The company now called Superhuman was, until late 2025, called Grammarly. It acquired Coda, a document workspace tool, and then Superhuman Mail, a premium email app valued at around $825 million, and rebranded everything under one umbrella. The pitch is a productivity suite where AI assists with writing, email, documents, and task management across every app a professional uses.
All of that depends on a custom AI model running in the background every time someone types. At peak, that model handles more than 200,000 requests per second. To put that in context, a mid-sized airline might handle a few thousand booking requests per second on a busy day. Superhuman's grammar correction alone runs at a scale most companies will never approach.
For years, the company managed its own infrastructure for this. They built their own system using open-source tools, managed their own servers, and handled everything from capacity planning to performance tuning themselves. It worked, but the cost was not just money. Every new version of the model required months of manual work before it could go live. A small team was spending serious time keeping the engine running rather than building the product.
That is the hidden tax of self-managed AI infrastructure. The hardware costs are visible. The engineering hours lost to maintenance, scaling problems, and outage responses are not on anyone's P&L line, but they compound.
The company eventually partnered with Databricks to hand off the infrastructure layer entirely. Databricks manages how the model is served, how traffic spikes are handled, and how the system scales up during busy hours and back down during quiet ones. Superhuman retains full ownership of the model itself: they decide how it is trained, what it optimises for, and how quality is measured.
This split is more significant than it first appears. It is not outsourcing. The company did not lose control of its core product. What it shed was the operational overhead that had nothing to do with the product. Engineering effort that was going into infrastructure is now going into the AI itself.
The broader pattern here matters to anyone thinking about deploying AI seriously. Running AI at scale is not one problem, it is two: building a good model, and keeping it running reliably under load. Most companies have expertise in their own domain. Very few have deep expertise in both simultaneously.
For the average business deploying AI, whether it is a claims processing system, a customer service tool, or an internal document assistant, the self-build path looks attractive because it feels like control. But the infrastructure problem keeps growing as usage grows. Managing traffic spikes, cutting startup times when demand surges, and tuning performance across different hardware generations is a specialised discipline. Handing that to a platform provider and focusing internal resources on the business problem itself is increasingly the rational choice.
Inference, which is the process of actually running an AI model to produce answers, is already estimated to account for the majority of AI compute costs in production. By 2030, it is projected to represent 80 percent of all AI-related computing load globally. The companies that figure out how to do it cost-effectively, whether by building in-house capability or through smart partnerships, will have a durable advantage over those still treating it as a side project managed by a stretched engineering team.