OpenAI Buys Ona (ex-Gitpod): Whoever Owns the Agent Runtime Owns the Stack

OpenAI said on June 11 it will acquire Ona, the cloud dev-environment startup formerly known as Gitpod. As Codex weekly users jumped from 3M in April to 5M in June, the runtime that executes agents became the bottleneck. Agent infrastructure is being pulled under the model layer. For founders built on thin wrappers, that’s a warning shot.

What Happened

OpenAI announced on June 11 that it will acquire Ona. Ona is the German company that launched in 2019 as Gitpod, and its core business is spinning up isolated development environments in the cloud. Terms were not disclosed. The reasoning sits inside Codex’s growth curve. Codex weekly active users went from 3 million in April to 5 million in June — a sixfold jump since the February desktop launch. But as agents turned more autonomous — spawning sub-agents, running multi-step refactors that take hours — the laptop-bound execution model broke. Close the laptop and the job dies. Ona fixes that. An agent runs for hours inside a cloud sandbox, spins up sub-agents, runs tests, and reports back when the developer reopens the app. On top of that sit enterprise guardrails built for regulated industries: role-based access control, audit logging, and deployment inside a customer’s own VPC. Ona’s enterprise usage grew 13x this year, with customers including a major U.S. bank, European pharmaceutical companies, and Asian sovereign wealth funds. Co-founder and CEO Johannes Landgraf and the entire team join the Codex group on close. For OpenAI this is the second acquisition in three months; in March it bought the evaluation tool Promptfoo.

What This Means for Founders

Read the vector, not the headline. For two years founders competed on “which model do I call.” This deal signals that the center of gravity is shifting from the model to the runtime that executes it. Models keep getting cheaper and more commoditized. What stays expensive and hard is the isolated compute that runs an autonomous agent safely for hours — plus the permissions, audit trail, and network boundary layered on top. That’s why OpenAI is starting to behave less like a model company and more like an infrastructure company, pulling the layer directly beneath the model — agent execution and orchestration — in-house. For Silicon Valley founders this cuts both ways. If your product is a thin wrapper selling “agentic automation” over the OpenAI API, the moment the platform absorbs that layer your differentiation evaporates. The flip side is sharp opportunity. The guardrails Ona was built around — provable auditability of what an agent executed and learned, isolation that runs inside the customer’s own infrastructure — are the entry ticket in regulated verticals. A horizontal platform shipping generic controls does not automatically satisfy a bank’s compliance team or a hospital’s data-governance review. That gap is where vertical depth wins.

What You Can Do Now

First, draw an honest line: is your product beneath the platform or beside it? If you sit on a thin feature OpenAI could absorb next quarter, this is your pivot window. Second, stake your moat on execution and data, not the model. Abstract the model so you can swap whoever’s API you call, but bury your workflow, customer data, and permission model deep. Third, build auditability in from day one. If you can’t prove via logs what an agent executed and how far its authority reached, regulated-industry buyers won’t sign. Half of why Ona sold at a premium lives here. Fourth, go vertical where the platform stays generic. Healthcare HIPAA boundaries, financial-services compliance, on-prem and VPC isolation — a global platform takes time to satisfy each vertical’s specifics. That lag is the window for founders who own one industry’s rules end to end.