Where supercomputing meets agentic AI: What’s new in Google Cloud HPC and AI at SC26

At SC26, Google Cloud demonstrated how agentic HPC systems can automate the deployment and optimization of high-performance computing workloads by combining AI agents with reusable, validated skills. These skills guide agents through infrastructure setup, workload configuration, and diagnostics, reducing errors and manual intervention in production environments. The approach leverages Cluster Toolkit blueprints paired with self-describing tools, ensuring deterministic and tested workflows before execution.

Domain-specific skills now support fluid dynamics, life sciences, and subsurface simulations, enabling agents to configure tools like OpenFOAM, GROMACS, and AlphaFold 3 with browser-based visualization. Governance features such as IAM, quota checks, and automated debugging tools like gke-kueue-debugging ensure compliance and troubleshooting within agentic pipelines. The system prioritizes human oversight, requiring approval before deploying clusters or running workloads.

AlphaEvolve, developed by Google DeepMind, introduces an evolutionary loop powered by Gemini to optimize algorithms and engineering parameters at cloud scale. Using Cloud Batch, it automates parallel evaluations across thousands of candidates, integrating with notebooks, CLI tools, and agent-driven workflows. The system evaluates candidates on production-scale infrastructure, including Cloud RDMA and Managed Lustre, to ensure performance aligns with real supercomputing fabrics.

Google Cloud’s infrastructure for agentic HPC includes tightly coupled HPC and AI compute, 8th-gen TPUs, and sub-millisecond parallel storage like Managed Lustre with up to 10 TB/s throughput. The hardware supports MPI simulations, AI training, and accelerated modeling, while Batch provides flexible provisioning. Live demos at SC26 highlight tools like Antigravity for HPC and AlphaEvolve, with validated templates available for immediate implementation.