OpenAI diversifies compute beyond Microsoft

OpenAI is turning compute from a dependency into a portfolio. For years Microsoft was both OpenAI’s largest backer and the main place its models ran. The October 28, 2025 reset kept Azure as a huge partner, with a new $250B purchase commitment, but removed Microsoft’s right of first refusal on new workloads. That opened the door for OpenAI to split training and inference across Oracle, AWS, and specialist infrastructure, which lowers supply risk and gives it more leverage when negotiating price, chip access, and build speed.1 2 3 4

• The clearest signal is contractual, not rhetorical. OpenAI and Microsoft said on October 28, 2025 that Microsoft no longer has first refusal on OpenAI compute, while OpenAI still committed to buy an incremental $250B of Azure services. This is not a breakup. It is a shift from exclusivity to multi cloud at very large scale.1 2
• The new providers are not interchangeable backups. Oracle is the cloud layer for Stargate sites like Abilene, where OpenAI said Oracle Cloud Infrastructure runs NVIDIA GB200 systems. AWS adds a different supply lane, including a February 27, 2026 expansion that raised the AWS relationship by $100B over 8 years and committed about 2 GW of Trainium capacity.3 4 5
• This looks more like the buying pattern of a hyperscaler than a software company. OpenAI had already reached an estimated $50B annualized revenue run rate by September 30, 2026, and its infrastructure commitments now resemble long dated capacity reservations made to secure future product growth, especially for model training, always on inference, and agent workloads that can spike unpredictably.2 3 4

The next step is a more modular OpenAI stack, with different clouds and chips handling different jobs. That should make OpenAI faster at bringing new capacity online, less constrained by any one partner’s roadmap, and more able to match each workload to the cheapest or fastest infrastructure available.2 3 4 5