OpenAI models arrive on Google Cloud. The cloud no longer picks your model, but it still sets the boundary

Picture a bank that standardized on Google Cloud three years ago. Its data sits in BigQuery, its identity in Google’s IAM, its perimeter in VPC Service Controls, and its budget in a multi-year spend commitment. Its developers want GPT models for some tasks. Until this week, saying yes meant a second vendor contract, a second security review, a second billing relationship and a data path that left the perimeter the risk team had spent years approving.

On October 21, Google Cloud and OpenAI announced that OpenAI’s frontier models are coming to Google Cloud’s Model Garden: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, bought through Google Cloud Marketplace, billed on the Google Cloud invoice and drawn down from existing Google Cloud commitments. For that bank, the second contract may disappear. The question of where the boundary sits does not.

What is actually available

Less than the headline suggests, for now. Google describes the models as “rolling out now in private preview to select enterprise customers.” OpenAI says general availability is planned “in the coming weeks.” Several of the controls a regulated buyer would ask about first are announced, not shipped.

OpenAI on Google Cloud, by status on October 21

	Status	What it means for a buyer
GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna in Model Garden	Private preview, selected customers	Evaluate now, plan production for after GA
Marketplace billing against existing commitments	Announced with the preview	The procurement change is the real news
Zero Data Retention option	Coming soon	Not yet a basis for approving sensitive workloads
Confidential Compute	Coming soon	Same: ask for a date in writing
OpenAI models in Gemini for business and Antigravity	Coming soon	Plan for it, do not design around it

Google also says customers can point local Codex CLI and desktop workflows at its platform for inference, and that prompts and outputs stay inside the customer’s data perimeter and are not used to train foundation models. Those are vendor statements; the contract and the data processing terms are where they become commitments.

One line in OpenAI’s post deserves attention: “Microsoft remains OpenAI’s primary cloud partner.” OpenAI models have been purchasable through Oracle cloud commitments since June, and AWS offers Bedrock Managed Agents, built with OpenAI, in preview. Google is now one more cloud reselling them, and it claims to be the only hyperscaler offering Gemini, Claude and GPT side by side.

The model is leaving the cloud that sells it

For years, an enterprise’s primary cloud quietly chose its model family. Azure meant OpenAI, Google meant Gemini, AWS meant whatever Bedrock carried. That coupling is dissolving. The same frontier model can now be bought through several clouds, each wrapping it in its own identity, network controls, logging and billing.

That is good for buyers, with one caution: the same model through a different cloud is not the same service. We saw this with Claude on Azure and AWS: identical weights, different data boundaries, different feature timing, different contractual paths. OpenAI on Google Cloud will follow the same pattern. Zero Data Retention may be available directly from OpenAI before it is available through Model Garden. A model version may reach one cloud weeks after another. Rate limits, regions and support escalation will differ.

So the decision moves. It is no longer “which model?” alone. It is “which model, through which path, under which controls?”

Procurement is doing the architecture

Google’s announcement leads with commercial plumbing for a reason: draw down existing commitments, count usage toward Flexible Savings Plans, one invoice. For many companies, that is what will decide adoption. A team that would have waited months for a new vendor agreement can start an evaluation inside a contract that already exists.

We have argued that marketplace commitments are starting to shape AI architecture more than benchmarks do. This announcement extends the logic to a model provider that, until now, mostly sold through its own API and one partner cloud. Expect finance teams to ask a fair question: if the commitment can buy any model, why are we paying list price somewhere else?

The risk is the mirror image. Commitment-funded access makes it easy to adopt a model without the evaluation that would have happened during a separate purchase. Cheap to start is not the same as approved to run.

Does this reduce lock-in?

Partly. It lowers model lock-in: if your workloads run through Model Garden, switching between Gemini, Claude and GPT becomes closer to a configuration change. It does not lower cloud lock-in; it may increase it. Every new model available inside the commitment is another reason to keep spend, identity and logging in one place.

That is not necessarily bad. It is a choice, and it should be made on purpose. The layers worth owning independently of any one cloud are the ones we listed in our model portfolio framework: your evaluation sets, your routing policy and your record of which model handled which task. If those live outside the cloud’s console, a second or third delivery path is an option you can exercise. If they live inside it, the commitment chose for you.

What to do while it is in preview

• Request preview access only for a named evaluation, with the same test set you use for the models you run today, so the comparison is about your tasks.
• Ask for the data processing terms for the Model Garden path specifically. Do not assume the terms of your direct OpenAI agreement carry over.
• Hold sensitive workloads until the “coming soon” controls ship, if your approval depends on Zero Data Retention or confidential computing.
• Record the delivery path in your model inventory, alongside the model name and version. The same model through two clouds is two entries.