In this morning's Agenda, we get into why it's hard for even a big lab like xAI to fully utilize its GPUs, and why AI researchers more broadly are faking their GPU utilization. theinformation.com/newslette…

This wraps up the 1st Workshop on Generative AI in Genomics (Gen²) @iclr_conf. We had so many speakers we wanted to invite, yet couldn't fit into our schedule. We'd like to acknowledge our organizers, advisors, and colleague, @SandeepKambham2, who helped us run the workshop!

This week at Google Cloud Next, we introduced 8th gen TPUs, a critical milestone in our accelerator roadmap. TPUs enable us to optimize the entire stack for AI (with 8t for massive-scale training and 8i for low-latency inference). Exciting breakthrough from our hardware teams! Read more on the systems architecture: blog.google/innovation-and-a…

We are hiring at @mithrilcompute! Access to high-performance compute (GPUs, TPUs, and the like) is fundamentally broken: 1. prohibitively expensive and 2. heavily under-utilized. The providers (mostly publicly traded companies) want financial guarantees that the bet will pay off: contracts with them must be large and long. Only a small set of players (i.e., well-funded "startups") can afford those terms. The demand from these players is incredibly high—Blackwell GPUs are essentially sold out everywhere—but if you look at the utilization numbers, something is way off. Why? At Mithril, we believe this is due to a fundamental fact: price control just doesn't scale with AI usage patterns. Small and medium players want access to these advanced chips in unpredictable patterns: "I need to immediately run this experiment for 48 hours and then again 15 days from now" said a genomics researcher somewhere. "My inference service explodes in demand every end of the month; I need 500 more chips just for 3 days every month" said the CEO of an Accounting AI platform somewhere else. This is an incredibly rich field to be part of! At Mithril, you’ll be working to enable millions of research institutions and companies worldwide to leverage one of the pinnacles of human ingenuity—without the usual hassle. From building our advanced Virtual Machine and high-performance network orchestration, to creating durable, resilient APIs and establishing state-of-the-art consumption principles... there’s just too many GOOD open problems to solve! Take a look at our openings at mithril.ai/company#roles. All positions are in-person at our offices in San Francisco and Palo Alto, CA.

Jared Quincy Davis @jaredq_, Founder and CEO of our partner @mithrilcompute, is a fascinating person to talk to. Through Mithril’s platform clients across sectors can access Nebius infrastructure through tools they already know and trust. We spoke with Jared about what inspired him at #NVIDIAGTC, how the industry is pushing itself to create new solutions and how the infrastructure layer needs to evolve to support emerging workloads. #GTC26

Computer use models shouldn't learn from screenshots. We built a new foundation model that learns from video like humans do. FDM-1 can construct a gear in Blender, find software bugs, and even drive a real car through San Francisco using arrow keys.