OpenAI's robotics hiring push is built around data infrastructure, not headcount.
The AMW Read
Extends yesterday's in-house robotics-hardware-team story with concrete data-center-scale facility specs and teleoperation methodology, meaningfully updating OpenAI's robotics vertical-integration baseline without resolving an open debate.
OpenAI's robotics hiring push is built around data infrastructure, not headcount.
OpenAI has 27 robotics job listings open this week, more than double the 11 posted four months ago, with a top salary of $500,000 for a machine learning engineer role focused on training-data pipelines. A $380,000 inference-engineer role targets real-time policy deployment, and $255,000–$325,000 Rust/C++ roles reflect robot-control latency needs. The roles map onto two sites: a San Francisco teleoperation lab running GELLO controllers on Franka arms since February 2025, now staffed by 400-plus contract workers versus roughly 100 at launch, and a newly leased 202,000-square-foot Richmond, California warehouse with over 14,000 amps of electrical capacity — power density comparable to a small data center — built to run the GPU clusters that turn teleoperation footage into trained robot policies.
OpenAI shut its first robotics program in 2021 because, per co-founder Wojciech Zaremba, it lacked enough real-world data to train capable models — a data problem, not an algorithmic one. This buildout answers that gap: teleoperation is the manual demonstration-collection method robotics has always needed at scale. Per the AI Market Watch index, OpenAI's name-matched news volume hit 339 items in our pipeline over the last 90 days, up from 287 prior (coverage of a name, not a census) — consistent with yesterday's report on its new in-house actuator and motor-design team, running in parallel with this data-infrastructure buildout.
For builders, OpenAI is now trying to own both the data-collection rig and the inference stack that consumes it, raising the bar for the humanoid startups it once funded and is now hiring away with $500K offers. For investors, the Richmond facility's data-center-grade power spec is the harder signal to underwrite than the headcount numbers: it's capital committed to a multi-year data pipeline, not a research bet that unwinds as easily as the 2021 team did.
