OpenAI says Habitat serves more than 500 petabytes for products used by more than one billion people across 40 regions. Its architecture moved from one shared Azure Cosmos DB container, to separate containers for critical workloads, and back to multitenancy by default with explicit data-movement capability.
The first shared design suffered logical-key collisions, noisy neighbors and “weak partitions,” where storage-driven splits reduced effective throughput while demand stayed stable. Separate containers improved isolation but made every service team responsible for capacity planning, which OpenAI says did not scale organizationally.
The revised shared system throttles the hottest keys first, uses quality-of-service tiers so bulk work yields before interactive traffic, and employs a serverless mode that can split partitions using request-unit demand as well as storage. OpenAI reports months without provisioning or noisy-neighbor incidents after the change; that is the company’s operational account, not an independent audit.
Habitat is more than 90% reads. OpenAI says three combined changes—near-user replication, in-region caching and a Rust rewrite—reduced global median read latency from roughly 120 milliseconds to 10 milliseconds. The number is median latency, not tail latency, and the improvement cannot be assigned to one change. An in-house exabyte-scale store remains in development. OpenAI engineering report.