Erik Bernhardsson and Patrick Chase discuss the fragmented market for GPU capacity and the differences in location, networking and reliability. Bernhardsson describes Modal’s approach to compute aggregation and infrastructureAI compute infrastructure is the hardware, facilities, networks, power, cooling, storage, and software used to train and run AI models..
The interview covers inference optimization, hardware selection and the tradeoffs between open-source tools and proprietary infrastructure. Bernhardsson discusses coding agents, the need for engineering supervisionHuman-in-the-loop AI keeps a person involved in reviewing, correcting, approving, or guiding an AI system's work. and changes to developer workflows.
The participants discuss infrastructure financing and future demand. Bernhardsson’s market expectations and company-specific claims are perspectives rather than established outcomes.
Watch on YouTube
