Eliza Ward: Hey. So we have something today.
Brian Reed: Yeah, I saw it come in — what's the confirmed shape of it?
Eliza Ward: Okay, the confirmed shape: OpenAI and Google Cloud announced today — October 21st — that OpenAI's frontier models and Codex are coming to Google Cloud's Model Garden. Enterprises can buy them through Google Cloud Marketplace, draw down existing GCP commitments. That's the deal.
Brian Reed: Google Cloud. Hosting OpenAI models.
Eliza Ward: On the same shelf as Gemini, yeah.
Brian Reed: So Alphabet is — hang on — Alphabet is literally selling a direct competitor's models through its own cloud infrastructure. That's the thing we're processing right now.
Eliza Ward: That's the thing. And — wait, worth flagging — the models aren't actually live yet. The announcement says availability is coming in the coming weeks. So this is the deal existing, not the product existing.
Brian Reed: Right — so it's an announcement of a future availability. And this completes, what, all three major hyperscalers now? Azure, AWS, and now Google Cloud?
Eliza Ward: All three, yeah. But that framing — 'expanding beyond Microsoft' — is actually six months stale. The expansion already happened. Today is number three.
Brian Reed: Wait, walk me through the sequence because I think the timeline is doing a lot of work here.
Eliza Ward: Okay — think of it like a restaurant that used to have an exclusive contract with one delivery app. That contract quietly expired in April. They already signed with a second app — AWS Bedrock. Today is just the third app. The headline is calling it a breakthrough. It's actually the last step of a thing that already happened.
Brian Reed: Right, and the April date matters specifically — April 27th, Microsoft and OpenAI formally restructured. That's when Microsoft lost exclusive cloud distribution rights. Not today, not when AWS launched. April 27th.
Eliza Ward: And there's a February version of the story that tried to paper over it. February 27th, there was a joint statement — Azure is the exclusive provider for stateless APIs, the AWS deal fits within existing contracts. That's the official line. Then April comes and the exclusivity just... formally ends.
Brian Reed: So they negotiated the end in phases. February was — let me see — February was basically damage control framing, and April was the actual surrender.
Eliza Ward: That's the inference. What's confirmed is the gap exists. And then — okay, here's the detail that reframes everything about what 'primary partner' even means now — Microsoft no longer pays OpenAI a revenue share under the revised deal. But OpenAI still pays Microsoft 20% through 2030.
Brian Reed: Hold on. OpenAI is paying Microsoft, not the other way around?
Eliza Ward: Through 2030. Azure OpenAI Service keeps first-ship rights, the model license runs to 2032 — but the 20% going the other direction, that's gone. So 'primary partner' at this point is... I mean, it's real in the sense that Azure ships first. But it's also a debt relationship dressed up as a privilege.
Brian Reed: Right, but that debt framing — I keep seeing something different floating around, and I think it's wrong. The take is that Google hosting OpenAI is this confident, 'Gemini is so secure we can afford to host the competition' move. I don't buy it.
Eliza Ward: No. I don't think that's what happened either.
Brian Reed: So what's the honest read?
Eliza Ward: Google looked at enterprises with fifty million dollars in GCP commitments who were routing to Azure anyway to buy OpenAI. And decided — losing the transaction entirely costs more than whatever cannibalization risk comes from listing a competitor in Model Garden.
Brian Reed: That's not confidence, that's — wait, that's actually a defensive posture dressed as an open platform story.
Eliza Ward: And it goes further than just the listing. Google isn't just putting Codex in Model Garden next to Gemini — it's integrating Codex into Antigravity and the Gemini Agent. It's taking on full data governance responsibility for OpenAI's model traffic. Authenticated identity, audit logging, VPC Service Controls. That's operational commitment to something they didn't build and genuinely compete with.
Brian Reed: So Alphabet — the same Alphabet competing with OpenAI on frontier models through Gemini — is now securing OpenAI's enterprise traffic. Neither company resolved that tension in the announcement.
Eliza Ward: Neither resolved it because I'm not sure it — I mean, can it be resolved? The structure is the tension. And the part we haven't touched yet, what this actually does to OpenAI's pricing leverage once it's on all three clouds — that's where the second-order effect gets uncomfortable.
Brian Reed: The distribution looks like strength from the outside and might be the opposite from the inside.
Eliza Ward: And that gap between 'looks like strength' and 'is strength' — that's actually the pricing story. Because the thing that let OpenAI charge premium rates wasn't just model quality. It was friction. Getting to GPT-4 used to mean going through Azure. That's gone now.
Brian Reed: Okay, let me make that concrete. Picture a procurement officer — she's got, let's say, fifty million in GCP commitments she needs to draw down by Q4. She opens Model Garden. And sitting right there, next to Gemini, are OpenAI's models. Two hundred plus other models on the same screen. Same billing account. She never calls OpenAI.
Eliza Ward: She never negotiates with OpenAI.
Brian Reed: Right — and she's comparing OpenAI's model directly against Gemini, against whatever else is in that garden, on the same interface, in the same checkout flow. The access friction is just... it's not there anymore. So what's OpenAI's pricing argument at that moment?
Eliza Ward: Model quality, theoretically. But — wait, that's actually the uncomfortable part — once you're in Model Garden running side-by-side evals, 'good enough' becomes a real competitive standard. It's not 'is GPT-5 better than Gemini,' it's 'is GPT-5 better enough to justify a price difference when I'm already drawing down GCP commitments for Gemini anyway.'
Brian Reed: That normalization is — I mean, that's the part that doesn't get named in the announcement. The same unified billing pattern runs across AWS Bedrock and Azure OpenAI Service now. Switching costs between hyperscalers for OpenAI access are lower than they've ever been. Which is good for buyers and genuinely terrible for OpenAI's margin.
Eliza Ward: And then there's the IPO number sitting underneath all of this. CNBC reported OpenAI is targeting around a $1.4 trillion valuation — that's one source, not confirmed across multiple — but if that's even directionally real, you need distribution volume now. You price for growth, not margin.
Brian Reed: So the trade is: flood every enterprise channel before the IPO window, even if that structurally weakens your pricing leverage afterward.
Eliza Ward: That's the inference. What I'd watch is whether OpenAI's model pricing inside GCP Marketplace diverges from what Azure OpenAI Service charges — because if those numbers start separating, that tells you the hyperscalers are already competing on OpenAI's behalf. And at that point 'primary partner' is just a title.
Brian Reed: The thing that would actually settle it for me — we don't know yet, because the models aren't live on Google Cloud. That's not a hedge, that's just the state of it as of today. The announcement exists. The product doesn't. So the real test — whether a GCP enterprise buyer actually draws down commitments toward OpenAI models instead of routing to Gemini on the same screen — we can't run that test yet.
Eliza Ward: Right. And that's — I mean, that's the only signal worth waiting for. Not the announcement. When the models go live in Model Garden and we can see whether the drawdown mechanism actually gets used, or whether enterprises just... stay on Gemini because it's already there.
Brian Reed: Because if the drawdown doesn't happen at scale — if buyers are in Model Garden and they're just picking Gemini anyway — then being on all three hyperscalers isn't distribution strength. It's just... everywhere and nowhere.
Eliza Ward: Ubiquity without pricing power. That's the version of this that's actually bad for OpenAI at a $1.4 trillion target. And we genuinely don't know which version this is yet.
Brian Reed: Watch the launch date. And watch the price.
