OpenAI Launches Robotics Division: Physical AI Is Now a Priority

From Language Models to Physical Machines

OpenAI shut down its first robotics project in 2021 to focus on language models. That bet paid off spectacularly — ChatGPT and GPT-5 redefined the AI industry. Now the company is making its second and far more serious move into physical AI.

"OpenAI Robotics is hiring, looking for exceptional full-stack hardware, ops, systems, and ML engineers to help us program and manufacture robots that are useful for society," Altman wrote on X on May 31. "AI should be able to help people in the physical world."

The timing is not accidental. OpenAI's world simulation research — which began generating internal momentum in 2025 — has now formally evolved into a standalone division with a mandate to ship.

The Figure AI Foundation

The acquisition of Figure AI — completed in May 2026 for $2.9 billion — gives OpenAI something competitors building purely software AI don't have: proven humanoid hardware in active commercial deployment.

Figure's bipedal robots are currently running in BMW's South Carolina manufacturing plant and Amazon warehouse operations. This isn't lab prototype status. These machines are handling real production workloads alongside human workers.

OpenAI's plan is to integrate GPT-5's multimodal reasoning capabilities directly with Figure's hardware stack, and to build a Robotics API that allows developers to program physical tasks using natural language. Beta access is expected in Q3 2026 — roughly the same timeline that the ChatGPT API opened the generative AI ecosystem to third-party developers in 2022.

Why Robots? Why Now?

Three converging factors explain the timing.

1. GPT-5 capability plateau argument Robotics expert Stefanie Tellex at Brown University argues that frontier language model development may be approaching a ceiling in purely digital environments. The next jump toward AGI likely requires models that can "see, understand, and interact" with the physical world — not just text and images. OpenAI's own job postings reinforce this, stating that the robotics team's "core goal is to unlock general-purpose robotics technology and advance AI's evolution toward AGI levels in dynamic, real-world physical scenarios."

2. Simulation infrastructure is ready NVIDIA's Isaac simulation platform — named in multiple OpenAI robotics job descriptions — enables training AI on millions of physical task iterations in virtual environments before touching real hardware. Combined with OpenAI's investment in xAI's Colossus 2 computing infrastructure (1.5 GW), the compute cost of robot training has dropped significantly.

3. Competitive pressure The humanoid robotics space is no longer emerging — it's accelerating. Tesla's Optimus is in factory deployment. Google DeepMind has active robotics partnerships. Boston Dynamics is commercially deployed. OpenAI watched competitors build physical AI infrastructure while it focused on language. The Figure acquisition and the Robotics division launch signal that it is done waiting.

The Competitive Landscape

Skilled Workers, Not Replacement

Altman was careful with framing: the near-term focus is robots that assist skilled workers on construction and infrastructure projects — not replace them. This is a deliberate positioning choice in a labor market where AI displacement anxiety is high.

AI Weekly's analysis cuts through the framing: "If humanoid robots first target skilled trades in infrastructure build-out, the automation pressure lands on a labor segment that has so far been considered relatively insulated from AI displacement." Whether that pressure becomes structural displacement or productivity augmentation depends on how these systems are deployed — and governed.

The OpenAI Frontier Governance Framework, also published on May 28, addresses cyber and CBRN risks from frontier AI but says nothing yet about physical AI labor displacement. That gap will likely attract regulatory attention as robotics deployments scale.

What to Watch

Q3 2026: Robotics API beta access. This is the moment that reveals whether OpenAI can replicate the developer ecosystem strategy in physical AI.

GPT-5 + Figure hardware integration: First public demonstrations of multimodal reasoning controlling physical robots will set the benchmark for what "language-driven robotics" actually looks like in production.

Regulatory response: The EU AI Act's Code of Practice and California's Transparency in Frontier AI Act currently cover digital AI. Physical AI with labor displacement implications is likely next.