Lambda | How we build GPU clusters for the age of superintelligence

What does it take to deploy GPU clusters that scale from one GPU to tens of thousands? We don't just deploy hardware. Our teams co-engineer with customers across GPU, networking, cooling, and power to size every layer of the stack for specific workloads. Every cluster is validated with a real ML training workload before it ships. In this video, Lambda's infrastructure team shares: • How scalable units are defined, up to thousands of GPUs per data hall • Why liquid cooling reduces thermal footprint by four times while enabling denser, lower-latency clusters • How CPO (Co-Packaged Optics) technology adds hundreds to thousands of GPUs that weren't possible before • What network topology looks like as GPU counts increase • How we validate every cluster end-to-end before it goes live Rich Underwood: "When you're working with massive data centers that have hundreds of megawatts of power, adopting CPO technology allows us to add hundreds to thousands of GPUs that we wouldn't have been able to." Learn more: https://lambda.ai/ai-infrastructure?utm_source=youtube&utm_medium=organic-social&utm_campaign=2026-03-pre-gtc&utm_content=description Join our community: X (Twitter): https://x.com/LambdaAPI LinkedIn: https://www.linkedin.com/company/lambda-cloud/ Facebook: https://www.facebook.com/lambdaai Reddit: https://www.reddit.com/user/LambdaAPI/