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

Runware unveiled a self‑contained unit designed to bring AI inference closer to where data is generated. The Sonic Inference Pod packs GPU‑accelerated compute into a format that can be moved on a standard truck and set up in under half an hour. The announcement positions the pod as a practical answer to latency and bandwidth concerns that arise when large models run far from the edge.

What You Need to Know

The pod is built around a 20‑foot ISO container (approximately 6.1 m × 2.44 m × 2.59 m). Inside, Runware installs up to twelve Nvidia H100 Tensor Core GPUs, each rated at 700 W, giving a peak compute capacity of about 8.4 kW before accounting for overhead. Power delivery is handled by a three‑phase 480 V input, and the unit includes an integrated liquid‑cooling loop that keeps inlet temperatures below 35 °C.

Networking consists of two 100 GbE uplink ports for connection to external backbone links and an internal 25 GbE fabric that links the GPUs to a shared NVMe storage pool. The storage tier offers up to 30 TB of NVMe SSDs, configured for low‑latency read workloads typical of inference tasks. All hardware is mounted on shock‑absorbent rails to survive transit vibrations.

Why It Matters

Running large language models at the edge reduces the round‑trip time between user devices and inference servers, which is critical for applications such as real‑time translation, autonomous vehicle perception, and interactive gaming. By placing compute inside a portable pod, operators can avoid the capital expense and lead time of building a permanent data‑center wing while still achieving comparable performance.

From an operational standpoint, the pod’s modular nature simplifies scaling. Additional units can be shipped to a site and plugged into existing power and networking infrastructure without major civil work. This flexibility helps companies match compute capacity to fluctuating demand, such as seasonal spikes in video‑streaming traffic or short‑term research bursts.

Key Details

• Container dimensions: 20‑ft ISO (6.1 m × 2.44 m × 2.59 m)
• GPU configuration: up to 12 × Nvidia H100 (700 W each)
• Power input: three‑phase 480 V, ~8.4 kW GPU draw plus cooling
• Cooling method: closed‑loop liquid cooling, target PUE ≈ 1.1
• Networking: 2 × 100 GbE uplinks, internal 25 GbE fabric
• Storage: up to 30 TB NVMe SSD, optimized for read‑heavy inference

What’s Next

Runware plans to begin field trials with select cloud‑service partners in Q4 2025, focusing on latency‑sensitive AI services. Feedback from those pilots will inform a second‑generation pod that may incorporate higher‑density GPUs and improved power‑management features. The company also intends to publish reference architectures that show how multiple pods can be interconnected to form a scalable, portable inference grid.

📌 Source: Techcrunch Ai