OpenAI’s pivot to physical AI is not primarily a hiring story — it is a data-infrastructure story. The company posted 27 robotics-related job listings this week, more than double the 11 it advertised four months ago, with base salaries reaching $500,000 a year. That salary ceiling is a tool for pulling specialized engineers away from the humanoid startups OpenAI once funded. But the deeper signal in the listings is structural: when you map the open roles onto OpenAI’s two operating robotics facilities, what emerges is less a research lab than an industrial pipeline for manufacturing the one thing that has held general-purpose robotics back since the very beginning — labeled training data at scale.
OpenAI Shut Down Its First Robot Program Because It Ran Out of Data
OpenAI ran an internal robotics team until 2021, when co-founder Wojciech Zaremba explained the team had reached a hard limit: the underlying approach was technically sound, but the company lacked enough real-world data to train capable robot models. That is not an algorithmic failure. It is a data failure — and it is a failure mode that is well understood in the academic literature.
To train a robot to do something as deceptively simple as placing a rubber duck in a cup, researchers need hundreds or thousands of recorded demonstrations: what the camera saw, what the arm position was, what the gripper did at each moment. That process — called behavior cloning or imitation learning — works the same way large language models learn from text, except that instead of scraping the internet, someone has to physically move the robot through the task, repeatedly, with a controller that records every joint angle and force reading. The academic term is teleoperation: a human operator wearing or holding a device that mirrors the robot’s kinematics drives the arm through the task while the system logs every state-action pair. Researchers at UC Berkeley codified a low-cost version of this approach in 2023 called GELLO (General, Low-Cost, and Intuitive Teleoperation), a 3D-printed controller that costs under $300 per unit and has the same physical structure as the robot arm it controls. The Berkeley group, led by Pieter Abbeel, showed that GELLO outperforms VR controllers and spacemice for collecting high-quality robot demonstrations, based on an extensive user study.
OpenAI chose GELLO controllers paired with Franka robotic arms for its San Francisco teleoperation facility, which launched in February 2025. The facility operates 24 hours a day across three shifts, with contract workers directing the arms through progressively complex household tasks — placing rubber ducks in cups, loading bread into a toaster, folding laundry — to generate the state-action datasets that robot policies learn from. Since its launch, the lab has grown from approximately 100 contract workers to more than four times that scale.
What OpenAI Is Building in Richmond
In March 2026, OpenAI signed a lease for 202,000 square feet (approximately 18,766 square meters) of industrial warehouse space at the Portside Commerce Center in Richmond, California — a 15.75-acre (6.4-hectare) waterfront facility at 1411 Harbour Way South. County property records confirm OpenAI OpCo LLC signed the Richmond warehouse lease on March 9, 2026; the site previously housed portable battery startup Moxion Power before that company’s August 2024 bankruptcy. The building’s electrical capacity exceeds 14,000 amperes, compared to the 200–4,000 amps typical of most industrial warehouses. For context, a standard American household operates on a 200-amp service. The Richmond facility’s power density is comparable to a small data center — which is precisely what it functions as, except that instead of processing tokens, it powers arrays of robotic test cells, actuation rigs, and the GPU clusters needed to turn raw teleoperation footage into trained robot policies.
OpenAI planned a second robotics lab at the Richmond site, in addition to its existing San Francisco Mission Bay footprint and a recently signed lease in Mountain View.
What $500,000 Buys — and Who It Is Designed to Attract
OpenAI’s pay structure has never been modest. Federal H-1B visa filings published earlier this year showed the company’s research scientists earning up to $685,000 and hardware engineers up to $555,000 in base salary. The new robotics listings represent an extension of that structure into a discipline where the talent pool is dramatically smaller.
The highest-paying robotics listing is a machine learning engineer focused on data pipelines and training infrastructure, advertised at $380,000 to $500,000 base. A flat $380,000 inference engineer role — focused on deploying trained policies to hardware — reflects a real technical challenge: robot control requires inference at 200-hertz loop rates, far faster than cloud-based language model serving. A robotics software engineer requiring five years of experience in Rust or C++ is listed at $255,000 to $325,000. Those language requirements matter: real-time robot control systems cannot use Python’s garbage collector or interpreter overhead. Rust and C++ are chosen for deterministic latency — the kind of guarantee that lets a joint controller respond within microseconds.
For comparison, industry-wide data from KORE1’s robotics compensation survey in May 2026 showed mid-level robotics engineers across the sector earning $150,000 to $205,000, with senior engineers in the $205,000 to $300,000 range and humanoid and foundation-model specialists reaching $280,000 to $475,000 in total compensation. OpenAI’s $380,000 to $500,000 base for ML infrastructure engineers sits above even the specialist ceiling — a deliberate signal to the market, not an accident of HR.
All 27 open positions are based in San Francisco.
How OpenAI’s $500K Offer Reshapes the Competitive Field
Several hardware roles — actuator designer, firmware engineer, PCB layout engineer, prototyping technician — do not carry publicly listed salary ranges. Those omissions are notable: they likely reflect positions where OpenAI is willing to negotiate beyond its posted ceilings, or where it is protecting its recruiting strategy from competitive benchmarking.
The combination of roles that do appear — actuator designers (building the mechanical joints), firmware engineers (handling embedded systems), PCB layout engineers (designing circuit boards), machine learning engineers focused on training pipelines, and a dedicated data-collection infrastructure specialist — describes a vertically integrated effort: hardware and software developed in tandem, controlled entirely in-house. One listing that mentions a laser rangefinder and battery hints at an autonomous mobile robot, though the specific form factor is not specified. The presence of a robotics attorney listing underscores how seriously OpenAI treats the legal scaffolding around a physical-world product category.
This vertical integration posture is a departure from the company’s prior strategy. From 2023 to early 2025, OpenAI functioned primarily as an investor and model supplier in the robotics space — leading the OpenAI Startup Fund’s $23.5 million Series A2 round for 1X Technologies in 2023, and joining the $675 million Series B for Figure AI alongside Microsoft, Nvidia, and Jeff Bezos in February 2024. That Figure AI partnership ended in early 2025, with Figure going on to develop its own in-house vision-language-action model, Helix, which powers its Figure 03 platform. Figure AI secured $1 billion-plus Series C funding in September 2025, at a $39 billion post-money valuation, led by Parkway Venture Capital.
The researchers and engineers who understand the hardest layers of humanoid robotics — calibration, sensor fusion, simulation realism, real-time control, and robot data pipelines — were already contested before OpenAI entered the same labor market. Analysts at humanoid.guide noted in May 2026 that compensation pressure would intensify for startups, warning that OpenAI entering the talent market “could push compensation higher and make specialized hiring more difficult” for companies operating with tighter margins. OpenAI’s willingness to pay at the top of an already-elevated band accelerates that pressure.
What the Industry Looks Like When OpenAI Arrives
OpenAI is entering a physical AI landscape that has transformed since 2021. Google DeepMind released Gemini Robotics 2 on July 30, 2026, a family of three vision-language-action models enabling whole-body humanoid control — covering legs, torso, arms, and hands under a single trained policy — demonstrated on Apptronik’s Apollo 2 robot. The model’s published benchmark shows 68.4% tabletop success on pickup tasks and 45.7% on floor-level pickup — numbers that illustrate both the progress made and the distance remaining for commercially reliable manipulation.
Tesla began pilot production of Optimus Gen 3 at Fremont on January 21, 2026, and formally ended production of the Model S and Model X there on May 10, 2026, repurposing the floor for Optimus scale-up, with a long-term production target of one million units per year at Fremont. Figure AI’s BotQ facility in San Jose achieved one robot per hour throughput by May 2026 — up from one per day at the start of the year — at its initial capacity of 12,000 units per year.
Physical Intelligence, Agility Robotics, Apptronik, and Unitree round out a competitive field in which every company is simultaneously building the hardware, training the models, generating the data, and scaling the factory — the full stack OpenAI is now assembling from scratch.
Sam Altman’s Stated Vision — and the Governance Gap It Left Exposed
In a post announcing the OpenAI Robotics hiring drive on May 31, 2026, Sam Altman described the near-term target as robots that support skilled workers building infrastructure, with data centers cited as one of the first application domains. Altman added that he envisioned a world where everyone has their own personal robot, doing anything they need. The post received 158,000 views.
The announcement arrived roughly 12 weeks after the most significant governance rupture in OpenAI’s short robotics history. On March 7, 2026, Caitlin Kalinowski — the former Meta AR hardware lead OpenAI had hired in November 2024 to lead its robotics and consumer hardware division — announced her resignation on X. Her reason: OpenAI had agreed to deploy its AI models on the Pentagon’s classified cloud network without, in her view, adequate guardrails in place. “Surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got,” she wrote. “This was about principle, not people.”
The Pentagon deal — reported as a $200 million contract — had emerged shortly after rival Anthropic declined to permit unrestricted military use of its models. OpenAI stated the agreement included safeguards on its technology’s use.
Kalinowski’s departure did not visibly slow the hiring push, but it introduced a governance question that sits alongside the technical one: what does it mean for a company building general-purpose physical robots — systems that will operate in warehouses, homes, and eventually workplaces — to also supply AI to classified military networks, and who inside the organization is responsible for ensuring those two use cases do not converge?
Why the Data Machine Is the Real Product
The shape of OpenAI’s robotics effort makes most sense when you read it as infrastructure for solving a specific problem rather than as a product launch. The three-shift teleoperation lab, the Richmond warehouse, the GELLO controllers, the Franka robotic arms, the firmware and PCB layout engineers, the ML engineers focused on data pipelines — together these describe an apparatus whose primary output is not a robot. It is labeled training data, generated at industrial scale.
That distinction matters because it is what separates OpenAI’s 2026 approach from its 2021 failure. The algorithms for robot learning have not changed fundamentally in the intervening years. What has changed is the recognition that building competitive robot intelligence requires the same kind of industrial data infrastructure that built competitive language intelligence — and that infrastructure requires human labor (contract teleoperation workers) running shifts around the clock to produce labeled physical demonstrations. The lesson from 2021 is not that OpenAI was building the wrong thing. It is that the company was trying to build the right thing without the factory to feed it.
Whether OpenAI can close the gap with rivals who have had years of lead time in the physical world is the question the industry will now be watching closely.
Frequently Asked Questions
Why does robot training require so much human labor?
Teaching a robot to manipulate objects — picking things up, opening drawers, loading a dishwasher — requires thousands of recorded examples of humans performing those same tasks while wearing or holding a controller that logs every joint position and force reading at each moment. Unlike language models, which can be trained on text scraped from the internet, there is no internet’s worth of robot arm demonstrations to download. Every example has to be physically enacted and recorded. OpenAI’s San Francisco lab uses GELLO controllers paired with Franka robotic arms; contract workers operate these systems around the clock to generate the datasets that robot policies learn from. That is why the facility’s workforce has grown from roughly 100 contractors at launch to more than 400, running three shifts daily.
How does OpenAI’s robotics salary of $500,000 compare to what humanoid startups pay?
The gap is substantial. Industry data from KORE1 (May 2026) showed that even the most specialized humanoid and foundation-model robotics engineers across the sector earn $280,000 to $475,000 in total compensation. OpenAI is offering $380,000 to $500,000 in base salary alone — before equity — for machine learning engineers focused on training infrastructure. That differential is a deliberate recruiting tool: OpenAI is competing for the same small pool of engineers as Figure AI, Agility Robotics, Physical Intelligence, and Tesla Optimus, and it is using its balance sheet to outbid them. Analysts expect this to put upward pressure on compensation across the entire humanoid robotics sector.
What happened to OpenAI’s first robotics team, and is this effort different?
OpenAI shut its original robotics team in 2021 after co-founder Wojciech Zaremba concluded that while the technical approach was sound, the company lacked sufficient real-world demonstration data to train capable robot policies. The current effort is structured to solve exactly that problem: rather than attempting to build robot intelligence directly, OpenAI built the data-generation apparatus first — the teleoperation lab, the industrial warehouse, the firmware and hardware engineers needed to maintain physical test infrastructure. The 27 current job listings, the Richmond warehouse lease, and the three-shift data collection operation are all elements of that infrastructure. The research hypothesis is the same; the resource commitment behind it is orders of magnitude larger.
How does OpenAI’s approach compare to Google DeepMind and Tesla?
All three organizations are pursuing what is broadly called a vision-language-action (VLA) approach: training large models that take visual input and natural-language instructions and output motor commands. Google DeepMind’s Gemini Robotics 2 (released July 30, 2026) demonstrated whole-body humanoid control on Apptronik’s Apollo 2, with 68.4% accuracy on tabletop pickup tasks and 45.7% on floor-level pickup — its own data, not independently audited. Tesla is converting its Fremont factory from Model S/X production to Optimus humanoid manufacturing, targeting one million units per year at scale. OpenAI’s strategic distinction is that it is building the training data infrastructure from scratch, which gives it direct control of the data moat but puts it years behind rivals in deployed hardware experience.
