OpenAI Goes Beyond Azure: Expanding AI Infrastructure

In January 2025, the exclusive contract between Microsoft and OpenAI expired, under which OpenAI was required to use only the Azure cloud platform to train and host its models, including ChatGPT.

Starting this year, the rules have changed: now, the “first refusal” principle applies: if Microsoft's resources prove insufficient, OpenAI has the right to engage third-party providers.

Transition to a Multi-Cloud Architecture

OpenAI quickly took advantage of the opportunity to diversify its infrastructure. In addition to its ongoing collaboration with Microsoft, the company has entered into agreements with CoreWeave, Oracle Cloud, Google Cloud, and a number of other computing power providers. This step allows it to scale model training to unprecedented volumes and reach the 1 million GPU milestone by the end of 2025.

However, expansion also has its technological challenges. OpenAI already uses distributed training across multiple data centers, which requires high synchronization and infrastructure stability.

Competitors Are Not Far Behind

OpenAI's plans to reach 1 million GPUs by the end of the year are not unique. Competitors are ramping up their capacity at a similar pace. Meta is aiming for 1.3 million GPUs in the same timeframe. Meanwhile, xAI, a company founded by Elon Musk, has purchased an old industrial facility in Memphis, where it plans to install up to a million graphics accelerators in addition to the 200,000 already in operation.