OpenAI Buys Ona to Give Codex Agents a Home in the Cloud

The agent needs a place to live

The most revealing detail in OpenAI’s latest acquisition is a closed laptop. On June 11, the company announced it will acquire Ona, the German startup formerly known as Gitpod, so that Codex users can “delegate work that may take hours or days to the coding agent without being tied to a single device or active session.” Translation: the coding agent has outgrown the computer it runs on, and OpenAI just bought it a house.

The deal’s terms went undisclosed, per CNBC’s reporting, but the strategic logic is legible from across the street. Ona spent eight years moving software development off local machines and into secure, reproducible cloud environments, accumulating 2 million developers along the way — including, by its own telling, America’s largest banks, Europe’s leading financial and pharmaceutical companies, and Asia’s top sovereign wealth funds. Once the transaction clears standard closing conditions, Ona’s employees will embed directly within the Codex team, Yahoo Finance reports. OpenAI is not buying a product line. It is buying the plumbing for a future where the agent, not the IDE, is the unit of software production.

The stakes ride on a single growth curve that has gone vertical. Codex now serves more than 5 million weekly active users, up from 3 million in April — a 400% increase from earlier this year, per OpenAI’s own figures. The same Yahoo Finance account adds a detail that should reframe how anyone thinks about this market: knowledge workers now comprise roughly one-fifth of Codex users and are growing at triple the rate of core developer adoption since the desktop app launched in February. The fastest-growing Codex user has never configured a dev environment in their life. They cannot install Docker. They will not manage SSH keys. If the agent is going to do real work for them, the environment has to materialize on demand, somewhere else, with the credentials and guardrails already in place.

That is precisely the machine Ona built. Its platform bundles sandboxed, API-first cloud environments pre-loaded with dependencies, authenticated against source control, databases, and secret managers — deployable in Ona’s cloud or inside a customer’s own infrastructure, with enterprise controls like role-based access, SSO, audit logging, and VPC isolation. Strip away the branding and Ona is a factory for disposable, pre-credentialed workspaces. For an agent that needs to clone a repository, run a test suite, and iterate for six hours while its human sleeps, that factory is not a convenience. It is the difference between a demo and a deliverable.

There is also a war on, and this is a wartime acquisition. The AI coding market has collapsed into a brutal two-front contest between OpenAI and Anthropic, whose Claude Code commands a 46% “favorite tool” rating among developers in JetBrains’ April survey against 19% for Cursor and 9% for GitHub Copilot. In May, the two labs began openly raiding each other’s users — OpenAI dangling two free months of Codex to enterprises that switch, Anthropic countering with a 50% usage-limit increase for Claude Code subscribers through July 13. Buying Ona hands OpenAI something promotional pricing cannot: an infrastructure moat under the agent itself.

Two million developers, eight years, one undisclosed check

Ona’s history reads like a parable about being early. The company spent its first act as Gitpod, known for browser-based cloud development environments — a category it helped invent and then waited years for the market to want. The conviction was technically elegant, beloved by platform-engineering teams, and stubbornly resistant to mass adoption for one human reason: developers, on the whole, liked their laptops just fine. A cloud workspace solved problems most engineers had learned to stop noticing, and no amount of onboarding polish could make a person abandon a local machine that already worked.

Then the customer changed species. In September 2025, Gitpod rebranded as Ona and repositioned its entire stack around AI agents rather than human developers. Co-founder and CEO Johannes Landgraf compressed the thesis into seven words: “IDEs defined the last era. Agents define the next.” The pivot was not cosmetic. Ona reorganized into three products — Agents, Environments, and Guardrails — and the company’s own rebrand announcement disclosed numbers that explain why OpenAI came knocking: Enterprise ARR quadrupled year-over-year, a near eight-figure contract landed in a single quarter, and Ona’s internal agents co-authored 60% of the company’s merged pull requests, contributing 72% of merged code lines.

Sit with that last statistic, because it is the quiet headline of this acquisition. A company whose own agents write nearly three-quarters of its merged code is not selling a hypothesis — it is selling an operating manual. Humans at Ona had already migrated up the stack to specification and review, exactly the workflow OpenAI is betting 5 million Codex users will adopt. The acquisition imports not just infrastructure but institutional knowledge of what breaks when agents do most of the typing: environment drift, credential sprawl, the security posture of a machine that no human is watching.

The enterprise arithmetic explains the urgency. OpenAI expects enterprise customers to grow from roughly 40% to 50% of its business by year-end, per PYMNTS’s account of the deal — and enterprises do not buy agents the way individuals buy subscriptions. They buy compliance artifacts: audit trails, scoped credentials, environments that die on schedule and leave a log behind. Ona’s Guardrails layer — role-based access control, single sign-on, audit logging, VPC deployment — is the piece of this acquisition that sales engineers will quote in security questionnaires for years. The agent demo wins the meeting; the guardrail documentation wins the contract. OpenAI could have built all of it eventually, but “eventually” is a luxury that a 50% enterprise mix target does not permit.

The deal also extends a now-unmistakable acquisition pattern. Yahoo Finance notes that Ona follows OpenAI’s purchases of Promptfoo, Torch, Software Applications, and the $6 billion-plus io deal — a string of buys that each fill a specific gap between model capability and deployed product. OpenAI has already shown what it does with this kind of vertical integration: just last week it shipped Codex Sites, which turns prompts into hosted web apps without the user touching a server. Ona is the same move aimed at a harder target — not hosting the output of agent work, but hosting the work itself.

Stitch the public numbers together and the arithmetic turns startling. Codex added roughly 2 million weekly users in the ten weeks between April and June 11 — which means OpenAI’s coding agent acquired the equivalent of Ona’s entire eight-year, 2-million-developer install base in about 70 days. That asymmetry is the deal’s actual price rationale: distribution was never Ona’s problem to solve, and infrastructure was never going to be OpenAI’s organic strength on an enterprise timeline. Each company held the asset the other would have needed years to build. The undisclosed check simply collapsed those years into a closing date.

The product trajectory had been pointing here for months. In April, OpenAI shipped a beefed-up Codex with deeper control over the user’s desktop, a release TechCrunch read straightforwardly as a strike at Anthropic’s Claude Code. But the desktop strategy contains its own ceiling: an agent bound to a local machine inherits the machine’s limits — it stops when the lid closes, it sees only what the laptop can mount, and it terrifies every security team that audits it. The Ona acquisition is the admission that the desktop was a beachhead, not the destination. The agent reaches through the laptop to demonstrate value, then moves to the cloud to deliver it at enterprise scale, under guardrails a CISO can actually sign off on.

The timing against OpenAI’s balance sheet is no accident either. The company filed a confidential S-1 on June 8, and every pre-IPO quarter now doubles as a narrative-construction exercise. An agent platform that runs hours-long autonomous tasks for enterprises — provisioned through infrastructure the company owns outright — is a materially better growth story than a chatbot with a coding mode. Meanwhile Anthropic closed a $65 billion Series H at a $965 billion post-money valuation in late May, overtaking OpenAI as the most valuable AI startup on the strength of $47 billion in run-rate revenue. When your chief rival overtakes you in valuation while citing enterprise agent adoption, buying the enterprise agent infrastructure company is as much signaling as strategy.

The ways this acquisition could misfire

Start with the most uncomfortable precedent: enterprises bought Ona specifically because it was not a model vendor. Ona Guardrails let customers run agents inside their own VPCs with models they select through AWS Bedrock or Google Vertex — Claude included. That neutrality was the product for the regulated banks and pharmaceutical companies on Ona’s customer list. The moment Ona’s roadmap reports to the Codex org, every one of those customers must ask whether their model-agnostic infrastructure layer just became a distribution channel for one lab’s models. Some will conclude yes and migrate. OpenAI’s history offers little comfort here; its acquisitions tend to dissolve into the mothership rather than persist as neutral platforms, and the announcement’s language about embedding Ona’s team “within the Codex team” reads like exactly that.

The competitive response will be fast, because the playbook is public. Anthropic already operates Claude Code’s cloud sandbox infrastructure and has shown — as covered when Fable 5 shipped behind safety classifiers — that it will spend heavily on controlled execution environments. Microsoft, which has been methodically reducing its dependence on OpenAI with its own MAI model family, owns GitHub Codespaces: a cloud development environment with vastly more enterprise distribution than Ona ever had. If the thesis is that agents need cloud homes, Microsoft already owns the largest subdivision in the neighborhood. OpenAI bought a head start measured in quarters, not years.

There is also a cost problem that nobody in the announcement mentions, and developers are currently rioting about it one platform over. When GitHub moved Copilot to token-based billing on June 1, the meter exposed what flat-rate pricing had concealed: agentic coding is priced like compute, not software. Developers projected bills jumping from $29 a month to $750, and “what a joke” became the quote of record. Now extend the task length from minutes to the “hours or days” OpenAI is explicitly designing for, and add metered sandbox time on top of inference. Long-running agents are a margin machine for the vendor only if customers tolerate the invoice. The Copilot backlash suggests tolerance is thinner than every lab’s revenue model assumes.

The deepest risk is that the premise itself is ahead of the evidence. Ona’s flagship statistic — agents co-authoring 60% of merged PRs — comes from Ona’s own engineering team: a few dozen experts who built the platform, operating it with maximal skill on a codebase they know intimately. That is the best-case deployment in the world. The median enterprise runs on legacy Java behind three layers of compliance review, where an unattended agent’s six-hour run can produce six hours of confidently wrong code. If autonomous multi-hour delegation stalls at the pilot stage in regulated industries — as so many AI initiatives have — OpenAI will have bought world-class infrastructure for a workload that arrives years late. Integration friction compounds the risk: acquired infrastructure teams bleed talent during exactly the migration quarters when their expertise matters most.

Geography adds a quieter complication. Ona is a German company whose enterprise pitch leaned hard on European data sovereignty — VPC deployment, customer-controlled infrastructure, and distance from the American hyperscaler stack. Those assurances read differently the day a US lab signs the purchase agreement. European financial and pharmaceutical customers operating under GDPR and the EU AI Act’s tightening obligations will now re-run vendor risk assessments they thought were settled, and some procurement departments will treat the acquisition itself as a triggering event for contract review. None of this is fatal — hyperscalers navigate it daily — but it converts Ona’s cleanest sales motion into its most complicated one, precisely in the territory where the customer list was strongest.

And lurking behind all of it is the antitrust shadow. OpenAI is pre-IPO, serially acquisitive, and consolidating the agent stack vertically — models, IDE surface, hosting, and now execution infrastructure — while lawmakers urge the FTC and DOJ to scrutinize exactly this style of AI consolidation deal. No single acquisition of an undisclosed-but-modest size will trigger intervention. But the cumulative pattern of a near-trillion-dollar company absorbing each layer of the developer toolchain is precisely the fact pattern that turns into a consent decree three years later, when unwinding it would hurt most.

What operators should do while the ink dries

The direction of travel is now unambiguous, even if the timeline is not. Every major lab is converging on the same architecture: a frontier model attached to a fleet of disposable, credentialed cloud workspaces where agents do multi-hour work under enterprise guardrails. OpenAI just paid to own that layer instead of renting it. The week’s surrounding news confirms the build-out — OpenAI simultaneously opened its models and Codex to Oracle’s enterprise customers through existing cloud credits, threading agent distribution through procurement channels enterprises already trust. The agent is becoming a line item in the cloud bill, which is exactly where the money normalizes.

Expect the next two quarters to define the categories. Ona’s integration will tell the market whether “agent infrastructure” remains a neutral middleware layer — the way Kubernetes stayed vendor-neutral — or collapses into vertically integrated lab stacks the way mobile did. The honest answer is probably both: labs will own the consumer and mid-market path end to end, while a handful of neutral players (and the hyperscalers’ own offerings) serve enterprises that refuse single-vendor dependency. The losers in that scenario are standalone coding-tool companies caught between lab-subsidized pricing below them and hyperscaler distribution above them. The JetBrains survey’s middle tier — the Cursors and Windsurfs of the world — now face competitors who own everything from the silicon contract to the sandbox.

Watch the meter, too, because long-running agents rewrite the unit economics of the entire category. A chat completion bills for seconds of compute; an agent that works through the night bills for inference plus sandbox-hours plus storage plus the orchestration that keeps it honest. That cost curve looks like cloud infrastructure, not software — and it explains why every lab now wants to own the execution layer outright rather than rent it back from a hyperscaler at margin-destroying markup. Vertical integration here is not empire building; it is cost-of-goods management for a product whose marginal cost grows with its ambition. The Oracle channel matters for the same reason: routing agent consumption through existing cloud commitments cushions the procurement shock when invoices start scaling with how much work the agent is trusted to do.

For teams making decisions this quarter, the checklist writes itself:

• Audit your dev-environment dependency now. If your organization runs on Ona, get written roadmap and data-residency commitments before renewal. If the answer is vague, price out GitHub Codespaces, Coder, or a hyperscaler-native alternative while you still have leverage.
• Pilot long-running agent tasks on bounded, low-blast-radius work. Dependency upgrades, test-coverage expansion, and migration scaffolding are the proving grounds where a six-hour autonomous run can fail cheaply and audibly.
• Model the unit economics before the invoice does it for you. The Copilot billing backlash is a preview: estimate cost per merged PR under token-metered, sandbox-metered pricing, and set per-team budget caps before adoption scales past finance’s line of sight.
• Demand the Ona-style guardrail stack from every agent vendor. VPC deployment, RBAC, audit logs, and scoped credentials are no longer enterprise nice-to-haves; they are the minimum viable trust for unattended execution.
• Track the knowledge-worker cohort, not just developers. One-fifth of Codex users already aren’t engineers, and that segment is growing three times faster. The governance problem of non-developers shipping agent-built software will land on platform teams within the year — write the policy before the first incident, not after.

The closed-laptop framing will be remembered as the tell. Every prior era of developer tooling assumed a human present at the keyboard; the entire stack — IDEs, terminals, local runtimes — inherited that assumption silently. OpenAI just spent real money on the proposition that the assumption is dead. Landgraf, for his part, called the sale the moment his “life’s work just got bigger.” The less sentimental version: the cloud development environment finally found its killer user, and it was never a person.

In other news

OpenAI threads Codex through Oracle’s enterprise billing — OpenAI announced that Oracle Cloud Infrastructure customers can apply their existing Oracle Universal Credits toward OpenAI frontier models and Codex, with availability beginning in the coming weeks. The deal lets enterprises adopt OpenAI under procurement contracts they already have, removing a separate procurement cycle — often the slowest step in enterprise AI adoption.

ChatGPT becomes the fastest app to 1 billion monthly users — Sensor Tower estimates put ChatGPT’s mobile app past 1 billion monthly active users, reaching the milestone in roughly three and a half years — faster than TikTok, Instagram, or YouTube. Weekly active users hit 900 million, more than double the figure from February 2025.

GitHub Copilot’s token billing sparks open revolt — GitHub’s June 1 switch to usage-based AI Credits replaced flat-rate premium requests with per-token metering, and developers projecting 10x–50x cost increases for agentic sessions flooded forums in protest. The removal of the free fallback model drew particular fury from individual developers on the $10 Pro plan.

DeepSeek makes its 75% price cut permanent — DeepSeek announced that the steep discount on its flagship V4 Pro model, originally set to expire May 31, is now permanent, with output pricing at $0.87 per million tokens — undercutting every Western frontier model. The company had previously tied pricing relief to volume shipments of Huawei’s Ascend 950 chips in the second half of 2026.

Anthropic banks $65 billion as the IPO queue forms — Anthropic’s Series H closed at a $965 billion post-money valuation with $47 billion in run-rate revenue, and TechCrunch reports it could be the company’s last private fundraise before a highly anticipated IPO. Memory-chip makers Micron, Samsung, and SK hynix joined the round as strategic investors, binding the supply chain to the lab.