OpenAI Says 2 Engineers Rewrote Its Storage Layer in Rust

OpenAI said two engineers rewrote Habitat, the storage system behind ChatGPT and Codex, from Python to Rust during the second quarter of 2026, using the company's own Codex and GPT-5.5 tools, according to a blog post published this month, the first of a two-part series.

Habitat now handles more than 70 million requests every second, supporting products used by over 1 billion people each week across almost 40 geographic regions, OpenAI said in the post. The system serves more than 500 petabytes of data, up from a single database when it first launched to support GPTs at DevDay 2023.

Why Python became the bottleneck

Python's asyncio model "helps Python execute I/O-bound workloads concurrently, but does not help work around the Python GIL," OpenAI wrote, referring to the Global Interpreter Lock, a Python mechanism that stops more than one thread from running Python code at the same time. OpenAI said "asyncio scheduling delay can easily dominate tail request latency" as traffic grew.

A second problem compounded the first. "Python's aiohttp TCPConnector defaults to LIFO connection reuse: the most recently returned connection is selected," OpenAI wrote, which meant "LIFO encourages more work to concentrate on those same slower servers" once any backend server fell behind, risking what OpenAI called a "thundering herd" against downstream dependencies.

What changed when the rewrite shipped

"In Q2 2026, with just 2 engineers, Codex, and GPT-5.5, we were able to rewrite the entire service in Rust," OpenAI wrote. The company said the Rust service is 6 times more CPU efficient and 15 times more memory efficient than the Python version it replaced.

1. CPU efficiency6 x more efficient
2. Memory efficiency15 x more efficient

Source: OpenAI, "Rapidly scaling online storage to serve over 1 billion ChatGPT users," accessed Sept. 15, 2026

"This new Rust service is now handling 95% of our production requests; we'll be deprecating Python entirely in the coming weeks," OpenAI wrote.

Three years of engineering ahead of one quarter of rewriting

OpenAI's post also states that Habitat's traffic "grown more than 10x year-over-year for the last three years," and that the company deliberately kept the slower Python version running during that growth. "By accepting the performance trade-offs of a Python service in the short term, we were able to prioritize more immediate challenges, establish our core APIs, and build out" the surrounding infrastructure, OpenAI wrote. That three-year process, not just the quarter the rewrite took, is what gave a 2-person team a stable set of APIs and a fully mapped request path to translate.

InfoQ China reporter Chu Xingjuan, covering the post the same week, reproduced OpenAI's figures without independently verifying them and did not name the two engineers either.

What OpenAI has not published yet

OpenAI did not name the two engineers or disclose a request-latency figure alongside the CPU and memory multiples. The company said a second post will cover "how we made multi-tenancy reliability at scale, our layered strategy for optimizing read performance, and how we scaled our partnership with Azure Cosmos DB to reliably handle unprecedented demand," the first time OpenAI has named Microsoft's Cosmos DB as a Habitat dependency, without saying when that post will run. The efficiency multiples rest on OpenAI's own before-and-after measurement, the same self-reported pattern behind the request-caching gains AWS reported for SageMaker HyperPod this month, and the same pattern behind the per-minute latency claims in OpenAI's own GPT-Live-1 API pricing post.