Technology

Can AI Data Centers Be Built Like Cars? One Startup Is Trying

Martin HollowayPublished 2w ago4 min readBased on 3 sources
Reading level
Can AI Data Centers Be Built Like Cars? One Startup Is Trying
source:runware.ai

Runware plans to build data center pods on a production line, much as carmakers build vehicles. The company is aiming for nearly 1 gigawatt of data center capacity per month. It expects to reach that by raising output at its Romania-based factory to 30 megawatts per day by 2028, according to TechCrunch.

Runware already runs one factory, which it has used for a few years to build pods. It started building a second factory earlier in 2026. That facility is expected to be finished within months and open early next year.

Build time is the number the company is trying to cut. Runware says a pod once took three weeks to build. It says pods will soon take one day as production speeds up.

The company's stated goal is to have pods running within 48 hours for installation and set-up. That window does not include build time.

The product is the Sonic Inference Pod, which the company has unveiled. It is a modular data center, built in a factory and shipped ready to install. "Inference" means running an AI model to produce an answer, as opposed to training the model in the first place.

The pods have a published specification. In an August 2026 blog post, Runware describes its Sonic Pods as modular 1MW inference data centers. Each holds 1,200 GPUs, the specialized chips used for AI work, packed into a 20-foot shipping container. They are cooled by a closed loop of liquid that circulates through the system.

Runware's website says its orchestration layer, the software that assigns work across machines, and the pods together form what it calls the Sonic Inference Engine.

The company states that its modular AI data centers run at half the cost. It also offers inference at up to 90% below market rates. Runware says its platform has served more than 10 billion inference requests across image, video, audio, language and 3D. More than 1 million developers build production AI features on it, and applications running on its infrastructure reach more than 300 million end users. The website says the platform kept 99.99% uptime, meaning the share of time a service stays available, over the last 90 days.

On the commercial side, Runware rents GPUs as a serverless service. Customers are billed by the second, and their capacity scales down to zero when they are idle. Listed pay-as-you-go pricing for RTX PRO 6000 GPUs is $1.99 per GPU-hour. Reserved serverless capacity is listed as low as $0.63 per GPU-hour.

Dedicated bare-metal clusters on Sonic Pods start at 72 GPUs. Bare-metal means whole machines reserved for one customer. The company says it offers more than 400,000 models through a single endpoint, which is one access point for software.

Runware has raised more than $60 million to date and plans to raise a Series B, according to TechCrunch.

In my view, the central question is whether this manufacturing pace holds. A production target for a factory still under construction is a forecast, not a track record. The move from a three-week build to a one-day build is a ramp the company has described, and the coming year will show whether it can deliver.

The broader context here is that most of the delay in getting AI computing online has come from preparing sites, securing power and installing equipment, rather than from the servers themselves. If a pod can be built, tested and shipped as a finished unit, the time between spending money and running AI tasks could shrink. That would matter to anyone whose projects are held back by how quickly capacity can be put in place.