Runware Packs 1MW AI Data Center Into Shipping Container

Modular AI infrastructure ships ready to run

Runware has begun deploying Sonic Inference Pods, modular AI data centers that fit one megawatt of compute and approximately 1,200 GPUs into standard 20-foot shipping containers. Each Pod arrives fully assembled and tested, requiring only ground preparation, power connection, and network access to begin serving inference workloads within roughly 24 hours.

The startup, founded in 2023 by Romanian developers Flaviu Radulescu and Ioana Hreninciuc, is targeting deployment of more than one gigawatt of compute capacity by 2027 across 160 sites. That figure matches the total US data center capacity currently under active construction for 2026 delivery, according to details first reported by Forbes contributor Gabriel Alin Zainescu.

Why it matters

Grid connection timelines have become the binding constraint on AI infrastructure expansion. In Northern Virginia, the world's largest data center market, new connections now require five to seven years, and Dominion Energy has indicated some large customers face waits as long as seven years. Of roughly 16 gigawatts of US data center capacity announced for 2026, only about 5 gigawatts are under construction, and industry analysts expect 30 to 50 percent of announced projects to face delays or cancellation. Runware's containerized approach bypasses the traditional construction and interconnection queue entirely, offering a path to deploy compute capacity on timelines measured in days rather than years.

Closed-loop cooling eliminates water consumption

Each Pod uses closed-loop liquid cooling with water blocks on every processor, recirculating the same 1.5 cubic meters of water continuously. The design eliminates external water consumption during normal operation, contrasting with conventional data center cooling that contributes to the industry's annual draw of more than 560 billion liters of water, much of it lost to evaporative processes.

The GPU mix includes Nvidia RTX PRO 6000 chips as the primary workhorses, with B200 and B300 processors handling larger workloads. Every server includes local NVMe storage, which Runware says keeps models loaded and eliminates cold-start latency.

Network-level resilience replaces facility redundancy

Traditional data centers build redundancy into each facility through backup generators, battery systems, and duplicate cooling infrastructure. Runware distributes resilience across its network instead: if a Pod loses power, cooling, or connectivity, inference requests automatically route to another machine running the same model in a different Pod. The company claims this architecture, combined with purpose-built hardware and software integration, enables inference pricing 30 to 90 percent below conventional alternatives, though those figures have not been independently verified.

Runware has operated its own image and language-model workloads on this infrastructure for several years, serving more than 10 billion generations for hundreds of thousands of developers through its platform.

Deployment timeline and funding

European sites are already operational, US West deployment is underway, and the company plans to bring the first 10,000 nodes online through 2026. Runware raised a $50 million Series A led by Dawn Capital in December 2025, with participation from Comcast Ventures, Speedinvest, Insight Partners, and a16z Speedrun, bringing total funding to $66 million.

The Pods are accessed through Runware Serverless, which offers two modes: Serverless Compute for customers deploying their own models or containers with per-second billing, and API Gateway for dedicated public or private endpoints billed per token or asset.

Details on the deployment and technical specifications were first reported by Gabriel Alin Zainescu for Forbes.