OpenAI Slips Microsoft’s Leash

The redrawn $135 billion alliance gives Sam Altman freedom to sell anywhere, and reshapes the economics of the AI boom.

For three years, the deal that defined the generative AI era had a simple shape: Microsoft wrote the checks, Microsoft ran the GPUs, and OpenAI ran on Azure. That arrangement, struck when ChatGPT was a research curiosity and Satya Nadella was the only major-cap CEO willing to bet his cloud business on a generative model, made Microsoft the indispensable partner. As of this week, it is merely a partner.

OpenAI and Microsoft have rewritten the terms of their roughly $135 billion alliance, giving Sam Altman’s company the freedom to sell its models through any cloud provider while preserving Microsoft’s equity stake and access to OpenAI’s intellectual property. The mechanics are technical. The implication is not. The most consequential commercial relationship in artificial intelligence has been loosened because neither side could afford to keep it tight: OpenAI’s revenue ambitions have outgrown Azure’s distribution, and Microsoft’s balance sheet, large as it is, can no longer underwrite the compute bill for a frontier lab on its own.

What actually changed

Under the original 2023 terms, Microsoft held exclusivity over OpenAI’s API on Azure and a contractual claim on a substantial share of OpenAI’s profits through 2030. OpenAI’s customers, in practice, were Microsoft’s customers too: Fortune 500 firms that wanted GPT-4 bought it as an Azure service. The new framework, as the Financial Times reports, unwinds the cloud exclusivity. OpenAI can now route inference and sell enterprise contracts through Amazon Web Services, Google Cloud, Oracle, or any other infrastructure provider that meets its needs.

What Microsoft kept matters too. The company retains its equity position, its IP licensing rights to OpenAI’s models and weights through the existing term, and a continuing revenue-share arrangement. Microsoft, in other words, did not get cut out. It got recategorized: from gatekeeper to shareholder.

That distinction is the whole story. A gatekeeper extracts rent from every transaction. A shareholder benefits when the underlying company grows, regardless of which pipe the revenue flows through. Microsoft is betting that OpenAI’s total addressable market is larger if Altman can sell freely than if every contract has to clear Redmond’s compliance and pricing apparatus first.

Why now

Two pressures forced the renegotiation. The first is compute. Training and serving frontier models has moved from expensive to almost unaccountable: industry estimates for the next generation of training runs cluster in the tens of billions of dollars per model, and inference costs scale with usage. OpenAI has already signed multi-year compute commitments, including a reported $300 billion arrangement with Oracle and a separate large-scale deal with CoreWeave, that Microsoft alone could not absorb on its own infrastructure roadmap. Azure has its own Copilot products, its own enterprise customers, and its own GPU shortage. Telling OpenAI it could only buy Microsoft’s compute had become a constraint on both companies.

The second pressure is revenue. OpenAI is reportedly targeting an annualized revenue run rate that would place it among the fastest-growing enterprise software companies in history, and it is doing so while burning cash at a rate that requires continuous fundraising. Every customer it could not sign because that customer was an AWS or Google Cloud shop was a real cost. Enterprises do not casually migrate clouds to access a single vendor’s model, especially when Anthropic is available natively on AWS and Google’s own Gemini is improving. The exclusivity was leaving money on the table that OpenAI needs to fund its next training run.

For Microsoft, the calculus shifted as its own AI strategy matured. Copilot is now a real business. Microsoft has hired Mustafa Suleyman to lead an in-house consumer AI effort. It has cut deals to access models from Anthropic and others. The company’s dependence on OpenAI as the sole source of frontier capability has declined, which made giving OpenAI more freedom less risky than it would have been even a year ago.

The Musk overhang

The renegotiation lands while OpenAI is fighting Elon Musk in court over a $180 billion lawsuit alleging the company betrayed its founding nonprofit mission by transforming itself into a profit-maximizing entity controlled, in Musk’s telling, by Microsoft. The suit is partly personal grievance and partly a serious legal challenge to OpenAI’s planned conversion into a for-profit public benefit corporation, a restructuring that the new Microsoft terms are designed to accommodate.

The looser deal cuts both ways for that litigation. On one hand, OpenAI can argue more credibly that it is not a Microsoft subsidiary in disguise, that it operates independently, sells across clouds, and makes its own commercial decisions. On the other hand, the very act of renegotiating equity and IP terms in a way that clears the path for a for-profit conversion is exactly what Musk’s complaint targets. Discovery in that case will now include a fresh set of documents about who agreed to what, and why, in the autumn of 2025.

Musk’s parallel project, xAI, is a reminder that the structural argument cuts further than one lawsuit. Every serious frontier lab is now either independent and capital-hungry (OpenAI, Anthropic), captive to a hyperscaler (Google DeepMind), or bankrolled by an idiosyncratic billionaire (xAI, Safe Superintelligence). The Microsoft-OpenAI marriage was the model for option two. Its loosening suggests option two may not be stable at scale.

What it signals about the industry

Read against the broader AI economy, the redrawn deal is the clearest sign yet that the hyperscaler-plus-lab pairing, the dominant structure of the past three years, is fragmenting into something messier and more multipolar. Anthropic already takes investment from both Amazon and Google. OpenAI now sells through everyone. Compute contracts are being unbundled from equity stakes. The neat diagram of the AI industry, with three hyperscalers each owning a model lab, is dissolving into a web of overlapping commercial relationships in which the labs court multiple infrastructure providers and the providers hedge across multiple labs.

The economic logic is straightforward. The capital required to stay at the frontier has crossed a threshold where no single corporate balance sheet, not even Microsoft’s, can fund it without partners. OpenAI’s compute commitments alone now exceed the annual capital expenditure of most Fortune 50 companies. The only way to finance that is to spread the obligations across multiple counterparties and to maximize revenue by selling everywhere. Exclusivity is a luxury the math no longer supports.

This has consequences down the stack. The wave of AI-driven tech layoffs at Microsoft, Google, Meta, and Amazon over the past year reflects in part a reallocation of payroll dollars toward GPU procurement and data center construction. When a hyperscaler commits to tens of billions in AI infrastructure spending, something has to give, and what has given is headcount in legacy product lines. The OpenAI-Microsoft renegotiation does nothing to slow this; it accelerates it, because Microsoft now has to compete harder on its own products rather than relying on exclusive access to OpenAI’s models as a moat.

Wall Street’s sorting

Equity markets have been working through a related question all year: which software companies benefit from AI, which get disrupted by it, and which are merely adjacent. The redrawn alliance sharpens the answer in at least one respect. Companies that depended on a thesis of “Microsoft owns the AI distribution layer” have less to stand on. Companies whose products integrate directly with OpenAI’s API, regardless of cloud, have more optionality. Oracle, an unlikely AI winner a year ago, looks better positioned than its market cap suggests. Pure-play application software vendors that have not yet shown durable AI-driven revenue acceleration look more exposed.

The deeper signal for investors is that the AI value chain is not yet settled, and bets on any particular structural arrangement, including the assumption that hyperscalers capture most of the economics, should be held loosely. The companies that looked like permanent winners in 2023 are renegotiating their terms in 2025.

The closing math

What Microsoft gave up was control. What it kept was upside. If OpenAI becomes the trillion-dollar revenue business Altman has reportedly described to investors, Microsoft’s equity stake will be worth more than any cloud-exclusivity arrangement could have generated. If OpenAI stumbles, Microsoft is no longer the sole counterparty bearing the cost.

That is a rational trade for Microsoft and a necessary one for OpenAI. It is also an admission, by the two companies that defined the current AI era, that the era’s foundational deal structure was not built for the scale of what comes next. The leash is off, because the dog got too big to walk.