OpenAI Discloses Habitat Storage Architecture: The Scale and Limits Behind Over 1 Billion Weekly Users

On September 11, 2026, OpenAI published a post titled “Rapidly scaling online storage to serve over 1 billion ChatGPT users” on its official engineering blog, describing how its internal storage platform, Habitat, evolved over the years from a Python client-side library into an independently operating service. The post is filed under the Engineering category and credited to three named technical staff, Jon Lee, Chaomin Yu, and Ben Ries; it closes with a link to an engineering job opening. It is a technical blog post, not an announcement of a new product or plan.

This article was checked on September 18, 2026 against the original OpenAI post, OpenAI's official news RSS feed, and the Engineering section of OpenAI's official sitemap. For several days beforehand, the post's page on openai.com would sometimes return only an access-restricted page; by the day of this check it could be read directly. Every figure below is OpenAI's own statement; this site has not run any tests of its own and offers no advice on purchasing or using anything. This article uses no invented everyday scenario.

What This Post Is: An Engineering Blog Post, Not a Product Announcement

OpenAI files this post under its Engineering category, describing Habitat as the online storage platform it built in-house so its products can retrieve the data they need quickly and reliably: a user logging in, checking their Codex settings, or opening a new ChatGPT conversation can all involve multiple data queries behind the scenes. OpenAI writes that when those queries are slow, the product feels slow; when they fail, the product simply stops working.

The stat block at the top of the post lists three numbers side by side: over 70 million requests per second, over 1 billion people a week, and over 500 PB of data. OpenAI explains that Habitat first launched at DevDay 2023 to support GPTs, starting as a simple Python client-side library connected to a single database, and has since grown into a complex distributed system.

The post closes with a link to an engineering job opening (the URL carries the name of Seattle); this is the first of two posts on the storage topic, and the second is expected to cover multi-tenancy reliability, a layered strategy for read performance, and how the partnership with Azure Cosmos DB is being scaled up. As checked against the official sitemap and news feed through September 18, 2026, no entry for the second part has appeared yet, and no date has been announced.

The Scale OpenAI Describes: A Few Numbers Worth Reading Together

OpenAI writes in the body of the post that Habitat now handles over 70 million requests every second, supports products used by over 1 billion people a week, and spans nearly 40 regions; the post does not list the countries or cities behind those regions. The closest thing to regional detail appears on a diagram the post itself labels “simplified,” which shows three example database codes, us0, us1, and eu0 — labels on a diagram, not a published list of regions.

Worth noting: OpenAI's official RSS feed and the page's own metadata separately state “22M requests per second.” Rereading the full body text on the day of this check, this article found that figure only in the summary and metadata; it never appears again in the body, and the post never explains how it relates to the 70 million or 20 million figures.

OpenAI also states that its scale has grown more than 10x year-over-year for the last three years; the post's own line notes that systems engineers typically design for 10x the current scale, hoping it holds for a few years while they prepare for the next 10x.

Why Users Can Feel a Single Slow Query

OpenAI explains that when the average user's request triggers hundreds of database calls behind the scenes, what the user feels is the slowest one. The post calls managing this tail latency — a small number of unusually slow queries — the main challenge of running a Python service at this scale.

The post gives one example: Habitat uses Statsig, a feature-flag tool, which by default polls for updated configuration once a minute with no jitter, and that configuration includes every production rule across every service. Layered on top of the architectural decision that a pod runs up to 8 Python processes, the result was that every minute, each pod would have some moment where all of its worker processes simultaneously stalled their in-flight requests to instead parse that giant configuration file. OpenAI says that once CPU profiling helped them find the root cause, the fix was simple: deploy narrower-scoped configuration, lengthen the polling interval, and add jitter to background work like this.

The other example involves connection pooling: OpenAI says Python's aiohttp library defaults to preferring the most recently released connection for reuse, which is normally a reasonable default. But during a burst of requests, slower, already-overloaded servers returned connections to the pool later and were therefore selected more often by subsequent requests, gradually concentrating more traffic on the pods already struggling. OpenAI calls this a kind of “metastable failure”: even after the client causing the overload stops, a subset of processes keeps degrading and only recovers after a restart. Switching to preferring the earliest-released connection not only broke the feedback loop but also reduced variance in request volume under steady state; this kind of connection management is now mostly handled by Istio and Envoy.

What OpenAI Does Not Explain in This Post

This article compared the original English wording term by term: the post does not contain the words durability, backup, or replica, and does not state how many copies of a piece of data are kept. OpenAI describes, just once, wanting to move its most critical datasets into an Azure Cosmos DB account distributed across regions, to narrow the blast radius of a single-region failure — and says nothing further about retention or redundancy mechanisms. The post also gives no availability figures or service-level agreement, describing only a few incidents its own systems have had — for example, one team rolling its own service back to an older version with a problematic client for unrelated reasons, which set off the very outage so much effort had gone into avoiding — without stating a date, duration, or the actual impact on users.

On regions, the post states only that it covers nearly 40 regions, without naming any country or city as a data location, and without mentioning Taiwan or Asia. The closest things to regional information are the three database codes on the diagram and the city name in the URL of the job link. The post likewise does not spell out what data Habitat actually stores — logging in, checking Codex settings, and opening a new conversation may all require querying data, but the post never states outright whether conversation content itself is stored in Habitat, saying only that this is where user data is protected, a key point guarding against unauthorized access by external, internal, and agent actors.

The numbers for the Rust rewrite are the same kind of claim: OpenAI says its own data shows the new version is 6x more CPU-efficient and 15x more memory-efficient, with both average and tail latency significantly reduced; it now handles 95% of production requests, and OpenAI plans to fully retire the Python version within the next few weeks. OpenAI also says it plans to publish another post sharing more learnings, and that follow-up likewise has no announced date. The post includes no testing methodology, test scenarios, or third-party verification, so for now these numbers can only be treated as internal measurements OpenAI has published itself.

How Readers Might Weigh This Post

This post announces no new product, plan, region, or setting, so ChatGPT or Codex users do not need to change anything because of it. OpenAI itself never claims in the post that users will feel anything faster or cheaper; for now, the benefits of the rewrite remain solely OpenAI's own internal account.

This post is best treated as a reminder: every service has its own storage and availability design behind it, and the scale figures an operator publishes about itself — whether request volume, user counts, or growth multiples — can only be taken as the vendor's own statement until there is third-party verification. Readers may want to notice whether a post like this addresses durability, service-level agreements, or incident details, which usually says more about a service's stability than any single eye-catching number.

Readers who want to check for updates can consult two free official channels themselves: OpenAI's news RSS feed and the Engineering section of the sitemap — this article used exactly those two channels to confirm that the second part has not appeared yet. Note that the time stamped on the sitemap does not match the publication date printed on the page itself, so it should not be treated as the post's publication or update date; that is an observation made on the day of this check.

2026 AI News Roundup: Highlights and Daily Life Applications from January to September

OpenAI and Broadcom Unveil the Jalapeño Chip: The Official Stage, Division of Labor, and What's Not Yet Announced

• AI tools
• Apps & software
• AI news