Swapnil Bhartiya: Hi, this is your subtle Bharati. And today we have with us Molly Presley, SVP of Marketing at Hammerspace. Molly, it’s great to have you back on the show.

Molly Presley: Yeah, thanks for having me again.

Swapnil Bhartiya: It’s my pleasure. We are going to talk about, of course, you know, a lot of milestones that Hammerspace, the, you know, the whole revenue results. We’ll talk about all of that. But before that I want to for our audience because the focus nowadays has totally shifted to AI, no matter who you talk to. So just tell them Hammerspace from the lens, very, very myopic lens of AI. Hammerspace. Yeah.

Molly Presley: In the building phase of AI, we were very involved as the data storage and data plane layer for helping to train models, foundation models, quite a few of them. We were essentially the storage to get the data to those models. But now, you know, it’s only 12, 24 months later, things have changed and there’s a lot of conversation around inference and RAG. And Hammerspace is really focused in delivering not just speed of data to GPUs, but aggregating and preparing the entire data estate. When you have data in multiple sources or multiple locations and getting that data to the GPUs.

Swapnil Bhartiya: Perfect. Awesome. Now this has become a very, very messy space. Of course it’s hard to find SSDs. If you do find it, the prices are same thing with GPUs. And I think a week ago OpenAI and Broadcom, they partner to create Jalapeno, a chip they have designed purely for inferency. Because the fact is, I always draw a parallel with the traditional software world. Writing software is more or less like training, but deploying software or SaaS or something, that is inferencing. And most of us, that’s what we do is inferencing. So that’s why the whole shift is happening towards inferencing. Whether it’s you, you are redesigning your own architecture, or if you look at hyperscalers the way they are looking at infrastructure. So can you talk about while the AI adoption is in production? And here we don’t talk about hype, we only talk about AI in production. Data, of course, has always been the oil. But people get obsessed with GPU capacity. Of course, token cost is becoming a big problem, but there are a lot of other factors. They play a very big role in getting the result that you want from your AI, no matter how good your agent is, no matter how. So talk a bit about the pain points when it comes to data. And that is something Hammerspace is basically to solve.

Molly Presley: I think that’s a great setup for the way we are seeing the conversations with our production customers going. We have a global Fortune 100 company that is doing massive scale inference and needing to use data that’s in many different countries, GPUs that are up in the cloud and organize this as one unified workflow. That is really the gold star of what most enterprises are trying to do. Use data sets that already exist in many different locations, maybe different storage systems, maybe different geographic locations, get visibility to their AI teams to that data and then be able to get it to the GPUs. And it’s really interesting if you think about how a lot of the market started, what people would do is go buy an entire new storage system, usually an SSD based storage system, curate a data set that was going to get copied to that storage system, and then they do AI training or whatever they’re going to do on that storage system. The SSD shortage has created a requirement to stop and look and say, well, I can’t buy SSDs affordably right now and I can’t get them fast enough for my mandate to get something rolling, maybe a new agent in the next month, what do I do? And it’s forced the function of, wait a minute, I already have data. I already have data storage systems. How can I just operationalize those without going and buying yet another storage system and creating yet another copy of data for my AI initiative? And that’s where Hammerspace comes into place. We ingest the metadata from all of those different storage systems. We have an MCP server so we can have good understanding of what is in a PDF or a video file or whatever the data is, and then present that information layer to the AI engineers, to the models, and move only the data, the GPUs that is really needed, but also have visibility to all the data that optimally would be used. So this data preparation piece has become critical and managing the data through another infrastructure silo, has become more difficult because the SSD and memory shortages that we’re all seeing.

Swapnil Bhartiya: No, the issue here is not exactly issue, but as you rightly mentioned, data is scattered across silos. And I mean, whatever they built in past was not designed for AI. Now AI want that data everywhere, anywhere, whenever it wants. Now in your organization there is different kind of structure, unstructured, very, very sensitive data. If you’re in a regulated industry, you don’t want to connect everything to the MCP server as well. So what about the whole data governance? How much control, Hammerspace gives to teams so they can vary sophistically what data they can get access to, whatnot. Does that question even make sense?

Molly Presley: It does make sense. And the way we approach it is, generally speaking, it’s better to bring a logical layer to all of your data where you can see, let’s say you have five storage systems, a couple in the cloud, a couple in a data center, whatever it is, and you can govern at one layer where you can set your policies on this type of data, can be used for training this type of data, cannot those types of things, it’s easier if you set one policy and that applies to all of your storage systems, versus if you try to take your legacy architecture where you’re putting policies on each storage system and then when you make a copy of the data, you lose track of where did that copy go. So there’s just some data management kind of pragmatic decisions companies need to make. Are you going to manage on infrastructure boundaries or are you going to have a logical layer that can see everything and everything is governed on the same rules? And I absolutely think for moving forward, if you’re going to be using data sets created by many different departments, many different users, doing that as a single layer makes sense. But going on to your question of then what? So great. Setting policies on my MCP server can get visibility into a spreadsheet and it has, you know, maybe pay information, great. That can’t be ingested, but a spreadsheet that has information about research results can. All of that is done with Hammerspace. And then some of our data platform governance partners who have specific platforms for HIPAA or GDPR or government security. So we do partner to have experts in governance apply policies to the data management and data movement layer and work tightly. And that’s part of being a data platform is being able to work with partners like that. And it is critical and it has to be automated because we’re talking about millions, billions of files, data moving around. So you want to be able to set the right policies and then see what’s happened. Okay, this file moved, this file went into this model. You want to be able to audit that.

Swapnil Bhartiya: Very well said and thanks for taking that question. Now, there are two ways we also talk about new cloud these days. There are two kind of approaches. One is, you know, one size fits all. You just sign up like regular chatbot and start doing. And then second is that given the nature of businesses and of course the whole global geopolitical sovereign AI is there, GDPR is there, CRA is coming, a lot of AI related policies coming. What are some of the unique business challenges that customers are bringing to Hammerspace? That is also helping you not only rethink how you are looking at it, but that will also very well tie in with your growth as well, that we are also seeing massive growth. So let’s talk about some of those critical business challenges that even you’re like, wow, this is actually a challenge for us.

Molly Presley: Yeah, absolutely. I think so. Maybe dissecting that question. Let’s talk about neoclouds first. Different neoclouds and different enterprises have different models. Some enterprises will use one neocloud for one thing, another for another thing. They’re very flexible and dynamic in their architectures, maybe just based on price or availability or the type of GPUs and often the customer in that case will say, I’m going to use Hammerspace to prepare my data and move it to whichever GPUs or neocloud I want to use. And so it’s a customer driven decision. On the other side, sometimes the neoclouds are working with us to say we’re building out not just our storage systems for the data that’s local to the GPUs, but we believe we can do more customer acquisition if we have a way to reach into and provide a data orchestration service. So thinking about neoclouds and our growth, some is customer demand driven. Give me the flexibility to move my data to whichever cloud, data center, neocloud I want and I’ll make that I have that control. The other side comes from the neocloud building a service that they need, a data orchestration service, not just a data storage service in order to continue to acquire more customers and get more data from their existing customers into their neocloud. And then on the other side you mentioned sovereign AI and that’s a completely different and yet really important, especially as we get further through Europe and Asia and a lot of the sovereign buildups that are happening, but we’re even seeing it here in the US where the US government has a deployment of Hammerspace and one of their own neoclouds for their own sovereign data. And really the piece there is having the ability to know where data went and where which data was used. And this is a really hard challenge in AI when your data is coming from everywhere, we call it AI anywhere. And you need to be able to prove to whoever your governance layer is on your sovereignty that the right data was used and not the wrong data and which data was used. And that’s a really core piece of what Hammerspace provides is not just the visibility into all the data and the ability to automate the movement, that’s all about efficiency and time to AI results, but the tracking and auditing and automation of putting boundaries around data that are logical, not physical because you can’t necessarily use it all behind a firewall in a data center, in a discrete system. We can’t just use physical boundaries so we provide that logical layer.

Swapnil Bhartiya: You also mentioned that of course there is shortage of SSDs and memories which is also affecting enterprises capability. Now I think last week OpenAI and Broadcom, they announced their own Jalapeno chip which is more designed for inferencing, not necessarily using general purpose GPUs which are used for training as well. Can you talk about what is happening in this space and how is it affecting customers and do you see a silver lining where we will finally overcome this hardware shortage?

Molly Presley: You know we definitely see not just OpenAI but other companies, Meta, like you mentioned, Apple, others are in the hardware business as well as the software business. And sometimes because they have a very specific technology requirement that they want something built for their very specific environment. Some is margins, revenue, supply chain control. I think there’s a lot of objectives underway. But your point of view and when we think about inference, when we think about RAG, you definitely don’t need the latest and greatest and biggest GPU for all of those. And they have different workflow requirements. So we’re seeing a huge amount of growth in the number of new technologies coming out. Memory processors, types of SSDs that are specific to these different workflows. And I think the next debate will become is proprietary good or is open standards based solution good? And there’s benefits to both. If you have a proprietary stack and you have a vendor you know you want to work with and you invest in their stack, that’s okay. But if you want to move out of that stack to something else, it also limits you. And so a lot of unknowns for the future. But what we are seeing is most customers ideally want flexibility that wherever their AI journey takes them 12, 24 months from now, they can run it on the hardware in the location with the model they want. They want flexibility but because we’re so constrained right now on so many of the components, on the hardware side, the processors, even some of the networking, the memory is so constrained they may have to choose is time to market versus flexibility a trade off they need to make. So I’m kind of not giving you a good looking in the crystal ball answer. But I think there’s upsides to vendors controlling their entire stack and giving a really controlled workflow, being able to guarantee the output. But the downside is you lose some flexibility there. And we’ve seen that even with SaaS tools today, that running AI with Snowflake, I mean, I’m sure there’s good things about it. Absolutely. But now you’re locked into Snowflake stack. Is that really what you want? And that can happen on the hardware side too. So there’s a lot of considerations to take into place.

Swapnil Bhartiya: If you look at AI as AI moves beyond model building. How is, you touched upon that, but I just want to have a discussion of that. How does the Hammerspace data platform help businesses prepare for the next phase of operational AI? You have done all the pilots, you have done all the testing. Now we are talking about AI in production. How do you folks help them?

Molly Presley: Yeah, absolutely. And this was actually what we launched at GTC and it was part of Nvidia’s reference architecture for AI data platforms. And what Nvidia’s goal and really all of us who are doing a reference solution on this is, AI was a lot easier when you had a curated data set that you were putting into a discrete and controlled system. And so launching those pilots tended to go pretty well and you could show some results. But then when you start to get into data sprawl where you have lots of data, you have new data being created constantly and you’re trying to figure out how do I have AI just run on its own day by day and continue to get the same results that I got out of my curated environment in the pilot. That’s really challenging and that’s what these AI data platforms are designed to do, is provide the same experience, the same SLAs on data performance, data visibility, but across a wide range of net new different siloed data being still able to be produced and driven into your AI factory essentially. So the touch point of your data to the AI factory is the same, but now all of a sudden your data is being aggregated across more than just a curated data set. And that’s what enterprises need to move into production.

Swapnil Bhartiya: Really based on your own growth, based on the use case that you’re seeing. We cannot really predict, even AI cannot predict what things will look like in a couple of years. But how do you see AI through your customers is enabling some of those use cases which were not possible otherwise. We have been covering AI for so long, but it was very specialized, only data scientists can get access to very, very expensive machines, but it has been democratized now. So from your perspective, what kind of use cases you have seen where you’re like, in the next few years, these use cases will continue to grow and Hammerspace will continue to play a very critical role in enabling use cases which were not possible earlier.

Molly Presley: Yeah, I think where we’re being adopted most right now is in places like customer support experiences, taking logs of every customer interaction and not just being the bot that interacts with the customer, but being able to help solve problems, troubleshoot faster response. And that takes a lot of insight into a lot of data and then making sure you have the governance to not share with customer A anything about customer B’s environment. So you’re training on your own data across a huge set of customers, but then providing a unique experience to each individual customer. And we’ve seen a lot of growth there across a lot of different types of industries. And then I think it’s probably obvious that new product development, R&D, when you think about how can you accelerate and gain a new product in the market, whether you build cars or drugs or any other type of solution that your business may be in, having the data sets, if you think through, you’ve had all these PhDs who have written papers, you have students maybe who did their own little research projects, you have your less documented products, but you want to aggregate all that information for your future product development. Being able to aggregate what used to be very isolated data sources and then move very quickly in developing a new product, we see lots of adoption there. So I think those are the two big use cases. And really almost every government or enterprise has both customer support and some sort of new product development. But those are the areas we see it.

Swapnil Bhartiya: Are you seeing the strongest demand, whether enterprises, governance, sovereign AI deployment, research institute or hyperscaler, which is also kind of driving the growth of Hammerspace, we’re seeing.

Molly Presley: Governments move quickly and with fairly well funded projects. And I think that where I came from, the supercomputing space, and there was a time 10 or 20 years ago where governments liked to flex their muscles as having the biggest supercomputer. And that was a big thing with the top 500 list of supercomputers. I think AI has become that. So governments want to own and show they have better AI capabilities and I think that’s important to national security, national power. So government is big. We see a lot of fairly substantial spend. When you think outside of the people building models and the people building up neocloud infrastructure, I mean, because that’s still, there seems to be a new neocloud every day. And getting their infrastructure built up is certainly a large revenue opportunity for companies like Hammerspace. But when you think about end users, it’s still the places that are using a lot of data to build products. So it’s still life sciences, it’s financial services and things like quantitative trading. Really data intensive businesses who understood 10 years ago how to use data, but now they can do it even faster. They’re moving fastest in AI deployments and I think they’re the ones that some of the verticals and industries who are moving a little slower will go and look at, what did you do? How did you build it? And they’ll be kind of the reference for others who maybe are moving a little bit slower.

Swapnil Bhartiya: Oh, you’re absolutely right and thank you for that. As you rightly said, of course, data is when it comes to AI, I mean that’s what it’s all about. Now if there is an AI architect that knows very well that data is their bottleneck, what is the first practical advice you would give them so that they can take steps to prepare for inference and physical AI? And because Hammerspace can only help, you have to do some preparation internally, you have to do some homework so that things are ready. So what would be your advice to them?

Molly Presley: Yeah, really two things. One is just have your policies. There’s the human conversation of what are our rules around data usage? And make sure you understand those. You don’t want to move too fast and assume you know the answer. So your company needs some kind of rules on your data usage. And beyond that, having the ability to then, we call it prepare your data state, but have your data state unified. Ideally, but certainly visibility, even if it’s in a couple places of what data do I have throughout my entire organization? And that can be automated with things like MCP servers and metadata assimilation. There’s technical things that can do that. Or you might have to go through the human practice of hey everyone, upload your files to this location so that we have visibility. And the problem with that is trying to get humans to do that is historically been impossible. People hold onto their data and their systems the way they want to and getting a policy in place where humans are doing it tends to not work very well. But either way you need to get visibility to your data state and have some rules. And then of course you have to have some kind of AI factory to load it into. But I think just making your data AI ready and then ensuring it will follow the rules of your company are the two basic things that are not easy. But the first two foundational things you need to do.

Swapnil Bhartiya: Since we are dealing with data, massive amount of data. One concern which is emerging, actually there are a lot of frameworks even at the government level is which is post quantum readiness. What is your advice for folks? Because once your data has been gathered by the bad actors, when the quantum is ready, they will start harvesting it and by that time it’s already too late. So what is also practical advice to protect the data as well?

Molly Presley: Yeah, I think that once you let your data out into an AI model, it’s already probably the horse has gotten out of the gate, so to speak. But this is a good moment to get ready for what the future holds. That technology is becoming increasingly powerful. We certainly know quantum computers put GPUs to shame on the speed of things they could do. Once they get stable and predictable, there’s some work to do on that side. But it has become where I have a podcast and I was talking to a woman who wrote a book on data governance and she said 10 years ago she sat at a bar and said I wrote a book on data governance. Nobody wanted to talk to her. AI comes about and all of a sudden she’s pretty popular. People are pretty interested in what she says and it’s going to be even more so in the days of quantum computing. So this is more urgency around getting your policies in place, getting your data secured. And don’t let people leave with copies of data on USB keys. Don’t let people make copies. Someone who is a DevOps practitioner might make a copy of the data up to the cloud for their own work, where all of a sudden that data is now in the cloud and it’s not within your parameters of control. So getting these policies in place for your AI strategies is just doing some homework and groundwork you’re going to need to do for the quantum era anyway. And it’s becoming increasingly critical and increasingly important that this is happening right now. So a lot of the tools and policies we’re putting in place for AI will help enforce the security you need once you get to quantum.

Swapnil Bhartiya: Last question. Since you did talk about something which was relevant 10 years ago, and certainly it is more relevant today. In general, we have seen a lot of culture shift. We saw the whole DevOps movement, DevSec, SREs, a lot of labels, a lot of personas evolved. Because what we have learned is that technology can only do so much if you don’t have right processes, right practices in place, it will not work out. When it comes to AI, when it comes to whole governance processes, what are you seeing with their organization? They don’t have any policies or practices in place yet. And also the last thing I want to hear is something some new labels coined SRE, OPS or AIOps. I mean of course those terms are there, but what I’m trying to say, I don’t want personas, but how much role do you see culture? I was at Cisco Live and one of the concerns, security concern, is that these models are multimodal. They can scan PDF, they can scan images and agents are taking action on your behalf. They will go scan a website and you will find whatever. Now those websites may have things that is human readable and then they may have content which is machine readable. So you will not even realize it, but there will be clear prompt for your AI to take action on their behalf. Same thing with the PDF, same thing with the images. We don’t even know how bad actors are going to do that. The point, the small question I have is how much role do you think culture is going to play in ensuring that you also don’t trust AI blindly because it will make mistakes, it will hallucinate. So you cannot put all your eggs in the basket. Talk about the cultural aspect. Are you seeing anything or you feel that companies should start building right culture also in place so that it’s not too late by that time?

Molly Presley: Yeah, I mean I think culture is a huge part of the answer to all of this. And we have seen so far the Fortune 500s who’ve been around for maybe 50, 100 years that are more conservative have embraced AI generally much more slowly, at least in business critical use cases. It’s for exactly all the reasons you just mentioned. Change is hard. Managing and monitoring and improving the change is hard, but there’s also the risk of falling behind if you don’t. And so I was reading an article recently on the CFO of the AI age is much less risk adverse. They have to invest in forward technologies with less proof points. So it’s become just faster and riskier. And I think cultures of the companies that are going to be successful will have to figure out a way to move fast enough and secure enough to embrace AI. So what does that mean? It means embracing humans. You were talking about the human versus augmented AI or AI augmented human. It’s going to be the humans who know how to use these tools and are bold enough to understand them and deploy them and do it in a way that’s safe. And you may not be able to have a year long corporate approval process to get that done. And so I think the types of humans are going to be successful in the future. And this comes back to culture. Who do you hire? There’s a certain person who will do well in an environment like this that maybe would have been way too much of a risk 10 years ago in a different model. So it’s understanding the models, understanding the tools, understanding the output of what you’re doing, then being bold enough to be able to lean in and get it done. So I do think the workforce is going to change based on AI and I think it’s exciting to see where we can go. But people definitely need to slow down enough to understand how do you put guardrails around it? What is that checkbox that you need to click in the AI tool that’s buried four layers below the top level GUI that makes sure your data isn’t saved? There’s some understanding needed. So continuous learning is absolutely critical.

Swapnil Bhartiya: Molly, thank you so much for joining me. And of course talk about Hammerspace’s growth and how you folks are solving when it comes to data, getting right data in the right hands, in the right way. Because this is going to be very, very, I think, critical discussion going forward. There’ll be more challenges as well, but that is really what is becoming a bottleneck. So thanks so much for sharing all these insights. I would love to have you folks back on the show because in this space so much is changing and so many new challenges are emerging that need to be addressed. So once again, thank you for your time.

Molly Presley: I really appreciate the conversation and I think you’re right. Every day is a new day in AI right now, so the conversations are always fresh.