OpenAI Just Showed Its Hand on Rapidly Scaling to 1bn Users

On 11 September 2026, OpenAI published an engineering post describing how its internal storage platform, Habitat, went from a small Python library to a globally distributed system rapidly scaling to serve more than 1 billion people each week. The headline figures: over 70 million requests per second, almost 40 geographic regions, more than 500 petabytes of data.

What this actually means for a SaaS growth team

Ignore the Python. The number that should land on your desk is 1 billion weekly users, stated by OpenAI itself in a primary source rather than an analyst estimate.

That is the size of the layer now sitting between your buyers and your website. When a VP of Engineering asks ChatGPT which observability tool handles high-cardinality metrics, your content either shows up in that answer or it does not. There is no position four to fall back on.

We have been telling clients for a while that AI search is a distribution channel, not a curiosity. This post is OpenAI quietly confirming the scale of it.

What OpenAI reported, and over what timeline

Habitat launched at DevDay in 2023 to support GPTs. It started as a Python client-side library talking to a single Azure Cosmos DB database.

By mid-2025, OpenAI says the client-side model had hit its limits, so the team pulled Habitat into a standalone service. The 11 September 2026 post is part one of two, with a second covering multi-tenancy and read performance still to come.

Habitat, 2023 to 2026, as reported by OpenAI on 11 September 2026

Measure	DevDay 2023	September 2026
Form	Python client-side library	Standalone distributed service
Storage	Single Azure Cosmos DB database	500+ petabytes
Throughput	Not stated	70m+ requests per second
Reach	GPTs only	1bn+ people weekly, ~40 regions
Growth rate	More than 10x year-over-year for three years

How rapidly scaling actually worked here

The mechanism is worth understanding because it is not what most engineering blogs describe. OpenAI did not architect for the end state. It sequenced.

The team kept Habitat in Python even though Python added latency and CPU cost, and even though they knew a rewrite would eventually be forced. They call this a strategic incursion of technical debt: accept the performance hit, unblock product engineers, stabilise the core APIs first.

They also bet that their own coding models would make the later migration cheaper. OpenAI says that bet paid off. We have written before about how OpenAI's own staff shifted their work to Codex agents, and this is the same pattern seen from the infrastructure side.

What scaling software teams rapidly teaches marketers

The transferable lesson is not about databases. It is that a rapidly scaling business wins by fixing the one constraint that blocks growth now, and naming the debt it takes on to do it.

Most in-house marketing teams take the same debt accidentally. Thin content shipped to hit a quarterly target is debt. Attribution stitched together in a spreadsheet is debt. The difference at OpenAI is that the debt was chosen, documented and scheduled for repayment.

What to do about rapidly scaling AI search this quarter

Nothing here changes a ranking factor, and we would not pretend otherwise. What it changes is your justification for budget.

Split ChatGPT, Gemini, Perplexity and Google AI Overviews out as named sources in your reporting, rather than letting them sit inside direct traffic. That is a measurement and attribution job before it is a content one. Then check whether your product, pricing and comparison pages answer the questions buyers actually type, which is the core of organic search across SEO, AEO and GEO. Rapidly scaling assistant usage only turns into pipeline if you can see it and be cited by it.

Cover photo by Levart_Photographer on Unsplash