OpenAI details Habitat storage platform behind 1B+ weekly ChatGPT users

OpenAI published how Habitat, its online storage platform, serves more than 70 million requests per second and more than 500 petabytes for products used by over 1 billion people each week — figures that are OpenAI’s.

This is OpenAI’s own write-up of how it stores ChatGPT data at huge scale, including a Python-to-Rust rewrite. Treat every metric as OpenAI’s until someone else measures them.

OpenAI published “Rapidly scaling online storage to serve over 1 billion ChatGPT users” (Part I) by Jon Lee, Chaomin Yu, and Ben Ries, Members of Technical Staff. Read it as an engineering how-to from the people who run the store, not as an independent capacity audit.

OpenAI says Habitat handles more than 70 million requests every second, stores more than 500 petabytes (a petabyte is a million gigabytes), spans almost 40 geographic regions, and supports products used by over 1 billion people each week. Those four numbers are the company’s scale claim.

How it grew, on OpenAI’s page: Habitat began as a Python client-side library on Azure Cosmos DB — Microsoft’s hosted database — for custom GPTs at DevDay 2023. OpenAI says that by mid-2025 it had become a standalone Python service to cut cross-team rollout friction. Then in the second quarter of 2026 two engineers used Codex and GPT-5.5 to rewrite it in Rust, a systems language chosen for speed and memory control.

OpenAI says the Rust service now handles about 95% of production requests and is roughly 6× more CPU-efficient and 15× more memory-efficient than the Python version, with lower average and tail latencies — meaning typical waits and the worst waits both fell, on OpenAI’s numbers.

The API is a constrained NoSQL design (inspired by Meta’s TAO objects-and-edges model): simple keyed reads and writes, not ad-hoc SQL. Complex queries go to a Rockset secondary view via change-data-capture — copying each write into an analytics store — which OpenAI states as the tradeoff that keeps online request cost predictable. See Sources for the schema notes.