How Parallel NFS Fuels Enterprise AI
Hammerspace Inc. just closed a $100 million Series B funding round led by Altimeter Capital, with participation from ARK Invest. We believe this sizable raise – following a $56 million round in 2023– signals growing investor confidence that solving data bottlenecks is a key stepping stone to building out AI infrastructure. Hammerspace, a startup founded by storage veteran David Flynn, is positioning itself not as another storage vendor but as a “global data platform” for the AI era. It’s mission to date has been somewhat confusing to observers and the company aims to focus its messaging on accelerating AI across all parts of the value chain.
To that end, the company’s technology aims to let enterprises access and move data wherever it’s needed, on demand, without creating duplicate copies. In the words of Hammerspace co-founder and CEO David Flynn: “How can you have your data in every place that you need it without ever having copied it?”
This research note is based on an analyst call with key Hammerspace executives and investors. We analyze the impact of the new funding and the company’s evolving market position with a critical eye – drawing on an exclusive theCUBE interview with CEO David Flynn by John Furrier and additional analysis – to understand where Hammerspace fits (and doesn’t fit) in today’s AI and data infrastructure landscape.
Big Money for Big Data Problems
The $100M round is a strong endorsement for a company in the traditionally hard-to-crack storage arena. Altimeter’s lead and ARK Invest’s participation (ARK also backed the 2023 round underscore that investors see Hammerspace as aligned with the explosive AI trend. ARK in particular is known for betting on disruptive tech, and here the bet is that AI’s next bottleneck is data access, not just model size or GPU count. Altimeter for its part is a “full cycle” investor, from startup to public company (e.g. Snowflake, Confluent, MongoDB). Hammerspace’s customer list – which it previously disclosed includes Meta (Facebook), the U.S. Department of Defense, and the National Science Foundation – gives further credence that some major organizations are leaning into its approach. These are marquee users with massive data needs, suggesting Hammerspace has proven its value in high-performance and large-scale environments.
Why is data access such a critical issue? In AI projects, especially deep learning and analytics, fast GPUs can sit idle waiting for data. As the we’ve observed, a cluster that can process 10 GB per hour but only retrieves 8 GB in that time is underutilized. In other words, the throughput of feeding data becomes the choke point. This has made storage and data management a priority in machine learning infrastructure. It’s not as flashy as model architectures or LLMs, but without rethinking how data flows to AI compute, companies won’t get the ROI from expensive hardware. Moreover, data is distributed. While much or most of AI today is being done in the cloud, over time, enterprises will want to apply AI to their proprietary data sets. These are problems space Hammerspace targets, and it helps explain why a storage-oriented startup can nab a nine-figure investment in 2025. The funding will fuel Hammerspace’s international expansion and presumably accelerate product development to stay ahead in the race to eliminate AI data bottlenecks.
The Hammerspace Magic: Global Data Orchestration
So what exactly does Hammerspace do differently? In a nutshell, Hammerspace’s software creates a unified data environment across disparate storage systems and locations, with emphasis on performance and low latency. The company uses the term global namespace – something others like Isilon have used in the past. The difference is Hammerspace approaches data access with very fine-grained granularity, and moves small files and objects as needed so that they are proximate to the GPU. According to a Hammerspace premise, the only thing with greater gravity than data is GPUs.
Hammerspace effectively abstracts multiple file systems and cloud storage buckets into one big namespace that applications can access. Under the covers, Hammerspace orchestrates how and where data is stored and retrieved, but to users and AI workloads it looks like a single pool of data. This addresses a bane of enterprises: data sprawled across on-premises NAS, cloud object stores, edge devices, etc., often requiring expensive copying or manual data wrangling to get the right data to the right place.
Hammerspace’s secret sauce is making this global data orchestration fast. A core design point is parallelism – instead of reading from one storage system at a time, Hammerspace can retrieve data from multiple storage arrays simultaneously, maximizing throughput. This parallel I/O capability is crucial for feeding “elephant” sized AI pipelines. It effectively treats all available storage endpoints as a collective source that can be tapped in parallel, rather than sequential silos. In high-performance computing, parallel file systems have long done this, but Hammerspace is bringing a similar idea to enterprise data by using NFS, a ubiquitous standard format.
Notably, Hammerspace built its platform on Linux’s built-in NFS capabilities. NFS is decades-old (dating to 1984) and widely supported, but Hammerspace claims to have enhanced it for today’s modern needs. By using NFS, Hammerspace avoids forcing customers to install proprietary client software or modify their applications. Competing solutions often require special software on each server or each GPU node to tap into their storage cluster. Flynn emphasized in his conversation with theCUBE that eliminating those extra steps was a priority – companies don’t want to refactor code or maintain new client software just to improve data access. Hammerspace’s approach means any Linux-based system (which is virtually all AI infrastructure) can use Hammerspace’s parallel data layer natively. In our view, this approach facilitates integration, in particular because the platform deploys in without agents, appearing as just another standard file mount, yet behind that mount it’s pulling data from everywhere.
Performant NFS? Really?
An obvious concern is how does a forty year old protocol like NFS deliver cutting-edge performance? Hammerspace claims to have made key and novel optimizations. For example, traditional NFS setups funnel traffic through a single NAS server that can become a bottleneck; Hammerspace says it eliminates that extra hop by letting clients talk directly to storage endpoints. It also enables data transfers to bypass CPUs on the storage path when possible, reducing latency – a concept similar to emerging techniques like NVIDIA’s GPUDirect Storage in spirit (moving data on the fastest path to GPUs). The result, Hammerspace claims, is that applications (like AI training jobs) can pull data from wherever it resides, at speeds previously hard to achieve without building a custom HPC filesystem. “We orchestrate data to the GPU faster regardless of where it is physically stored,” Flynn said, underscoring the mission. “We instantly assimilate data from third-party storage so it’s ready to process faster. We deploy and scale easily and quickly so our customers can achieve their business outcomes faster.” In effect, Hammerspace acts as an overlay across existing storage that accelerates and simplifies data access.
NFS Purpose-Built for High Performance
To clarify, this implementation is parallel NFS or pNFS. Long time Linux maintainers from firms like NetApp and Red Hat helped co-author the latest pNFS v 4.1 and v4.2 specs and their implementation. According to technical documentation, Git commit logs and our research of publicly available information, significant portions of the v4.1 and 4.2 client and server-side code bases were contributed by developers who became Hammerspace employees and continued their work (e.g. Trond Myklebust, Hammerspace CTO, et al.). Parallel NFS (pNFS) is an extension to the NFSv4 standard that breaks the classic “single‑head” bottleneck by separating metadata management from data placement. In practice, the client still talks to one or more metadata servers for namespace and security, but receives a layout map that lets it read and write data directly—and in parallel—to multiple data servers. Our inference is that this architecture delivers near‑linear throughput scaling and dramatically higher concurrency, because I/O paths fan out across many back‑end nodes rather than funneling through one controller.
The point is, Hammerspace didn’t merely adopt pNFS—they co-authored the NFS v4.2 pNFS specification in a working group which developed the reference implementation. While the working group defined the reference implementation, most of the upstream Linux code was written by contributors now at Hammerspace. Version 4.2 added more flexible “layout types” for file, object, and block storage, fine‑grained sparse‑file and copy‑offload commands, and richer snapshot and ACL semantics, making it a more cloud‑ and container‑ready version of NFS. By contributing to the spec and the code, Hammerspace can tune performance, add data‑services hooks, and upstream innovations without waiting on third‑party roadmaps. In our view, that tight coupling of protocol stewardship and product execution positions Hammerspace to be a leader the next wave of scale‑out file systems where latency, parallelism, and data mobility are table stakes.
Eliminating “Copy Creep”
Another aspect of the secret sauce is the so-called zero-copy architecture that Hammerspace touts. Typically, if a workload in one data center needs data from another, you’d copy the dataset over the network, leading to multiple redundant copies stored in different places. Hammerspace’s global file system is designed to avoid that – i.e. it lets the remote data be accessed in place, without creating duplicate copies, saving time and storage costs. Flynn likes to frame this in paradoxical terms – i.e. how to have your data everywhere without copying it. By virtualizing file access, Hammerspace can present data that lives in New York to a job running in San Francisco, at high speeds. This addresses data gravity issues in large organizations – instead of moving petabytes around or maintaining synchronized silos, Hammerspace aims to make one authoritative copy accessible from anywhere (with caching and clever behind-the-scenes movement of small objects and files to optimize as needed).
It’s important to note Hammerspace is software-only and hardware-agnostic. It doesn’t sell storage devices; it leverages whatever storage an enterprise already has (flash arrays, cloud storage, GPU node local NVMe, etc.). In theory, this means Hammerspace can complement traditional storage systems by federating them rather than replacing them outright. Enterprises can layer Hammerspace on top of, say, their NetApp filers, Dell/EMC Isilon (PowerScale) clusters, cloud buckets on AWS and Azure, and treat it all as one giant file system. The underlying systems remain in place – Hammerspace acts as the platform to orchestrate their data. Flynn describes this as unlocking a level of scalability and global access that was previously not possible. In the interview with Furrier, he acknowledged the skepticism they face given the checkered history of such ideas: “People don’t believe us when we say that we make it work… We have to fight the ghosts of dead companies past because of this.” Indeed, past startups (and even Flynn’s own prior venture) attempted global data virtualization and fell short, leaving behind skeptical enterprise buyers with “scar tissue.” Hammerspace says it’s determined to prove this time is different – that a combination of modern cloud thinking and old-school NFS reliability can finally crack the code on ubiquitous data access.
Complement or Compete? Positioning in a Crowded Landscape
Hammerspace’s value proposition straddles the line between augmenting existing storage and disrupting it. On one hand, Hammerspace clearly complements traditional file storage – i.e. it needs those storage endpoints to virtualize. Rather than storing all data itself, it indexes and orchestrates data across other systems. In that sense, it can make legacy storage more useful – for example, allowing an old NAS appliance in one office to participate in a global data fabric alongside cloud storage, without forklift upgrades. Hammerspace can give a second life to existing infrastructure by abstracting it into a modern, AI-friendly data layer.
On the other hand, by inserting itself as the global control plane for data, Hammerspace inevitably competes with features provided by storage vendors and even cloud providers, which will create “enemies.” If Hammerspace handles replication, high-availability, and performance optimization across sites, a customer might rely less on the proprietary software features of their storage arrays. In some cases Hammerspace could even let organizations mix and match cheaper storage hardware underneath e.g. white box Linux servers and storge), rather than paying a premium for a single vendor’s top-tier system, because Hammerspace is providing the intelligence on top. It’s a similar dynamic to software-defined storage or virtualization in the past (i.e. Hammerspace is to storage what VMware was to servers)– the layer that abstracts resources tends to commoditize the resources underneath. This won’t be lost on traditional vendors, and it means Hammerspace in some deployments might replace the need for certain global file system products or cloud data management tools.
The AI Imperative Creates a New Competitive Dynamic
The AI angle intensifies this coopetition. Take high-performance file systems: companies like Weka, VAST, DDN, IBM with Spectrum Scale, and others offer their own solutions to feed AI workloads with massive throughput. Many enterprise IT shops also simply rely on cloud storage or big SAN/NAS systems for AI data. Hammerspace will need to demonstrate that its approach is simpler or more flexible than going with those more established options. The Linux kernel/NFS-based approach is a double-edged sword here – i.e. it’s accessible and standard, but some might question if it can truly outperform purpose-built, proprietary systems that were designed for the AI era. Hammerspace is essentially saying “we can achieve HPC-grade performance using open standards and software intelligence.” If true, it’s a compelling story – but it invites comparisons and skepticism from incumbents touting benchmark numbers.
Messaging is an area of potential confusion for Hammerspace as it navigates this landscape. Is Hammerspace a storage solution or a data management platform? The company’s own description leans toward a data orchestration platform, emphasizing unification of data silos and policy-based management, rather than calling itself a file system vendor. Yet, much of what it delivers (high-performance file access, a global namespace, NAS features like snapshots or tiering) overlaps with what many would consider the domain of storage and file systems. This can lead to confusion for customers – in other words, do they budget for Hammerspace from their storage infrastructure funds, or is it a data management/software purchase? Does the storage team own it, or the data architecture/cloud team? Hammerspace’s marketing around AI and global data is smart to tap into broader business outcomes (around speed to value across the cycle), but they must also assure the technical buyers (who often think in terms of performance, capacity and reliability) that it fits into their stack.
From what we’ve observed, the company is navigating this tension by framing Hammerspace as an enabler for the entire data lifecycle in AI projects, rather than a competitor to any one storage box. In theCUBE interview and the analyst briefing, Flynn and Furrier discussed how Hammerspace’s technology is a layer that could work with your existing storage investments and cloud, providing a missing capability (global, high-speed access) that none of those by themselves can tout. Hammerspace uses terms like “global data platform” and “data orchestration” to plant the idea that this is something new – not just another NAS. At the same time, as indicated, Flynn openly acknowledged that convincing the market involves overcoming the ghosts of past failures. It requires careful education to show why Hammerspace’s approach (e.g. leveraging the in-kernel NFS and their enhancements) is fundamentally different from prior attempts at global filesystems that struggled with performance or complexity. The recent Tech Innovation Award Hammerspace won for data storage innovation is more validation in their messaging (Disclosure: theCUBE Research and a panel of forty judges selected the winners, which paid a nominal fee to provide detailed submissions). But awards and funding aside, the company’s long-term success will depend on balancing that complement vs. compete narrative – i.e., assuring customers that deploying Hammerspace will play nicely with what they have, even as it potentially disrupts how they’ve managed data historically.
Gotchas: Hard Truths on the Road to AI-Ready Data
Despite the optimism around Hammerspace’s technology and the fresh $100M in the bank, enterprises should keep several “gotchas” in mind when evaluating this solution (or frankly, any similar data platform) for AI:
- AI Adoption Beyond Technology: A faster data layer won’t automatically make an organization AI-driven. Many enterprises still struggle with talent, data quality, and defining AI use cases. Hammerspace can accelerate data access, but companies need data science processes, data harmonization approaches (semantic layer) and strategies to truly capitalize on AI. It’s easy to buy infrastructure; it’s harder to foster an AI culture. This means Hammerspace is a piece of the puzzle – not a silver bullet that guarantees AI success by itself.
- Data Integration Complexity: Hammerspace’s promise is to unify disparate storage, but achieving that requires integration work. Companies have to connect Hammerspace to potentially dozens of data sources across cloud and on-prem, set up policies, and ensure security controls carry over. Heterogeneous environments are messy, and deploying a global data fabric will surface inconsistencies (differing file permissions, network latencies, etc.). In practice, the initial rollout may be complex and time-consuming. Organizations should be prepared for a careful planning and migration phase – it’s not as simple as flipping a switch to instantly federate everything.
- Blast Radius: Infrastructure Risk & Performance Validation: Introducing a new global data layer means adding a critical piece of infrastructure that everything else might rely on. This raises questions of resilience and risk. If Hammerspace were to go down or malfunction, could it choke off access to data across the board? Enterprises will need to architect for high availability and test failure scenarios (just as they would for any storage system). Additionally, performance claims need real-world validation. Hammerspace will be deployed in environments with unpredictable workloads – it must prove that it can consistently deliver the promised throughput and low latency, and not become its own bottleneck. Early customers and prospects should benchmark it with their data to avoid unpleasant surprises.
- Messaging and Positioning Confusion: As discussed, Hammerspace sits at the crossroads of storage and data management. Customers may initially be confused about what budget or team should own the project, and what exactly Hammerspace replaces or augments. Is it a backup for NAS? A cache? A primary storage system? The answer is that it’s a bit of all of those (and also neither, as it doesn’t physically store data long-term). Such nuance can be hard to communicate and might lead to internal friction on the customer side (e.g. storage admins vs. cloud architects debating solutions). Hammerspace and its buyers will need to align on clear use cases – e.g. “we need to feed our GPU farm from multiple data silos, and Hammerspace will handle that.” Without clarity, there’s a risk of overhyping and underutilizing the platform due to simple misunderstandings of its role.
In short, Hammerspace is not a panacea for all AI data woes, it’s an infrastructure building block. Companies still need to do the hard work of data preparation, governance, data harmonization, agent orchestration and aligning IT teams. The technology addresses a specific pain point – albeit a significant one – but it doesn’t eliminate the need for sound data strategy and architecture.
The Bottom Line: A Building Block for AI Infrastructure, Not a Silver Bullet
Hammerspace’s $100M funding round and the buzz it’s creating are testament to the growing importance of data orchestration in the AI era. As enterprises race to implement generative AI and advanced analytics, they are hitting the wall of data silo complexities and I/O bottlenecks. Hammerspace offers a compelling vision to grease the data pipelines, making data available wherever the compute needs it, quickly and efficiently. It has demonstrated technical feats in using standard technology (Linux NFS) in novel ways to achieve performance that impresses demanding, high-end users. The conversation between David Flynn and John Furrier revealed a confident CEO who believes his team has solved an “intractable” problem that has stymied others. When Flynn says they built “a truly unprecedented level of scalability in a very robust, feature-rich NAS system” (dubbed “hyperscale NAS”), it reflects a bold claim that Hammerspace can marry the scale of cloud with the reliability and features of enterprise storage. Flynn has a proven track record with FusionIO, and learnings from failed endeavors that he is applying to Hammerspace.
In our view, it’s critical to view Hammerspace as a building block in the bigger AI infrastructure puzzle. It doesn’t replace the need for good data engineering practices or make all other storage technologies obsolete overnight. In fact, Hammerspace’s success will likely ride on how well it works in tandem with existing systems – complementing them – while offering enough unique value to justify its overlay. In some cases it will compete at the edges with incumbent solutions, but its broader narrative is enabling something new – i.e. a global data environment for AI and analytics. That’s an intriguing prospect for enterprises that today spend inordinate effort wrangling data; and we would advise that enterprises include Hammerspace on the short list when building out your AI stack. If Hammerspace can execute, it could become an important data substrate for AI, analogous to how virtualization became a substrate for compute.
In Breaking Analysis we often convey that no technology is a cure-all; the winners are those that address real pain points and integrate into the fabric of how companies operate, with minimal friction. Hammerspace is targeting a very real pain point – the gap between data and increasingly distributed AI compute – with a solution that appears technically elegant and is well-funded. The next 12-18 months will be telling. With fresh capital, Hammerspace needs to educate the market, scale its go to market/customer base, and fend off both legacy vendors and upstarts vying for the same opportunity. Enterprise tech buyers should keep an eye on this space as it is moving quickly. At the very least, Hammerspace’s rise is a signal that AI-centric infrastructure innovation is shifting toward the unsexy but vital domain of data plumbing. And for the industry, that’s a positive development in our view because solving the less glamorous bottlenecks will enable AI to thrive.
For deeper insights and a more detailed technical narrative, check out the interview John Furrier conducted with CEO and technical visionary David Flynn:
