Google Cloud Will Sell OpenAI's Models and Plans Up to 960,000 Rubin GPUs, While NVIDIA Says 'Capacity Is the Constraint' and Pays Suppliers to Add It

Three announcements on October 21, 2026 describe the same market from different layers: who sells frontier models, who hosts them, and who can build the hardware.

OpenAI models in Google Cloud

OpenAI and Google Cloud are "partnering to bring OpenAI's frontier models and Codex agentic workflows into Google Cloud's Model Garden". Customers will buy through Google Cloud Marketplace, and enterprises "will be able to draw down their GCP commitments for these purchases". OpenAI models "will soon be available in Google Antigravity and the Gemini Agent." General availability "is planned in the coming weeks", and the post ends: "Microsoft remains OpenAI's primary cloud partner." Thomas Kurian framed it as choice: "the right model for any given job".

Vera Rubin arrives at Google Cloud

NVIDIA's post is headlined general availability, but its text says "General availability of the instances is coming next month", with Europe to follow. A5X combines Vera Rubin NVL72, ConnectX-9 SuperNICs, Google's Virgo network and fourth-generation liquid cooling. Google Cloud "plans to scale its Vera Rubin-based infrastructure to as many as 960,000 Rubin GPUs across multiple sites." The first customer is Ineffable Intelligence, David Silver's London company, which is building "superlearners that continually improve through their own experience". WPP says that in its initial tests, workloads "that typically required a multi-node cluster ran on a single A5X instance".

Confidential computing for the model builders themselves

NVIDIA introduced Confidential Bare Metal, which lets "frontier model builders run sensitive workloads on dedicated third-party infrastructure while retaining control of their models and data". Support extends to Grace Blackwell NVL72 and to HGX and DGX Rubin NVL8. Google Cloud will offer it on A4X Max (GB300 NVL72) and made confidential G4 VMs with RTX PRO 6000 Blackwell generally available. Cohere plans to use it in Model Vault, and ten providers, including CoreWeave, Lambda, Nebius and Azure, support confidential deployments.

Paying for supply

Jensen Huang: "Demand is extraordinary. Capacity is the constraint." NVIDIA says substrate deals with IBIDEN, Samsung Electro-Mechanics and Unimicron will together "drive a 2.5x increase in capacity over the next three years". Amkor targets more than 300% more interposer wafer capacity by the end of 2028, a $1.1 billion advance to KYEC funds US testing facilities aimed at April 2027 volume, and SPIL targets 15x optical-engine capacity for co-packaged optics by mid-2027.

What remains uncertain. No prices, model list or regions were given for OpenAI on Google Cloud. The 960,000-GPU figure is a plan with no date, and supplier timelines run to 2028.

Analysis: Model access is turning into a procurement question. When an OpenAI model can be paid for out of a Google commitment, the vendor decision moves from "which cloud" to "which contract", and multi-model routing becomes a finance feature as much as an engineering one. Confidential Bare Metal answers a different buyer, the lab that wants third-party GPUs without exposing its weights. That market will grow if supply stays tight for years. Teams planning 2027 capacity should treat NVIDIA's packaging and testing dates as the real schedule, not product launch dates.