OpenAI on Google Cloud Is a Capacity Play, Not a Partnership: Google Takes 1.5–3% of a $50B Run Rate as Azure's Exclusivity Premium Disappears

Three things are new. OpenAI's frontier models and Codex reach a third hyperscaler storefront, purchases settle through Google Cloud Marketplace, and enterprises can spend down existing Google Cloud commitments against them. One thing is not new, and it is the load-bearing point: Microsoft remains OpenAI's primary cloud partner, OpenAI products still ship first on Azure, OpenAI's first-party products still run on Azure, and the commitment to buy at least $250 billion of Azure services through 2032 is untouched. Azure lost exclusivity in April and kept first-launch rights — a downgrade in bargaining power dressed up as continuity.

The line that matters for enterprise budgets is the third row. Cloud spend commitments are how large companies actually pay for AI now: a multi-year Google Cloud commit is a pool of prepaid dollars, and Model Garden purchases draw that pool down. That converts a locked infrastructure contract into an AI consumption budget, and it is why Google needs a rival's models on its own metal. OpenAI inference booked on Google Cloud turns a fixed TPU fleet into billed revenue at a cost to Google of a fraction of a percent of the transaction.

Alphabet's June 2026 quarter was the strongest cloud print it has ever posted: revenue of $24.8 billion, up 82% year over year, with segment operating income of $8.8 billion against $2.8 billion a year earlier. That lifts segment operating margin to 35.6% from 20.7% — nearly 15 points in twelve months — while the same quarter produced negative free cash flow of $5.9 billion on $44.9 billion of capital spending. Alphabet raised its 2026 capex guidance to $195–205 billion from $180–190 billion, after starting the year at $175–185 billion, which implies roughly $57–62 billion a quarter in the second half against $44.9 billion in the second. Google is buying demand for capacity it has already paid for, and a rival's frontier models are the cheapest demand available.

Here is the part that is not obvious. A business selling a competitor's product normally trades margin for reach. Google Cloud's margin expanded by nearly 15 points in a single year while the segment was compounding at 82%. That is the signature of a fixed-cost asset with tenants, not a reseller. The inference: Alphabet converts a capital program into a rental business, and OpenAI is one of the tenants. Whether that is durable depends on a number neither company has published — how much of a metered model-inference dollar Google actually keeps.

When a frontier model is sold through a hyperscaler's storefront, the bill is settled by the cloud and the cloud keeps a slice. Google's published Cloud Marketplace schedule takes 1.5% on private offers of $10 million or more, 2% on private offers between $1 million and $10 million, and 3% on standard offers or on sub-$1 million private offers. The rate that applies to metered model consumption through Model Garden is not disclosed. That gap matters more than it looks, because the published floor gives every model lab a number to price against when it decides what to charge, and it sets a ceiling on what the labs can ever negotiate away.

This is the first hard evidence of where the margin settles in a world of interchangeable frontier models. Anthropic's prospectus shows nearly half of 2025 revenue arriving through Amazon and Google Cloud storefronts, while the company burned $7.3 billion on compute and $42 billion overall and committed itself to $518 billion of future cloud obligations. OpenAI is now running the same play on a larger revenue base and a smaller cost plan. Multi-cloud distribution buys reach and surrenders the spread — which is why the take rate, not the benchmark score, is the number to watch.

Two layers sit outside the model and the marketplace, and both are paid in cash that originates in the same place. Upstream are the silicon and the memory that Google's accelerator fleet is built from. Downstream are the channels and capacity providers that carry the model and the compute behind it — and they are not equal partners in this arrangement.

CoreWeave is the cleanest read on what is happening underneath. Its largest customer fell to 36% of quarterly revenue from 71% a year earlier, and a second customer reached 26%. A neocloud built to sell one anchor tenant's capacity is now serving three, and the anchor's share is halving. Multi-sourcing compute and multi-channeling the model are the same move — and both shrink the toll road's pricing power. The remaining performance obligations of $103.7 billion are about 14 times trailing revenue, financed with $45.2 billion of net debt against a $34.3 billion market capitalisation, so any re-pricing of the top contract re-prices the whole structure.

Three risks break this thesis. First, price: if OpenAI discounts models on Model Garden to win enterprise share, Google subsidizes inference and the 35.6% segment margin is the first casualty. Second, substitution: Model Garden gains can be Gemini losses, and at 82% growth the mix matters more than the dollars. Third, concentration: the take rate is a contract term, and the only party that has published the real number so far is a lab that filed to go public and told the market it earned 47% of its revenue through cloud storefronts. Expect the next real disclosure of this spread to come from an IPO prospectus rather than from a cloud earnings call, because the clouds have no reason to publish it.

Fact: on October 21, 2026, OpenAI put its frontier models and Codex on Google Cloud's Model Garden, sold through Google Cloud Marketplace, with general availability planned in the coming weeks, while Microsoft's first-launch rights and the $250 billion Azure commitment through 2032 stayed in place. Inference: Google is monetising a TPU fleet it has already financed, and the commercial price of that monetisation is a published 1.5% to 3% marketplace take rather than an exclusive partnership. Speculation, offered as such: if the labs become pass-throughs of cloud-billed consumption, the durable margin in this industry settles with whoever owns the meter, and the labs keep the price per token rather than the relationship with the buyer.

What would change the read: a published take rate for metered model consumption on Model Garden, an OpenAI disclosure of revenue split by channel, or evidence in Alphabet's third-quarter numbers that third-party models lift cloud revenue without denting Gemini's mix. Until one of those lands, the take rate stays the least-observed number in AI's largest revenue pool, and the most valuable disclosure of the next year will be a footnote in a prospectus rather than a line in a cloud segment.