OpenAI CEO Sam Altman: The Company Will Definitely Build Humanoid Robots
Sam Altman stated on the Sources podcast: "We will definitely build humanoid robots, as well as other forms." The reason is that real-world facilities like doors, computers, vehicles, and kitchens are designed for human bodies, making humanoid robots easier to navigate physical environments.
The current focus is not on household servants, but on machines that can assist skilled workers in constructing and maintaining infrastructure, with data centers using specialized non-humanoid machines. The long-term goal is for everyone to have a personal robot that can take on tasks they are unwilling to do. OpenAI Robotics has publicly recruited full-stack hardware, operations, systems, and machine learning engineers, covering custom electromechanical actuators, gears, electronics, firmware, inference, data collection stations, simulation, and control software, aiming to program and manufacture socially useful robots in real environments. The team evolved from a world simulation research project led by Aditya Ramesh and incorporated personnel from the original Sora direction.
The company closed its early robotics group around 2020 to 2021, judging that real-world data was insufficient and iterations were slower than text code models. Early achievements included the dexterous hand Dactyl solving a Rubik's Cube. Since then, it has invested in startups like 1X Technologies, Figure AI, and Physical Intelligence through its venture fund, planning to hire again starting in 2024, with job descriptions in 2025 to 2026 indicating experience in general, adaptive, real dynamic environments, and million-scale mechanical systems. The collaboration with Figure later publicly broke down, and hardware lead Caitlin Kalinowski resigned due to a Pentagon contract. Altman also mentioned that if intelligence approaches generality but cannot perform real-world tasks, it is a limitation.
Who is buying and who is selling: The event is driven by the need to translate large model capabilities into physical bodies, not by already mass-produced humanoids. Buyers are looking for internal computational power expansion for data center operations and construction site assistance, while sellers are focused on actuator and remote operation data collection positions. Beneficiaries are labs with existing world models and multimodal stacks, while those under pressure are purely simulation routes lacking real contact data. Funding is flowing from chat subscriptions to electromechanical, collection stations, and production line trials.
Source: Public information
ABAB AI Insight
OpenAI's first foray into robotics hit a data wall: there is infinite text online, but the physical world lacks equivalent action data. Dactyl proved that a hand can solve a Rubik's Cube, but it does not demonstrate that factories and kitchens can be scaled. After closing, the shift to Figure and 1X essentially outsourced the body while retaining the model. Now publicly stating "we will definitely build humanoids" suggests that the model side believes it has crossed the perception threshold and wants to reclaim the body to prevent humanoid companies from starting anew with its models.
The capital path is to first stack computational power and data centers using chat and code proxies, then use robots to build the next data center, creating a self-expanding closed loop. Money is flowing into actuators and collection stations because the bottleneck of embodied intelligence has shifted from parameter quantity to contact force and failure samples. Personal robots are a long-term narrative; if there are any near-term orders, they are likely in server rooms and construction sites.
The comparative objects are Tesla's Optimus production line narrative, Boston Dynamics' hydraulic tradition, NVIDIA's Isaac and Jetson Thor toolchains, and the pressure for self-research after the split with Figure. The industry position is shifting from investing in others' bodies to expanding self-manufactured bodies, with control lying in who possesses action data rather than just language model weights.
Structural changes indicate technological replacement. The mechanism is that environments are designed for humans, and humanoids reduce the cost of transforming the world; however, if intelligence is sufficient, non-humanoids are more effective in server rooms. The coexistence of both forms illustrates that morphology follows tasks: humanoids serve established human spaces, while specialized machines serve newly constructed machine spaces.
ABAB News · Cognitive Laws
The world is built for humans, so general robots must first grow into humanoids.
If models cannot reach hands and feet, general intelligence remains confined to screens.
What is lacking is not parameters, but real-world data from failures.