OpenAI pivots toward Amazon partnership over Microsoft

The headline figure of Amazon’s $50 billion investment in OpenAI obscures a financial structure that is far more circular than a traditional equity investment. Of the $50 billion, only $15 billion is upfront Series C Preferred Stock. The remaining $35 billion is tied to performance-based triggers including AGI benchmarks or an IPO — milestones that may or may not materialize. Meanwhile, OpenAI expanded its existing $38 billion AWS cloud agreement by an additional $100 billion over 8 years, committing to consume approximately 2 gigawatts of Trainium capacity. In practical terms, a significant portion of Amazon’s investment flows directly back to AWS as cloud spending.

This circular dynamic has not gone unnoticed by market observers. Social media commentary has characterized the arrangement as ‘Amazon gift cards’ cycling back as AWS consumption. Tech investor Hiten Shah highlighted the contrast: Microsoft invested $13.8 billion starting in 2019 and secured 23.5% ownership with exclusive IP and API control, while Amazon paid $50 billion for just 6% ownership and distribution rights. The deal validates Trainium as enterprise-grade silicon — Amazon needs a marquee customer to prove its custom chips can compete with Nvidia — but the financial architecture ensures that Amazon’s cash outlay is substantially offset by guaranteed cloud revenue. For OpenAI, the arrangement provides both capital and compute infrastructure; for Amazon, it is as much a cloud customer acquisition strategy as an investment.

The Dresser memo represents a remarkable public fracture in what was the defining partnership of the AI era. When Dresser wrote that the Microsoft partnership ‘has also limited our ability to meet enterprises where they are,’ she was articulating a distribution problem that had been building for over a year. Many large enterprises run on AWS, not Azure. By being locked into Microsoft’s cloud ecosystem, OpenAI was effectively ceding the AWS-native enterprise market to Anthropic, whose Claude models were already available on Bedrock. The competitive cost was becoming visible: Glean CEO Arvind Jain described enterprise enthusiasm for Claude as reaching the level of ‘a religion.’

The strategic consequences for Microsoft are severe. The company’s stock is down 23% year-to-date, and it has taken the extraordinary step of listing OpenAI as a competitor in its annual filing — a company it invested over $13 billion in. AWS is now the exclusive third-party cloud distributor for OpenAI Frontier, OpenAI’s enterprise agent-management platform, and the two companies are co-building a Stateful Runtime Environment on Bedrock that enables AI agents to maintain persistent memory and context across sessions. This is not a surface-level distribution deal but a deep infrastructure integration. The era of exclusive AI lab partnerships appears to be ending, replaced by a multi-cloud model where AI companies distribute through whichever platform reaches the most enterprise customers.

Perhaps the most striking element of the Dresser memo was not the Microsoft criticism but the direct attack on Anthropic’s financial reporting. Dresser accused Anthropic of inflating its $30 billion revenue run rate by approximately $8 billion through gross revenue accounting of cloud partner rev-shares, stating they ‘use accounting treatment that makes revenue look bigger than it is, including grossing up rev share with Amazon and Google.’ This is an unusually aggressive public salvo between private AI companies, and it reveals how central revenue narratives have become to the AI competitive landscape.

The accounting dispute points to a structural issue in how AI companies report revenue when distributing through cloud marketplaces. When an enterprise purchases Claude through AWS Bedrock, Amazon takes a rev-share. Whether Anthropic reports the full transaction value or only its net portion after the rev-share meaningfully changes the reported revenue figure. OpenAI, by raising this issue internally (and now publicly through the leaked memo), is attempting to reframe the competitive narrative ahead of a potential IPO. With OpenAI’s enterprise revenue now exceeding 40% of total revenue and paying business users nearly doubling from 5 million to 9 million in six months, the company is positioning its growth metrics as more organic and reliable than Anthropic’s. The revenue war is no longer just about signing customers — it is about controlling the narrative of who is actually winning.

Buried beneath the financial headlines is a technical development that may prove more consequential than the investment itself: OpenAI and AWS are co-building a Stateful Runtime Environment on Amazon Bedrock that enables AI agents to maintain context, remember prior work, and act across multiple systems over extended periods. This moves enterprise AI beyond the stateless chat-and-response paradigm into persistent, multi-session autonomous workflows. An AWS executive described inbound demand for this capability since the partnership announcement as ‘frankly staggering.’

This technical integration explains why the partnership runs so deep. OpenAI is not simply listing its models on another cloud marketplace — it is building infrastructure that ties its agent capabilities directly into AWS’s enterprise stack. The 2-gigawatt Trainium commitment and $100 billion expanded cloud deal are the compute foundation for running these persistent agents at scale. OpenAI plans to reach 30 gigawatts of compute by 2030, compared to Anthropic’s target of 7-8 gigawatts by end of 2027. The gap in compute ambition reflects fundamentally different bets on how quickly enterprise AI shifts from on-demand inference to always-on autonomous agents. For enterprises already running on AWS, the Stateful Runtime Environment creates a switching cost that pure model quality comparisons do not capture — it locks OpenAI into the operational fabric of enterprise workflows in a way that a simple API endpoint never could.