OpenAI's storage platform now handles 22 million requests per second to serve 1 billion ChatGPT users, according to the company's engineering blog published September 11 S¹. The system, called Habitat, grew from a Python library into a globally distributed storage architecture. What it took to get there, and what breaks when a code library has to become planetary infrastructure, is the part that matters for anyone building systems that must survive their own success.

My read: This is the second storage scaling post from OpenAI in 2026. In January, they described stretching PostgreSQL to 800 million users P³. Nine months later, they are at a billion and talking about a different layer entirely. The jump from tuning a database to building a distributed storage platform tells me the Postgres approach hit a wall somewhere between 800 million and a billion users. The 22 million requests per second figure is striking but self-reported, and OpenAI does not clarify whether that is a sustained rate or a peak spike. I would treat both the user count and the request rate as directional until a third party verifies them.

From 800 million to a billion in nine months

In January 2026, OpenAI engineer Bohan Zhang wrote about scaling PostgreSQL, the open-source database that has powered ChatGPT and OpenAI's API for years, to serve 800 million users P³. That post described years of work tuning a familiar database to handle explosive growth. The September post tells a different story. Habitat is not a database. It is a storage platform, and it serves a billion users S¹. The 200-million-user jump in roughly nine months tracks with the trajectory OpenAI has been on all year, as free ChatGPT access expanded in August and enterprise use deepened, with companies like NVIDIA using ChatGPT Work to scale internal expertise. Even niche research applications, like ChatGPT searching extinct genomes for new antibiotics, add to the storage load that Habitat has to absorb.

ChatGPT users served by OpenAI storage infrastructure

The shift from PostgreSQL tuning to building a distributed storage platform suggests the bottleneck moved. When you have 800 million users, your problem is making a database go faster. When you have a billion, your problem is that no single database architecture can be in enough places at once.

What Habitat actually is

OpenAI says Habitat started as a Python library and evolved into a globally distributed storage platform S¹. The company has not published the source code or detailed architecture diagrams, so the description stays at the level of a storage layer that spans the globe and serves every ChatGPT user.

The name carries some baggage. Facebook Research has maintained a project called habitat-lab on GitHub since 2019, a library for training embodied AI agents in simulated environments P⁶. OpenAI's Habitat is a different thing entirely, serving file and object storage for ChatGPT. The name collision is a reminder that Habitat is a popular word in AI engineering circles, not a unique brand.

What we can verify: OpenAI published the blog post on their official domain on September 11, 2026 S¹. The post describes a platform that handles 22 million requests per second across a global footprint. The evolution from library to platform implies the team outgrew the original design and rebuilt around it, rather than starting from zero.

Why 22 million requests per second is a number worth pausing on

22 million requests per second is a rate that very few systems on earth have to sustain. For context, Hugging Face, which hosts hundreds of thousands of AI models and datasets, spent much of 2025 migrating its Hub storage from Git LFS to a new backend called Xet, and described that migration as a major infrastructure undertaking P⁵. That was for a platform serving model downloads, not real-time chat inference at planetary scale.

The OpenAI figure is self-reported and comes with caveats. The blog post does not specify whether 22 million requests per second is a sustained average, a daily peak, or a burst capacity number. The 1 billion users figure also lacks a clear definition: it could mean total registered accounts, monthly active users, or some other measure. OpenAI has not clarified either point, and no third party has audited the claims.

What to do about it

If you run a team building applications on top of ChatGPT or the OpenAI API, the storage layer is not your problem to fix. But the pattern OpenAI describes, a library that grows into a platform because the original design cannot handle the load, is one that repeats at every scale.

Consider a startup that builds a document-analysis tool on top of the GPT API. At 1,000 users, storing conversation history and uploaded files in a single database instance works fine. At 100,000 users, that database becomes the bottleneck. The team faces the same fork in the road that OpenAI describes: keep tuning the database, or build a storage layer that distributes across regions. The OpenAI post is a useful data point for engineering leaders thinking about when to make that jump, because it shows that even a company with effectively unlimited engineering resources chose to build a new layer rather than keep stretching the old one.

One practical thing to check this week: if your application stores user session data or conversation history, look at your read-to-write ratio and your geographic distribution. A system where 80% of traffic comes from one region can survive on a single well-tuned database far longer than one where traffic is spread across three continents. The storage architecture question only becomes urgent when geography, not query volume, is the constraint.

What we don't know yet

The biggest gaps are definitional. OpenAI has not said what 1 billion users means in terms of active versus registered, daily versus monthly. The 22 million requests per second figure has no context for whether it is sustained or peak. The blog post is titled part one, which suggests more technical detail is coming, but no date for part two has been announced.

The architecture itself remains opaque. We know Habitat is globally distributed and evolved from a Python library, but the post does not describe the underlying storage engine, the replication strategy, or how it integrates with the PostgreSQL systems described in the January post P³. It is possible that Habitat sits on top of PostgreSQL, or that it replaces parts of it, or that the two systems handle different workloads. OpenAI does not say.

The next signal: OpenAI's part two of this storage series, which the blog post title promises. We will check it against the claims here, particularly whether it clarifies the sustained-versus-peak question and the user-count definition.

Sources: S1 — Rapidly scaling online storage to serve over 1 billion ChatGPT users · S2 — Rapidly scaling online storage to serve over 1 billion ChatGPT users - · P3 — Scaling PostgreSQL to power 800 million ChatGPT users | OpenAI · P4 — openai/GABRIEL · P5 — Migrating the Hub from Git LFS to Xet · P6 — facebookresearch/habitat-lab

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.