How OpenAI Scaled Habitat to One Billion ChatGPT Users

OpenAI has rebuilt Habitat, its online storage layer, from a Python library into a globally distributed platform. The company says that platform now backs more than one billion ChatGPT users and absorbs 22 million requests per second.

Why a library stops working

A library ships inside every application that uses it. That is cheap at the start: no extra service to operate, no network hop, no separate team. It stops being cheap once many services depend on the same assumptions about where data lives and how failures behave, because changing any of those assumptions means changing every caller.

Turning the library into a platform moves that logic behind an interface. In the summary available to us, OpenAI does not spell out which components it chose, how many regions the platform spans, or what it replaced.

The number that matters is the load

Twenty-two million requests per second measures system load, not people typing. A single conversation can trigger several reads and writes depending on which features are in play, so the two figures OpenAI quotes describe different things and cannot be divided into one another.

In the material we could access, the company does not say whether 22 million is a peak or a sustained average, or over what window it was measured.

What we could not verify

The source URL marks this post as the first in a series. The page itself did not load when we tried to retrieve it, so the engineering detail behind the headline figures - storage engine, replication strategy, latency targets, failure behavior - remains unchecked on our side.

Every number here comes from OpenAI's own announcement. No outside party has confirmed it, and self-reported infrastructure figures carry no external audit.