OpenAI Is Building Exabyte Storage; Its Reads Got 12x Faster

OpenAI said on 23 October 2026 that it has been building a custom in-house exabyte-scale storage system for the data behind its products, in an engineering post on Habitat, the storage platform that serves ChatGPT. The same post says the platform’s global median read time fell from roughly 120 milliseconds to roughly 10 milliseconds this year, which OpenAI calls “a 12x improvement”.

Habitat is the system OpenAI products use to fetch data, for example when someone logs in, checks Codex settings or starts a new ChatGPT conversation, according to OpenAI. OpenAI says it has to respond in milliseconds “to over 1 billion users across 40 geographic regions” and serves “over 500 petabytes of data”.

Business Pill · PERCENTILES

A short Business Engineer explainer of percentiles, such as the p50 median. It teaches the general idea only and says nothing about OpenAI or Habitat.

The Numbers OpenAI Published

Source: OpenAI engineering posts, ‘Rapidly scaling online storage to serve over 1 billion ChatGPT users’, Part I (11 September 2026) and Part II (23 October 2026). The figures are OpenAI’s own and were not independently verified.

The System OpenAI Is Building

In order to scale Habitat and conversation storage for the next billion users, OpenAI wrote, it needs to address the difficulty of replicating its data across its infrastructure on an even larger scale. “To preempt this, we’ve been building a custom in-house exabyte-scale online storage system, optimized for bulky Habitat data types,” the post says.

OpenAI says the in-development system combines the RocksDB Cloud core from Rockset with Raft-based replication. It implements the same document model as Azure Cosmos DB but uses disaggregated storage and physical replication to reduce the amount of flash capacity needed, “in exchange for a lower peak qps,” according to the post.

An exabyte is 1,000 petabytes. The post does not give a launch date, a size target or a cost for the new system.

How OpenAI Runs Azure Cosmos DB Today

Part II describes how OpenAI scaled the database underneath Habitat, Microsoft’s Azure Cosmos DB. OpenAI says it went through a circular journey: from one shared container, to single-tenant containers, and back to a shared container by default, after working closely with the Azure team on fixes.

• Scale. “Today, our largest Azure Cosmos DB account stores 10s of petabytes of data and serves more than 20M QPS,” OpenAI wrote. “To our knowledge, we are the only company which has deployed Azure Cosmos DB at this scale.”
• A custom mode. OpenAI says the Azure Cosmos DB team built it a new serverless mode that lets every physical partition scale up to 10K RU/s, against 5K RU/s in the earlier serverless mode.
• An outage. In late 2024, a power outage struck the primary Azure region for one of OpenAI’s largest accounts, and OpenAI realised its largest accounts would take several hours to fail over to a secondary region, according to the post.
• Co-management. The post says that when companies grow to OpenAI’s scale, managed services become less “managed”. It describes shared tests, shared dashboards and direct engineering channels with Azure.

How Reads Got Faster

Reads make up over 90% of Habitat’s requests, OpenAI says. Its changes included storing data near users, with some accounts replicating to all 40 Azure regions where it serves traffic, adding caching, and rewriting the Habitat service in Rust. “Together, these changes have reduced our overall global read p50 from roughly 120 ms to roughly 10 ms,” OpenAI wrote. The p50 is the median: half of reads are faster than that figure.

Part I, published on 11 September 2026, says that in Q2 2026, “with just 2 engineers, Codex, and GPT‑5.5,” OpenAI rewrote the entire Habitat service in Rust. The post says the Rust service handles 95% of production requests and is 6x more CPU efficient and 15x more memory efficient than the Python version.

Connecting the Coverage

Our GPT-6 piece reported OpenAI’s figure of more than 1.2 billion people using ChatGPT each week on 7 October 2026; the storage posts use “over 1 billion”. Our Rust rewrite piece covered an earlier case of an AI agent porting a codebase to Rust, and our token spend piece covered OpenAI’s Peter Steinberger putting his token spend for last month near $10 million. What’s new: OpenAI has published its storage scale and says it is building its own exabyte-scale system alongside Azure Cosmos DB.

The Structural Read

As we read it, the cost lever OpenAI names is flash: the post says the in-development system uses disaggregated storage and physical replication to reduce the amount of flash capacity needed, in exchange for a lower peak qps.

As we read it, the 12x figure describes the typical request: OpenAI reports the global read p50, the median, and says reads are over 90% of Habitat’s requests.

As we read it, the Rust rewrite is also an AI-coding data point: OpenAI says 2 engineers with Codex and GPT-5.5 rewrote the service in Q2 2026, and that the Rust version is 6x more CPU efficient.

Three Implications

FOR MICROSOFT OpenAI describes joint engineering with the Azure Cosmos DB team, including a new serverless mode built for it, and says its in-house system implements the same document model as Azure Cosmos DB.

FOR OPENAI The post ties the in-house system to scaling Habitat and conversation storage for the next billion users.

FOR OTHER AI COMPANIES OpenAI wrote that at its scale managed services become less “managed”, with workloads becoming outliers in the cloud provider’s infrastructure.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of The Inference Economy.

The framework says “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query”.

As we read it, storage lookups are part of that per-query work: OpenAI says Habitat handles more than 70 million requests every second. The framework was published in March 2026 and does not discuss Habitat or these posts.

What Is Not Established

All figures come from OpenAI’s own posts and were not independently verified. The posts do not give costs, a timeline for the in-house system, or how much of Habitat’s data would move to it. They do not say whether OpenAI’s use of Azure Cosmos DB will change. We did not contact OpenAI or Microsoft.

The Bottom Line

OpenAI says Habitat serves over 500 petabytes to over 1 billion people a week, that its median read time fell about 12x this year, and that it has been building a custom exabyte-scale storage system for the next billion users.