OpenAI's Storage Layer Now Handles 70 Million Requests a Second

OpenAI's storage system, called Habitat, now handles more than 70 million requests a second, the company said Sept. 11. That is up from a peak of 20 million a second on the Python service it retired in 2026.

Habitat launched at OpenAI's DevDay in 2023 as, in the company's own words, "a simple Python client-side library connected to a single database" supporting GPTs. It became a standalone service in mid-2025 to centralize security enforcement across dozens of internal products, according to the same post. It now spans nearly 40 regions and stores more than 500 petabytes of data, supporting products OpenAI says are used by over 1 billion people each week.

The rewrite that took two engineers one quarter

In the second quarter of 2026, two OpenAI engineers rewrote Habitat's service layer in Rust, using the company's own Codex and GPT-5.5 as coding tools, according to OpenAI. The Rust version is six times more CPU-efficient and 15 times more memory-efficient than the Python service, with lower average and tail latencies, the company said. As of the post, the Rust service handles 95 percent of production requests, with full Python deprecation planned within weeks. OpenAI has not published an independent audit of those efficiency figures; they are the company's own comparison of its old system to its new one.

Metric	Python service	Rust service
Peak requests per second	20 million	70 million-plus
CPU efficiency	baseline	6x
Memory efficiency	baseline	15x

The layer underneath Habitat has the same problem

Habitat is not OpenAI's only storage system straining under growth, and it is not even the first one the company has written about publicly. An earlier OpenAI post described scaling a single-writer PostgreSQL cluster to power roughly 800 million users, offloading reads to about 50 replicas while routing new, shardable workloads to Azure Cosmos DB instead of the aging Postgres primary. Reviewing that architecture, engineer Matthew Penaroza wrote that OpenAI's approach reports "only one SEV-0 PostgreSQL incident" a year, which he called a sign of strain rather than an achievement, adding that continuous growth in queries and connections "almost certainly cost the company millions of dollars in completely unnecessary infrastructure pain." Habitat and the Postgres cluster solve different problems, one for blob-style storage and one for relational data, but both show the same shape: OpenAI's user growth keeps outrunning whatever architecture it shipped the year before.

The number that doesn't quite line up

OpenAI's Sept. 11 figure of "over 1 billion people each week" sits ahead of the timeline suggested by the company's own earlier disclosures. The Information reported on July 29 that ChatGPT was still only approaching that same 1-billion-weekly mark, seven months behind an internal target of roughly December 2025, according to a summary of that reporting published by PYMNTS. Six weeks separate the two figures. That is not a contradiction on its own, since weekly active users can grow quickly, but neither OpenAI's storage post nor its earlier user-count disclosures name the exact week the threshold was crossed, which leaves the round number doing more work than the underlying data supports.

Why the throughput number is not the one to watch

The number that matters here is not the peak throughput figure, but what it costs per request to sustain it. OpenAI has not published a dollar figure for Habitat's infrastructure, so the efficiency gain from the Rust rewrite is the only proxy available. A service six times more CPU-efficient can, in principle, either serve six times the traffic on the same hardware footprint or serve the same traffic on a sixth of it, and those are very different capacity-planning outcomes. OpenAI's post does not say which one is happening, and it is the more interesting number of the two.

The pattern is not unique to OpenAI. AWS's SageMaker HyperPod took a similar bet in September, cutting cold starts by about 60 percent using local NVMe caching instead of re-downloading model weights on every restart, and AWS priced its new Graviton5 chips at a 9 percent premium for a claimed 25 percent speed gain on Lambda the same week. Each case trades engineering effort now for a lower marginal cost of serving later, and each reports the gain against its own prior baseline rather than against a competitor's system.

What would change this read

The clearer test is Python's actual retirement date, which OpenAI has given only as "within weeks," and any dollar or hardware-count figure the company eventually attaches to Habitat's footprint. Until either appears, the six-times and 15-times numbers describe an internal comparison against OpenAI's own prior system, not against any competing storage architecture, and the 1-billion-user figure they are built on has not been independently pinned to a date.

The Postgres cluster's history is a reasonable guide to how much weight to put on that gap. OpenAI wrote that its Postgres primary logged "only one SEV-0" incident in the 12 months before that post, tied to the ChatGPT ImageGen launch, alongside a longer history of lower-severity overload events the same post referenced without a total count. A storage layer can look stable in its own incident log and still be the thing an outside engineer points to as costing the company money it does not have to spend. Habitat's Rust rewrite is OpenAI's attempt to get ahead of that pattern before it repeats, rather than after.