The Compute 100: Roman Chernin, Co-Founder of Nebius
1 hr 13 min
Today’s guest is Roman Chernin, Co-Founder of Nebius. Roman has been at Nebius since day 1. Since being on the journey from Yandex, to launching Nebius just over 2 years ago, the company has been on a historic rise as a dominant player in the neocloud market, known for training and inference offering, vertical integration of their hardware stack, and selling compute to some of the biggest buyers in the world. Few people question demand for AI compute, which shifts attention to whether a neocloud can finance its buildout and own more of the customer experience. We get into how Nebius builds in-house, spreads projects across many sites and how they’re thinking about funding new capacity. From financing to GPU reliability, we discuss:- Why building data centers, racks and software in-house changes Nebius’ costs- How short-term capacity sales and prepayments help fund new buildouts- Where demand is coming from across labs, AI-native products and enterprises- Why agentic workloads are driving demand for CPU compute and sandboxes- What makes GPU capacity reliable at scale, from failing nodes to storage speed Timestamps:(00:00) Introduction(01:04) Market swings and the DeepSeek stock drop(04:59) Vertical integration and financing discipline(08:50) From Yandex to Nebius(12:32) Data-center delays and revenue worries(17:02) Where Nebius builds its data centers(19:47) Financing the buildout and the Meta backstop(24:31) Short-term capacity and customer prepayments(29:30) Where AI demand is coming from(38:02) Nebius’ product layers and Roman’s role(40:38) Agentic search and open models(45:05) Why agents are driving CPU demand(49:22) One compute pool for training and inference(55:10) What makes compute high quality(01:02:09) Customer advisory board and the compute rebate(01:08:09) Culture, hiring and an underdog mindset The Compute 100 is led by Brian Schechter and Gaby Lorenzi, compute-focused investors at Primary. Primary is a pre-seed and seed-stage venture firm that back...