The Information reported on 26 August 2026 that SoftBank Group is in talks to buy a majority stake in 1X Technologies at about US$6 billion. Reuters carried the same report that day and said it could not independently confirm it. SoftBank declined to comment. 1X did not immediately respond. The talks are still open and the terms can still move. That is the news event: a control conversation around a home humanoid that has a public spec page and no public proof a paying customer has one in a house.
Call it a capital story, not a close. Majority is undefined. Pre-money versus post-money is unreported. Primary cash into the company versus a secondary cash-out is unreported. Whether OpenAI sells any of its 2023 Startup Fund position is unreported. Founder Bernt Børnich’s operating control is unreported. Treat every headline that says SoftBank “buys 1X” as ahead of the filing.
The number sits under the figure 1X was chasing last fall, when it was reported to be seeking up to US$1 billion at US$10 billion or more and, per The Information, landed less than half the raise it wanted. A majority purchase and a minority primary round are not the same instrument, so do not flatten them into “1X is now worth 40 percent less.” Do note the obvious: control usually prices at a premium, and a US$6 billion control talk still has to live next to a robot that is late on its own delivery promise.
66 lb robot, 154 lb lift, unpublished hip torque
1X’s own NEO page is more complete than most humanoid marketing sheets, which is why the gaps on it matter. Height 5 ft 6 in. Mass 66 lb (about 30 kg). Arm payload 18 lb. Carry 55 lb. Lift 154 lb. Hands 22 degrees of freedom each, arms 7, neck 3, spine 2, legs 6 each. Walking speed 1.4 m/s. Claimed max run 6.2 m/s. Pack 842 Wh and a 4-hour runtime, with a 6-minute charge per hour of runtime and a self-dock. Hands IP68, body IP44. Audible noise 22 dB. Compute is a 1X “NEO Cortex” on Nvidia Jetson Thor, rated up to 2,070 FP4 TFLOPS. Those are 1X’s numbers, not mine.
The lift figure is the one I would mark in red on a purchase order. A 30 kg tendon-driven machine claiming a 70 kg lift is a peak, whole-body, or both-arms number until 1X publishes the test. The 18 lb arm payload is the only figure that looks like a working load. Carry at 55 lb needs a two-hand, torso-close condition next to it. None of the three come with a continuous rating, a reach, or a thermal soak. Hip, knee, and ankle peak and continuous torque are still not on the page. Tendon Drive is specified for 2 percent torque accuracy, 95 percent backdrivability, 2 million nominal load cycles, and 100,000 peak (3× nominal) cycles, with a “1 day service replacement” note. That is a vendor endurance claim, not a published joint curve.
Hands are the same split. The product page still lists 22 DOF per hand. On 9 July 2026, 1X announced a 25-DOF tendon-driven hand — 22 actuated in the fingers and palm plus 3 at the wrist — with about 5:1 to 15:1 quasi-direct-drive ratios, 3.5 Nm at the thumb CMC, 2.6 Nm at the finger MCP, 45 N distal flexion, and 17.75 Nm at the wrist. Until the order page and the July brief describe the same SKU, treat 22 and 25 as two documents, not one shipping hand. Tendon hands fail on stretch, routing, heat, and contamination long before they fail a DOF count. 1X says it can build 10,000 of those hands this year. Building hands is not the same as a home robot finishing a laundry cycle without a remote operator.
The 6.2 m/s run claim is 22 km/h on a 30 kg, 22 dB, knit-covered home machine. I will believe a lab peak. I will not size a house around it without a duty cycle and a fall-recovery clip that is not a cut.
Expert Mode is the home test, not the knit suit
NEO is sold as a US$20,000 Early Access purchase with priority 2026 delivery, or US$499 a month later, with a US$200 deposit on the site. 1X said the first year of capacity sold out in five days after the October 2025 open — 10,000 units on paper. Late August 2026, no customer has been named with a unit in a house. 1X has not announced that shipments reached buyers. Internal employee home testing is not a delivery. Preorder sold-out is demand. It is not logistics, service, or a mean time between assists in someone else’s kitchen.
The honest part of the 1X pitch is Expert Mode. NEO is supposed to run Redwood AI autonomously on chores it already knows. For chores it does not know, a scheduled 1X Expert can remotely supervise — VR, mobile, the same path as the “pilot your NEO from anywhere” remote-control feature. That is teleoperation with a learning story attached. It is also a privacy and uptime problem: a home robot that needs a scheduled stranger on the joint bus is not a dishwasher. Four hours on 842 Wh is a real pack number. It is also a short domestic shift once you add walking, manipulation, and the Thor board. Self-charge only helps if the dock is reachable after a fall.
Hayward, California, is the factory 1X points at — 58,000 square feet, vertical, aiming at 10,000 units a year now and 100,000 by the end of 2027 with a San Carlos site. Those are capacity slides. SoftBank money, if it is primary, would fund lines, compute for the World Model Lab, and the service network a home fleet actually burns. If it is mostly secondary, 1X gets a new owner and the same delivery problem.
OpenAI already has a 2023 Startup Fund position. The Information said OpenAI and 1X talked about an outright purchase last year and did not close. SoftBank’s robotics book already includes a US$5.4 billion agreement for ABB’s robotics business. A consumer tendon humanoid is a different cell than an ABB arm. That is the strategic picture. It does not make NEO a shipped product.
What a $6 billion talk does not publish
A majority stake at US$6 billion, if it closes, buys time. It does not publish hip torque. It does not reconcile 22-DOF versus 25-DOF hands. It does not convert a 154 lb lift claim into a continuous payload at reach. It does not name a customer with a NEO in a hallway. It does not replace Expert Mode with a published autonomy rate on unstructured chores. Reuters still cannot confirm the talks. Until there is a filing, the engineering test is the same one we use on factory bipeds: did the unit that leaves Hayward match the page, and did it work a shift without a pendant.
FIRGELLI does not make 1X’s tendon drives or the NEO hand, and this is not a claim that we do. Linear actuators belong on the dock, the cabinet, and the fixture around a cell — not as a substitute for an unpublished hip number. If you are comparing NEO to Optimus, Atlas, or a Walker-class industrial biped, hold 1X to the same bar: peak and continuous joint torque, rated payload at reach, pack draw under manipulation, and a delivery that a buyer can photograph. The SoftBank talk is real reporting. An unshipped home robot next to it is also real.