OpenAI details Habitat storage upgrades for ChatGPT scaling

News

OpenAI says Habitat improvements including multi-tenancy, global caching and Azure Cosmos DB delivered 12x faster reads for ChatGPT.

OpenAI said it has further scaled Habitat, its internal storage platform for ChatGPT and other services, by adding reliable multi-tenancy, global caching and Azure Cosmos DB support to improve read performance as usage grows. In a post published on Oct. 23, 2026, the company said the changes delivered 12x faster reads for ChatGPT.

Evidence and context

The post is the second part of OpenAI’s account of how it is scaling online storage for what it describes as more than 1 billion ChatGPT users. Habitat supports ChatGPT, the API, Codex and internal services, and OpenAI has described it as a central platform for routing, schema lookup, authorization, encryption, caching, request shaping, connection pooling and data-residency placement. The company says its storage model is intentionally constrained: it represents objects and edges, colocates each object with its outgoing edges to support partitioning, and excludes arbitrary SQL and general graph traversals.

OpenAI’s earlier Sept. 11, 2026 disclosure said Habitat was handling over 70 million requests per second, storing more than 500 petabytes of data and operating across almost 40 regions. Those figures were presented as publication-date metrics. The company also said it rewrote the service in Rust during Q2 2026; on Sept. 11, Rust handled 95% of production requests, with reported gains of 6x CPU efficiency and 15x memory efficiency. OpenAI has not said in the material provided whether the planned Python retirement has been completed.

The latest post adds that change data capture feeds isolated Rockset instances for more complex queries, and that Habitat’s responsibilities now include provisioning, developer experience, privacy enforcement and first-class change data capture. The material provided does not include independently verified live measurements for the new read-performance figure, nor does it provide a date for any next catalyst.