Jordan Nanos

Jordan Nanos is a member of the technical staff at SemiAnalysis, where he leads ClusterMAX, a research project evaluating the GPU clouds used to train and run AI models. His work brings experience designing large computing systems to the practical question of which rented clusters are reliable, secure and usable.[1][4]

Nanos co-authored ClusterMAX 2.0, which reviewed 84 providers and examined areas including reliability, networking, security, storage and support. The 3.0 report covered 77 providers in its review while tracking 323 in a wider market view. Its ratings concern managed clusters: the numbers describe different scopes, and the ranking does not assess every business or service offered by those companies.[3][4]

In research on GPU-cluster cost, Nanos and his co-authors argue that the advertised price per GPU-hour is only part of the bill. Setup, debugging, storage and downtime also consume money and engineering time. Their goodput framework counts useful training work: when a job fails, the cost can include work lost since the last checkpoint as well as the time needed to restart.[5]

His team’s security research examines the software and isolation around rented compute. It also introduced a free ClusterMAX command that checks a cluster or machine against software-version baselines associated with known vulnerabilities. That check covers only a subset of the team’s wider security testing.[6]

Nanos joined Dylan Patel on Latent Space to discuss ClusterMAX 3.0 and the state of GPU-cloud infrastructure.[8][7]

Nanos joined Hewlett Packard Enterprise as a summer intern in 2016 and went on to hold six roles there. As a distinguished technologist, he worked on the design and implementation of large supercomputers before joining SemiAnalysis. He holds a bachelor’s degree in electrical engineering from Queen’s University.[1][2]