Realtime AI News
OpenAI details Habitat storage: scaling to 1 billion ChatGPT users and 22M requests per second
On September 11, OpenAI published an engineering post explaining how its online storage system, Habitat, evolved from a Python library into a globally distributed storage platform. The post says the system now serves more than 1 billion ChatGPT users and handles 22 million requests per second.
On September 11, OpenAI published an engineering post titled Rapidly scaling online storage to serve over 1 billion ChatGPT users, describing how its online storage system, Habitat, evolved over time. The article is not about model capabilities; it focuses on the storage layer that keeps ChatGPT running.
According to the post, Habitat now serves more than 1 billion ChatGPT users and handles 22 million requests per second. For any online system, that is an extreme level of read and write pressure.
OpenAI says Habitat began as a Python library. As ChatGPT's user base expanded, it was turned into a globally distributed storage platform that handles online data access.
The storage layer is nearly invisible to users, but it directly shapes latency, reliability and cost. At tens of millions of requests per second, even a brief hiccup can turn into a large-scale experience problem.
Publishing these engineering details is itself a signal: in the competition to serve large language models, infrastructure capability now matters as much as model capability. The teams that get the unglamorous work of storage and scaling right are the ones that can hold up under real user volumes.
The post is labeled as part one, meaning OpenAI plans to share more about the architecture and scaling path behind the system. For people who follow large-model engineering, it offers a rare official look at how ChatGPT is actually operated.
What to watch next is whether the following installments add more detail on Habitat's design trade-offs, and whether that engineering work has to be rethought again as ChatGPT keeps growing.
Why it matters
The disclosure shows the battle to power ChatGPT is not confined to models; storage and other infrastructure are equally decisive, and the 1 billion users and 22 million requests per second figure give outsiders a rare official yardstick for the platform's real scale.
Nearby Updates
All09/11, 18:08
Zhipu's GLM-OCR Model Update Tops OmniDocBench V1.5 at 0.9B Parameters
Zhipu's zai-org updated the GLM-OCR model card on Hugging Face on September 11, presenting a 0.9B-parameter multimodal OCR model that scores 94.62 on OmniDocBench V1.5 and ranks first overall. Built on a GLM-V encoder-decoder with a PP-DocLayout-V3 layout pipeline, it ships under the MIT license with documented vLLM, SGLang, Ollama and Transformers deployment paths.
09/11, 16:59
Amap Releases ABot-Earth 0.7, a 3D-Native City World Model, as Its Store Ranking Goes Fully AI
On September 10, Chinese mapping platform Amap said its year-old store ranking is going fully AI, using algorithms to weigh different navigation and in-store visit behaviors. The upgrade rests on ABot-Earth 0.7, which the team calls the world's first 3D-native city world model and which can generate a 3D city with realistic streetscapes from one satellite image or a short prompt on a consumer GPU in about ten minutes.
09/11, 22:05
Anthropic Posted a $450K Sales Role Devoted Solely to Meta — Then Closed It in a Week
Anthropic briefly listed a job titled Mega Account Executive, Meta, a role serving only Meta, with on-target earnings of $380,000 to $450,000, or roughly 3.2 million yuan at current rates. The posting went live on September 3 and stopped accepting applications on September 9, as Meta engineers already lean heavily on Claude Code.
09/11, 13:53
Anthropic says it stopped AI misuse for cyberattacks and bioweapons
Anthropic says it has stopped its AI from being used for cyberattacks and bioweapons-related misuse, according to a Euronews report. The disclosure puts frontier-model abuse monitoring back in the spotlight.