Hi, everybody. In Hebrew, we say, know where you came from and where you are going. And this gives us an opportunity for the first time in front of our employees, our customers, our partners, to give a brief history of time. Where did VAST start? Where are we going? What are we doing today? And so those will be the three sections of my talk. And then after me, you'll see Jeff, and then we have a really talented speaker joining us, Tim Urban.
Let's start with the first half. We started almost exactly 10 years ago. At the beginning, it was just me, at home, watching YouTube videos, reading books, trying to find a problem so hard that we could never solve it, a challenge that we will never outgrow. And one thought kept coming back. We needed to try and build a thinking machine, a machine that could then go to solve all of the great challenges that humankind was not able to solve, to uncover the great mysteries. And so let's go back in time and see what that looked like back in 2016.
In 2016, early 2016, AI was not yet a thing. Google recently acquired DeepMind. OpenAI was just founded. Jensen did not yet give the first DGX1 to Elon Musk. But it was already clear that a new thing is forming. What we realized was that these neural nets, that forever were thought to be useless, were showing promise. By feeding them massive amounts of information, we were suddenly able to recognize cats in pictures. And without understanding how the human brain works, we were able to mimic it in this very, very basic way. And so we asked ourselves, what if we could feed these algorithms even more data, even faster? What would happen then? But we are not AI experts. We didn't know anything about hardware design. What we knew was software infrastructure. We knew data. And data seemed to be at the core of enabling this next revolution. And so, VAST Data was born out of the amalgamation of fast and big, and with the intent of reimagining the software infrastructure layer with the dream of building the AI operating system.
Rabbi Tarfon said, "You're not expected to finish the task, nor are you allowed to shirk from it." And so, in the process of taking this grand challenge, we knew the odds of success were slim. We knew that failure was likely. But we also knew that for us to succeed, we had to build the greatest company in the world, because only the greatest company could hope to attain this dream. A great company is not measured in numbers. It's not revenue, it's not growth, it's not valuation. A great company is one that is always there for its customers, for its employees, one that makes the hard decisions, that will do the right thing. It's like a great person. It's about character. And the company's character is about its culture. We are mission-oriented builders. We're innovators. We're hungry and curious. We always try to learn new things. Eric encouraged you to learn from each other. We are here to learn from you as well. We follow principles, not rules. We assume nothing, and we challenge everything. We have unreasonably high expectations of ourselves and an overgrown sense of urgency, myself specifically. We love to break trade-offs, but most importantly, we focus on you, our customers. You tell us what to build. You give us purpose. You keep us alive.
And so 10 years ago, we started by engaging with you. The only way for us to succeed was to listen, to listen to what you actually needed. We only build what you ask us to, and our best ideas came from all of you. I'd like to give you a few examples of how that happened. So in 2016, when we were just getting started, one of the hedge funds we spoke to that is in this room, but still won't let us use their name, complained about needing to segregate new and old data, that they only had fast access to a tiny subset of the most recent information. What we realized after speaking to a bunch of other potential customers in other fields, that this trade-off between high capacity and high performance was holding back every large-scale analytics project. And if we didn't solve it at its core, it would become the obstacle that will inhibit AI. And so what we built was a new architecture. You all know this, disaggregated, shared everything. We built it from the ground up to eliminate those tiers and break that trade-off. We needed to organize metadata in a completely different way. We needed to make forward-looking bets on parts from Mellanox and Intel that didn't really exist yet. And they weren't in production. We needed new algorithms. We needed new protocols. We needed to develop new math in order to break this fundamental computer science trade-off. And we did. The architecture has proven, 10 years later, to be the foundation required for the most intensive workloads of today and tomorrow. The fact that it scales super linearly as we give it new underlying resources means that it will provide performance, efficiency, resilience, ease of use for the AI workloads that we have now and far into the future. We could not have done any of this without all of you helping us in those early days.
And so once we had the architecture, it was time to build a product. And the first product was what we call the VAST DataStore. 2017, HHMI, which is also here in the room, I met Gorin last night for drinks, was one of our first alpha testers, and Ken at HHMI was secretly our most prolific QA engineer. He found the most bugs in those early days. What we realized is, after talking again to a bunch of other potential customers, we didn't have customers at the time, like NASA and Boston Children's, we were able to leverage this new architecture to reimagine what a storage system could look like. The VAST DataStore was the first to provide the scale of object, the speed of block, and the convenience of file with true multi-protocol access across all the interfaces, limitless independent growth of performance and capacity, the ability to mix and match hardware across multiple generations and provide deterministic, consistent performance, regardless of the workload patterns. It was truly universal storage. It enabled the consolidation of all of your data into one platform, which today makes all of that data accessible to AI workloads. But that wasn't enough.
2018, Eric at Pixar, who I also met last night and is also here in this room, was our first beta customer for the media and entertainment space. He wanted a way to access all of his data across geographies, for a way for his teams to be able to collaborate around the globe. And so again, we took that idea, and we tried to see what could be built. We talked to a bunch of other early beta testers, like the National Hockey League and CZBio, both with us here today. And we realized that a truly consistent global namespace with low latency had never been done before. We like these types of challenges. And so our DataSpace, it didn't quite break the PAC theorem, but it came close. And we needed to leverage the system's understanding of access patterns to create a seamless experience where data is always available across public cloud, private cloud, edge, and device, solving for both global data access and data protection. And these abilities over the years proved crucial for distributed inference at the edge, for reinforced learning workloads, and for many points of presence across telcos and clouds. And so we couldn't have done any of this without you. But again, that wasn't enough.
2019, we were speaking with Down Under Geo, DUG, who's also here in this room, Phil from Down Under Geo, kept complaining about how long it takes to scan a file system, and how by the time you go through a billion files, something changed at the beginning, and it's no longer consistent. And so we took that problem, and we started to collaborate with a lot of other customers. We had customers at this time, like Brown University and Aquatic. And what we realized was that we really needed to build a new type of database. And for the first time ever, we combined the ideas of a data store and a database based on your feedback. It allowed us not just to query the metadata of the file system, but also user metadata. It allowed us to give structure to unstructured data, and then ultimately to give it meaning in this new era of AI. We broke the same trade-offs in the world of rows and columns that we broke in the world of unstructured data. And now you can have a trillion vector embeddings in one database and you can have a billion agents querying that database in parallel. We found more room for consolidation at this higher layer and we eliminated the need for many, many different types of databases that we had before. And of course, as with everything, when you put all of this structure data in the same place, undoing decades of complexity enables AI to now make use of all of that information.
At that point, we felt like we weren't a storage company anymore. We had a data platform. But for us to make that next leap, we needed to have a data engine. And here, it was a genomics company, PacBio, one of our first customers that in 2020 asked us for their genomics pipelines to become data-driven. We didn't know what that meant. And so we went to Agoda, and we went to Deloitte and NYU, and we asked them, what does this mean? What would you like to see? All of them were trying to rethink the traditional interface between applications and data. And together, we defined a new language for data-driven computation. It's based on triggers and functions and streams and filters, and we flipped the old dichotomy of logic and data on its head. These abilities, the DataEngine abilities, are now being leveraged for pre-training and post-training pipelines for agentic reasoning, for agentic communications. And all of this is enabling reinforced learning at very, very large scale. For the first time, this operating system is enabling us to have not just data, but also a compute runtime, to have everything vertically integrated in one stack.
And so over the years, we kept partnering and we kept asking questions and we kept listening to the feedback that you had to give us through 2021 and 2022 and 2023. And we realized that this architecture enabled a lot more than we originally thought. As we collapsed silos, as we consolidated infrastructure, you asked us to design things like large-scale multi-tenancy and quality of service into this platform. And these capabilities proved invaluable when we started working and partnering with the largest AI clouds in the world, like CoreWeave and Lambda and Core42. Having so many workloads on one system required very fine granularity monitoring, and that turned into agentic observability, leveraged by all of the AI labs. Again, some of them let us use their names, and some of them do not. But anyone who's doing AI at extremely large scale is building on top of the VAST Data Platform. We needed access controls. We needed audit logs, which we originally added to make you, our customers, compliant with regulators. But over time, we've realized that that foundation is the best thing to build on top of for agentic AI security. So time went on, 2024, more AI clouds, more supercomputers, more AI labs. I'm not supposed to say that name, but it's on the screen. And 2025, this last year, we found that a lot of the innovation is starting to happen around enterprises like Lowe's and CrowdStrike and Adobe. And a lot of the innovation, our innovation, started to be in collaboration with a great company called NVIDIA. As we build these massive AI factories together, we realize that there are more and more problems that need to be solved at that AI operating system layer. We did joint solutions like our InsightEngine, which gave enterprises a way to safely and easily have a chatbot interface to all of their proprietary data. More recently, we built a global KV cache that reduced inference latency on the one hand and made GPUs a lot more efficient on the other. And we have so many more of these projects that we're working on and that we'll share with you today and tomorrow.
These last 10 years, as you can tell, they've been your story as much as they have been ours. Our destiny is tightly intertwined with your destiny. And we want this conference to be a place for you to share opinions, a place for us to all debate the hard problems and brainstorm solutions. It's not about VAST, we're just the facilitators. And so I want to thank you all for building with us, and thank you for attending our first ever user conference.
Okay, so now it gets fun. In parallel to listening to all of you people tell us what to build, we've always tried to follow a North Star. And from the very beginning, it was AI. Back in 2016, this is actually from 2015. You can tell by how ugly this slide is that I made it myself. And I didn't show it to anybody in 2015 because it was crazy. How can anybody think of building a thinking machine 10 or 11 years ago? But that's what guided us, and that's what we tried to solve for. And over the years, it became clearer and clearer what that North Star is. And today, I think the answer is yes, we can build a thinking machine. We haven't done it yet, but we're in the process. Five years ago, in 2021, we had Allen, who knows how to make pretty slides, and we put these ideas to paper. We talked about building the natural data platform. We discussed things like machines adapting to humans and human-friendly interfaces and recursive perpetual learning machines. This was before anybody talked about things like training and inference and reinforced learning. But that was our North Star. That's what we shot for five years ago. And I think today it's proving itself. We're not there yet, but the world is moving in our direction.
And so today, 2026, let's try and go forward in time. Let's hit the VAST Forward button and see where it can take us. Where will the world be 10 years from now? The short answer is we don't know. There's two options. One option is that the only thing these models can actually do is guess the next word of what we're thinking about, and in some cases, maybe the wrong word. But in that case, there's no real intelligence behind it. It stays similar to what we have today. The other option is a lot more interesting. 2036. We just hit that VAST Forward button. The other option is that choosing the next word proves to be core to the creative human process. And over the next 10 years, we're going to have agents that will get closer and closer and closer in their abilities to what people can do today. And so let's try to imagine that future. In that world, we have extreme personalization. We have unimaginable abundance. If you need a piece of software developed, you spin up a few thousand agents, developer agents and software, QA agents and project management agents, and they'll code it for you in seconds. And you use it once, and you throw it away. Extreme personalization means that our medicine will be personalized and designed especially for us. Our entertainment, every movie will be generated once and viewed once, and then thrown away. Extreme abundance means that everything will be disposable. Agentic scientists will be able to test more theories in a second than the entirety of the human race was able to generate in a generation. They'll find new ways of generating clean energy and curing disease, new ways of fast and safe travel, and they will reduce the marginal cost of nearly everything down to zero.
From a software infrastructure perspective, remember, that's the part that we know. Enabling and coordinating and controlling all of these activities is far from being trivial. When everything is a million times bigger and needs to happen a million times faster, the current ways of maintaining order break down very, very quickly. And so what we need is a way to observe what agents are actually doing. What we need is a way to see what they're thinking about. We need a way to actually manage all of these agentic workloads. And for that, we need an operating system. See? It took a second. The AI operating system, it needs to keep track of every image and every sound bite that every agent is exposed to. It triggers that inference function for the agent to make sense of what it sees. It triggers the fine-tuning function for the agent to adapt its model of the universe and become smarter over time. More than that, it needs to sit in the middle of every arm movement of every robot that happens, in the middle of every thought creation. It remembers every agentic experience. It curates every new idea that gets formed, and it keeps track of every old memory. The AI operating system needs to provide a persistent, queryable way for us to understand what agents are doing and what they're communicating to each other about. Not only that, it needs to expose secure APIs for these agents to handle money and sensitive information on our behalf. It needs to do identity verification for these agents. It needs to build some type of kill switch in case something bad happens and we want to shut them down. For all of those reasons, we need an operating system. But it's not the operating system that we have today. It needs to function at the scale and the speed that these agents require across geographies, managing all of the aspects of agentic life. It needs to run within the robots and the devices at edge locations, in the cloud, in AI factories. It needs to maintain all forms of unstructured and structured data to orchestrate cognitive pipelines and to trigger actions in real time. It needs to make judgment calls to enforce policy. It needs to enable new discoveries and the mastery of new skill sets. All of that while maintaining strict consistency around the world and 100% uptime.
So that's the future, 2036. That's what it needs to look like. But we don't like musing about the future too much. We like to build and we like to build today. Thomas Carlyle famously said, "Our main business is not to see what lies dimly at a distance, but to do what lies clearly at hand." And so today, what must we do, and over the next few years, in order to enable this future that we just described? Now we go back in time to 2026. First and foremost, we're not allowed to use this picture, by the way. Don't tell the lawyers. We must enable the frontier model builders. They need to keep pushing the state of the art. They need to go from the current state to general intelligence to super intelligence. And they need the ability to keep training new models and keep serving existing models. And so for training, we need exponential growth in capacity and performance. We need much more data and many, many, many more GPUs accessing this data in parallel. For model serving or inference, we need distributed systems at the edge with very, very low latency and high levels of resilience because model serving is a production workload. If it goes down, none of us can use ChatGPT. But in parallel to pushing the state of the art and enabling the model builders, what is more important is we start to enable enterprises. We need the enterprises and everybody else to start using all of this great technology that is already available today. For that, the enterprises need us to focus on simplicity. They don't have 100 PhDs in computer science, definitely not in AI, and on security. They have regulators breathing down their neck. They need to comply. They need to make sure that everything is safe before they deploy these AI workloads into production.
And so a new stack is getting built. Jensen, the CEO of NVIDIA, who, by the way, we'll be hearing from later today, gave the analogy of a five-layer cake. He said energy is at the bottom, and then chips, hardware, cloud services, AI models and applications are at the top. And every layer of this five-layer cake needs to be redone. We went from 10-kilowatt racks to 500-kilowatt racks, from CPUs and 10-gig Ethernet and very small hard drives to fast networking and large SSDs and GPUs and TPUs and DPUs in the new stack. From old enterprise applications that every one of us is still using today to new agentic pipelines that we're all going to start to use in the very near future. And for all of this to happen, we need that middle part, that software infrastructure layer that I call the AI operating system to also be redone. That's our part. That's where VAST Data fits in. We partner with all of you up and down this stack in order to build the full solution, and we focus on the software infrastructure layer. As we shift from training to inference to reasoning to reinforced learning, the weight shifts between data and compute. It shifts from compute to data. This was supposed to be a Turing machine. Probably not many people in this room know what that is, but it was the best way that I thought of to depict compute and data, so that's what we used.
So every technological revolution has a defining characteristic. For the PC era, it was compute. For the Internet, it was networking. For AI, it's data. And natural, unstructured data has to be analyzed and understood. We need to give it structure and meaning and context. We need to manage weights of specialized models that are fine-tuned by various agents. We need to generate new information by combining real-world experiences with newly created intelligence. And all of that needs to be audited and observed and queried such that we can monitor these new environments. And so the next phase of AI is all about the data. Another way to look at this is through operating systems. As we shift from serving the frontier labs into serving everybody else, the weight again shifts from the hardware layer underneath to the operating system on top. Every revolution, in this case we're showing the mobile revolution, starts with hardware and a killer app. In the case of Nokia phones, it was cell phones, that was the hardware, and the killer app was people wanting to talk to each other. As we progress, of course, it's the operating system that allows us to have a plethora of applications on top of this new technology. And for AI, the hardware is GPUs, the killer app is ChatGPT, and now is the moment that this operating system will enable many, many new applications for all of us to enjoy.
So for the last decade, we've been building the foundation for this operating system. And we built the foundation for unstructured data management through our DataStore, for structured data through our DataBase, for compute through our DataEngine. And all of these enabled the new revolution. Now that we have the kernel of this operating system, we have a lot of work that we need to do in order to truly enable AI. And so I'll take you through a few of these features that we're missing that we need to add. And Jeff will, of course, dive a lot deeper into all of these. We're adding database functionality to consolidate all forms of structured data and make it all accessible to AI. We're creating a lightweight VAST instance. I saw a robot walking around here that had a VAST instance within it, such that we stretch our DataSpace beyond the edge. We're enabling very, very large scale efficient reasoning through the addition of KV cache capabilities up and down the memory stack. But we need to incorporate that into the security aspects that are important so that one agent doesn't leak information to another agent. We need to route those prompts using our DataEngine to where those caches are full as we make inference and model serving a lot more effective. We need to enable every agent to fine tune its own model. We announced the TuningEngine. Everything is engine now, earlier today. And we need these agents to be able to create their own IP in the form of these models such that the enterprises that own them can monetize that IP. We're enabling secure distribution and monetization of these models through confidential computing so that you can lease out your model and get paid for it without being worried about somebody stealing your weights. We're enabling the use of agents in the enterprise by providing observability and reproducibility mechanisms alongside several levels of agentic identity verification and data access controls. This AI operating system is key for setting and for enforcing policies, and it will continue to grow in capabilities as this new era of agentic AI unfolds.
A little bit about VAST. So we are perfectly situated to build this layer of the stack, not just because of our unique data architecture, but because of the company that we're trying to build. It's a company of pioneers. We love to forge our own path. We are determined to create this new category, to keep tackling hard engineering problems and to build what's never been built before. The fact that we're tripling our business year over year and the fact that we're generating more than $100 million of cash every quarter allows us to be the foundation of this next era. The fact that we've already attached ourselves to millions of GPUs and are a part of all of these very, very large AI factories allows us to build the foundation for this next era. It enables us to focus all of our attention on you, on our customers, on serving the builders and the innovators that are in this room today that will continue to build and innovate into the future. And for many, many decades to come, it enables us to fulfill the mission of building this AI operating system in collaboration with all of you.
I want to end with that which is most important. Not AI, but people. Every one of you here in attendance, every one of you at home watching, you are the most important part of this journey for us. A future of limitless abundance is great, but it's not more virtue. More abundance does not mean that we will have more virtue. My wish as we embark upon this next decade together is very, very simple. As we find ourselves needing to spend less time on wealth accumulation because everything will just be available for us and everything will be free, I hope that in proportion we gain more space to tackle the really important problems, that we acquire the wisdom to instill the proper values in our children, the next human generation, and also in our machines, the first generation of machines that can think. As I look out into this group of people, into this growing community, I can't think of anyone better with whom to share this journey, with whom to share this responsibility, and indeed, to share this great adventure. Thank you.
