Agent Inference at the "Speed of Light" — How NVIDIA moves like a $4.3 Trillion Startup

Swyx and Vibhu chat with Nader Khalil (https://x.com/naderlikeladder) and Kyle Kranen (https://x.com/KranenKyle) from NVIDIA about NVIDIA'S DX mission - Brev’s origin as a one-click way to access GPUs, and how the Brev team (https://brev.dev) integrated after acquisition. They discuss security risks of agents that can access files, the internet, and execute code, emphasizing limiting agents to only two capabilities and using isolated environments like Brev VMs for tools such as OpenClaw. Kyle explains NVIDIA Dynamo as a data center scale inference engine that optimizes serving by scaling out, leveraging techniques like prefill/decode disaggregation, scheduling, and Kubernetes-based orchestration, framed around cost, latency, and quality tradeoffs. They also cover NVIDIA’s “SOL” first-principles urgency concept, long-context limits and model/hardware co-design, internal model APIs (build.nvidia.com), and upcoming Dynamo and agent sessions at GTC. Timestamps 00:00 Agent Security Basics 00:39 Podcast Welcome and Guests 01:26 Surfboard Booth Story 03:45 What Brev Does 05:09 GPU Gift Card Stunt 07:34 Acquisition to Nvidia 09:53 Developer Experience Shift 11:48 DGX Spark on Brev 14:12 Jensen SOL Principle 19:12 Meet Kyle Background 23:08 Nvidia Culture and Email 25:01 Zero Billion Markets 26:55 From Recs to Dynamo 27:34 Test Time Scaling 28:04 Why Dynamo Exists 29:44 Scale Up vs Scale Out 31:18 Serving Big Models 34:08 Quality Cost Latency 35:59 Just Try Again 37:58 Nitron Release Stack 39:25 Disaggregated Inference 42:15 Kubernetes Scaling Logic 44:36 Context Length Limits 50:25 Unhobblers and Breakthroughs 54:33 GTC Dynamo Sessions 57:03 Agents and Codex Rollout 58:10 Email Agent Workflow 59:46 Agent Security Boundaries 01:00:35 Running Models In House 01:00:55 NVIDIA Model Playground Story 01:02:28 Enterprise NIM And Dynamo 01:03:43 Hackathons And Self Driving 01:06:11 Zero Minute Agent Challenge 01:08:04 CLI First Agent Workflows 01:09:41 Why CLIs Beat APIs 01:15:32