Janhavi Giri

I am a principal architect, field CTO at NetApp, based in the San Francisco Bay Area. My field of expertise is in semiconductor manufacturing, EDA, and AI. And at NetApp, my responsibility is to partner with the semiconductor customers and figure out what kind of data strategy they have and help them build their data solutions.

Oh, it is the foundation of everything we do these days. So, you know, the semiconductor industry, traditionally, if you look at it, has been very heavy on transistors, in terms of how you can scale smaller and smaller so that you get better performance. So that is the conventional notion of how the semiconductor industry has been, and that is what Moore's law is also. But over the decade or so, this industry has grown tremendously. If you just look at the market size, the revenue of the semiconductor industry has grown like 188 percent from what it was 10 years ago. So there has been a significant growth in the semiconductor market. And people usually think that, oh, it is maybe because we are shrinking the transistors, we are increasing the number of transistors on the IC chip, and maybe that is what is leading to all this growth, and maybe that is what is contributing to this data boom that we are seeing. But it is not, essentially. It is because of what everyone is talking about these days, which is AI. So when I say AI or anyone says AI, people start thinking about models in general. They talk about OpenAI, they talk about Gemini, they talk about Claude. But what is essentially fueling the AI? What is enabling the AI? That is the silicon. The silicon is the foundation of the technological advancements we are seeing through AI. And at the center of all that, to enable AI, to make sure that we are able to manufacture and design those advanced AI chips, is the infrastructure. So industry-wise, if you look at what is the core enabler for this AI revolution, it is the infrastructure layer. And when we talk about infrastructure, especially in terms of AI, there are these five main components. If you've seen NVIDIA GTC and Jensen Huang's keynote, he addressed that there are five layers of the AI cake, and one of them is storage. Now, storage has mostly been the silent foundation, but nobody actually cares about it, because it's just in the background. But now, with the kind of performance we need, the kind of context preservation we need with all these large language models, storage is becoming both the critical enabler and the bottleneck for the advancement of these technologies. In the conventional model, storage was something that, okay, you do it once and you forget about it. But not anymore. Storage is now becoming the bottleneck for all these AI workloads that we are seeing in the industry.

Yeah, that's capacity.

That's throughput, yeah.

Exactly, yes, that is exactly it. So whenever you start looking into what your storage strategy should look like, you're always thinking in terms of capacity. Capacity is the first thing that comes to everyone's mind. And then comes the latency: how quickly you can access the data. And that is so important these days because, especially in the semiconductor industry, everyone pretty much is operating in a hybrid environment, where some of the workloads may be running on CPUs and some of the workloads may be running on GPUs. And you know how expensive the GPU resources are. So in order to make sure that their GPUs are not waiting for the data to arrive, you need to have storage that can make that data quickly accessible and quickly replicable. You don't want to make copies of the data; you want to make the data readily accessible to the GPU, or anywhere else you are running your simulations or your workloads. So capacity is the conventional way of looking at storage, but these days, with the kind of AI workloads we are running and the kind of performance requirements we are looking at, latency, throughput, and how you can make the data more secure are becoming even more critical.

So first of all, yeah, you touched upon a couple of things: data security, data accessibility, data connectivity, and storing high-volume data as well. And the other thing, which we didn't talk about, is bringing intelligence to the data layer itself. That is, how do you bring AI to where you actually store the data? The conventional model is that you have your data residing somewhere, maybe in your file system, object system, wherever your data is. Then you bring the data somewhere else, maybe to your Databricks platform, any platform you have, your in-house analytics platform. So what you do, essentially, is you make copies of the data, you move the data somewhere else, and data has gravity. So what if you can do all those functionalities at the data layer itself? Because, as is known industry-wide, 80% of engineering effort goes just into the data preparation and data discovery step itself. So all the new enhancements that are coming in are focused in this direction. How do you make the data layer itself intelligent, so it takes away all that burden and helps with engineering performance?

Absolutely, absolutely. You put that very well. And you're doing all that at the data layer. You don't have to migrate your data to another platform.

Processing time, and it is true for... so my area of focus is the semiconductor industry, but this problem that we are talking about is applicable pretty much everywhere. And nowadays, we're not just looking at structured data. We are looking at unstructured data as well. All this automation that we are building with agentic AI is generating more and more data, and that is mostly coming in the form of text files. So how do you identify the relevant files that the agentic framework needs, right? You don't want to have an additional step that will go and look around and fetch that data. So if all that can be done at the data layer itself, then you are eliminating the extra effort that goes into it, which adds to your processing time and slows down the entire engineering process. That is where NetApp is focused these days. There's a lot of effort going into how we bring that intelligence to the data layer, and the company will be announcing new products at our upcoming conference, which is our INSIGHT conference. So there will be some announcements in that direction.

It is an AI company, absolutely. The conventional definition was that NetApp may be a data storage company, but now we call it an AI data company, which is basically bringing intelligence to the data layer and making your data accessible and secure, because security is at the top of everyone's mind these days.

Right, that's a great question. So the company has tools that provide the customer visibility into their entire data infrastructure landscape, so they get a single view of where their data is, and everything is installed in their environment. NetApp is not accessing or pulling their data at all. Even the data infrastructure is critical IP of the customer. They don't want to share that information with anyone else in the industry. So we do have tools in our product portfolio that give them visibility into how their storage infrastructure is being used and where the issues are. And the agentic AI solutions that we are building, which help them get that visibility into their data infrastructure through an agentic framework, will be running in their environment. So we are not pulling any information from them. We are giving them the tools and the capabilities that they install in their own environment. It is all secure within their own security guidelines.

We have both, cloud and on-prem. And that is one of NetApp's core capabilities as well. NetApp ONTAP is our flagship software product, which is one unified platform for both cloud and on-prem.

Actually, especially when it comes to semiconductor customers, they have a lot of concern in terms of where their IP is going. So compared with any other industry, from the semiconductor customer's perspective, they want to do everything on-prem. The only reason they want to go to the cloud is so that they can leverage the compute, because the needs are just growing exponentially, and there's a limit to how quickly you can scale if you're doing things on-prem. So that's where they are seeing the value of the cloud, but their preference is always on-prem.

Yeah, so that's a great question, and I have a really interesting, funny, I would say life-saving story. My connection to NetApp, even prior to joining NetApp, happened when I was a computational lithography engineer at Intel. One fine day I was doing my job: pulling data, running simulations, dispositioning the outcomes, and everything. In our world, there is always this one golden file that everybody has access to. That is the file we basically run our model on and simulate, and then we disposition the result, and then we say, okay, this is the final output that we then move to the next layer. And that golden file is your actual design file that everybody has access to. And what happened is, I had it open in my tool, and I probably did something, and I corrupted the file. And I was freaking out, because that file was probably accessed by 10 more teams in my company, and I had just corrupted it. So I had to go back in time and get the original file. How would I do that? Because that was just one file. So NetApp has something called a snapshot. A snapshot, just as the name implies, is a copy of that file at a particular point in time, but made in such a way that you are not essentially consuming any space in the storage. That is NetApp's proprietary technology. What it allows you to do is recover your data back to the point in time when that snapshot was made. So I was actually saved because of NetApp's snapshot technology. And that's when I realized, okay, storage is not just about archiving your data and putting it somewhere and forgetting about it. It is actually life-saving. It helps you recover your data, especially in scenarios like that, where any engineer can accidentally corrupt a data file. So that's where I was in awe of the technology, because it was a critical process technology data file that I corrupted by accident, and I was able to recover it.

It has changed significantly, because it depends upon which segment of the industry I'm talking to. If I am talking to, for example, the chip design or fabless customers, maybe like AMD or Apple, it depends on what stage of technology they're working with. So if I'm talking to chip designers, for example, the AI adoption over there is... especially when I say AI, AI can mean a lot of things. So let's backtrack a little bit. AI is not new to the industry. AI has been there since the early '70s in the semiconductor industry, and manufacturing has been way, way more advanced in terms of adopting AI for defect detection, metrology, doing simulations, all that. So AI is not new to the industry. What is different now between manufacturing and chip design is that the chip design folks have moved ahead in terms of adopting agentic AI solutions. Manufacturing is not so far ahead. But when I am talking with either of them, AI is on everybody's mind. Everybody has this agenda: they want to show how much they can leverage AI to increase engineering productivity and save cost. So they want to learn from us how we can help them do that at the data level itself. Based upon what their challenges are, the conversations usually always start with data discovery, because everybody wants to accelerate the process of root cause analysis. When any issue happens, how quickly can you debug it, and how quickly can you identify the source?

So if you just think about my personal background, my background is in physics and applied mathematics. I don't classify myself as either a hardware or a software person. I classify myself more as a domain expert.

Yeah, I am more of a math and physics person. So whenever I look at a problem, I'm more like a problem solver. I always follow a framework for how I'm going to solve the problem. So I'm more like a framework person; I think in frameworks. Whether it's a hardware problem or a software problem, I always approach it like, what kind of framework can I build to provide a solution to that problem?

Yeah, absolutely. Like, how can I model it, and how can I ensure that my data flows smoothly from one stage to another stage? Because often that's not the case. The data is so fragmented that the majority of the time is spent just pulling together the essential data you need to answer that problem.

Right, absolutely. So, generative AI. And this is my personal opinion: folks of my generation are the ones who actually know how to effectively use AI, because we have gone through that conventional learning path. We have learned from our mistakes. We know how to do research ourselves. We know how painful it is to write a PhD thesis. We know what writer's block means, right? Which, unfortunately, the generation that is just growing up with all these tools at their fingertips is not getting to experience. So when I am leveraging generative AI, I'm actually leveraging it as a tool to increase my own productivity. I'm not dependent on it. If you don't give me OpenAI, or if you don't give me access to Claude, it's not that I cannot do my work. It will just take me a little bit longer than what I currently do. If you ask me for any information and I use any of those generative tools, I can probably give it to you in less than 10 minutes, maybe in a minute. But I know how to make sure that whatever information is generated is not hallucinated, and whether the references are credible or not, because when I look at information, I know what is the correct information and what is the hallucination. That is coming from my conventional learning. I have gone through all those disappointments and those failures, which have taught me that. But folks who have these tools from the very beginning of their learning are becoming more dependent on them. They don't have an opportunity to be wrong, because AI gives them all the answers. So how are they going to learn?

I might have said some controversial things here.

Right. No, I do agree with your point that whenever these kinds of technological revolutions happen, the prior generation tends to think that their way was the best way, and that you can only be successful if you have gone through hardships and had failures. As a prior generation, we may think that this generation has too much comfort and doesn't know how to deal with failure. But that is one way of looking at it. The way to address it is that it will require us to introduce some systematic changes in the way education is conveyed to them. Even the conventional career path that we all had while growing up, I don't think it will be applicable anymore for this new generation coming up. So if we keep the conventional curriculum, and this generation has to perform with respect to that curriculum while they are using all these AI tools, then I don't think we are doing justice to them. We have to revise our curriculum. We have to revise our assessment methods. As part of my role, I collaborate with academic institutes like universities and research centers, and this is at the top of everybody's mind: we need to change the way we assess the students. It's no longer about a right or wrong answer. It is more about looking at their thought process. Because we cannot avoid it; we want them to adopt AI tools, so we cannot completely say stop. There are places where countries are putting restrictions on AI usage at the elementary or educational level. But I guess it requires us to think very differently about how we are approaching this curriculum and learning process, and we have to adapt along with that.

I am both anxious and hopeful. Hopeful in the sense that I see a lot of conversations actually happening at the ground level. In our education system, folks who have the opportunity to shape the curriculum are thinking about it, and it is becoming the focus of their conversations. So that makes me hopeful, in the sense that yes, it is something that society recognizes, and even the parents are all aware of it. There are a lot of conversations happening around it, and genuine actions are being taken as well. Anxious is more in terms of, of course, I would like to see my children be successful in this changing society, but I really don't know what kind of landscape they will be entering once they are ready to enter the job market, or even in their own personal lives as well. Because how do we reach out to new people these days? That has also changed significantly.

There was what you did. Correct. And if you did this thing, then you would have success, whatever way you defined your success.

My older one is 14, the younger one is seven. So I am more concerned about my seven-year-old. My 14-year-old, I think, will still be in the transition phase.

I think he's excited about the technology, and he loves technology. He likes to build things himself, and he's way more proficient in leveraging OpenAI tools than I am. All the kids are way, way more tech savvy, as you would expect them to be. But at the same time, I think he's very aware of what he's using the tools for and how to use them effectively. So it makes me hopeful that he will be able to find a path for himself in this changing situation. The youngest one, I don't know.

And society will look very, very different when they grow up. I don't know.

They have to. There is no other option.

Right. And I think that's where what is happening in the schools is so important. Right from the elementary level, just like we prepare them for an earthquake, what to do if an earthquake happens, what to do if you have some unwanted person on the campus, they have those drills, right? So right from elementary school, they are being taught how to deal with those kinds of difficult situations. Just like that, again, I'm thinking in frameworks, there has to be some framework given to those kids right from the elementary level, so that from early on they're aware of what this kind of exposure to this technology is going to lead them into and how they can be better prepared for it. Of course, families matter; what we teach them at the family level matters. But the major chunk of their time is spent in school. So whatever foundation we want to build, it has to happen at that level. That's what I would personally like to see.

Exactly. And it has to start right from a very young age.