Hammerspace Inc., an eight-year-old startup that provides high-speed access to distributed data, has introduced an artificial intelligence-focused data platform that prepares and delivers data to AI systems without requiring large-scale data migrations or new storage infrastructure.

The company announced general availability of its AI Data Platform at the Nvidia GTC conference this week, positioning the offering as a turnkey solution for enterprises struggling to move AI projects from pilot to production. The platform is built on an Nvidia reference architecture and designed to automate the discovery, preparation and delivery of enterprise data to graphics processing units for AI workloads.

“The primary obstacle we are hearing about is having enough data or access to the data and being able to govern it properly,” said Molly Presley, senior vice president of global marketing at Hammerspace. She said enterprises often have data scattered across business units, storage systems and cloud environments, making the process of preparing it for AI time-consuming and manual.

“Enterprises have different requirements for what they want to feed into a training model and what they don’t,” Presley said. “That’s been manual, and it’s slowed things down.”

The new platform is designed to address those bottlenecks by automatically identifying enterprise data wherever it resides and making it available for AI processing without first copying it into a separate storage environment. By leveraging data in place, the technology eliminates the need to purchase large amounts of new flash storage, the company said.

The system also monitors enterprise storage environments for changes and automatically prepares new files for AI pipelines. It uses a Model Context Protocol server to coordinate with AI tools and applications so only needed data is moved.

Near-real-time processing

“From the second a file lands, we detect it and process it,” said Sam Newnam, vice president of AI and business development at Hammerspace. “We’re only talking seconds from the file creation or file change.” The system doesn’t process streaming data but can catalog data as soon as it’s committed to a file.

Hammerspace said the platform works with existing storage systems and moves only the data that is needed to GPU resources.

The approach also addresses a growing issue of limited GPU availability in enterprise AI deployments. By some accounts, the average enterprise uses less than 30% of available capacity of costly GPU’s.

“Having the ability to automate the placement and movement of data wherever it is, is a big problem they’re looking to solve,” Presley said.

Hammerspace’s architecture creates a unified metadata layer that spans multiple storage systems and allows the platform to automatically locate relevant files and move them to computing resources when required. It works with existing enterprise security controls and file permissions.

“We’ve got a lot of tooling built in to make sure that we adhere to the same security primitives that exist on original data,” Newman said.

Presley described the system as a “turnkey automated AI data pipeline. Once it’s up and running, it is continuously updating the pipeline to make it easy for whoever’s using data to have a single place to find it.”

Image: rawpixel/Freepik