OpenAI Didn't Buy an IDE. It Bought the Runtime

Call this the "AI IDE wars" if you want, but that framing misses what actually happened. On June 11, OpenAI announced it's acquiring Ona — the German company you probably still know as Gitpod — and folding it into the Codex team once the deal clears regulators. Terms weren't disclosed. What was disclosed is the tell: OpenAI says Codex now has over 5 million weekly users, up 400% this year, and the tasks people hand it keep getting longer. OpenAI didn't buy an editor. It bought the place where agents run.

That distinction matters, because OpenAI already tried the IDE play and lost. Its roughly $3 billion deal for Windsurf collapsed in mid-2025 when Google hired away the leadership and Cognition scooped up the remains. A year later, OpenAI's answer isn't another editor — it's infrastructure. That's the right lesson to have learned.

Gitpod spent seven years selling this to the wrong customer

Gitpod pioneered cloud development environments: ephemeral, prebuilt, defined-in-code workspaces you opened in a browser. Technically impressive, and commercially a grind. Developers mostly refused to give up their laptops. Latency was annoying, local tooling was sticky, and "your dev environment, but in our cloud" never cleared the bar for most teams. The whole CDE category struggled — AWS quietly deprecated Cloud9, and GitHub Codespaces settled into a useful-but-niche existence. Gitpod itself ripped out Kubernetes in late 2024 and rebuilt its runtime from scratch, then rebranded to Ona in September 2025 and repositioned entirely around AI agents.

That pivot looked desperate at the time. It turned out to be the trade of the decade, because the customer who genuinely needs a cloud dev environment was never a human. It's an agent. A coding agent that works for six hours can't live in a terminal tab on your MacBook — close the lid and the run dies. It needs a persistent sandbox with the repo, the toolchain, network policy, and scoped credentials already in place. That's a CDE. Gitpod built the right product a half-decade early and finally found the user it was designed for.

The runtime is the new moat

Every serious player has converged on the same shape. GitHub's Copilot coding agent executes inside Actions runners. Cognition built Devin cloud-first from day one. Cursor added background agents that run off your machine. Anthropic's Claude Code pushes sandboxed, long-running sessions. The models are increasingly interchangeable week to week; where the agent executes — and who controls that sandbox — is where lock-in actually forms.

Ona's specific edge is the enterprise version of that story. Its environments deploy inside the customer's own VPC, with audit trails, role-based access, and kernel-level policy enforcement. Its customer list — banks, pharma companies, sovereign wealth funds, per Forbes' reporting — is exactly the segment that won't ship source code to a vendor's cloud, and exactly the segment OpenAI needs for the enterprise revenue story. Ona CEO Johannes Landgraf's line — "agents need more than intelligence; they need a trusted workspace" — is marketing, but it's accurate marketing. Slot this next to OpenAI's Promptfoo acquisition in March and the pattern is unambiguous: OpenAI is buying the enterprise deployment layer, not model capability.

What this means if you ship code

If you're an Ona/Gitpod customer, model neutrality is now on a clock. Ona's platform ran Claude and Gemini through Bedrock and Vertex alongside OpenAI models. Nobody has announced that changes, and until the deal closes the companies stay independent — but you don't acquire a runtime to keep routing tokens to Anthropic forever. If multi-model flexibility is why you picked Ona, start evaluating Coder, which just became the largest independent self-hosted CDE vendor, or Codespaces plus devcontainers if you're already deep in GitHub.

If you're building agent workflows, environment-as-code is the contract that matters now. The devcontainer.json / Gitpod-style spec — declared toolchain, prebuilt images, scoped secrets — stops being a developer-experience nicety and becomes the interface between your infrastructure and someone else's agent fleet. Repos with reproducible, declarative environments will get materially more value out of long-running agents than repos where setup lives in a wiki page. That's worth engineering time this quarter regardless of which vendor wins.

The unsolved problem is review. A two-day agent run producing a sprawling pull request has no mature tooling behind it: no good way to review incrementally, no clean rollback story, no established norms for what "approval" even means for work no human watched happen. Persistent VPC environments also mean persistent cost and persistent attack surface — a long-lived sandbox with repo access and network egress is a juicier target than a 90-second CI job. Buying Ona gets OpenAI the execution substrate; it doesn't get anyone the governance layer, and that gap is where the next generation of dev-tools startups should be aiming.

The verdict

This is a genuine shift, not hype — but it's a shift in the stack, not the editor. The IDE fight was always a fight over a surface, and surfaces get commoditized; VS Code forks are cheap. Execution environments woven into a bank's VPC, with compliance sign-off and audit history, are not. OpenAI just paid to skip several years of enterprise infrastructure work and enterprise trust-building in one move.

The losers are model-neutral infrastructure players, who just watched the strongest independent in their category get absorbed by a lab, and enterprises who wanted the agent-runtime layer to stay Switzerland. The winner, besides OpenAI, is oddly the CDE idea itself: after years of "why would I move my dev environment to the cloud?", the answer arrived. You won't. Your agents will.