Most of the noise about AI infrastructure is about GPUs, and the limit operators hit first is usually power. A data center has a fixed power budget, and every watt that goes to storage and networking is a watt the GPUs don't get. Storage sized for AI is now a large share of that draw.
Inference is now adding pressure in a new place. As models take on longer context and agents chain many steps together, the working state of a session outgrows GPU memory. Operators are starting to keep computed context, such as KV cache, on flash where inference software can retrieve and reuse it rather than asking GPUs to recompute it. Fast flash storage, close to the cluster, is becoming a standard part of an AI build.
The server in the middle
In most flash storage today, each node is a general-purpose x86 server with drives attached. The server runs the storage software and handles the network traffic. It also adds a hop to the data path, along with its own processors, memory, power draw, and rack space. Multiply that across a storage tier sized for AI and the overhead adds up.
The Open Flash Platform initiative, which Hammerspace and Xsight Labs launched in 2025 with Los Alamos National Laboratory, SK hynix, ScaleFlux, and the Linux community, asks whether the server needs to be there. Linux and NFS have matured enough that storage services can run on a DPU next to the flash, as long as the DPU has the compute and network bandwidth to keep up.
From initiative to product
Today Hammerspace introduced the Hammerspace Data System, and with it Hammerspace OFP, a flash storage system for AI training and inference. Within the Hammerspace Data System, OFP provides high-performance flash storage that can be managed alongside existing storage in a unified global namespace. Hammerspace OFP runs on the Xsight Labs E1 in place of the x86 server. Hammerspace says the design delivers more performance in less rack space, with lower hardware cost and infrastructure overhead.
In a more modern architecture, the E1 eliminates the need for a separate general-purpose x86 storage server, improving performance and power efficiency. Most DPUs grew out of a network card, with a fixed pipeline and a few Arm cores off to the side for the work the pipeline can't handle. The E1 puts 64 Arm Neoverse N2 cores in the data plane, with 800G of Ethernet, and runs standard Linux. Hammerspace's software runs on it the way it would run on a server, and the flash connects straight to the network.
What to look for
As infrastructure teams plan the next generation of AI storage, they need to look closely at what sits between the drives and the network, because it sets the power, the rack space, and the number of hops between a GPU and its data.
We see as much interest in the E1 for storage applications as for network infrastructure, and Hammerspace OFP proves the benefit of that interest. Details on Hammerspace OFP are available at hammerspace.com. E1 specifications are at xsightlabs.com.
