Three things happened this week that the coverage is treating as unrelated news items. They're not.
One: Microsoft and OpenAI end their exclusive revenue-sharing arrangement. Two: GitHub Copilot moves to usage-based billing. Three: OpenAI is apparently building a phone.
Read individually, these are business news. Read together, they're a distribution strategy.
OpenAI Is No Longer a Model-Layer Company
The exclusive deal ending means OpenAI can now go direct — to enterprises, to developers, to whoever they want — without routing value back to Microsoft. Copilot moving to usage-based billing means Microsoft is repricing its own AI layer to survive margin compression from a partner that just stopped being a partner in the traditional sense. And a phone with agents replacing apps is the clearest possible signal that OpenAI wants to own the surface where users actually touch their product.
This is not a pivot. It's a completion. The model was always the foundation; the question was always who would own the relationship with the end user. For four years the answer was "Microsoft, mostly." Now the answer is changing.
The blast radius taxonomy I've been developing assumes the foundation model provider stays in the model layer — that the risk surface is commoditisation, access revocation, auditor capture, toolchain capture. All of those assume OpenAI is upstream infrastructure that you access through some distribution partner's abstraction.
That assumption just became load-bearing in a way it wasn't before. OpenAI building a phone isn't a hardware story. It's an "I want to be the distribution layer" story. The phone is the most aggressive possible move in that direction: not just a ChatGPT app, not just an API, but a device where the model is the OS.
What Teams Got Wrong
A lot of teams made a bet that went roughly: "OpenAI capability, Microsoft safety. If something goes wrong with the OpenAI relationship, Microsoft is the backstop. Enterprise agreements, compliance coverage, Azure credits — that's the stable path."
That calculus assumed the exclusive arrangement was permanent infrastructure. It wasn't. It was a time-limited commercial arrangement between two parties whose interests were always going to diverge once OpenAI had enough scale to go direct.
The "Microsoft = safe OpenAI bet" framing confused distribution stability with relationship permanence. Microsoft hasn't lost access to OpenAI models — but it has lost preferred status. It's now competing for the same enterprise relationships using the same underlying technology, with a pricing model (usage-based Copilot) that signals it's already feeling the margin pressure of that new reality.
Microsoft is becoming a reseller. That's not catastrophic — resellers can be excellent channel partners. But you wouldn't model a reseller as a strategic moat.
I don't have a clean resolution here. The phone is two years from mass production if the rumor is right, and rumors about OpenAI hardware have been wrong before. The exclusive ending is real and documented. The usage-based billing is real and documented.
What I'm sitting with: the blast radius taxonomy most teams use needs a new axis. Not "will the model get absorbed?" but "is the distribution layer I'm depending on still the primary relationship, or is it a channel that can be bypassed?"
This week suggests the latter is now a live question.