Google Sells OpenAI Models | Future Intelligence #32
Google Cloud has started selling models from its direct competitor—GPT-6 Astra, Sol and Luna are heading to Model Garden, separating enterprise AI distribution from model ownership. In Berlin, Nvidia unveiled DSX POD: 504 Rubin GPUs in a single container, deployment in fourteen days, with deliveries starting only in 2027. But the limit on the new cycle isn’t chips; it’s electricity—Microsoft walked away from leasing 23 000 chips in Essex because the utility says a grid connection won’t be available until the 2030s, while AWS capacity is largely sold out through 2027, according to Andy Jassy.
Future Intelligence is an exclusive analytical report published once a week and available only to Bulios Black members. Members receive it automatically every Sunday morning by email—in full, including specific scenarios and market implications. Get ongoing access to the report by joining Bulios Black.
Key points
Why Google has started offering top OpenAI models, including the Codex coding agent, in its catalog
What Nvidia’s containerized AI factory with 504 graphics chips means for the pace of data center construction
How Amazon’s new tool moves models from the PyTorch environment to its own chips without rewriting code
Under the radar: Ultragenyx has its first gene therapy approved in years—and what that opens up for the company
Google Cloud outage in London, Microsoft’s retreat from chip leasing, and AWS capacity sold out
OpenAI in Model Garden: Google sells its rival’s models
Google Cloud and OpenAI announced on October 21, 2026, that top OpenAI models and the Codex coding agent will be coming to Model Garden, the model catalog on the Google Cloud platform. The first wave includes GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, currently in private preview for select enterprise customers, with full availability planned for the coming weeks.
OpenAI technology is now available on all three of the largest U.S. cloud platforms. For Alphabet $GOOG, that means revenue from the technology of its most direct Gemini competitor, running on its own infrastructure—and thus a separation between the role of capacity provider and that of model creator.
How selling a model from someone else’s counter works
Model Garden is a catalog from which enterprise developers select a model and call it through an interface within the Google Cloud environment. Purchases are made through Google Cloud Marketplace, where customers can apply their existing cloud spending commitments—pre-agreed spending volumes—to OpenAI models.
That detail is economically significant. Large enterprises have multi-year commitments with cloud providers, and buying through the marketplace lets them draw on those commitments without a new vendor contract, security audit or billing relationship. The friction that had protected each platform’s homegrown models is disappearing.
The agreement also includes Codex workflows: the local Codex CLI and Codex Desktop tools are to connect to Google’s platform for inference—the actual running of the model. The developer stays in their own environment; the compute bill goes to Google.
The announcement, however, includes no pricing, regional availability or technical model specifications. Without a price list, it’s impossible to calculate what margin Google will retain from selling someone else’s model—and that is the only number that will determine whether this is meaningful revenue or merely a customer-retention tool.
The end of exclusivity as a prerequisite
The move became possible only after April 27, 2026, when Microsoft and OpenAI rewrote their 2019 partnership: exclusive cloud distribution ended, the AGI clause disappeared, and Microsoft’s revenue share was capped. OpenAI products will still debut on Azure, but they are no longer exclusive to it.
At the same time, OpenAI is diversifying away from a single capacity provider. Its contract with Oracle is worth 300 billion USD over five years, or about 60 billion annually starting in 2027; its AWS deal is worth 38 billion USD, and above both sits the Stargate project with SoftBank, targeting ten gigawatts.
Distribution and compute capacity are thus becoming separate. OpenAI buys silicon from several suppliers and sells models through all three hyperscalers; the hyperscalers, in turn, are moving from the role of exclusive partner to interchangeable infrastructure providers with catalogs that put their own and outside models side by side.
Alphabet ahead of quarterly results on October 28
Cloud is the engine of Alphabet’s story. In the most recently reported quarter, Google Cloud revenue grew 82 % year over year, the segment’s operating profit reached 8,8 billion USD, and its margin expanded to 35,6 % from 20,7 %. According to management, Cloud’s backlog reached 514 billion USD, compared with 462,3 billion as of March 31, 2026.
The market nevertheless values the core business more cautiously: Citizens lowered its price target to 450 USD from 515 USD while maintaining a Market Outperform rating. That corresponds to roughly 28 times estimated 2027 earnings of 16,24 USD per share, compared with 32 times previously.
The reason is AI competition in search—precisely the part of the business that selling outside models won’t replace.
Alphabet will report third-quarter results on October 28. What will be interesting isn’t Cloud’s growth rate, which has been predictable after a run of quarters, but anything that sheds light on the margin from its catalog of outside models and whether Microsoft $MSFT retains its premium position in enterprise distribution after its share of OpenAI revenue was capped.
GTC Berlin: Nvidia packs 504 Rubin GPUs into a container and cuts deployment time to 14 days
At its first GTC in Berlin, Nvidia unveiled DSX POD on Wednesday, October 21—a reference design for a complete AI factory packaged in a modular container. Inside are seven Vera Rubin NVL72 racks, for a total of 504 Rubin GPUs and 252 Vera processors, connected by a Spectrum-X network.