On the left is Satya Nadella, on the right is Sam Altman. For years, Microsoft executives have been accustomed to disputes with OpenAI over the intricate terms of their collaborative agreement. The original contract granted Microsoft exclusive rights to sell OpenAI's technology to cloud customers.
Two months ago, OpenAI, the developer of ChatGPT, suddenly announced a major partnership, planning to sell its AI technology to enterprise clients through Microsoft's primary competitor, Amazon.com.
OpenAI believed the existing contract with Microsoft permitted such collaboration with Amazon.com, but Microsoft disagreed. According to insiders from both companies, as OpenAI's revenue surged and their interests diverged, the former allies—Microsoft CEO Satya Nadella and OpenAI CEO Sam Altman—engaged in multiple rounds of private discussions in recent weeks.
The talks involved Microsoft's President of Business Development, Jon Tinter, along with both companies' CFOs, Sarah Friar and Amy Hood. The two CEOs ultimately finalized a plan that averted a potential lawsuit and reached a mutual compromise addressing their long-term demands.
On Monday, the companies announced a revised agreement: OpenAI can now pursue deeper collaboration with Amazon.com and, for the first time, is permitted to sell its large AI models to enterprises through other cloud providers, including Google.
This grants OpenAI a level of business freedom comparable to its competitor Anthropic, which already offers AI services across Amazon.com, Google, and Microsoft's cloud platforms.
In exchange, OpenAI relinquished a key counterbalance clause from the original contract. The previous agreement stipulated that if OpenAI determined it had achieved the ambiguous technological milestone of Artificial General Intelligence (AGI), it could terminate Microsoft's exclusive rights to sell its most advanced AI models.
This AGI clause, included in Microsoft's initial 2019 agreement with OpenAI, had long been a source of anxiety for Microsoft executives, who were concerned about the unpredictability of when OpenAI might invoke it.
OpenAI also agreed that until 2030, even if it later declares AGI has been achieved, it will continue to share 20% of its total revenue with Microsoft, subject to an undisclosed cap.
Insiders stated this revenue share covers all enterprise revenue generated by OpenAI through Amazon.com, Google, and all other cloud channels. Concurrently, Microsoft will no longer be required to share 20% of the revenue it earns from selling OpenAI models to its own cloud customers with OpenAI.
The old agreement stipulated that the revenue-sharing mechanism would terminate immediately upon OpenAI declaring AGI.
The details of the revised agreement indicate both parties were willing to compromise on critical terms, which, if left unresolved, could have hindered OpenAI's path towards a public listing.
Microsoft is one of OpenAI's largest shareholders, with a stake currently valued at least at $135 billion on paper. A successful OpenAI IPO would yield substantial financial returns for Microsoft.
The new agreement also signifies a loosening of the once deeply intertwined relationship between Microsoft and OpenAI. Originally intended to give Microsoft an insurmountable lead in the AI race, the partnership now sees OpenAI fully engaging in business with Microsoft's major competitors.
Commenting on the updated agreement, Amazon.com CEO Andy Jassy posted on social media Monday, "This morning's announcement is quite interesting."
He revealed that Amazon Web Services (AWS) plans to soon make OpenAI's large models available to customers and will also launch jointly developed AI agent products focused on enterprise automation.
OpenAI projects its 2030 revenue will reach $284 billion. Based on the 20% share, this could mean payments to Microsoft exceeding $56 billion that year, far higher than the approximately $6 billion estimated for this year.
However, due to the cap, the final actual payout may not reach that level. Another insider revealed the new agreement removed OpenAI's option for deferred revenue share payments—a mechanism previously used to help optimize its cash flow management.
One clause that remains unchanged: Microsoft will continue to have access to OpenAI's intellectual property licenses, even if OpenAI acquires other software or application companies in the future.
It was this intellectual property sharing obligation that reportedly hindered OpenAI's plan to acquire the AI code company Windsurf last year.
A Turning Point
This new agreement marks the third revision of their partnership terms in nearly a year and represents a significant turning point in their relationship since the advent of ChatGPT.
Their collaboration began in 2019 when Microsoft, aiming to catch up with Amazon.com in cloud computing, invested $1 billion in OpenAI, which had just established a for-profit entity.
When OpenAI was founded as a non-profit research lab in 2015, Elon Musk was an early donor. Musk believes that when OpenAI created its for-profit arm, both Microsoft and OpenAI violated the original agreement's terms. His lawsuit against both companies is scheduled for trial this week.
In early 2023, OpenAI's release of ChatGPT ignited a global AI boom, prompting a rush across industries to develop similar models and applications. Microsoft subsequently invested an additional $10 billion in OpenAI.
In return, the parties signed a complex revenue and technology-sharing agreement that year, deepening their ties: Microsoft gained preferential access to cutting-edge AI models for resale to enterprise customers; Microsoft became OpenAI's exclusive cloud provider; and Microsoft accepted OpenAI's unique equity structure, ensuring investors like Microsoft had priority profit-sharing rights.
But tensions soon emerged: OpenAI's demand for data center servers surged, exceeding Microsoft's capacity for compute expansion, alongside disagreements over other terms in the 2023 agreement.
Renegotiations began in the fall of 2024, including discussions about converting OpenAI's profit-sharing mechanism into a traditional equity structure.
Between 2024 and 2025, Microsoft gradually allowed OpenAI to rent cloud servers from Amazon.com, Google, and Oracle, and to build its own data centers.
Last fall, when OpenAI announced the conversion of its profit participation units into traditional stock, Microsoft simultaneously agreed that OpenAI could rent servers from other cloud providers without prior approval.
Around the same time, Microsoft began paying to use Anthropic's models within its Copilot product and started reselling Anthropic's large models on the Azure platform.
However, Microsoft retained the exclusive right to sell OpenAI's technology via cloud channels. According to individuals involved in the partnership, some OpenAI executives were dissatisfied with this term, noting that competitor Anthropic faced no such restrictions and was gaining ground in the enterprise market by being available on multiple cloud platforms.
Microsoft later issued equal incentive policies to its sales teams, encouraging the promotion of both Anthropic and OpenAI models.
'Stateful' AI Creates an Unexpected Divide
When OpenAI revised its agreement with Microsoft in the fall of 2025, it added a new clause that later became contentious: it allowed OpenAI to co-develop certain products with third parties.
According to an OpenAI blog post from October of that year, Microsoft retained exclusive rights to sell OpenAI's stateless general-purpose models via API, but non-API products could be offered on any cloud platform.
Some Microsoft executives had previously overlooked this clause, while OpenAI management saw it as an opportunity to begin collaborating with Amazon.com on developing new products fully based on AWS, aimed at helping enterprise customers build custom AI agents.
On the same day OpenAI announced the Amazon.com partnership, it jointly issued a statement with Microsoft affirming that Microsoft Azure held exclusive rights to sell OpenAI's stateless general-purpose models via API.
Simultaneously, OpenAI announced it would jointly develop 'stateful' AI products with AWS to help businesses build custom agents and automate workflows.
Unlike stateless API services, stateful AI can retain customer-specific information, such as business data or conversation history from previous interactions.
Insiders said Microsoft executives were taken by surprise by the product being defined as 'stateful' by OpenAI and AWS, as this meant it could operate completely independently of Azure's hosted model infrastructure.
Furthermore, Amazon.com committed to investing up to $50 billion in OpenAI, expanding server rental cooperation and, for the first time, including provisions for OpenAI's use of Amazon.com's Trainium AI server chips.
Several Microsoft executives initially struggled to interpret the partnership announcement, concerned it might violate the existing agreement, and expressed strong internal reservations.
Weeks later, Microsoft stated publicly, "We believe OpenAI fully understands and respects the importance of fulfilling its legal obligations."
As tensions escalated, Altman and Nadella initiated multiple high-level meetings to resolve the potential legal dispute.
A key factor enabling the smooth progress of negotiations was Microsoft's decreasing reliance on OpenAI's technology. While Microsoft continues to use AI models freely in its own products, it has also become one of Anthropic's largest customers. Several Copilot features utilize Anthropic models, which in some scenarios perform even better than OpenAI's.
Nadella has recently downplayed the importance of any single model, emphasizing instead the value of applying models in real-world applications.
In a podcast interview last week, Nadella stated, "The industry has entered a phase of model oversupply. AI capabilities are mature; the key now is how to effectively use these capabilities. That's why I place more importance on product demonstrations than mere model parameters."
Comments