At the 70th IT Press Tour in Silicon Valley, Hammerspace Vice President of Product Marketing Floyd Christofferson described this familiar problem as one reason enterprise AI projects struggle to progress beyond initial testing.
In his presentation, he referenced a Berkshire Hathaway group CTO who had suggested that around 70% of enterprise AI projects fail before reaching production. That figure came from a conference discussion rather than a published study, but Christofferson's explanation of the problem was interesting.
Companies have spent decades building storage infrastructure around individual applications. But AI now needs information from across those systems, often in different locations and clouds. Hammerspace believes businesses should reconsider how they access existing data before committing to another expensive storage deployment.
The storage problem that AI has inherited
Most leaders will admit that they have accumulated too many storage systems over the years. Finance might use one platform, engineering another, and research teams a third. Some information lives in the cloud, while older records remain on equipment purchased years ago. Welcome to the very real effects of technical debt.
Each system may work perfectly well for its original purpose. Problems arise when an AI application needs data from several of them. Companies purchase a new storage environment designed for AI and copy the relevant information into it. The project gets access to its data, but the business now has another storage system to maintain, another set of files to protect, and potentially another version of information changing elsewhere.
Vanderbilt found nearly 10PB of storage it already owned.
The most convincing example from the presentation came from Vanderbilt University's Advanced Computing Center for Research and Education. The facility had approximately 40 petabytes of storage spread across different systems, including high-performance equipment, archive capacity, and storage installed inside its GPU servers.
Researchers needed additional fast storage, but funding was limited. Hammerspace discovered that a substantial amount of existing capacity was sitting inside individual compute servers where it could not easily be shared.
Using its Tier 0 technology, Hammerspace brought that local NVMe storage into a shared environment. Christofferson said the university recovered nearly 10 petabytes of capacity that had previously been isolated across its servers. That changed the purchasing calculation. Instead of acquiring another large pool of expensive flash storage, Vanderbilt could use capacity it had already paid for.
Researchers could also allocate resources to different projects without repeatedly reorganizing the underlying hardware. For a university handling changing research requirements and grant-funded projects, that flexibility has practical value.
Hammerspace has reported storage cost reductions of approximately 48% in the Vanderbilt deployment. The example raises a question worth asking in any organization preparing an infrastructure budget. How much expensive storage is already installed but unavailable to the applications that need it?
Why local GPU storage deserves another look
GPU servers often contain high-speed NVMe drives. They are useful for workloads running on the individual machine, but their capacity can remain isolated when hundreds of servers operate together.
Hammerspace wants to make that storage available across the wider computing environment. Christofferson presented a comparison involving a 500GB AI checkpoint, the process of saving a model's progress so training can resume after an interruption.
In Hammerspace's tests, writing that checkpoint to local NVMe took approximately 4.5 seconds. Sending it across a 400Gb network took around 11.1 seconds. The difference shows why data location matters.
Hammerspace's approach keeps frequently accessed information close to the processors, while less demanding data can reside on conventional flash, hard drives, or cloud storage.
Three clocks that change how AI costs are measured
One of Christofferson's more interesting ideas concerns how companies measure AI performance. The industry frequently discusses time to first token, which measures how quickly an AI model starts generating a response.
Hammerspace argues that organizations should consider two additional measurements.
INSERT CLIP
What businesses should consider before buying more storage?
The Hammerspace presentation raised a practical issue that also deserves attention as AI budgets grow. Businesses may already own much of the storage capacity they need. But the challenge spans disconnected systems, locations, and applications, making it expensive to access and manage.
Vanderbilt's experience shows what can happen when previously isolated capacity becomes available to a wider group of users. Other examples discussed during the briefing involved organizations connecting existing data to GPU resources in public clouds while controlling where information was stored.
Before signing off on another storage purchase for AI, perhaps the first conversation should be about what the business already owns, where its information resides, and why existing applications cannot access it. This was the part of the Hammerspace presentation I found most relevant to businesses.
There may be considerable value sitting inside infrastructure that has already been paid for.
I will be speaking with the Hammerspace on my podcast. If you have any questions you would like me to ask on your behalf, please let me know, and I will bring them into the conversation.
Special thanks to Alex Zaharov-Reutt, iTWire's Global AI and Technology Editor, for filming the presentation and allowing me to use the footage.