OpenAI's Habitat grew from a Python library to 22M req/sec

The storage layer under a billion ChatGPT users started life as a Python library. OpenAI's engineering blog now describes that system, Habitat, as a globally distributed storage platform serving 1 billion ChatGPT users and handling 22 million requests per second.

At a glance

• Habitat is OpenAI's in-house storage platform, and the company's write-up traces it from a Python library to a globally distributed system standing behind ChatGPT's entire user base.
• The two headline figures are 1 billion ChatGPT users and 22 million requests per second, the second reported only by OpenAI itself and not corroborated anywhere else.
• OpenAI's summary gives a user count and a throughput figure but no latency, hardware or cost numbers, and the write-up itself is filed as part one.

If you missed the earlier steps: ChatGPT passed 1 million users within five days of its November 2022 launch, reached about 30 million weekly users by January 2023 and 100 million by November 2023, according to Exploding Topics. By December 2024 it was at 300 million weekly users, and the line kept climbing from there.

ChatGPT went from 400 million weekly users to 900 million in twelve months

Weekly active users went from 300 million in December 2024 to 400 million in February 2025, then 700–800 million in October 2025 and 900 million in February 2026, a gain of roughly 500 million since February 2025. In June 2026 ChatGPT crossed 1 billion monthly active users.

OpenAI's own disclosures track the same climb. Per Exploding Topics, a September 2025 OpenAI report cited 700 million weekly active users, and COO Brad Lightcap gave 400 million in February 2025; per DemandSage, a February 2026 report from OpenAI put the figure at 900 million.

Habitat serves 22 million requests per second

OpenAI describes Habitat as a globally distributed storage platform that serves 1 billion ChatGPT users and handles 22 million requests per second. The Python-library beginning is the part worth holding onto: the first version was code that other services imported and ran in their own process, not a platform they called across the network.

The demand behind those requests is documented outside the post. ChatGPT users send about 2.5 billion prompts a day. According to DemandSage, citing Sensor Tower estimates reported by Reuters, ChatGPT was the first app to reach 1 billion active users within three years of launch, outpacing the early growth of Google Maps, TikTok, Instagram and YouTube.

Per Business of Apps, it was also the most downloaded app of 2025 with 770 million installs, and it generated $8 billion for OpenAI in 2025, up 128 percent on the previous year.

What does spreading storage across machines actually buy?

A distributed key-value store keeps data on many servers rather than one. Algomaster's explainer on the pattern makes the point plainly: that spread is what lets a system scale horizontally, absorbing large volumes of data and user requests while keeping latency low and reliability high.

Think of a coat check. One counter with one attendant works until the queue reaches the door; many counters work only if the ticket number tells you which counter to walk to, and only if a spare attendant can take over when one of them steps away.

The ticket rule is partitioning and consistent hashing; the spare attendant is replication and leader election. Per Algomaster, the standing questions for a store like this are how to persist data through node crashes, how to spread it evenly, how to keep it consistent across nodes, and how to detect and recover from failures.

The 22 million figure comes from OpenAI alone, and the summary attaches no conditions to it: no time window, no read-write split, no latency alongside it. OpenAI also does not say what Habitat stores or which consistency model it settled on, which is exactly where the trade-offs in a storage rewrite live. In our view the Python-library origin is the more interesting claim of the two, and it is the one the summary spends the fewest words on.

When part two arrives OpenAI has not given a date for the second installment or said which parts of Habitat it covers. The open questions are concrete: the partitioning scheme, how replication and failure recovery work, and what moving off the original library cost in migration work. The other checkpoint is OpenAI's own user reporting, which has served as the public yardstick for this system's load since the 400 million weekly figure in February 2025.