What's Up with Tech?
Tech Transformation with Evan Kirstel: A podcast exploring the latest trends and innovations in the tech industry, and how businesses can leverage them for growth, diving into the world of B2B, discussing strategies, trends, and sharing insights from industry leaders!
With over three decades in telecom and IT, I've mastered the art of transforming social media into a dynamic platform for audience engagement, community building, and establishing thought leadership. My approach isn't about personal brand promotion but about delivering educational and informative content to cultivate a sustainable, long-term business presence. I am the leading content creator in areas like Enterprise AI, UCaaS, CPaaS, CCaaS, Cloud, Telecom, 5G and more!
What's Up with Tech?
Defeating Data Gravity: How Hammerspace Powers AI Anywhere
Interested in being a guest? Email us at admin@evankirstel.com
The digital transformation journey toward AI implementation presents enterprises with a fundamental challenge: leveraging decades of data stored across disconnected silos. Unlike early AI pioneers who built purpose-designed infrastructure from scratch, most organizations must work with existing investments while meeting the unique demands of artificial intelligence workloads.
Hammerspace tackles this challenge head-on by flipping the traditional "data gravity" paradigm. While IT has historically moved compute to where data resides, AI requires the opposite—bringing distributed enterprise data to GPUs and models that may exist elsewhere. Through an innovative approach that synchronizes metadata across environments, Hammerspace creates a unified data plane that makes information queryable and accessible without initially moving massive datasets.
What sets this solution apart is its integration with Linux kernels, eliminating dependence on proprietary systems. When data movement becomes necessary, it happens intelligently at the file level, allowing organizations to selectively transfer only what's needed based on specific attributes rather than copying entire directories. This granular control preserves governance while dramatically reducing unnecessary data movement costs.
The approach has gained significant traction, evidenced by Hammerspace's recent $100 million funding round led by Altimeter Capital—investors known for backing innovative technology disruptors like Uber, NVIDIA, and Meta. Even cloud providers with substantial GPU investments recognize the value of technologies that enable data access for AI workloads, marking a shift in how these platforms approach customer data.
Looking toward future challenges, Hammerspace has launched the Open Flash Platform Initiative, bringing together industry partners to develop infrastructure solutions addressing the looming power and cooling constraints that threaten sustainable AI growth. By reimagining how data flows in the enterprise, Hammerspace is helping organizations transform existing information assets into fuel for AI innovation.
More at https://linktr.ee/EvanKirstel
Hey everybody, fascinating discussion today around AI and moving data around at light speed to where it needs to be, a challenging topic with Hammerspace today. Molly, how are you?
Speaker 2:I'm doing great, Evan. Thanks for having me.
Speaker 1:Well, thanks for being here Really intrigued by the work you're doing. Before that, maybe introduce Hammerspace the name, even kind of intriguing name. And what's the big idea?
Speaker 2:Yeah, absolutely so. Hammerspace, from a technology perspective, provides a data platform to simplify AI anywhere and we can talk more about what that means. But you know, ai is using data sets generated in a lot of places with GPUs and models in other places, and we make putting that data set together and having it where you need it simple. But the name itself is kind of fun and memorable If you think about, depending on your generation or genre of interest Mary Poppins, purse or Spidey's. And the latest Spidey verse he pulls a they actually mentioned a hammer first when he pulls his big mallet out of his little like tight pocket. You can kind of go through. But anywhere that these comic or kind of genre types of areas come from, where you're pulling something very large from a very small kind of magical space, that is what the term Hammer Space means and we take it to a technology level by talking about, you know, really large data sets that you can use anywhere, potentially without even moving the data around, and that's kind of the tie to the company name.
Speaker 1:That's fantastic, and you're at the center of a storm of sorts around AI, with growth and funding and new customers all very exciting. Maybe let's start with the basics. For folks who aren't familiar with what you do exactly, the problems that you're solving in the enterprise, how would you describe it?
Speaker 2:Yeah, absolutely so.
Speaker 2:As enterprises are embarking on having their data ready for AI, the challenges they tend to come across are they've, over the last 10, 20, 40 years however long the enterprise has been around created a bunch of data silos for specific reasons.
Speaker 2:Maybe one was a database, one was high performance storage, one was home directories, another was the archive. There was reasons why these data silos were created, and what that really means is specific infrastructure was purchased that had the attributes they needed of speed or capacity or redundancy, whatever it might have been, and now they have all this enterprise data sitting in these different locations and different storage systems and are trying to do AI, so use the data they have, perhaps in a private model or in their agentic endeavors, and they don't really know how to figure out for the AI systems what data do I have, how do I move it, how do I keep it compliant when that happens? And that's really the problem. Statement of where HammerSpace comes in by creating a single data plane across all of those storage systems that's queryable and has the ability to move data as appropriate, but really presenting a unified data plane to AI of all these different data sources that exist.
Speaker 1:Fantastic, and everyone talks about preparing for AI and the data challenge that you just described. And what are the you know IT infrastructure hurdles, roadblocks, challenges that you're seeing today with customers?
Speaker 2:What we generally are seeing, and it's different with enterprises than the early language model folks. Enterprises than the early language model folks the early language model folks who we worked with also pretty much had a new budget and went out and bought purpose-built AI everything. They selected the network, a storage system, you know the compute, placed it all in a certain place and built their language models and they had the luxury of designing exactly what they want and having doing everything brand new. Most enterprises aren't in that situation designing exactly what they want and doing everything brand new. Most enterprises aren't in that situation. They know they want to do AI and they may have a large budget or a small one, but they generally in the IT side, want to use as much of what they have today. So they've already invested in a corporate network and invested in storage systems, and maybe they're not even going to buy GPUs. They're going to use the ones up in the Neo cloud, and so they're trying to figure out how do they take advantage of both the infrastructure that they have today and the data they have today and put it to work in their AI environments.
Speaker 2:And most of the AI systems that are associated with data or storage say, well, just buy a new thing and copy all your data into it and you'll be AI ready. And for a lot of enterprises that's not practical. They don't want two copies of their data. How do you manage the gold copy? How do you keep them in sync? It doubles the budget. There's a lot of challenges with that if they want to use all of their enterprise data. So those are the kinds of things we see folks grappling with and trying to figure out what's a path forward. Do I have to create my own new AI silo and just copy the data into it that I'm going to use for AI, or is there a way to kind of use what I have today? And that's kind of the fork in the road of nearly every enterprise AI infrastructure conversation we're a part of.
Speaker 1:Got it, and so AI is creating something I see on your website called GPU gravity data gravity. What is that exactly? How do you define it? Why is it such a big deal?
Speaker 2:Sure. So any of us who've been around in this space for a while have kind of come to agree that there's a almost like a physical law of data has gravity, so you have to move your compute to the data, and that's kind of how IT has been designed and cloud has been designed for as long as I've been in the business. It's just too hard to move a petabyte over a little network, so you have to move the compute to the data. With AI. That's just too hard. To move a petabyte over a little network, so you have to move the compute to the data. With AI. That's just not possible.
Speaker 2:Organizations may not own enough GPUs to do the job they want to do. The models may not be sitting in their data center with the GPUs where the data is. That paradigm, while we've considered a physical law, doesn't work, and so we've had to come up with technologies and solutions to get your data to the GPUs. The data you know when you really think about that's the digital component. It should be movable. Gpus are things they're, you know, physical things that have to sit somewhere with power of networking. So that's what we're talking about is how do you design when you need to get data to available GPUs or models that aren't local to the data, and we have sleep technology, which we could talk for hours. We probably won't about how we do that.
Speaker 1:Yes, that would put a number of people to sleep, but so we'll short-circuit that. I've had dozens of data storage vendors on the show to talk about their perspective and approaches, and they're all doing very well, it seems, because we're in this boom period, so good for them. But it all seems very complicated when you feel back the onion that can get very expensive as well. You're talking, you know, petabytes. This isn't cheap. How is your approach different from the you know, status quo?
Speaker 2:Yeah, absolutely so. Hammerspace isn't a storage company. What's happening in our space is we started at this data plane layer. How do you make data and metadata universally accessible to multiple applications and make it where your data is not bound to the infrastructure stored? And if you kind of just stop and think about that for a second, generally data is bound by the rules of the infrastructure that sits in that file system, its replication rules, and what we do is we raise that out where your data is not tied to specific infrastructure.
Speaker 2:And the reason I bring this up is most of the storage companies who are doing very well in the AI space, started out with how do I make high-performance storage that goes fast enough to feed a GPU. And now they're starting to meet in the middle with us of interesting data services, kv caches and those types of things. So we're kind of meeting the storage companies in the middle and we provide the rich data intelligence and help and data preparation and then we use whichever capacity the customer wants. That capacity could be S3, it could be Blob, it could be NetApp, whatever it is. So we're kind of seeing data intelligence and storage starting to come together with the storage company's roadmaps. And then what we do.
Speaker 2:We have customers who say, well, I don't want to have to go buy capacity from someone else, why won't you provide that to me? And so we're starting to do that too. So we differ by. A lot of storage companies made their money in those greenfield. Deploy more capacity, and they made their money there. We've been much more involved in providing the data intelligence layer in AI factories and in hybrid cloud environments, even if you don't think about just AI, but research computing, where you have a hybrid cloud environment, you know and you're trying to bring your data sets together.
Speaker 1:Got it. That makes a lot of sense. So enterprises have the time held tradition of either going with proprietary, you know kind of systems for performance reasons. Maybe Others go for more of the open, best-of-breed, open-source kind of approach. Maybe talk about those two options for this space and what customers are telling you.
Speaker 2:But yeah, so what we've seen is that standards always win.
Speaker 2:It's just a matter of how long.
Speaker 2:We've seen that in server virtualization, network virtualization and we're just now starting to see that in storage, and the things that had to happen are most of the storage services thinking about intelligent clients that enable parallel file systems or data replication or the bits you know that are used to do incremental backups All those types of things in the data services have been built into proprietary kind of infrastructure-bound systems, whether it's a big NetApp file or whatever that is.
Speaker 2:And now what HammerSpace has done it's been in the works for a couple of decades is built all those capabilities into Linux. So the actual Linux kernels now have those capabilities. Any of the modern versions of the Linux kernels, whether it's Red Hat or Ubuntu that's running in the GPU nodes, have those services. So now all of a sudden, anything that's running Linux can be a client of our data platform, of our file system, instead of having to have that be tied to a proprietary system. And the benefit of this in AI is you have pretty much if Linux hadn't won the operating system war if you know, Solaris and all those DAOs, all these different operating systems were still-.
Speaker 1:They're really well.
Speaker 2:Yeah, yeah, absolutely. The standardization at this layer could not have happened but because Linux is one for all practical purposes, especially in high-end computing. Now we have the ability to let Linux be, have the intelligence of the data services, and then our platform just has to provide that unified data plane. And it's an interesting battle because, no matter which storage company you're thinking about, they're saying no, no, no, use our system, put your day in our system, you'll get all that capability. We're saying anything that generates data off of a Linux machine can be a member of our system, and that's how we can offer this universal data plane that many applications and locations can connect to, instead of it being isolated to a single infrastructure deployment location.
Speaker 1:Got it, so you can get data to the GPUs and the AI models without moving data around, just through this.
Speaker 2:Generally. Eventually you may have to move data, but the visibility is done through metadata. So we have synchronized metadata across any location in any site and metadata is teeny compared to the data itself, you know, as a fraction of the size. And then if you decide that you do want to move data, we do it in a really intelligent file, granular. We're very aware of the data. What's in the data? You can say I'm a genomics company and I want to use a model that only looks at genome sequences with certain attributes. They can query our metadata and say, okay, these are sitting at these five different systems. Just grab the genome sequences that make sense, move only those instead of copying an entire directory or an entire system, and you have access to those files while they're in flight, because you work with the metadata for the applications and the users and even if the data is in flight, it doesn't matter, because they can still make updates or queries or whatever they want while the data is in flight.
Speaker 1:Fantastic, so customers seem to like your approach. You had a big fundraising round earlier this year. Care to share any feedback from customers or anecdotes or stories on using HammerSpace.
Speaker 2:Yeah, I mean our fundraising was really focused on the company early on was really being deployed and used by the big language models that are being the early ones the metas and others that are in that caliber but we're not supposed to use their names publicly so we were, early on, paying for the run rate of the company through very large hyperscale and language model developers. Now, as we're moving into enterprise AI, where our go-to-market model and our sales motion is field teams in different regions selling to enterprises who are doing AI, we needed to expand out the size of the company, the size of the go-to-market team, and that's what our fundraising round was about. We brought in $100 million of outside investment just from equity companies we don't have any VCs involved in the company and they're really interesting.
Speaker 2:The folks who led that round was Brad Gershner and Jamin Ball from Altimeter Capital, and what we love about them is they were very early investors in Uber, nvidia, meta companies that kind of have this innovative approach to technology where they're doing something new and innovative, not just a slightly different approach to an existing market. That's the kind of companies they invest in. So we were excited to have them behind us and it certainly has helped as we continue to grow the company and the enterprise. That having the right financial backers is important. As you know, a financial services company is looking to adopt you. They don't see themselves as the next meta. They want to see you have the ability to be around and supporting them for the next 10 or 20 years of their enterprise, and that's really what that round was focused on.
Speaker 1:Got it Well. That's quite an accomplishment. And how are the hyperscalers, the cloud vendors themselves, reacting here? I mean, they typically would like to lock in A customer into their cloud, their environment, and multi-cloud seems to be the trend or private AI, private cloud how do you see that playing out over time?
Speaker 2:In our kind of small part of the world of what clouds care about. What we are finding is all the things you just said are absolutely true all the things you just said are absolutely true, but the cloud companies that are really investing in having a lot of GPUs.
Speaker 2:so Oracle any of the Neo clouds AWS also want to make sure those GPUs are being rented and used and they've made this big capital investment. They want to make sure they're being used and you need data to use those GPUs and they really see that as so important. That's like having a GPU and no data is not very helpful, and having a bunch of data and no GPUs also is not helpful helpful. So they've taken a little bit of a change in their approach.
Speaker 2:I would say with their GPU computing models that they are very engaged with the companies that are able to provide data into the workflows that use those GPUs and they know they need access to provide data into the workflows that use those GPUs, and they know they need access to more data than what's already sitting in their cloud. So they seem to be embracing technologies like ours pretty heavily.
Speaker 1:That's great. Well, well done there. So it's a busy summer in the AI world. No rest for the weary. I understand you have a lot of events and shows and a roadshow. What's on your radar the next couple months?
Speaker 2:We're really focused in enterprise AI right now because I think it's really evolving what enterprises are going to do with AI, how they want to deploy it in their environments, technology partners that are enabling enterprise AI and building out AI factories, and then being at the events to talk to the customers about the best practices, what technologies are available to help them. It's fun. It's an early stage situation right now, where most folks are still in the learning mode I think all of us are and it's very collaborative because discussions of different approaches an enterprise could take with their AI. So we're at AI summits and there's a million of them happening.
Speaker 2:But, the big ones that are kind of infrastructure focused, are where we're showing up right now. We also just to mention it probably is worth mentioning just launched an initiative called the Open Flash Platform Initiative, which is a bunch of vendors from system-wide chip companies to DPU providers to SSD providers, working together to build the next generation of infrastructure that is incredibly power, efficient on the storage side, and this is leaning towards that AI conversation of what are we going to do as an industry two or three years from now when this incredible demand for AI continues and there's not power to support it or there's new hot data centers and all the infrastructure we've built over the last 20 years can't run in those data centers.
Speaker 2:they're too hot, and so we launched an initiative as a company to build a storage system designed for those hot, low power environments but still fast enough to feed GPUs, and that just went into the market about a week ago and has been receiving lots of attention. It may end up getting absorbed by OCP, the Open Compute Project, but right now it's something that Hammerspace championed. This early stage announcement for Sounds amazing.
Speaker 1:We need to have a whole other discussion on that, but thanks for joining. Really appreciate the insights and, onwards and upwards, good luck with the mission.
Speaker 2:Thanks and thanks for having me on the show, Evan.
Speaker 1:Thank you and thanks everyone for listening, watching, sharing the episode, and be sure to check out our new TV show now on Fox Business and Bloomberg at techimpacttv. Thanks, Thanks, everyone.