Two engineers with Codex rewrote the storage layer behind a billion ChatGPT users. The cheapest line in that budget was the code.
On September 11 OpenAI published a deep dive on Habitat, a storage platform that grew from a small Python library at DevDay 2023 into the layer behind 1 billion weekly ChatGPT users: 70 million requests per second, 500+ petabytes. In Q2 2026 two engineers with Codex and GPT-5.5 rewrote the service in Rust: 6x better on CPU, 15x on memory, 95% of production traffic already on the new build, Python on its way out. The trigger was mundane: asyncio scheduling delay on CPU-heavy background tasks dominated tail latency.
The headlines will say "an agent replaced a department." My reading: the platform did the department's work. A 6x CPU gain at 70 million requests per second is a fleet of machines handed back; 15x on memory cuts that bill by an order of magnitude. But for two people to walk a rewrite to 95% of traffic, something had to be standing under them: percentage-based traffic migration, observability at every step, rehearsed rollbacks. The agent wrote the code almost for free. None of the rest is free, and none of it is getting cheaper.
Here is the claim you can argue with: "how many people for how many quarters" is dead as a unit of estimation for infrastructure projects. The new unit is "two people and an agent," and the bottleneck has moved from writing code to shipping it safely. The platform team does not shrink in this world. It is the foundation: invisible until it trembles.
Takeaway for the team lead: redo your estimates. Cut the coding line by multiples; keep migration, observability and rollbacks at full price, because they are the project now. And if you have no platform where moving 95% of traffic is routine, two people with an agent will not do a department's work for you.
📎 openai.com/index/scaling-storage-one-billion-users-part-one · daily.dev