The one-screen version: on October 21, OpenAI and Google Cloud said OpenAI's frontier models — and its Codex agentic coding workflows — are coming to Google Cloud, buyable through the Google Cloud Marketplace with your existing committed Google spend applied to the purchase. It starts in private preview for selected enterprises, broader access expected within weeks. The founder's read: OpenAI's models now sit in the same 200-plus-model catalog as Gemini and Claude, which means the three frontier families you'd actually choose between are buyable inside one cloud account — and "which model" just stopped being an integration project and became a procurement line.

What actually happened#

Until now, the only OpenAI presence in Google's Model Garden was the open-weight gpt-oss models, added as managed APIs in 2025. Those you could already run anywhere. What changed on the 21st is that OpenAI's proprietary frontier line — the models you otherwise buy only from OpenAI directly — and Codex, its agentic coding workflow, are being distributed through a rival's cloud.

The mechanics are the whole story:

Private preview first, selected enterprise customers, general availability "in the coming weeks" — so this is a direction, not a switch you flip today. But the direction is the point.

The one real idea: model choice is collapsing into cloud procurement#

For two years the honest friction in using more than one model vendor wasn't the model — it was everything around it. A second vendor meant a second account, a second set of keys, a second billing relationship, a separate data path, and a separate compliance review before your security team would sign off. That overhead is exactly why teams minimized the number of model vendors they integrated and then felt stuck with the one they chose.

A marketplace listing deletes that overhead. When OpenAI's model is a line item in the cloud you already buy from, adding it costs about as much as enabling any other managed API. And here's the thing founders keep getting backwards: the cheaper it is to add a model, the cheaper it is to swap one. The integration tax was the lock-in. Remove it and the model layer becomes genuinely fungible — which is why the labs are fine commoditizing it. The margin was never in the model shape; it's in the compute and the cloud underneath.

So the lock-in doesn't vanish. It relocates — the same move the rival labs made last week when they donated their agent plumbing to a neutral foundation. The thing that's now hard to move isn't your model choice. It's your committed cloud spend — the discount contract you signed to get decent pricing, which only pays off if you concentrate your consumption in that one cloud. You can change models in an afternoon. You can't change a three-year spend commitment in an afternoon.

The labs just showed you their own answer#

If you want to know whether multi-cloud is paranoia or discipline, read the other filing from the same week. Anthropic's draft IPO paperwork shows it has committed roughly $450 billion to cloud compute over the next decade — and it split that deliberately across three clouds: about $111B with Google, $110B with AWS, $31B with Microsoft, against $16.3B of revenue. The company whose entire business is one model family does not single-source the compute that family runs on.

That is the tell. The frontier lab with the most leverage any buyer will ever have still runs multi-cloud. For a solo founder with a thousandth of that leverage, single-sourcing your compute to chase the deepest committed-spend discount is a bet the lab itself declines to make.

What it changes for you this week#

Nothing urgent — this is a standards-of-procurement shift, not a migration emergency. But three moves follow, and they're the same discipline from both directions now:

The clouds just made it trivial to try every model. They're betting you'll stop shopping for the model and start locking in the cloud. Don't give them the second half.