OpenAI rewrote ChatGPT's storage platform from Python to Rust to keep up with 1 billion weekly users

OpenAI has published a detailed account of how it scaled Habitat, the internal storage platform behind ChatGPT, from a simple client library into a distributed system now handling tens of millions of requests per second — and why it ultimately rewrote the service's core in Rust.

What's new

According to OpenAI, "Habitat now handles more than 70 million requests every second, supporting products used by over 1 billion people each week," and the platform now serves "more than 500 petabytes of data" across roughly 40 geographic regions. Habitat first launched to support GPTs at DevDay 2023 as a straightforward Python library connected to a single database; growth since then forced it into a fully distributed system.

The most striking detail is the migration itself: in Q2 2026, OpenAI rewrote Habitat's core service from Python to Rust using "just 2 engineers, Codex, and GPT‑5.5." The payoff was substantial — OpenAI reports "the Rust service is 6x more CPU efficient and 15x more memory efficient than the Python version, with significantly lower average and tail latencies," and the Rust implementation now handles 95% of production requests.

This is the first of a planned two-part series on OpenAI's approach to scaling storage infrastructure.

Context

As ChatGPT's weekly user base has climbed past a billion, OpenAI has increasingly published engineering deep-dives on the infrastructure straining to keep pace — mirroring similar scale-focused posts from other hyperscalers. Small-team, AI-assisted rewrites of core infrastructure are becoming a recurring theme in OpenAI's own engineering narrative: the company is explicitly using its own coding models (Codex, GPT-5.5) to execute the kind of systems rewrite that would traditionally require a much larger engineering team.

Why it matters

The efficiency numbers are notable on their own — a 6x CPU and 15x memory improvement is a meaningful cost reduction at hyperscale. But the more significant signal is methodological: OpenAI is holding up a 2-person, AI-assisted rewrite of business-critical storage infrastructure as a proof point for how far agentic coding tools have come. That's a claim other engineering organizations evaluating AI-assisted infrastructure work will want independent verification of, but it's also a preview of the kind of case study OpenAI is likely to keep making as it sells Codex and its API models for serious internal engineering work, not just prototyping.