Hammerspace Now Sells Its Own Hardware, but Won't Publish Inference Numbers Yet

Hammerspace Now Sells Its Own Hardware, but Won't Publish Inference Numbers Yet

BackerLeader ●58 ●371 ●552
calendar_today ago • schedule4 min read

Hammerspace has spent years selling software that runs on someone else's servers. Now it sells the hardware too.

On Tuesday, the company announced the Hammerspace Data System. Customers can now buy it as one package: Hammerspace software, metadata nodes and Hammerspace's own flash storage, installed and supported by Hammerspace. The software-only option on third-party hardware or in the cloud isn't going away.

"Nothing changes for existing Hammerspace customers," said Eric Bassier, senior director of product marketing at Hammerspace. "The software remains the same whether deployed on Hammerspace hardware, third-party servers, or in the cloud, with all environments supported and managed as a unified system."

He said the change came from customers. "In particular, companies building new storage infrastructure for AI want the convenience of purchasing it as a system, with comprehensive support."

The new storage is Hammerspace OFP. It's named for the Open Flash Platform initiative Hammerspace has led since July 2025. The design removes the x86 storage server from the data path. High-density flash connects directly to the network through DPUs, in this case the Xsight Labs E1, and metadata runs on a separate path. Hammerspace OFP is available to quote now and ships in the first half of 2027.

Read the performance numbers carefully. The launch announcement claims 390 GB/s of write throughput, 365 GB/s of read throughput and 15 million IOPS per rack unit. Hammerspace says that's 4x, 2x and more than 2x the nearest competitor. The announcement doesn't name those competitors or describe the test setup.

Bassier gave me a different figure, based on preliminary single-sled measurements: "70 GB/sec write throughput per sled, so 350 GB/sec write throughput per 1U OFP tray." He said results will vary by deployment, and Hammerspace plans to publish fuller figures in the first half of 2027. Until then, none of these numbers is a full-system result.

KV cache without a custom connector

The piece most inference teams will care about is Tier 0. It pools the NVMe drives already sitting in GPU servers into shared storage for context, including KV cache. If an inference engine can reuse context it already computed, it doesn't have to compute it again.

"Tier 0 makes server-local NVMe available as shared storage across the cluster," Bassier said. "KV blocks are accessed locally when present on the consuming server, or over the network when stored elsewhere."

The integration story is simple. Hammerspace verified storage and retrieval in its lab through the existing file backends of LMCache, SGLang HiCache and NVIDIA NIXL. Developers point those backends at a Tier 0 share. "No Hammerspace-specific connector is required," Bassier said.

But that's as far as the testing goes. Hammerspace has completed functional testing. It hasn't finished end-to-end inference validation on GPUs, so it has no numbers yet on what Tier 0 does to GPU utilization or cost per token.

Bassier was direct about why. "The baseline is physics," he said. "When KV blocks are on the server that needs them, Tier 0 reads them from local NVMe with no network in the path, and that is faster than reading them from external storage. How much that improves GPU utilization and cost per token has to be measured in full inference runs on GPUs, which is the next step."

He also had a warning about how other vendors present these numbers. "I would press on claims like '20x faster time to first token,'" he said. "Many compare a 100 percent cache hit on a very long context against recomputing everything from scratch, which is best-case and not a production workload."

His advice is to ask which model was tested, at what context length, at what cache hit rate, with how many concurrent users and against what baseline. Then run it on your own workload. "That applies to us too," he said, "which is why we are not publishing inference numbers until we have completed real inference benchmarks on GPUs and can publish the model, workload, and configuration alongside the results."

Natural language queries with an approval gate

The other new piece is Data Intelligence. It catalogs and queries data across sites and clouds, and teams can turn the results into orchestration policies. Users can ask questions in natural language.

Letting an AI-assisted query trigger data movement raises an obvious question: what stops it from tagging or moving the wrong files? Bassier said the default setting blocks any action until someone with the right permissions approves it. A data user can query and explore. If that user asks to set a policy, it goes to the system administrator for manual approval first.

That's a default, not a hard rule. Teams should check that it stays on and decide who owns the review.

The Lustre case

Hammerspace points to a hyperscale research lab that replaced Amazon FSx for Lustre with a 30 PB Hammerspace cluster in AWS. According to Bassier, Lustre generated more support tickets and friction for the lab's AI researchers than anything else in the cluster.

The lab first cared about standard NFS access with no proprietary client. In testing, Hammerspace showed faster checkpoint loads, faster distributed saves, and faster large-file and list operations. The lab is now starting to move workloads to other data centers, which Bassier said "is much more difficult with Lustre." Hammerspace hasn't named the lab.

Hammerspace isn't for everyone. Bassier said a small, single-site team with basic NAS needs and no performance pressure "is well served by what it has today." The fit is strongest where GPU workloads need data that's spread across storage systems, sites and clouds.

There are two things to watch. One is GPU inference benchmarks for Tier 0, published with the model, workload and configuration. The other is full-system OFP numbers when the hardware ships next year. Hammerspace has said it will publish both. Bassier's own checklist is a good way to judge them when they arrive.

Want more #aiinfrastructure & #storage deep dives?

Join 5,953+ developers sharing real benchmarks, code breakdowns, and projects. Free forever.

More Posts

Local-First: The Browser as the Vault

Pocket Portfolio - Apr 20

Breaking the AI Data Bottleneck: How Hammerspace's AI Data Platform Eliminates Migration Nightmares

Tom Smithverified - Mar 16

AAIF Hits 10 Projects. Its CTO on What to Adopt Now and the Security Gaps Left Open

Tom Smithverified - Oct 23

What Developers Already Know About Data Center Delays

Tom Smithverified - Sep 28

Fine-Tuning and Inference in LLMs: From Custom Models to Production Deployment

Derrick Ryan - Jan 13
chevron_left
21k Points • 981 Badges
289Posts
171Comments
143Connections
LLM Training & Evaluation Specialist with hands-on experience building major AI models. As one of th... Show more

Related Jobs

View all jobs →

Commenters (This Week)

1 comment
1 comment
1 comment

Contribute meaningful comments to climb the leaderboard and earn badges!