VAST Data and Google Cloud have announced a full-managed VAST AI OS service – a unified global namespace across hybrid environments supporting Google TPUs.
VAST’s AI OS is basically its entire software stack: DataCatalog, DataBase (universal name space), DataSpace, DataEngine, InsightEngine, and AgentEngine, conceived as an operating system layer using GPU server and networking hardware to provide a training, inference and agentic interaction environment for AI models and agents. Google’s TPU (Tensor Processing Unit) is Google-designed GPU hardware for ASI training and inference workloads, available to run such workloads on the Google Cloud Platform (GCP). VAST says that enterprise-level customers can now “seamlessly connect clusters running in Google Cloud and on-premises locations, eliminating complex migrations and making data instantly available wherever AI runs.”
This is an important point. VAST and Google have connected US TPU and Japanese GPU processing clusters, more than 10,000 kilometers apart, by using VAST’s DataSpace. This setup “delivered seamless, near real-time access to the same data in both locations while running inference workloads with vLLM, enabling intelligent workload placement so organizations can run AI models on TPUs in the US and GPUs in Japan without duplicating data or managing separate environments.”
VAST co-founder Jeff Denworth said: “Through our partnership with Google Cloud we’re meeting customers where they are with a fully managed AI OS. By extending our global namespace with intelligent streaming, Google Cloud customers can auto-deploy a VAST-managed cluster via Google Cloud Marketplace and start production in minutes, delivering integrated governance and billing, elastic scale, and full operations handled by VAST, making enterprise data instantly usable for agentic AI.”
This VAST Support for GCP uses technology from VAST’s acquired Red Stapler business. Data is fed to TPU VMs over validated NFS paths with optimized model-load and small-file/metadata-aware I/O.
VAST says that, in testing with Meta’s Llama-3.1-8B-Instruct model, AI OS, connected to TPU virtual machines in GCP, “delivered model load speeds on par with local disks, while maintaining predictable performance during cold starts.” In a little bit more detail we learn that that the setup achieved warm-start load times on par with local NVMe, but not during cold starts, where there was “predictable, steady behavior. The data load time, if slower, was, at least, consistent.
Blocks&Files notes that there is no Nvidia GPU Direct facility for GCP TPUs. However, GCP does have Hyperdisk ML, a block storage service optimized for AI inference/serving workloads, and caching capabilities in Cloud Storage FUSE (filestore) and Parallelstore (parallel filesystem), which improve training and inferencing throughput and latency.
Think of VAST’s AI OS as maintaining a virtual central datastore, streaming subsets of data to the locations where AI Models are executing on Nvidia GPUs or GCP TPUs, either on-premises or in the Google cloud or both. There’s no replication or copying of full data sets across the network linking sites. VAST says its customers “can run production AI workloads on Google Cloud today against existing on-premises datasets without migration planning, transfer delays, or extended compliance cycles.”
Customers can choose what data to migrate, replicate, or cache to Google Cloud, while keeping a single namespace, with consistent governance and compliance by applying unified access controls, audit, and retention policies everywhere.
VAST can be deployed today in Google Cloud. Joint validation and reference guidance for establishing a VAST DataSpace spanning Google Cloud and external clusters are available to qualified customers and partners.
Comment
This ability to supply data from a virtual central location to data centers, distributed globally, is a feature of Hammerspace’s Global Data Environment. This supports the inclusion of a GPU server’s locally-attached storage drive, known as Tier Zero, and therefore the equivalent of VAST Data’s “local disks.”
Arcitecta also has the capability to stream AI-relevant data to distant data centers with its Mediaflux Real-Time offering. VAST, Hammerspace and Arcitecta are competitively colliding in this globally distant, intelligent AI data streaming market.
Where VAST has as an edge is the Google TPU and managed service support, as well as momentum with NeoCloud GPU server clouds, led by CoreWeave and private GPU clouds such as X/AI’s Colossus. We think that VAST is intent of setting up similar fully-managed AI OS services with AWS and Azure to form an AI OS public cloud service trifecta, with simultaneous support for enterprise customers running hybrid AI environments across the on-premises, AWS, Azure and GCP environments.
We might imagine that, if CoreWeave developed AI data service offerings, it could be included too. An AI public cloud is a public cloud if it is a NeoCloud too.
