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

Google Cloud announced a suite of agentic supercomputing tools, evolutionary optimization services, and infrastructure hardware ahead of the SC26 conference in Chicago, according to Google Cloud Blog's report. The expansion integrates Cluster Toolkit with deterministic agent skills available through the Google Skills and DeepMind Science Skills repositories to automate workload deployment. Alongside software automation, the company detailed AlphaEvolve on Cloud Batch, new compute shapes powered by AMD EPYC and Google Axion processors, NVIDIA Blackwell and Rubin GPUs, eighth-generation TPUs, and parallel storage capabilities.

This reporting examines the operational architecture, system integrations, and hardware specifications designed for research institutions, engineering organizations, and enterprise high-performance computing teams. It explains how specialized agent skills replace manual provisioning, how evolutionary algorithms benchmark code at scale, and how coupled simulation fabrics handle frontier artificial intelligence workloads.

01Blueprints and domain agent skills

Google Cloud expanded Cluster Toolkit by pairing declarative YAML blueprints with composable, deterministic agent skills tested through end-to-end automation. Before executing any deployment command, an infrastructure skill audits project quotas, conducts regional capacity canary checks, matches requirements to validated blueprints, and halts for mandatory human approval. This pre-flight validation prevents unguided models from pairing mismatched storage tiers with compute instances or generating multi-node MPI fabric errors.

Once a cluster operates, workload-specific domain skills direct autonomous setup across specialized scientific disciplines. The system supports computational fluid dynamics in OpenFOAM and Siemens STAR-CCM+, life-sciences pipelines involving GROMACS and AlphaFold 3, and subsurface modeling through Devito and Open Porous Media. All autonomous actions run within strict Identity and Access Management policies and Kueue quota boundaries, backed by audit trails in Cloud Logging and specialized gke-kueue-debugging routines.

02Evolutionary algorithm optimization on Cloud Batch

To accelerate mathematical solvers without dedicated optimization specialists, Google DeepMind introduced the AlphaEvolve on Cloud Batch solution. The system runs an iterative evolutionary loop driven by Gemini models to mutate, benchmark, and refine source code alongside complex three-dimensional engineering parameters. A single Cluster Toolkit blueprint provisions the controller environment, allowing organizations to migrate proof-of-concept algorithms from local workstations directly into distributed execution environments.

Cloud Batch manages the concurrent evaluation of hundreds of candidate algorithms across isolated single-machine sandboxes and multi-node MPI clusters. Research teams steer running experiments through hosted Python notebooks, local command-line interfaces, or integrated agent environments. Evaluating candidates across Cloud RDMA networks, Google Cloud Managed Lustre filesystems, Spot VMs, and the Dynamic Workload Scheduler ensures discovered algorithms excel on real supercomputing fabrics.

03Compute instances for coupled simulation workloads

The hardware layer relies on Cluster Director to manage execution across mixed simulation and machine learning environments. For traditional MPI modeling and electronic design automation, Google Cloud introduced H4D instances powered by fifth-generation AMD EPYC processors with Cloud RDMA networking. These configurations feature high-memory profiles reaching 3 TB of RAM to accommodate massive memory-bandwidth-bound engineering datasets.

For energy-efficient scale-out tasks, the platform offers Arm-based C4A instances driven by custom Google Axion processors to improve performance-per-watt metrics. Systems dynamically route compute jobs between traditional physics engines and foundation models such as AlphaFold 3, AlphaGenome, and WeatherNext. This coordinated orchestrator maintains consistent networking and execution parameters across disparate processor types without manual cluster repartitioning.

04Blackwell accelerators and eighth generation TPUs

GPU acceleration expands with A4 and A4X instances powered by NVIDIA Blackwell B200 and liquid-cooled GB200 NVL72 platforms for multi-node simulation pipelines. The company also introduced A5X bare-metal instances based on NVIDIA Vera Rubin NVL72 architectures for frontier model training and reasoning workloads. For computer-aided engineering, graphics processing, and visualization, G4 instances incorporate NVIDIA RTX PRO 6000 Blackwell GPUs to lower per-core software licensing expenses.

Custom silicon additions include eighth-generation TPUs split into two specialized hardware designs. TPU 8t accelerates large-scale pre-training, reinforcement learning, and mathematical modeling simulations, while TPU 8i focuses on low-latency inference and high-throughput agentic reasoning. Cloud Batch now provisions single-host TPU slices, enabling flexible compute allocations for coupled artificial intelligence and numerical simulation workflows.

05Managed Lustre storage and conference demonstrations

High-throughput input and output operations utilize Google Cloud Managed Lustre, which delivers up to 10 TB/s of aggregate read performance across clusters. The parallel storage system incorporates GPUDirect Storage over RDMA alongside an 80 PB Dynamic Tier featuring Cloud Storage Mirroring. Workloads requiring extreme localized bandwidth can combine Managed Lustre volumes with Cloud Storage FUSE local solid-state drive caching.

Demonstrations will take place November 15–20 at SC26 booth 3418 inside McCormick Place West in Chicago. Attendees can view Antigravity for HPC, the HPC Prompt Playground, and AlphaEvolve for HPC during dedicated workshop sessions and an Agentic and Hybrid HPC Birds of a Feather gathering. Production templates are currently available through the Google Skills and DeepMind Science Skills repositories.