Google Cloud bundles supercomputing hardware and DeepMind models into one science platform

The hardware layer: H4D, Titanium, and a managed parallel file system

The concrete infrastructure news is H4D VMs, described as Google Cloud's most powerful CPU-based VMs, built on the latest AMD CPUs and paired with Titanium network acceleration. The stated goal is letting scientists scale HPC applications to thousands of processors. H4D is in preview now.

Two supporting pieces matter for anyone who has actually managed a cluster. Cluster Toolkit promises repeatable deployments, and Cluster Director (renamed from Hypercompute Cluster) lets a large cluster be operated as a single unit. Google is acknowledging that the deployment and management burden, not just raw FLOPs, is what stops researchers from using cloud HPC.

Storage gets its own answer: Google Cloud Managed Lustre, built with DDN and based on EXAScaler Lustre. Rather than reimplement a parallel file system, Google is packaging an established one as a managed service to meet the extreme storage demands of simulation and AI training workloads.

DeepMind models arrive as deployable Cloud products

The more distinctive move is turning DeepMind research into shippable Cloud offerings. AlphaFold 3, from DeepMind and Isomorphic Labs, is offered as a High-Throughput Solution for non-commercial use, deployable through Cluster Toolkit, with batch processing of up to tens of thousands of sequences and autoscaling to control cost.

WeatherNext models are being made available through Vertex AI Model Garden, where they can be customized and deployed. That places a research forecasting model into the same catalog customers already browse for general-purpose models — a routing decision, not just a research release.

Having access to AlphaFold on Google Cloud can help our researchers rapidly predict and explore the structure and interactions of all biomolecule classes, accelerating our understanding of diseases.Montana Labs

Two research agents attached to the platform

Alongside infrastructure and models, Google is adding two agents in Agentspace, both in preview. The Deep Research agent synthesizes external and enterprise data into research reports; the Idea Generation agent produces novel hypotheses for scientists to test.

These are framed narrowly around the front end of the scientific workflow — reading and hypothesizing — rather than running experiments or validating results. That scoping is honest about what generative agents can reliably do today, and it keeps the human researcher as the one deciding what to test.

What ties this announcement together: vertical integration of the research stack

The specific implication here is that Google is trying to own every layer a computational scientist touches — the CPU VMs, the cluster orchestration, the parallel file system, the domain models, and the agents that suggest what to work on — and sell them as one platform rather than separate SKUs.

For applied teams evaluating this, the value depends on whether the layers actually compose. AlphaFold 3 deploying through the same Cluster Toolkit that provisions H4D clusters is the kind of integration that reduces friction; WeatherNext living in the general Vertex Model Garden is a lighter form of the same idea. The open question is coupling: how much of this only works well when you buy the whole stack, versus how portable each piece is if you already run part of your pipeline elsewhere.

Notably, the announcement is light on pricing and hard performance numbers — most items are in preview, and AlphaFold 3's high-throughput solution is non-commercial only. The strategy is clear; the terms under which research groups can actually depend on it are not yet.