This AI Infrastructure Startup Predicts GPU Needs Before Code Runs | Product-X

What does an AI infrastructure startup do while everyone else is buying more GPUs? To avoid out-of-memory crashes, AI teams routinely ask for three to five times the GPU and VRAM a job needs, and more than half of that capacity often sits idle. In this episode, Product-X breaks down Expanse, a San Francisco and London startup that predicts a job's GPU compute, VRAM, and CPU memory milliseconds before it runs. Expanse announced a $5.3 million Seed round. The round was led by Crane Venture Partners, with Y Combinator (P26), PXN Ventures, Standard Capital, Scribble Ventures, and angels including former DeepMind researchers. The four co-founders, Ismaeel Bashir (CEO), Nikodem Bieniek (CTO), Eren Mendi (CPO), and Yafet Melake (COO), come from HPC at the University of Edinburgh. Hosted by Rena, with Vivian Liu (angel investor, PreAngel Fund) on the investment case and Nolan (former Big Tech CTO) on the engineering. What you'll learn: why post-hoc monitoring tools like Datadog can't stop GPU waste; how Expanse reads code before it runs and balances hardware cost against crash risk; how it found nearly $8 million of idle compute in one production deployment test; and why dynamic agent workflows are the hardest problem left. Product-X decodes AI agents, AI infrastructure companies, and frontier startups for builders and investors. Subscribe for evidence-based breakdowns of what a company built, why it matters, and whether it's worth watching. #AIInfrastructure #GPUUtilization #Expanse Chapters 00:00 What an AI infrastructure startup does about idle GPUs 01:43 The investment case: the Edinburgh HPC team and the $5.3M Seed 03:16 How it works: reading code before it runs, cost vs crash risk, hardware routing 05:13 Where it goes next: utilization and dynamic agent workflows 06:09 Wrap-up: how much of your GPU capacity gets used