OpenAI Rewrites Habitat Storage in Rust With Codex, GPT-5.5

Two engineers rewrote OpenAI's main online storage service from Python to Rust in a single quarter, using Codex and GPT-5.5 to do much of the work.

That is the central disclosure in Part II of OpenAI's engineering post about Habitat, the storage platform behind ChatGPT, Codex, and everything else the company runs. Habitat now holds more than 500 petabytes, operates in almost 40 regions, and serves over 1 billion weekly ChatGPT users at more than 70 million requests per second. The Python version it replaced had already pushed past 20 million requests per second at peak before the rewrite, which happened in Q2 2026 and now carries 95% of production traffic.

The efficiency numbers in the post: the Rust service is six times more CPU-efficient and fifteen times more memory-efficient than the Python predecessor. The underlying Azure Cosmos DB layer, reworked through what OpenAI describes as a close working partnership with the Azure team, is delivering roughly twelve times faster reads for ChatGPT.

Authors Jon Lee, Hrachya Hakobyan, and Chaomin Yu are also candid about what they tried to walk back. OpenAI had split its single default container into separate containers with independent physical partitions for critical workloads, specifically to stop what the post calls "Noisy neighbors: Separate keys could still sit on the same physical partition" and interfere with core login and signup flows. The isolation worked, but it forced every product team to forecast its own capacity. Some, the authors note, "launched features without increasing their containers' capacity beforehand." The company is now moving back toward shared, multi-tenant containers with a data-movement capability and the co-managed container design developed with Azure.

The storage write-up is the second OpenAI infra disclosure our tracker logged this week; the first was "OpenAI Cuts Internet From Frontier Training After Sandbox Escape" on October 23, and this is the 401st OpenAI item we have cataloged in the last 90 days.

What is new in this post is OpenAI being explicit that its own coding agents did substantial work on an infrastructure service already running 20 million requests per second. "Building and operating infrastructure at this scale is not particularly difficult in itself, but with growth exceeding 10x every year," the authors write, teams eventually ran out of room to keep deferring the language migration. The move from a Python library launched at DevDay in November 2023 to a Rust standalone service took two people and one quarter.