The Microsoft Decoupling, Read Slowly

OpenAI and Microsoft announced an amended commercial agreement this morning. The press releases on both sides are studiously bland, "evolving the partnership", "next phase", "continued collaboration", which is the register companies adopt when something has actually changed and they would rather you not notice the shape of the change. I spent the morning reading both releases against the partnership terms as previously disclosed, and the shape is worth tracing.

Four moving parts:

1. Microsoft's licence to OpenAI IP becomes non-exclusive through 2032.¹
2. OpenAI can serve its models on any cloud, with Azure retaining a right of first refusal as primary cloud.
3. OpenAI continues to pay Microsoft a revenue share through 2030 at the same rate, but subject to a new total cap.
4. Microsoft no longer pays OpenAI a revenue share on its end.

Take these one at a time, because the structure is not symmetric and the asymmetry is the story.

The exclusivity unwind, priced

The original 2023 agreement gave Microsoft an exclusive licence to OpenAI's models and a first-look on compute. That exclusivity was the load-bearing wall of the deal, it is what made the $13bn in committed Azure credits and prepaid compute legible as an investment rather than a vendor relationship. Microsoft was buying a moat.

The moat is now non-exclusive. OpenAI can serve GPT-class models from Oracle, from CoreWeave, from Google Cloud if it likes, from its own Stargate capacity as that comes online. Azure gets a right of first refusal, which is a polite way of saying Azure has to compete on price and availability for incremental OpenAI workloads. The first-refusal mechanic is the same structure used in venture term sheets to give an existing investor optionality without commitment, and it carries roughly the same weight: useful at the margin, not determinative.

This is the bit the inference economics frame predicts. When inference becomes the binding cost, and at OpenAI's reported run-rate, with Sora-2 and the agent products live, it is, single-cloud dependency is a margin problem. Every GPU-hour that has to clear through Azure's gross margin is a GPU-hour OpenAI is overpaying for relative to a competitively-bid alternative. Sam Altman has been telegraphing the Stargate / Oracle / SoftBank multi-cloud posture for over a year; this amendment is the legal instrument that lets the telegraphed strategy actually run.

The revenue-share asymmetry

Here is the part I find slightly strange and want to walk through carefully.

OpenAI continues paying Microsoft a revenue share through 2030, same rate as before, reported at 20% of OpenAI revenues, now subject to a new total cap.² Microsoft stops paying OpenAI a revenue share on Microsoft's own AI revenue (the Copilot and Azure OpenAI Service flows that were, under the original deal, partially shared back).

Read as a netting exercise: Microsoft has converted a two-way revenue-share into a one-way revenue-share with a cap. OpenAI pays Microsoft until the cap is hit; Microsoft pays OpenAI nothing. If you assume Microsoft's AI-attributable revenue is now larger than OpenAI's share of it would have been, which, given Copilot's enterprise penetration and Azure OpenAI's growth, is plausible, Microsoft has just bought out a contingent liability. The cap on OpenAI's payments back is the consideration OpenAI got for letting Microsoft out of its share obligation.

This is a settlement, dressed as an amendment. Both sides had a contract that was producing outcomes neither liked: Microsoft was paying OpenAI a share of revenue from products (Copilot) where Microsoft had done most of the integration work; OpenAI was constrained from multi-cloud and from competing freely with Azure-hosted versions of itself. They've traded.

What this is a case of

The model weight lineage frame applies here in an interesting way. The original Microsoft deal was structured as an IP licence, Microsoft licensed access to weights and to a pre-AGI generation of models, with a famous carve-out for whatever OpenAI's board declared to be AGI. That structure assumed weights were the asset and exclusivity over weights was the moat.

What the amendment concedes, implicitly, is that exclusivity over weights is not, by itself, a durable moat when the licensee can't keep up with the licensor's capex needs. OpenAI needs more compute than Microsoft alone can or will fund. Microsoft, having watched the capex curve, has decided it would rather not be the sole financier of OpenAI's compute, and the price of stepping back is giving up exclusivity. The weights lineage continues; the commercial fence around it has been lowered.

This fits a broader pattern. The AI performativity frame, capex commitments as the determinative variable, predicts that single-counterparty financing arrangements break under the weight of the spend they have to support. Stargate is reported at $500bn over four years across partners.³ No single hyperscaler underwrites that. The original Microsoft–OpenAI structure was built for a world where the binding constraint was access to a frontier model. The binding constraint is now access to enough power and silicon to serve the model, and that constraint is too large for one cloud to absorb.

What the cap is worth

The total cap on OpenAI's revenue-share payments to Microsoft is the most consequential undisclosed number in the amendment. Neither party has filed it. Microsoft is a 10-K filer; OpenAI is not. If the cap is, say, $50bn cumulative, roughly what Microsoft's original investment plus credits would project to under reasonable revenue assumptions through 2030, this is essentially Microsoft taking its money back through 2030 and then walking away. If the cap is $150bn, it is closer to a continued partnership with a defined end.

I would watch Microsoft's next 10-Q for the disclosure. Material modifications to a partnership of this size are reportable, and the cap figure is the one that will tell you what each side actually thinks the relationship is worth from here.

What to watch

• Microsoft's 10-Q segment disclosure, particularly any change in how the OpenAI investment is carried and any disclosed cap figure or impairment.
• OpenAI's compute mix. If Oracle and CoreWeave shares of OpenAI inference grow visibly over the next two quarters, the first-refusal mechanic is doing what I suspect it is doing, which is, not much.
• Copilot pricing. With Microsoft no longer paying OpenAI a revenue share, the unit economics of Copilot improve at the Microsoft level. Whether that shows up as margin expansion or as price aggression against Google Workspace will tell you what Microsoft thinks competitive pressure looks like.
• The AGI carve-out language. The original agreement's most-discussed clause was the AGI declaration trigger. Whether that survived the amendment, and in what form, is the single most important thing not in the press release.

The partnership isn't ending. It's being repriced for a world where the deal that made sense in 2023 doesn't make sense in 2026. Most large partnerships eventually get repriced. This one took slightly less time than I expected.

Footnotes

¶