On April 27, two days after Google committed $40 billion to Anthropic and three days after Amazon expanded its own Anthropic stake, OpenAI and Microsoft announced something that got buried under the Anthropic coverage: a fundamental rewrite of the partnership agreement that has governed the most important relationship in AI since 2019.

The timing was not accidental. And the framing in the press release, "simplifying our partnership," was doing a lot of work to obscure what actually changed.

Spend a week watching Anthropic announce a 5 gigawatt TPU deal with Google and Broadcom, followed by a $100 billion AWS commitment, followed by a $40 billion Google equity round, and you start to understand the organizational pressure building inside OpenAI's leadership.

Anthropic was signing infrastructure deals with every major cloud provider simultaneously. AWS, Google Cloud, and Microsoft Azure were all racing to be Anthropic's partner, writing enormous checks and committing compute capacity, because having Anthropic as a customer and partner is a credibility signal in the enterprise cloud market. The independent AI lab with the best models gets to run a multi-cloud strategy that maximizes its leverage with every hyperscaler at once.

OpenAI, by contrast, was locked into Azure exclusivity. Microsoft had the right of first refusal on all of OpenAI's cloud capacity needs. OpenAI products shipped first on Azure, and in many cases only on Azure, by contractual obligation rather than by choice. OpenAI's revenue chief Denise Dresser said it plainly in an internal memo earlier this year: the partnership had "limited our ability to meet enterprises where they are."

The subtext of that statement is that enterprise customers running on AWS or Google Cloud were having friction adopting OpenAI products because the products were built around Azure infrastructure. Anthropic, with Claude available natively on all three major clouds, was easier to deploy in a multi-cloud enterprise environment. OpenAI was losing enterprise deals not because of model quality but because of distribution constraints baked into a seven-year-old partnership agreement.

That is the problem the April 27 amendment was designed to fix.

The press release language was carefully diplomatic. The reality underneath it is a meaningful structural shift in who has leverage.

Microsoft's license to OpenAI's intellectual property, previously exclusive, is now non-exclusive through 2032. OpenAI can now serve its products to customers across any cloud provider. AWS, Google Cloud, and every other platform can now build native OpenAI integrations without Microsoft's structural veto. That is a direct response to watching Anthropic operate freely across all three hyperscalers while OpenAI was constrained to one.

The revenue share arrangement also changed in a way that benefits both parties differently. Microsoft will stop paying OpenAI a revenue share. OpenAI will continue paying Microsoft a revenue share through 2030, but the payments are now subject to a total cap rather than being open-ended. OpenAI gets cost predictability and an upper bound on what it owes Microsoft. Microsoft gets a guaranteed revenue stream through 2030 regardless of OpenAI's technical progress, which was a specific concern given the original agreement's AGI clause that gave Microsoft certain rights contingent on OpenAI's definition of whether AGI had been achieved.

The AGI clause deletion is worth noting separately. The original partnership contained language that would have required Microsoft to determine its response if OpenAI declared it had achieved artificial general intelligence. That clause created genuine uncertainty about what Microsoft's obligations and rights would be in an AGI scenario. Removing it is a clean simplification, but it also signals that both parties have quietly agreed to stop treating AGI as an imminent event that requires contractual preparation.

Microsoft remains OpenAI's primary cloud partner. OpenAI products still ship first on Azure unless Microsoft cannot or chooses not to support the required capabilities. Microsoft continues to participate in OpenAI's growth as a major shareholder. The relationship is not ending. It is being renegotiated from one of dependency to one of preference, which is a different thing.

Here is where the story gets more complicated than the amendment coverage suggested.

OpenAI has been building the most aggressive infrastructure independence play in the industry, quietly, for the past 18 months. The total infrastructure commitments across all partners now exceed $1 trillion in planned spend through 2035. Oracle at $300 billion over five years. Microsoft at $250 billion in incremental Azure services. Nvidia at $100 billion in a cross-investment. AMD at $90 billion in GPU commitments. AWS at $38 billion. CoreWeave at $22 billion. Broadcom at $350 billion in custom silicon.

Stargate, the joint venture with Oracle and SoftBank announced at the White House in January 2025, was supposed to be the centerpiece of this independence strategy. Nearly 7 gigawatts of planned data center capacity. Over $400 billion in investment across five sites. A 10 gigawatt commitment by end of 2025, ahead of schedule on paper.

The reality has been messier. OpenAI has effectively abandoned the original Stargate structure of directly owning infrastructure alongside its partners. After more than a year of disagreements about control, ownership, and responsibility, the company now prefers to lease compute capacity on long-term agreements rather than own physical data centers. Stargate has been reframed internally as an "umbrella for our compute strategy" rather than a specific joint venture with hard ownership positions.

The practical implication is significant. OpenAI is not building owned infrastructure. It is building a portfolio of long-term leases and compute commitments with partners who own the physical assets. That is a different risk profile than Anthropic's position, where compute is reserved on hyperscaler infrastructure that the hyperscalers built and maintain. OpenAI is trying to have it both ways: the independence of a multi-cloud strategy without the capital burden of owning data centers. Whether the partners building those data centers under the assumption of OpenAI's long-term leases remain committed through 2030 and beyond is an open question that nobody is asking publicly.

The Microsoft partnership's primary cloud commitment runs through 2030. The revenue share payments run through 2030. The Oracle Stargate contract runs through 2031. AWS runs through 2032. Most of the major infrastructure commitments have end dates in the 2030 to 2032 window.

This matters because the commitments were structured in an era of extraordinary AI investment enthusiasm, pre-IPO valuations, and the assumption that frontier AI revenue would grow at the pace it has shown in 2025 and 2026. Renewing those commitments in 2030 requires OpenAI to be a public company with strong financials, or a private company with equally strong financials, negotiating from a position of demonstrated long-term commercial viability.

The infrastructure partners signing these deals are not doing it out of charity. AWS committed $38 billion to OpenAI because it wants the revenue and the competitive positioning of hosting the most widely deployed AI models. That logic holds as long as OpenAI's models remain the most widely deployed. If Anthropic closes the gap further, if open-weight models commoditize more of the workload, or if OpenAI's commercial growth slows, the terms of 2030 renewals will look different from the terms of 2025 signings.

Anthropic's position looks different on this dimension. The dual hyperscaler structure, with AWS and Google both financially and strategically invested in Anthropic's success, creates a situation where the infrastructure partners have equity skin in the game rather than just contract revenue. They want Anthropic to grow because their equity appreciates when it does. OpenAI's infrastructure partners, with the exception of Microsoft as a shareholder, are primarily vendors rather than co-investors.

That is a subtle but important structural difference as the industry heads toward a capital environment that will eventually be less forgiving than the one that produced trillion-dollar infrastructure commitments inside a single year.

Read together, the Microsoft amendment and the Anthropic infrastructure deals tell a coherent story about where the industry's center of gravity is moving.

The era of exclusive cloud partnerships, where an AI lab traded distribution freedom for guaranteed compute and capital, is ending. Both Anthropic and OpenAI are executing multi-cloud strategies, for different reasons and from different positions, but the direction is the same. The hyperscalers that bet on exclusivity as their primary AI strategy are watching that leverage evaporate.

Microsoft handled this gracefully. It traded an exclusive license it was increasingly unable to enforce for a guaranteed revenue stream, a continuing primary partner relationship, and the release of contractual obligations that were becoming a liability. It remains OpenAI's largest shareholder and primary cloud partner. It gave up a lock it could not maintain in exchange for terms it can actually collect on.

The more interesting question is whether the multi-cloud independence both labs are building actually delivers the commercial flexibility they need, or whether it creates a new kind of dependency on a distributed set of infrastructure partners whose interests are aligned only as long as the commercial trajectory justifies the commitments they made at peak enthusiasm.

That question does not have an answer yet. The 2030 renewals will answer it.