Begin Capital reposted this
Today we’re launching Sonic Inference Pods, our network of modular data centers built for AI inference. Compute demand is growing faster than new capacity. Only about 30% of the US data center capacity planned for 2026 is under construction, and 30 to 50% is expected to be delayed or canceled. In the largest markets, connecting a new facility to the grid can take four to seven years. We built a different kind of data center. Each Sonic Inference Pod delivers 1MW of compute from a 20-foot shipping container. It arrives fully assembled and tested, and needs ground, power and a network connection. Installation takes about a day. Inside each Pod: - Around 1,200 GPUs, with 2 to 8 GPUs per node - RTX PRO 6000 as the workhorse, with B200 and B300 for larger workloads - Local NVMe on every server - our proprietary hardware make the models stay warm and cold starts disappear - Closed-loop liquid cooling, with no water mains and no water consumed in normal operation Pods are small enough to deploy next to wind, solar and hydro generation, reducing the cost and loss of transporting electricity. They can also run closer to users or at a customer’s premises when dedicated local capacity is required. The network is distributed by design. If a Pod loses power, cooling or connectivity, requests move to another machine running the same model in another Pod. Resilience is handled across the network instead of being duplicated inside every facility. We have been running our own image and language-model workloads on this infrastructure for years. By designing the hardware and software together, and removing infrastructure inference does not need, we can offer inference at 30 to 90% lower cost. Sonic Inference Pods are available through Runware Serverless: - Serverless Compute: bring your models, Docker containers, services or code. Scale from zero and pay per second. - API Gateway: bring your model and run it behind a dedicated public or private endpoint. Pay per token, frame or asset. We are deploying across 160 sites, bringing the first 10,000 nodes online through 2026 and targeting more than 1GW of compute in 2027. Learn more and access the network through Runware Serverless. Link in the comments.