Inside the Storage Layer Serving More Than One Billion ChatGPT Users
OpenAI described how Habitat evolved from a Python library into a global storage service. Its most useful lesson is the order in which a platform should absorb complexity.
OpenAI published an engineering account of Habitat, the online storage platform supporting its products, on September 11, 2026. The company says the platform serves more than one billion weekly users, handles over 70 million requests per second and manages more than 500 petabytes across almost 40 regions.
These figures are reported by OpenAI. The more transferable part of the story is how the architecture changed as product demand grew.
From library to platform
Habitat began as a Python client library connected to a single database. Adoption spread because it solved an immediate product need without requiring every team to operate a new service. As scale and product diversity increased, the storage access layer became a service responsible for routing, caching, access policies, placement, isolation and request shaping.
OpenAI delayed rewriting the Python service while more urgent reliability work took priority. In the second quarter of 2026, the company says two engineers working with Codex and GPT-5.5 rewrote it in Rust. The Rust version was handling 95% of production traffic when the article was published, with claimed gains of six times CPU efficiency and 15 times memory efficiency.
Why it matters
The story argues against premature platform perfection. The original library created adoption and exposed the real workload. Centralization came when shared reliability and policy became more valuable than local freedom. The language rewrite came later, when resource efficiency had become a material constraint.
Practical lessons
- Introduce a platform through a narrow interface that product teams want to use.
- Centralize routing, policy and isolation before every team invents its own version.
- Measure tail latency, connection pressure and downstream overload—not only averages.
- Delay expensive rewrites until profiling shows where capacity is being lost.
- Separate vendor-reported scale figures from architecture principles that can be tested locally.
BronzeDev view
Most businesses do not need frontier-scale storage. They do face the same sequencing problem at smaller scale: standardize access, add visibility, then optimize the proven bottleneck. A modest integration layer with ownership and telemetry is often more valuable than choosing a more fashionable database or language too early.
Primary source
OpenAI Engineering: Rapidly scaling online storage to serve over one billion ChatGPT users
