Google Cloud adds Google Skills for agentic HPC runs

Google Cloud's SC26 message is that the bottleneck in high-performance computing has shifted from raw hardware to the expertise needed to configure complex scientific software. The two new repositories, Google Skills and DeepMind Science Skills, encode that expertise as reusable agent instructions that plug into Cluster Toolkit's declarative blueprints.

The company is also packaging AlphaEvolve, its Gemini-driven evolutionary optimizer, as a Cloud Batch solution. A single blueprint provisions the controller environment, and Batch fans out hundreds of candidate evaluations concurrently across single-VM sandboxes or multi-node MPI task groups, with Cloud RDMA, Managed Lustre, Spot VMs and the Dynamic Workload Scheduler available to tune cost against turnaround.

On hardware, Google is broadening its menu rather than making a single bet. Eighth-generation TPUs split between TPU 8t for pre-training, reinforcement learning and mathematical modeling, and TPU 8i for low-latency reasoning and agentic inference. GPU options include A4 and A4X on NVIDIA Blackwell B200 and GB200 NVL72, newly available A5X bare metal on Vera Rubin NVL72, and G4 shapes with RTX PRO 6000 Blackwell for CAE and CFD. The authors frame the combination as letting teams spend their time on discovery rather than infrastructure plumbing.