Hello, I'm John Furrier with the CUBE here in the CUBE studios in Palo Alto. Of course, we have our NYC CUBE wired studio connecting Silicon Valley to Wall Street. It's a CUBE conversation with Sanjay Poonen, CEO of Cohesity, fresh off his sales kickoff event for his company. A lot of news this summer with Maestro, which kind of points to the whole Anthropic open models, foundation models, headless systems, and a ton of activity that's changing the game in cybersecurity and cyber resilience. Sanjay, we're living in a great time. Welcome back to the CUBE for your nth appearance. I don't know how many—
I'm great friends with you and Dave, and congratulations on all your success.
We appreciate you, and you've been a great sharer on the CUBE, and we really appreciate that. We've covered Cohesity and the combination, and the momentum at this point in time, we're seeing significant momentum. We were just talking before we came on camera around some of the things Salesforce is doing called ClaudeForce. You guys had Maestro announced in the June-July timeframe. I think it was June. A similar thing, but headless has been a thing, right? Because you're seeing the user interface change. GUIs have been around since I've been in college in the '80s. Now you're starting to see the interface become the models and the agents, programmable. This is kind of a major user shift. You have Maestro. Let's start there. What's the state of Maestro? What does that mean? Because that announcement ties into some of the activities going on with agents. Certainly, Black Hat did a full postmortem on the OpenAI, Hugging Face, off-the-rails agents gone rogue. A lot of governance conversations, a lot of agent adoption on the horizon. Take us through what's on Maestro.
Yeah, quick context: Cohesity has now emerged, it's about 12-18 months since we did the acquisition of Veritas, we emerged as the top player in our space in size. 12,000-odd customers, significant presence in the Fortune 2000, 70% of the Global 500. In certain industries like banking, public sector, healthcare, telco, we just have an enormous amount of share, with the biggest brands relying on us for resilience. Now, what's really happening is the entire world of security, cybersecurity, has taken a whole new frontier because of Glasswing. We'll get to that in a second. What we started to do, as we got closer to the AI players—we've always been pretty close to the cloud players: Amazon, Microsoft, Google; AWS and Google are investors in Cohesity—and as we were talking to them, it was very clear that these two AI frontier models, independent ones, OpenAI and Anthropic—initially it was OpenAI, but later on we saw Anthropic come into the world, and then of course Gemini, Google's model—and we saw the most interest in the enterprise from Anthropic, and then second was OpenAI and Gemini. So we started getting closer to Anthropic, and we became the first company in our space to get inside Glasswing and get our code tested. Binaries were tested by a couple of our customers already in Glasswing.
Explain Glasswing. This is a super important thing that happened, and it was kind of a catch-the-chief moment for security.
Yeah. Around March or April is when Dario, the CEO of Anthropic, announced Mythos. I saw him at a conference and said we need to be inside it. There were only 10 companies in it, but a couple of customers were in it, and there were two security companies: CrowdStrike, Palo Alto. We're not as big as them, but they're very good partners of ours. A few weeks later, we got in, and it was largely because many of our customers, some of the biggest banks and some of the biggest agencies of the US government, said, "We depend on Cohesity as the last line of defense for all the data that sits there." So if a big bank gets attacked, their last line of defense for all those immutable backups is Cohesity. So in some sense, it was good to have our customers lobby for us, and we got into Glasswing. That allowed us to—
And Glasswing was what? Just explain it for the folks that aren't tracking inside the ropes, because this is a super important thing.
Yeah, it's what Tom Friedman called the most nuclear model of all: Mythos running in a way—it's highly expensive. The US government shielded it from the entire world from getting it, but a few companies—I think it was 100 or so; initially it was 10, then 100—
And you can then use that to test both your binaries and source code. We began immediately checking what we found, and we found vulnerabilities we had not seen before, and we immediately got to fix them. We're now the first company in our space to be Glasswing tested and fixed, and we're doing great with them. The Anthropic partnership—we have a Catalyst user conference coming up; they're going to be a keynote speaker there. We're very pleased to host them, but not exclusively. We're also working with OpenAI—they have an effort called Daybreak, GPT-5.5 Cyber—and Gemini. Now you put this all together: two years ago, we announced the ability for our product to act as an agentic platform. We had an agent called CoPilot that allowed you to command it, but it was our UI. We took it a step further, actually inspired by the headless momentum that Salesforce had announced a year ago at Dreamforce. We said, "What if we were to make the entire command interface of Cohesity be one of these LLMs?" Think about 20 or 25 years ago when search tools first came out. People said Windows and macOS are going to be replaced by search as the default desktop. Now, I think prompts are going to be the default desktop of how people command and control their entire IT experience, and maybe even their cowork life. So that's what we did. We're a very API-rich company. We allowed all of our cluster data, our agentic flows, our telemetry, everything to flow into these LLMs. In fact, we showed at our sales kickoff a demo where your command and control was a prompt, and then all of the actions were happening in our product through that prompt. Now, there are certain things, for example, a lot of us build reporting tools for our products: very sophisticated charts and graphs. It's amazing what Anthropic, OpenAI, and Gemini can do with that. You don't have to; you just send them the telemetry, and they do the reporting for you. So there are many functions like that that get significantly easier when you do it. So we announced Maestro. Maestro is just that orchestration initiative across all these LLMs, and we were the first in our space to do it. It is another example of pioneering innovation in an API-rich platform.
Talk about the productivity gains, because you hit a couple things that we've been covering, which is—I mentioned the GUI thing. I've been on this rant, but there really hasn't been a change to the graphical user interface since the '80s when SIGGRAPH started and you had Windows machines, you had SaaS applications, apps you click and you have an interface. Now, this new interface is the business logic is actually in the agents and the data. The prompts are outcome requests: "Hey, give me the third-quarter forecast," or some prompt. Now there's pre-filled decode—Nvidia gets into all that mathematics stuff that goes on in AI—but the user experience, the cat's out of the bag, Sanjay. People love it.
It's not even a question. But this is a significant shift in user behavior and expectation that I've never seen before. This almost makes the GUI concept and how people present systems to users—
I think most of us lived in the workflow of the tools we use most commonly day to day, whether it's Slack, email, messaging-type traffic, or search. And now with prompts—prompts that are very intelligent, and that, by the way, you can use with voice too.
You can start with typing, and then you can continue going on your hike with a voice prompt. This is the new world, and I do think that the future will be prompt-like interfaces that are voice- and text-driven, and you allow the brain to do all the work. We live in a world where inquisitive people ask questions. Think about what a child does when they ask the earliest questions. They start asking why, what, how—open-ended questions. I think all of us are going to have to go back to those early days of our lives and ask these brains open-ended questions. The most innocent question a child wants to ask is, "Why is the sky blue?" If we can get comfortable on the data and the telemetry we produce, asking this brain to go and figure it out without us having to do all that work, the brains of these AI frontier models are extremely powerful. I gave one example: charting and graphing. We'd have people say, "Could you show us this chart and graph with these schemas?" I'm like, make the entire telemetry schemaless, send it into a frontier model, let it discover the schema. I lived my early life at SAP, as you knew me in the analytics business. It's going to revolutionize that. You take raw amounts of data, send it in, it will figure out the schema, and it will produce a better chart or graph than you'd ever been able to do. That's an example of the power of these brains these days.
Well, Maestro's integration—because back in the cloud days, you had to have connectors, you needed to have integration, maybe some engineering. How are you viewing these models? Are you directly integrated with them? Talk about the relationship and how that impacts your customer base. Do they have to do integrations or special connectors?
The way we think about the word connectors: first off, you want to be an API-rich platform. We were built in the modern era, the last 10-12 years, and Gartner recognizes us as the best platform with APIs. That's one. What's the modern equivalent of an API today? MCP: Model Context Protocol. If you don't know what an MCP is, think of it as a modern-day API for the world of AI. We took our product and made the entire thing MCP-capable to then tie into any application that allows you to talk MCP. Salesforce could be one, but every one of the SaaS applications and even traditional applications have opened up their world to MCP. What we're now doing is taking the Cohesity platform and making it MCP-callable, so to speak, by Anthropic, OpenAI, Gemini, whoever you have. Then we expect their richness to take over in a lot of the things that people commonly ask of our platform to do. One example was charting and graphing, but you could think about anything where all you're doing is commanding our platform to do something, and you're changing the way in which you command it to be in words. Those words get translated into a command call that's ultimately an API call, and our platform does the work.
APIs have been great; MCP has been great. Gartner—since you mentioned Gartner—they predict that 68% of agent work through MCP will fail if they don't have the right context, reasoning, and knowledge layer. So talk about the context, because again, about 1% of enterprise data is unlocked. There's a lot of value in your customer base; they want to protect that and make that cyber-resilient. How do you view that? Because now you've got the frontier models as general intelligence, but you've got the movement towards specialty intelligence. You guys fit on both sides because you've got the model relationships, great, but now what about specialism?
Context is key. The key word you mentioned in your entire question there is context. Think of us as the Databricks of secondary data, right? All of a company's data backup is just a time series copy of everything you have. We have 200 exabytes because the largest companies in the world rely on us. They back up their data years and years and years: banks, public sector, healthcare, telcos, so on and so forth. So that 200 exabytes of data has a lot of context in it. It has all the historical data for a large bank. When you run a query—let's say you have 100 billion PDF files in our backup—you want to be able to unveil that context through what's called a catalog. What we did is a product called Gaia. This is the product that we built with Nvidia; Nvidia invested in us to build this out. It's the world's first retrieval-augmented generation, RAG, solution with a vector database on backup data. In fact, we just got a patent for that, John, the first company in our space to do that. So you connect that catalog of that information either directly through our UI, you could query it, or you can interface to it with Glean, Anthropic, or OpenAI. But the bigger idea is even connecting that catalog to Databricks. So now you could be inside Ali's Databricks, or Sridhar's Snowflake, or Microsoft Fabric, and you might be querying unstructured data that sits in our system. We've connected the Unity—
The knowledge layer on top of Databricks.
No, we're underneath Databricks. You should think of Databricks as the primary layer of unstructured data, and where they need some historical context because those documents don't exist in Databricks, they come to us. So maybe in a bank, the current 30 days' worth of unstructured data is sitting in Databricks, but all their 10 years' worth of data is sitting in Cohesity. We link those two catalogs, and now someone can conduct their query inside Databricks, but the actual data is in our data catalog, and you don't have to actually move that data. You just connect the catalogs. That catalog has the context that you need to be able to allow the user inside Databricks or Snowflake. The reason we've been multipurpose here is we don't know who's going to be the frontier model five years from now. It could be Anthropic, it could be OpenAI; you have to work with all of them.
There'll be a lot of them.
We don't know who's going to be the data lake of choice. Is it going to be Databricks? Is it going to be Snowflake? Is it going to be Google BigQuery? Is it going to be Microsoft Fabric? So you connect with all of them. I don't think there'll be hundreds of them; there'll be 5 to 10, in the same way there are four or five clouds. So we built our platform to be open, and this whole project around Gaia and connecting its semantic layer and its catalog—we announced that in March; it's really gaining momentum. Again, these are all first and only in our space. Most of our competition are backup companies. They've really not made the transition to security and AI companies, and we're at the frontier of doing that.
And I like the Glasswing, because that's a super important initiative. Open weights is popular. You mentioned open source—
MCP is open. Open weights is open.
We should talk about that. You have a lot of platform capabilities with open. This is what customers want. Open tables changed the game on data lakes.
It was very clear to me that there were certain companies that were at the front of pushing this. Nvidia is one of them; they invested in us. Anthropic is another; we're pretty close to them. OpenAI is going to be another one. But it's not clear to us that the frontier models are going to be the only ones. So when Nvidia and Microsoft began to announce open weights, we got on that very early. I approached Jensen and Satya, and we're the only company in our space that's part of the open weights initiative that Nvidia, Microsoft, and Dell are involved in. Who knows where this is going to go, right? Are the winners going to be the weighted models, or is it going to be open weights?
I don't think there's a winner takes all. I think it's a power law. Because what we're seeing in our data is, in the enterprise, there's going to be a mix of frontier general intelligence and there's a real shift towards specialized intelligence. If you look at the success of companies like Fireworks AI, for instance, they're allowing AI-native developers to get all the benefits of the CapEx. You have all the benefits of your install-base data you mentioned. Now you have this AI-native front end—call it a developer prompting, could be anybody—they're going to leverage a lot of that CapEx. They need that frontier-like experience for smaller datasets. So specialized datasets come out.
Yeah, I think that outside the US, there's a lot of concern that either because of the US CLOUD Act or these frontier models being US companies, what are they doing with the data? I think there's a lot more willingness outside the US to be friendly to open-weights models. Obviously, we've got to make sure—
