CONFIRMED: scale has become part of the product

OpenAI published an engineering account on September 11 about Habitat, a system that grew from a Python library into a globally distributed storage platform. The company says it serves more than 1 billion ChatGPT users and sustains 22 million requests per second. This is not a new model or a feature released in the interface. It is a look at how the infrastructure had to change as the product became widely used.

From library to distributed platform

The central story is the transformation of a software component into an operational layer at massive scale. Habitat had to move beyond the scope of a library used by applications and take on global distribution, availability, and heavy traffic. The official publication does not turn the system into an open product, disclose pricing, or promise that outside companies can adopt it. Its value is in the architecture and engineering decisions OpenAI chose to make public.

The number is impressive, but it does not tell the whole story

Twenty two million requests per second is an aggregate capacity measure. It does not mean that every user receives that volume or that every operation has the same latency. OpenAI also does not provide enough detail in the public summary about the read and write mix, replication, consistency, costs, or failure rates. That matters because real storage systems are judged by their full operating profile, stability, and recovery, not by peak traffic alone.

Independent context supplies the necessary brake

On the same day, PlanetScale published a test of its distributed Postgres database Neki, reaching 118.5 million queries per second with 512 shards and 1.22 pebibytes of data. The company explicitly notes that the workload used simple key based reads, with no writes, cross shard queries, replicas, or failover. The comparison does not decide a winner. It shows why scale numbers only make sense when the workload and test conditions are visible.

Why this matters to businesses

For a small or midsize company, the lesson is not to copy OpenAI’s architecture or start buying infrastructure. It is to recognize that agents, automated support, CRM, and content products depend on reliable storage for history, files, permissions, and execution state. Before scaling, teams should separate hot and cold data, define retention limits, watch cost per operation, and test recovery. Most businesses do not need millions of requests per second, but they do need to stop growth from turning a database, queue, or storage layer into a silent bottleneck.

The MaxAssistant reading

This release matters because it moves the AI conversation from the visible layer, the model, to the layer that supports the entire experience. ChatGPT may look like a conversational product, but its scale requires a global data platform with hard choices around distribution, latency, and reliability. My verdict: Habitat is not a tool the market can try tomorrow, and the figures still need technical context. Even so, the publication confirms an important direction. The competitive advantage of AI labs increasingly lies in the complete system, not only in the smartest model.

Sources

OpenAI, Rapidly scaling online storage to serve over 1 billion ChatGPT users, September 11, 2026: https://openai.com/index/scaling-storage-one-billion-users-part-one | PlanetScale, 118 million queries per second on Neki, September 11, 2026: https://planetscale.com/blog/118-million-queries-per-second-on-neki