Hammerspace at AI Field Day 8: Unification, Not Centralization

Disclosure: I attended AI Field Day 8 as a Tech Field Day delegate. The views here are my own.

The foundation piece laid out the capabilities a distributed filesystem brings to AI work. The clearest way to read Hammerspace’s AI Field Day 8 session is as an argument about one of them: where your data should live and whether you should move it at all. Most storage pitches start with a box: speed up your data, buy a faster system, move everything into it, and the problem is solved. Hammerspace argued the opposite, and the reason it can make that argument now comes down to scarcity.

The session opened with the current reality: you can’t buy your way out of a capacity problem like before. Power has lead times. GPUs are quoted in years. SSDs are scarce or priced out of reach, a point the presenter noted other storage sessions had made. Even cloud providers now cap how much customer workload they accept because they need capacity for their own AI. When you can’t add resources on demand, the move is to get more out of what you already have. That framing ran through the talk and explains why Hammerspace built the session around customer case studies rather than a bits-and-bytes architecture tour.

The presenter drew a distinction worth holding onto: unification is not centralization. Centralizing means wheeling in a rack of drives and copying everything into one repository, which used to simplify life but is now nearly impossible. Unifying means giving everyone one access point to data that stays where it already lives. Hammerspace enters an environment, presents unstructured data scattered across NAS, object storage, and commodity hardware as a single namespace without moving any of it, then orchestrates movement by policy. It accelerates data into GPUs through its parallel file system or extends the namespace into the cloud only when something actually needs to move.

The mechanism is pNFS on NFS 4.2. A shop running Linux can get parallel-file-system performance without installing a proprietary client. This separates it from HPC file systems that require an agent on every node. Hammerspace stays out of the data path, so the client talks directly to the storage. Regarding whether a namespace can span what you have without forcing a migration, the answer is direct: you build on top of what is already there. This avoids new silos, copy sprawl, and idle GPUs waiting on a migration that the buy-a-box approach creates.

Tier Zero is the part of the platform I was most interested in. The session and questions around it were more candid than a benchmark slide would have been. Tier Zero takes the NVMe local to your AI servers — capacity that usually sits stranded and unused for shared work — and folds it into the global namespace as a fast, low-latency shared tier. The pitch is performance and efficiency at once because the storage is physically next to the GPUs and you use drives you already paid for instead of buying an external array. It does not have to consume the whole drive either. A delegate asked whether Hammerspace takes the local drives over completely, and the answer was you can partition and leave capacity behind.

The more useful exchange was about what Tier Zero is and is not. Asked where it fits relative to GPU memory tiers, the presenter described it as semi-ephemeral — a fast part of the shared file system where you tier data on and off, rather than a permanent home you read and write continuously. It is not, today, a memory extension or KV cache, though that is being explored. An application-side delegate suggested it sounded like a fit for semantic caching, and the presenter agreed it might be while being open that they are still learning how customers will use it. That candor matters more than a throughput number. Hammerspace's published MLPerf results claim up to roughly ten times the performance of external storage. Still, the session stayed on customer outcomes rather than benchmark figures, which is the right instinct.

The questions did the work of locating the honest edges of the pitch, which is what a good Field Day room is for.

The sharpest came from a delegate asking who the product is not for. The answer was direct: you need a big enough problem to require Hammerspace at a petabyte-to-multi-petabyte scale, in environments sophisticated enough to have storage administrators and capable AI staff. The presenter used a phone analogy and was honest — this is closer to an Android than an iPhone, powerful once configured but expecting you to do some configuration to get the full benefit. Another delegate pressed on the tension that the smaller end of the market has just as much storage pain and more storage diversity, with less competition serving it. The presenter acknowledged this rather than spinning it. The platform’s value rises with the complexity of the environment, and so does the expertise it requires.

A delegate with a management background asked how to reduce data movement and the operational burden on teams without dedicated storage staff. The answer started where the session began: do not move the data at all. Find out where it is and what it is, unify it in place, and move only what genuinely needs to move. The presenter was clear about the limit because some movement is unavoidable — bits over a wire obey the laws of physics, and no magic erases that. The mitigation is automation and a natural-language interface in the AIDP product. You describe the project and the outcome you want, and the system carries out the data work. This aims to make it approachable for someone who is not a storage specialist.

Deployment came up the same way. Standing up Tier Zero alone is straightforward: you point Hammerspace at existing shares, and it manages them. Fully assimilating an environment — taking over both reads and writes — takes longer, depending on how much data must move. On cost, a delegate asked what lies behind the multi-million-dollar savings on one slide. The answer was concrete: deployments can start around $50,000, but most customers engage at the million-plus level for multi-petabyte problems. Pricing is based on data under management.

A few topics were deferred to the demo or to future sessions rather than answered in depth — what Flex Files adds on top of pNFS, how you discover unstructured data in the first place, and how data sovereignty rules are enforced. Hammerspace has core NFS contributors on staff and clearly has depth on all three; the deferrals mostly mark where a single session runs out of room.

The most forward-looking moment did not come from the deck. A delegate connected the surge in Claude-based agentic coding this year to a pattern they were watching — teams running many systems at once, spinning environments up and tearing them down within hours for point solutions, and hitting a storage wall as they do so. The presenter was visibly interested and wanted more of those conversations, but it is worth being precise about what that means. The case studies Hammerspace actually showed were classic AI infrastructure: a semiconductor company’s training cluster sitting behind a patchwork of NAS, an HPC shop adding Tier Zero alongside its existing parallel file system, a team that bought object storage and then needed file semantics for its data scientists, a NeoCloud unifying three data centers for inference, and a customer bursting from Azure to a NeoCloud for more GPUs. The agentic-development workload, the foundation piece described as the bleeding edge, was the audience’s contribution, not a deployment on the slides.

That gap is the honest read. Hammerspace’s architecture maps cleanly onto the agentic case — keep the repository and its caches in one namespace, place them near the compute, move them live as the work shifts across clouds — but the demonstrated traction is in training and inference, with agentic development still an emerging conversation rather than a shipped story. For a team feeling the repo-scale version of this problem, that is the thing to probe in a proof of concept.

Against the capability set from the foundation piece, Hammerspace leads on the namespace itself, spanning sites, clouds, and vendors without a migration, and on policy-driven orchestration that moves data only when needed, including live, across on-prem and multiple clouds. Tier Zero is its proximity answer, with the honest caveat that the team is still mapping its best uses. The standards-based pNFS client keeps you off a proprietary path. What you take on in exchange is everything the Q&A surfaced: it is built for large, sophisticated, petabyte-scale environments; it expects capable hands to configure it; it is priced for problems of that size; and it does not pretend that data movement is free.