OpenAI Habitat storage now serves products used by more than 1 billion people each week while managing more than 500 petabytes of data. The company published the second part of its Habitat engineering series on October 23, 2026, expanding on a September 11 overview of how the platform evolved from a Python library into a global service.
How OpenAI Habitat storage reached global scale
Habitat provides storage access for ChatGPT, the OpenAI API, Codex and internal services. It handles routing, authorization, encryption, data residency, rate limiting and caching while hiding the underlying infrastructure from product teams. The first engineering note said the platform processes more than 70 million requests per second.
The storage layer originally relied on a shared Azure Cosmos DB container. That simplified access but exposed workloads to partition-key collisions, noisy neighbors and weakening per-partition capacity as physical partitions split. OpenAI then isolated critical workloads in separate containers, a step that improved fault boundaries but forced product teams to plan capacity and created operational overhead.
The newest design returns most workloads to shared containers while retaining tools to move data when isolation is necessary. OpenAI said Azure added controls that throttle the hottest keys first, prioritize interactive requests over bulk work and let each physical partition scale to 10,000 request units per second. The company reported no provisioning or noisy-neighbor incidents for several months after the change.
Azure co-engineering and lower latency
OpenAI said its largest Azure Cosmos DB account stores tens of petabytes and serves more than 20 million queries per second. At that scale, the company and Microsoft’s Azure team use shared tests, dashboards and direct engineering channels. The collaboration followed outages during a 2024 regional failover and the subsequent return to the original region, events the October article describes as historical lessons rather than new incidents.
Read traffic accounts for more than 90% of Habitat requests, so the 2026 performance work emphasizes data placement and caching. Smaller, latency-sensitive datasets can replicate broadly, while a distributed Valkey cache holds recently accessed objects near serving infrastructure. Cosmos DB change feeds propagate invalidations within seconds to keep cache entries current.
The design choices show why infrastructure has become a competitive constraint for AI companies. Fusion Market News has also covered the OpenAI Decisions API’s speed and cost targets, the Physical AI Growth Lab and Southeast Asia’s expanding digital economy. Each story points to the same underlying requirement: reliable compute, storage and networks must scale with product adoption.
What OpenAI is building next
OpenAI said broad replication becomes difficult when datasets reach hundreds of petabytes. Its in-development system will use disaggregated storage and physical replication to reduce flash-capacity requirements, accepting lower peak query throughput for bulkier Habitat data types. It will preserve the document model used by Cosmos DB to simplify migration.
The company did not provide a launch date or capital-cost estimate for the new system. The October update is therefore an engineering roadmap, not a product announcement or financial forecast. For investors following the AI infrastructure cycle, the disclosed scale illustrates both demand and the operating complexity behind consumer AI services.
Sharron Kendi
Sharron Kendi is a seasoned crypto market analyst and writer with over three years of experience decoding blockchain trends, price movements, and market dynamics. She holds a Bachelor’s Degree in Commerce (Finance) from Kenyatta University, blending a solid academic foundation with a sharp eye for technical analysis and a deep understanding of on-chain data. Her work delivers clear, data-driven insights that empower investors to navigate the fast-evolving digital asset space with confidence.
