Case Study: Harvard accelerates AI safety research with IBM Cloud

Accelerating AI safety research with a scalable cloud infrastructure

Harvard’s Calmon Lab at the Harvard John A. Paulson School of Engineering and Applied Sciences was advancing AI safety research, but its progress was slowed by GPU bottlenecks and limited access to the NVIDIA H100 resources needed to run and test large language models. The team needed a more scalable infrastructure to continue work on aligning models like DeepSeek-R1 and Llama with human values and safety standards.

IBM helped Harvard by deploying IBM Cloud infrastructure with NVIDIA HGX H100 8-GPU servers, secure VPC, high-performance storage, IBM Cloud Object Storage, and software including Red Hat Enterprise Linux 9, Anaconda, and vLLM. As a result, Harvard was able to train and deploy LLMs without wait times and run inference at over 2,000 tokens per second, enabling faster experimentation and new AI safety breakthroughs.

Harvard

Professor Flavio du Pin

Associate Professor