OpenAI scales Habitat storage platform to serve 1 billion ChatGPT users

Per OpenAI's engineering blog, Habitat began life as an internal Python library and has grown into a globally distributed storage platform underpinning ChatGPT. The headline numbers are staggering: 1 billion users served and 22 million requests per second, figures that quantify just how large ChatGPT's footprint has become and the infrastructure required to keep it responsive worldwide.

The post is framed as 'part one,' suggesting a series that will detail the architectural choices — data distribution, replication, consistency and latency management — needed to operate at this scale. For a company that just had to pause its $200/month Pro plan because GPT-6 Astra overwhelmed capacity (a separate story), publishing storage-scaling internals doubles as a signal that the constraint is compute for inference, not the storage layer for user data.

The engineering community reads posts like this the way it read Google's early Bigtable and Spanner papers: as a window into how a hyperscaler solves problems most companies never face. The '22M requests/sec' figure in particular contextualizes ChatGPT's reach relative to other internet-scale services and underscores OpenAI's transition from research lab to full-blown infrastructure operator.

Strategically, owning and articulating this stack matters for reliability commitments to enterprise customers (the financial-services vertical, government contracts) who need assurances that ChatGPT can scale without data-integrity or availability failures. It also positions OpenAI's infra know-how as a differentiator against rivals who lean on cloud providers' managed storage. Caveats: a company blog is self-reported, light on the hardest tradeoffs, and offers no independent benchmarks. Watch the follow-up parts for concrete architecture and whether OpenAI open-sources or productizes any of Habitat.