Harvard John A. Paulson School of Engineering and Applied Sciences

Solution

Harvard’s Calmon Lab partnered with IBM to move AI safety experimentation onto a high-performance IBM Cloud environment. The lab provisioned two NVIDIA HGX H100 8-GPU servers inside a secured virtual private cloud, each backed by large memory, high-IOPS block storage, fast network file sharing and IBM Cloud Object Storage for data transfer. Red Hat Enterprise Linux 9, the Anaconda platform and vLLM were used to stand up model-serving and training environments quickly. This infrastructure removed the scheduling bottlenecks that had limited the team’s work on model alignment and chain-of-thought reasoning. Instead of waiting for oversubscribed internal cluster resources, researchers could train and deploy models such as DeepSeek and Llama on demand, collaborate through shared storage and push experiments faster. IBM’s role here was not application AI but the cloud platform and storage architecture that made frontier AI-safety research operationally practical.

Data flow

Research datasets and models are stored and transferred through IBM Cloud Object Storage, loaded onto GPU-backed IBM Cloud servers, and served or trained through the lab’s Linux and vLLM environment for AI safety experiments.