarchy. On the hardware side of a humanoid, nothing matters more than the hands, and this is, without exaggeration, something new in robot engineering. Put a capable brain behind hands like these and you get the most functional robot in the world, and by extension the market leader.
But note where the claim actually lands: 1X is not arguing it has better AI. It's arguing that the hardware ceiling is gone, so data is now the only barrier. That reframes the whole race. If dexterity stops being the constraint, competitive advantage moves to whoever can (a) manufacture dexterous hands at volume and (b) run the most contact-rich experiments per dollar. Both are supply-chain and throughput questions (actuators, tendon materials, tactile sensor stacks, yield, end-of-line test), not model questions.
Hence the 10,000-hands number. A hand that can't be built at scale can't run experiments at scale, and without data at scale there's no embodied AGI. That's the bet, stated plainly.
The TT desk thoughts
This does not change our position from Is physical AI investable?; it sharpens it. Don't bet on the humanoid; bet on what every humanoid must buy. A 25-DoF, force-transparent, tactile hand needs precision actuators, harmonic-free low-ratio transmissions, tendon materials, tactile sensor stacks, IP68 polymers and, above all, a high-yield line. 1X vertically integrated all of it, which tells you exactly where the scarcity is.
Two things to watch from here. First, whether anyone else ships comparable dexterity at comparable volume in the next 12 months; if not, 1X's data flywheel compounds alone. Second, whether the hands survive contact with customers rather than demo reels: millions of test cycles is a lab claim until fleets are in homes. Vertical integration is the moat while it holds and the cost structure when it doesn't.
▸ The full engineering story — 1X's argument, condensed
Condensed from 1X's technical write-up, "NEO's Hands" (July 9, 2026).
1. The hand is a perception stack, not an actuator
Watch a person meet an unfamiliar object and notice how little they look at it. They press it to find hardness, slide a fingertip to read texture, heft it for weight, squeeze to feel it give. Touch isn't a passive channel like a camera; it's an experiment. The hand asks a question with force and reads the answer back through the same joints that asked it.
- Writing the question takes precise force control.
- Reading the answer takes backdrivability and force transparency, so the world's reaction reaches the motor instead of dying in a gearbox.
- Posing the question takes degrees of freedom and precision.
- Feeling the fine print takes skin. Catching a fast answer takes bandwidth. Asking millions of times takes robustness, because probing is contact and contact is wear.
2. Read-write vs write-only
At the 100:1 and 200:1 gear ratios common in the field, friction swallows contact forces before they ever reach the motor. The hand is numb through its own joints, so builders wrap it in external sensors and infer what's happening at the fingertips: a camera pointed at a hand that can't feel. The 1X Tendon Drive runs quasi-direct-drive at ~5:1 to 15:1, making all 25 DoF natively force-controlled and fully backdrivable.
Proprioception
Because every joint is closed-loop, the hand always knows its configuration without looking. Pose plus effort, through the same 25 joints, all the time.
Tactile skin
Tactile data is an image: dynamic range, resolution, channels, field of view. Fingertips and surfaces measure normal force, contact location and shear, so slip is detected as it begins. The skin is co-designed with the sensors inside it and the tendons behind it: a functional material, not a cosmetic one. Vision alone fails on small, transparent, deformable or occluded objects.
3. Safe by construction
Extremely low gear ratios plus tendon drive and low distal inertia let external impacts safely backdrive the fingers. 1X's slow-motion footage shows the hands yielding when slapped, hit with a hammer, pinched in a closing drawer, or slammed into Styrofoam. A machine that perceives by touching had better be gentle by construction; the world it needs to touch has people in it.
4. Manufacturing is the strategy
Hundreds of hands have already come off a dedicated production line, built end-to-end in-house: tendon materials, 1X Motors, soft polymers, skin, tactile stack, hand-specific firmware. High-yield processes with full end-of-line testing give capacity for 10,000 hands this year.
The bottom line
For seventy years, robotics worked around the hand problem. The humanoid bet is the reverse. And it lives or dies at the fingertips.
A hand that perceives by acting turns every task into an experiment: every joint reports force, every fingertip reports contact and shear, every pose is known. Every grasp a policy takes arrives pre-labeled. It's the same loop children run: act, feel, update, act again.
What would change our mind
A humanoid maker bringing hands, actuators and tactile stacks in-house at volume. The position here is that the body is commodity and the components are not; vertical integration by the assemblers is what would move the margin back up the stack.
Keep reading
Is physical AI investable? — the full framework · Robotics · All alpha