Runware unveils AI factory to build modular AI data centers just like cars

Runware today announced a modular AI data center factory that will build 10 gigawatts of new AI compute capacity every year, producing complete, AI-native data centers on an assembly line similar to car manufacturing. Located on a 538,000-square-foot campus (50,000 sqm) in Europe, the 237,000-square-foot (22,000 sqm) manufacturing facility will begin production in the first quarter of 2027 and reach that annual production rate by year-end.

NVIDIA's DSX POD announcement, introduced at GTC Berlin, validates the approach Runware has been developing from the start: manufacturing complete AI data centers in shipping containers. The factory's configurable production line will accommodate NVIDIA reference architectures for Vera Rubin, including DSX POD, 16-rack NCP, DGX POD and NVIDIA DGX SuperPOD configurations, alongside liquid-cooled GB300, B300 and NVIDIA RTX PRO 6000 Blackwell Server Edition.

The factory will deliver Sonic Pods in two formats: a 1-megawatt data center in a 20-foot shipping container and a 2-megawatt data center in a 40-foot shipping container. Each unit will leave the production line assembled, tested and ready to ship.

The factory will expand the infrastructure behind Runware's inference platform, the Sonic Inference Engine®, while supplying Sonic Pods for customers' dedicated AI deployments.

Inside the factory

The production line is designed for the equivalent of up to 30 megawatts of compute facilities per day. Robotic assembly and specialist integration will bring together compute, cooling, networking and power systems.

Runware's in-house hardware design team develops the pod enclosures, liquid-cooling systems, power distribution units and controls as one integrated platform. Owning these designs lets the team rapidly adapt the pods for new GPU generations as they launch and control component, manufacturing and integration costs.

Racks, equipment, internal cabling and pipework will be installed on the line. Each unit will undergo electrical, cooling and connectivity checks before shipment, with cooling tested at full load. Runware's software verifies the installed equipment and cabling and prepares the compute hardware for provisioning.

The new facility will scale a process already in use at Runware's assembly operation in Europe, where 41 Sonic Pods are being assembled by EOY. Runware has been running the Sonic Pods in production for its own inference workloads, with the current version that has been in production the longest running from Q1 2026 with 100% uptime.

"Demand for AI is outpacing the industry's ability to build data centers," said Flaviu Radulescu, founder and CEO of Runware. "We have turned the data center into a manufactured product. The factory will let us build thousands of complete units and deploy them wherever suitable power and connectivity are available."

Pricing

Runware will offer factory-produced pods for Vera Rubin deployments at $1 million per usable megawatt per year under its pod-only service model. The rate is based on usable IT capacity and excludes compute hardware and electricity.

For context, Cushman & Wakefield's 2026 Data Center Development Cost Guide puts average upfront greenfield development costs for modern facilities in the United States and Canada at $17.6 million per megawatt, excluding chips and GPUs. Runware's annual service model gives customers an alternative way to procure data center capacity and manage upfront capital requirements.

Building data centers this way cuts the capital cost of bringing AI compute online, and Runware passes those savings on. Customers get GPU capacity on Sonic Pods at the lowest rates in the industry: from $0.63 per GPU-hour for NVIDIA RTX PRO 6000, $3.62 for GB300 NVL72 and $5.07 for Vera Rubin NVL72, around half the cost of a conventional neocloud.

Configurable for different compute deployments

Sonic Pods can operate independently for distributed inference or combine into Scalable Units and larger training and inference deployments. Rack layouts, networking and power configurations are adapted to the customer's hardware, workload and availability requirements.

Each pod carries its own closed-loop, direct-to-chip liquid cooling, which rejects heat to outdoor air. It requires no permanent water connection and consumes no water in normal operation.

From the factory to ready-to-use sites

Alongside the factory, Runware is working with powered-land and energy providers to develop a network of deployment sites, initially targeting locations of 5 to 50 megawatts. Sonic Pods can also be deployed across larger customer-provided sites, including 2-gigawatt campuses. Runware's current site pipeline covers 397 megawatts, with expansion to 1.2 gigawatts in the next 6 months.

Site preparation can proceed in parallel with factory production. At a prepared site, individual pods are designed to be installed and commissioned in under 48 hours.

Runware serves as the operating partner, handling commissioning, expansion and ongoing operation to agreed performance and availability standards. Services include 24/7 monitoring and observability, remote hands, incident response and maintenance.

How to get capacity

Customers can bring their own site or deploy at a Runware location. They can supply their own compute hardware and GPUs or have Runware provide them, with the system configured to the agreed rack specifications, network architecture and delivery schedule.

Organizations can also access Runware-operated compute through its model API, managed inference and Runware Serverless. Dedicated Sonic Pod deployments are available to reserve now at runware.ai/sonic-pods.

About Runware

Runware develops and operates AI infrastructure for inference and training. Its technology combines proprietary inference software with modular data centers to serve AI developers and enterprises.