Is the future of data centers portable? Runware builds a pod to find out

Runware, an AI infrastructure company, unveiled something it calls the Sonic Inference Pod on Tuesday, and the pitch is refreshingly literal: a single transportable unit that does the work of a data center without needing to be one. Rather than pouring concrete for months, Runware wants companies to just add another pod when they need more capacity. CEO and co-founder Flaviu Radulescu told TechCrunch the company sees this as inevitable, arguing that distributed compute placed closer to end users, for faster inference, is what ultimately wins.

The pitch rests on speed and flexibility more than raw scale. Radulescu says the pods can be deployed anywhere there's power, adapt quickly as new hardware comes out, and use a closed-loop cooling system that skips water entirely and can be built in days rather than the months or years traditional facilities require. Runware already has 10 pods running across the U.S., Europe, and Asia-Pacific, with 160 sites lined up to host more. Customers so far include Higgsfield AI and Wix, and the company backed this expansion with a $50 million Series A raised in December, originally aimed at infrastructure for image generation.

The timing is notable because everyone else in AI is thinking bigger, not smaller. OpenAI is reportedly closing in on a $500 billion arrangement to build a data center in Ohio, and SpaceX and other labs keep racing to build their own massive facilities. Radulescu doesn't see those projects as competition. He frames the pods as a network rather than isolated boxes: requests route to wherever there's spare capacity, and if one pod fails, traffic just shifts elsewhere instead of taking down an entire facility. Customers wanting dedicated hardware can also get a whole pod exclusively for themselves.

He's also unbothered by the idea of rivals copying the approach, pointing to the slow, unglamorous mechanics of hardware design as the real moat. A circuit board mistake, he said, can cost months once you factor in redesign, simulation, fabrication, testing and delivery, and the talent pool capable of catching those mistakes is thin.

The bigger backdrop here is the growing unease around data centers themselves, especially their appetite for power and water, with nearby communities already reporting rising utility costs. Radulescu doesn't pretend Runware has solved that yet, but argues that AI's power demand is going to climb no matter who supplies it. His argument for the pods is that they avoid transmission losses, skip water-based cooling, and lean on power that already exists instead of requiring new grid capacity, meaning more inference for the same environmental footprint rather than less demand overall.