OpenAI just pulled back the curtain on one of the least glamorous but most critical parts of running ChatGPT: where all the data lives. According to OpenAI, the company rebuilt an internal system called Habitat, taking it from a humble Python library into a globally distributed storage platform that now serves over 1 billion ChatGPT users and handles 22 million requests per second. That’s the kind of number that stops you mid-scroll.
What stands out here is the trajectory. OpenAI reports that Habitat didn’t start as some grand infrastructure project. It began as a Python library, the sort of tool an engineer writes to solve a specific problem, then grew into the backbone holding up one of the most-used products on the planet.
What OpenAI actually built
Storage sounds boring until it breaks. Every chat you send, every memory ChatGPT keeps, every file you upload has to be written somewhere, found again fast, and served back without a hiccup. Do that for a handful of users and almost any database works. Do it for a billion, at 22 million requests per second, and the ordinary tools fall apart.
OpenAI’s account frames Habitat as the answer to that scaling wall. The team evolved it from a single library into a platform spread across the globe, close to users wherever they are. The goal is simple to say and brutal to execute: keep latency low, keep data consistent, and never fall over when traffic spikes.
Why the numbers matter
Let’s put 22 million requests per second in context. Most large consumer apps measure their peak load in the tens or hundreds of thousands of requests per second. OpenAI is operating two orders of magnitude beyond that on the storage layer alone.
A few things worth pulling out:
- 1 billion users puts ChatGPT in the same league as the largest platforms ever built, and the storage system has to match that reach.
- 22M requests/second is a sustained firehose, not a one-time benchmark. Systems at this scale live or die on their worst moment, not their average.
- Python library to global platform shows a pattern OpenAI keeps repeating: start small, ship, then re-architect once the problem is real.
What practitioners can take from this
If you build software, there’s a practical lesson buried in OpenAI’s story, and it isn’t “go build a global storage platform.” It’s the sequence. OpenAI didn’t over-engineer on day one. They shipped a library, learned what they actually needed, and rebuilt only when scale forced their hand.
That’s the opposite of how a lot of teams operate. Engineers love designing for a billion users before they have a thousand. OpenAI’s path suggests the smarter move is to solve today’s problem cleanly, then let real demand tell you what to build next.
The other takeaway is that infrastructure is now a competitive moat for AI companies. Model quality gets the headlines. But if your storage can’t keep up, your product goes down, and users don’t care why. OpenAI treating this system as worthy of a public write-up signals how central reliability has become to staying ahead.
The bigger picture
This is a window into what running frontier AI at consumer scale actually costs in engineering effort. The models get the attention. The plumbing that keeps them online for a billion people rarely does. OpenAI publishing the Habitat story is a reminder that the race isn’t only about smarter models. It’s about who can serve them, reliably, to everyone at once.
Expect more of these infrastructure stories as AI products keep growing. Full technical details are available in OpenAI’s original write-up for those who want to go deeper.