LONDON – AI storage vendor Vast Data is partnering with Google Cloud on a hybrid cloud solution to meet rising demand in AI inference and training.

The deal develops Vast’s existing relationship with the hyperscaler by deploying Vast AI Operating System (AI OS) as a fully managed service on Google Cloud Platform (GCP). The service has been expanded to operate a single addressing instance across on-prem and cloud, as well as AI-native cloud provisions from neocloud vendors.

In an exclusive media briefing, Vast CEO Renen Hallak told SDxCentral the deal was the firm’s first hyperscaler partnership. While Vast’s Data Platform is available on Amazon Web Services (AWS) and Microsoft Azure, its relationship with Google goes one step further, providing Vast access to its tensor processing units (TPUs), as well as leveraging Vast’s AI OS, which was launched this May.

The operating system extended on Vast’s flagship data platform by evolving its storage system base into a full operating environment for AI workloads, meaning GCP users can connect clusters running in the platform with other forms of distributed infrastructure.

“It’s the first time that we're using their TPUs, which is a big part of it,” Hallak explained. “We're using those and enabling abilities within GCP and in hybrid environments for Google customers to use our OS.”

The CEO added that customers demand the same experience across various environments as part of their hyperscaler setups.

“They're asking the hyperscalers for the same, and when the customers demand it, then we have to do it, and they have to do it, and so we start building together.”

Working alongside Google, Vast touted a demo using Vast DataSpace, a core component of its data suite that provides a single, global namespace for data stored across multiple locations.

Vast claimed the demonstration connected clusters more than 10,000 kilometers (6,213 miles) apart, with one in the U.S. and another in Japan. This potentially provided near real-time access to the same dataset in both locations without managing separate environments, supporting inference workloads with vLLMs (virtual large language models) where AI models are run on U.S. TPUs and Japanese GPUs without any duplication of data.

In another test with Meta’s Llama-3.1-8B-Instruct model, Vast claimed its OS delivered model load speeds comparable with local NVMe disks, while maintaining consistent performance during cold starts, referring to an initial delay that occurs when an AI model is deployed after a period of inactivity.

This enables users to serve TPU virtual machines through validated network file system (NFS) connections, benefiting from optimized model loading and small-file or metadata-sensitive input/output, as well as warm-start load speeds one would expect in an NVMe capacity.

“We're definitely architecting for this new AI era of training and inference… All of these things run on Vast in ways that they can't run on old infrastructure," said Hallak, who pointed to financial services and life sciences as leading the demand in inference and training.

GCP users will also be able to select which data to migrate, replicate, or cache according to their needs, while using DataSpace to streamline compliance processes and lower operational risk.

With its first hyperscale partner in Google, Vast is adding to its current ecosystem of neocloud vendors such as Crusoe, GMI Cloud, and Nebius Group.

This week, the company also signed a $1.17 billion deal with CoreWeave to supply storage and software to the neocloud's AI Cloud infrastructure.

“We love those guys,” Hallak enthused. “CoreWeave, they're the biggest AI cloud. They're very aggressive. They're servicing all of these new AI startups, as well as now starting to service enterprises for their production workloads. Customers like those keep us ahead of everybody else.”