Micron on Why Memory Has Become AI's Strategic Foundation

Micron has raised capital spending aimed directly at closing the memory supply gap, but still has no timeline for when supply will catch up with AI demand.

At the Six Five Summit 2026, Patrick Moorhead sits down with Sumit Sadana, Senior Advisor to the CEO at Micron, for a Semiconductor Spotlight session.

Sadana describes a structural imbalance decades in the making. DRAM manufacturing has consolidated among a handful of companies, while more than 20 companies now design processors—and every new customer forecast Micron receives points to even greater memory demand.

Technology transitions alone can no longer produce enough bit growth to keep pace. Closing the gap now requires new fabs. Micron has increased its planned U.S. investment from $200 billion to $250 billion and accelerated the timeline, but Sadana says meaningful new supply will not begin ramping until 2028.

Key Insights:

🔹 Memory has become a core performance differentiator. AI system performance now depends as much on memory bandwidth and capacity as it does on processor design.

🔹 The market faces a structural supply imbalance. A small group of DRAM manufacturers must support a processor ecosystem with more than 20 chip designers, making it difficult for supply to respond quickly as demand rises.

🔹 Long-term agreements are replacing transactional purchasing. Micron’s Strategic Customer Agreements establish multi-year supply and demand commitments, giving both Micron and its customers greater visibility for capacity planning.

🔹 Micron is dramatically expanding its investment. Capital spending is expected to roughly double from just over $13 billion in fiscal 2025, with spending projected to exceed $45 billion in fiscal 2027.

🔹 Humanoid robotics could become a major new memory market. Sadana expects each device could require hundreds of gigabytes of DRAM and terabytes of NAND storage.

The challenge is not only building more capacity—it is manufacturing increasingly complex products. A single HBM stack combines 12 layers of DRAM bonded to a base die and connected to a GPU, all within a production cycle that takes nearly five months from wafer start to shipment.

Sadana describes the result as close to “a miracle that it actually works.”

Watch the full video at sixfivemedia.com, and subscribe to our YouTube channel so you never miss an episode.

Explore more sessions from Six Five Summit: AI Unleashed 2026 at sixfivemedia.com/summit.

Disclaimer: Six Five Media is for information and entertainment purposes only. Over the course of this video, we may discuss companies that are publicly traded, and we may reference their equity share prices. Nothing discussed during this webcast should be considered investment advice or a recommendation to buy or sell any security. We are not investment advisors, and you should not rely on this content as financial advice. Six Five Media collaborates with technology companies and industry leaders to produce research-driven interviews and multimedia programming for enterprise technology audiences.

‍

Sumit Sadana:
AI system performance today is dependent to a first order on memory subsystem performance and memory capacity. These are the two things that really define how well an AI subsystem can do.

Patrick Moorhead: 

Welcome back to The Six Five Summit 2026, AI Unleashed. It is amazing how important and strategic memory has become. I mean, listen, I always thought it was important, but given the age of AI, it's incredible what it has enabled us to do. Because quite frankly, without enough memory, without the right kind of memory, all the amazing stuff that you see around us cannot happen. To have this discussion here, I am pleased to announce Sumit with Micron. Sumit, great to see you.

Sumit Sadana: 

Great to see you, Pat. Thank you for having me.

Patrick Moorhead: 

Yeah, you know, it is just amazing. I've been in the industry 35 years, a consumer of memory, a partner with memory, and now analyzing the memory market and all the corresponding ones around it. But it is amazing how far we've come here. And, you know, last year when we talked, I actually categorize memory as strategic. I'll take a little bit of credit for that. But quite frankly, you deserve a lot more credit because you actually do it. I just talk about it. But today, that perspective really seems a lot more relevant and a lot more people are talking about it. What has changed in the last year, in this age of agentic AI, that the industry thinks more about memory, and why has AI actually accelerated this shift?

Sumit Sadana: 

Well, great question, Pat, and that was such a good discussion that we had, and I do agree with you. A year has gone by, but it seems like a very long time in terms of how the world has changed in this last year. Certainly from the perspective of our customers, a lot has changed. We have gone into a situation where memory is in significant shortage across all market segments, And we also, despite our best efforts on supply, don't have line of sight today as to when we will be able to have the supply catch up with demand, because demand signals from our customers, and again, across all segments, continue to escalate in every year that we get these projections from our customers. And we have a situation where customers are looking at memory in a different way for a few reasons. One is, of course, this big gap between supply and demand. But the other, of course, is also that when you think about AI, the performance requirements from the systems are going up. And as you think about what makes a system perform at peak levels, at peak capability, it's not just about the processor. A lot of it is now driven by the performance of the memory, the bandwidth between the processor and the memory, and the capacity of the memory itself. Because these AI models need to be stored in memory, a lot of data has to move back and forth. So that performance, that bandwidth between processor and memory becomes a key bottleneck. So customers are realizing that they need to plan their roadmaps, define their roadmaps with a different approach to memory. Now, it is all about how can you design memory to gain a competitive advantage. how can each customer differentiate their system using memory? If it's going to be the same JEDEC-compliant off-the-shelf memory that everyone else is using, as used to be the case a few years ago, then you cannot so easily get that differentiation. So we are working with customers embedded now in their multi-year roadmaps, and they are now focused on how can we eke out a differentiation? How can we change the game on our competition? And that means thinking about memory in a different way, partnering with a company like Micron to design a different type of capability, that then can be used and leveraged. And one example I'll give you there is what we have done with NVIDIA on LPD RAM in the data center. We were the first company to introduce that capability in the data center, and we were sole sourced in that segment for a long period of time. And that is now starting to proliferate. More customers are starting to understand that it's so much better to use lower power DRAM. It helps them through numerous benefits of density, size, performance, and of course, significantly lower power consumption, which is a critical part in the data center. And that's just one of many examples that are now starting to proliferate as we look out for the next five years.

Patrick Moorhead: 

Yeah, it's interesting. When I try to explain why we've seen such an increased interest in memory, I get back to a college course that I took, which was explained to me. This was in the late 80s. I'm old. Where really a basic fundamental is, if you can stay inside of memory, you're going to increase performance. But this is a little bit even different. By the way, it's true. But you get better results the more memory that you can have and the closer it is to what you're doing. Because anytime you go outside of memory, things slow down. And in this age of AI, that's a challenge. It's interesting. I've seen nine memory cycles in my career. And again, I worked as an OEM. I worked at a chip company and now an analyst firm for the last 15 years. It's been, seemed to be the most cyclical, even though I know there were other curves, some other industries, but also looked at as quite commoditized. You answered a little bit of that in the first question, but why is AI fundamentally changing that perception?

Sumit Sadana: 

Yeah, it's a great question. And if you go back to the 90s, in the early 90s, there were upwards of 20 companies that were doing DRAM. And DRAM went through a lot of consolidation in the industry. And back then, they were like just a couple of processor companies, right? And you fast forward to today, and now you have more than 20 processor companies. I mean, if you actually sit down and count all the companies that are doing processor design, you'll get to that number pretty quickly. but a very small number of companies were doing DRAM. So when you think about AI design, AI subsystem design, not just in the data center, by the way, you're talking about in the data center, of course, training workloads versus inference workloads. And even the inference workloads are going to become more and more fragmented into different types of inferencing. Some can use SRAM in a small category of those, but a large portion still very much needing as much DRAM capacity as we can design into these systems. Then you go from there to even autonomous cars, industrial systems. Look at smartphones and PCs, and you had this whole open-claw movement with so many of the Mac minis flying off the shelf because of the unified memory architecture. where it all comes down to AI system performance today is dependent to a first order on memory subsystem performance and memory capacity. These are the two things that really define how well an AI subsystem can do. And so when that becomes a core design point in the system, then you have to start thinking about memory in a different way. And then it all becomes, how can you get to market fast with a differentiated product? And this is where I think the memory business of tomorrow is going to be very different than the memory business of yesterday. And combine that with the idea that the whole industry, as it turned out, strangely, got to a point where greenfield capacity expansion was needed. Because the demand growth driven by AI has been so significant, there is not a way to satisfy that demand just through technology transitions, which was how it was done for several years. That tech transitions used to produce enough bit growth to satisfy the demand growth. Now, dramatic amount of wafer increase is needed in the industry. That means building out new clean room space, And when you run out of clean room space in existing manufacturing sites, which most companies seem to have reached that point simultaneously, you have to build out capacity in greenfield sites, and that just takes a very long time. It's a really time-consuming process. And that process is bottlenecked by a lot of physical aspects of how fast construction can be done in various parts of the world. Regulatory approvals can be taken in various parts of the world. How do you take a place which is barren land or just trees and bring up all of the infrastructure necessary to bring in power, water, and water treatment plants? chemicals and the whole infrastructure needed to set up fabs. It takes a long time and that's what the industry is going through. So we are now in a situation where we are going to struggle to meet demand despite the best efforts from our side and what we are able to see from others in the industry. it's going to take quite a long time for supply to be able to catch up. And that is another factor that's also changing customer behavior. And this is where we have been talking about these strategic customer agreements. And so these strategic customer agreements are fundamentally changing our business model and our approach to business. And this is where we are doing these multi-year deals, committed supply to our customers, and they give us committed demand. Unlike previous years, where it used to be like a one-year agreement, they would buy if they feel like it, we would supply as long as the supply was there. These are like committed supply agreements. and they have really robust ROI so we can underwrite the CapEx needed for a multi-year time frame. So that's how we're looking at it. Of course, this doesn't even get to the point of agentic AI, which is the next phase of AI growth. After that, you get to even physical AI, which is the next big wave of demand. So if you look out these different waves of demand, they compound on each other, they build on each other. It's not like one replaces the other, but they sort of build on each other. And that's where comes the big challenge of how do you bring so much supply physically online in this period of time to feed all of that demand.

Patrick Moorhead: 

So I kind of see your SEAs as a proof point that memory is more strategic. I mean, I know it's easy to lump it into, hey, we need to have SCAs to lock in future capacity. It is that, but it's also co-planning technologically. I was really struck, and I'm glad you brought up AI on the edge. The most interesting things, whether it's with OpenClaw, are some of the new designs that the client computing folks are doing with their processors. Very tightly coupling memory with the CPU and the GPU. to enhance the bandwidth to be able to do that. So, and that is a, you don't see architectural changes a lot in client computers. They've been pretty static for a long time. I want to move on to, we talked a little bit about hyperscalers, and that's where a lot of the action's happening, just if nothing more, the investment in the CapEx, and at least right now, the innovation on models and applications. I'm curious though, when it comes to autonomous vehicles, we talked a little bit about PCs and devices, robotics, intelligent edge, how are memory requirements changing those environments?

Sumit Sadana: 

Yeah, it's a great question. And it all started in the data center, of course. But as we think about how AI will progress, it's not going to just be limited to the data center. That intelligence wave is going to move towards the edge. And ultimately, intelligence will be everywhere in all devices. So it's all kinds of consumer devices, your car, and of course, even the industrial revolution phase two, where you'll have all of this intelligence distributed around the world in enterprises. And of course, your personal devices, your PC, whether it's at work or at home, as well as your smartphone, and even companies thinking of new devices that may not be the same smartphone or PC form factor, just completely different form factors. Because when you think about building a native AI-first device, you can think of different ways in which it can be done versus the traditional PC and smartphone approach. In all of these devices, again, it goes back to with the lowest power consumption envelope that some of these devices that need to be battery powered, for example, can get. How to maximize performance? How to be able to run the small models that need to be big enough to deliver value to the consumer without frustrating them, and yet need to be small enough that they can be run on device and don't need to go back to the Cloud. That is something that is the new frontier of AI where that experimentation will continue those models that are small enough to fit on a device like a PC or a smartphone will get better and better with time. As that happens, you get into the big positives of privacy and confidentiality of data. So consumers value that highly and those companies who can then use these on-device models and give consumers the comfort that everything that they interact with that device is not going to the Cloud for somebody to pry on. then that is something that will unleash a whole new set of applications that will bring a lot of value. And then you can think of autonomous driving, which is starting to really do well now and is really being aided by AI. And the next frontier, which is all of the robotics-driven capabilities. And you think about humanoids, That's going to be one of the largest markets of any product that has ever been built. Of course, it will take some time, but we are getting closer and closer to that singularity when you'll be able to talk to a humanoid just like any other human, almost indistinguishable in terms of cognitive capability. And then over time, these robots will get more and more sophisticated with their physical capabilities, and that would be a revolutionary outcome. So initially, it'll start with areas where they're in a factory automation type of environment, very specific set of tasks, fewer degrees of freedom. And once they graduate from there, the most complex environment is in the home because it's very unstructured. and you'll get to that sort of a situation. But I'm confident that you get into latter part of this decade and into the 30s, robotics is going to be a massive growth driver. But the thing about memory that's so exciting with robotics is it's going to be a huge amount of content per device. Each humanoid is going to have hundreds of gigabytes of DRAM content and terabytes of SSD NAND content. And so they're going to be massive drivers of growth in demand for both DRAM and NAND.

Patrick Moorhead: 

Yeah, I think a lot of the times people can't fully appreciate that it has to be autonomous and it can't phone home every time. That's right. And also, as you were going through that, I was struck by thinking that Micron is enabling, you know, picowatts to gigawatts. That's right. And the thought of having to invent the technology To do that just scrambles my brain a little bit, but I think it's something that people need to keep in mind when they think about Micron. So as you've gotten more strategic in the age of AI, whether it's from a cool device that hasn't been invented yet, released in a year to tap directly into AI and models, to gigawatt data centers, this idea of co-design or co-invention comes up. And I've heard, I've talked to some of your partners about it, and they talk about some ideas of what they're doing. I talked a little bit about the changes in architectures already. And those were traditionally through, I would call it more of a JEDEC standard type of thing, but now we're looking at a serious co-designer. Can you talk about that changing relationships that are going on with your partners?

Sumit Sadana: Absolutely. Yeah, I mean, our partners are looking to figure out how they gain an edge in their markets against their own competition. And the memory processor subsystem has become such a critical portion of the AI subsystem because that memory processor interaction is essentially what defines so many critical parameters of an AI system. it defines how much power is going to get consumed. It defines what kind of models can you run, what size of models, how fast can you run them. And when you think about all of these advances that are happening in LLMs of all shapes and sizes, for all different market segments, To create that differentiation in the hardware, memory cycles on the design side and processor design cycles are so long that it is imperative that the design start in very close collaboration. This is not a plug-compatible JEDEC-compliant part that you can plug in after the fact and just do some qualification and run with it. There is some part of the market that will be that. But there is increasingly a need from customers to say, how do I differentiate? I need to be able to create some special capabilities that off-the-shelf products just don't provide to me. Let's partner with a memory company, get into a long-term, five, seven-year type of roadmap discussion on R&D. And we are getting embedded in the R&D roadmaps of a lot of our customers. They're coming up with very interesting ideas. Some of them can work. Some of them take longer to create the inventions and innovations needed to create those breakthroughs. But we are very, very excited about the diversity of ideas that we see. And like you said, these span such a broad range of systems. And just like you said, I mean, some of these devices consume very, very low power because they're all battery operated all the way to gigawatt scale data centers. And we have seen innovations come out of several of these segments and some of the leading customers in each of these segments. And we'll continue to see that. And we are working actively on very exciting projects that stretch the capabilities that we have to deliver some fascinating design concepts that our customers are eager to pursue. So deeper discussions, they are not able to partner in that deep fashion with multiple companies. it's typically just one or two and then maybe others would be able to catch up years later. So this is another benefit of this model is you really can partner that closely with just one or two players. So oftentimes, when you go down that path, you become single sourced for a period of time or there is just maybe one other source of supply. and it changes the relationship to more like an ASIC model of engagement where the customer takes more of an ownership of the design in terms of the resources and then there is that longevity that you have on the demand side because you are in that special ASIC-like relationship and it's not the type of memory type of relationship that existed years ago.

Patrick Moorhead: 

Yeah, and the big difference that I see, too, what's changed so much is you can do this at scale.

Sumit Sadana: 

That's right.

Patrick Moorhead: 

Meaning the volumes are so big. And quite frankly, the dollars enable investment on both sides to be able to do that. And you hit it when you kick this off, which is about differentiation. That's right. People are looking for a way to, whether it's performance, efficiency, or TCO, pick your metric. They want to be different. And I think that's important. So I'd like to look ahead here from an investment point of view. You did talk a little bit about on the capacity side investments. It's funny, if I look at some of what's written, there's almost this expectation that you can vibe code a memory factory. But, listen, I used to work for a chip company that had a fab, and even today looking, it's a very, very difficult capital intensive, and you have to make a bet three or four years before anything is coming on the other side. Can you talk about the investments that you're making that really solidifies your role in the next generation? Of course, it's going to be AI. Maybe it's going to be AI or something different or a different variation of AI. But again, investments do take a long time. Audience, you got that out there? Don't make that mistake.

Sumit Sadana: 

That's a very good point. And by the way, if anyone knows how to wipe code DRAM factory, we would certainly like to know that because it'll make our life a lot easier. Our customers definitely want to have the supply quickly and we are really doing everything we can. We have pulled out all the stops and being able to accelerate all of these investments. If you look at just our CapEx profile, we have gone from just over 13 billion of CapEx in fiscal 25, fiscal 26, we'll do double of that. And then we have pointed to fiscal 27 being above 45 billion even. And so that number has escalated through time. And we have just recently announced that in the U.S., we have taken our $200 billion investment up to $250 billion and even pulled in the timeline of by when we expect that $250 billion to be invested. And this is across the board, we are investing in Idaho-1, what we call ID-1 fab. That's going to have first wafer routes in mid next year. ID-2 just follows right after that. It's going to have first wafer routes close to the end of calendar 28. We have the Tonglu fab that we purchased in Taiwan. That's going to come out with wafer starts also in 2027. We're making further expansions there. We're making expansions in Japan. We are making expansions in Singapore. So we have like 20 investment projects going on around the world of all different shapes and sizes to bring on front-end capacity, back-end capacity. But it takes a long time to do that. We have to, for every Greenfield fab, for example, whether it's in Idaho or in New York, New York, we have a four fab cluster. First fab, we expect to have output in 2030. We broke ground on it in January. But ahead of schedule, we've had this concrete pour milestone as well. Between that and the Idaho fabs and the Manassas fab in Virginia, we have a lot going on as well around different parts of the world. But these projects, you have to do all the permitting, you have to build out the entire infrastructure. There is a lot of construction talent that is in short supply. So if you think about all the construction that's happening around the US and in many parts of the world, you have all these data centers getting built out. These data centers need power, so power plants are being built out. They need semiconductors, so semiconductor facilities, both front end and back end, are being built out. So you look at a big crunch in terms of people who are skilled at their job to be able to do all of this very sophisticated construction. And there is not enough technicians to be able to man all of these fabs. And so we are investing in communities to bring up talent to really ensure that there is that pipeline, not just helping us to make all of these projects happen, but ultimately will be a help for the whole ecosystem. So we are making all these investments and human talent in our community colleges, working with towns and communities where we do business. And this is a long-term endeavor for us. So we expect that there will be a continuous increase in the scale of our facilities over time as we bring up. As you know, these are like clusters of fabs, right? You can't just build one fab. You need that scale and you need to get past the knee of the cost curve quickly. So you have to keep building till you get past the knee of the cost curve, which means you're talking about expansion that will continue for the next decade and beyond, right? These are long-term projects for us. And we really wish we could snap our fingers and just bring capacity up online, but it just doesn't work like that. So we think that meaningful supply will start to ramp only in 2028. And that too, it'll be only an initial ramp, right? It'll gather momentum only in the years after. And so it will take a while for the industry to find a new equilibrium point, but when exactly that happens, we don't have line of sight to yet.

Patrick Moorhead: 

Smith, that's great example, great information about the investments you're making. One thing, though, that I think some people might get wrong, I've seen it written out there, is the complexity. We know logic is difficult, okay? Talk to me about the complexity of memory.

Sumit Sadana: 

Memory is one of the most complex semiconductor technologies out there. We know we talk about logic being on the leading edge, but memory for sure is also very much a leading edge technology. And it uses EUV machines, which are massive. There are like 2,000 process steps in the fab. and you start to run a wafer in a fab, and by the time you can ship a product to a customer, it's like a five-month cycle time. That's amazing.

Patrick Moorhead: 

People have no idea.

Sumit Sadana: 

Yeah, exactly, right? I mean, so it takes three and a half plus months to go through the fab, approaching four months, and then a month on the assembly packaging and test, sometimes a month and a half for more complex products. And if you think about the complexity of products like an HBM product, right? I mean, it is been called by different people like a Marvel, you know, it's sort of a miracle that it actually works. And, you know, think of stacking 12 high die of DRAM, one on top of the other, and having a base die at the bottom that connects to a GPU. And that whole thing has a very significant level of thermal challenges and power consumption issues. And how do you cool that thing while it's operating at such high bandwidths, moving data back and forth from the processor? The packaging technology alone is like a clean room environment. Very, very different from the type of flip chip packaging that we have done in the industry for a long period of time in the logic domain. So no matter which way you look at it, whether it is sophistication on the packaging side or on the front end process technology with EUV and each node of DRAM now becoming more and more difficult to scale. And same thing on the NAND side, we are piling up these layers, over 200 layers, over 300 layers, over 400 layers, and then being able to do these devices through that, the process complexity of that is just enormous. And then to be able to have these products on the NAND side store data and do this, a single SSD now can be like 245 terabytes, right? I mean, just enormous amount of capacity that we can fit into such a small form factor is just nothing short of a miracle that the engineers and the manufacturing team members perform every day. So this is just an incredible amount of complexity And then to continue to scale this in the years ahead, to go to finer and finer pitches, and to actually make this technology work at scale that is needed is just another marvel, right? So it's just something that takes a lot of time, effort, but a huge amount of technological innovation as well.

Patrick Moorhead: 

I'm amazed at how complex it is, and I'm also amazed that you can do it across such a scale of power planes. That really blows my mind. I'm unaware of many people on the planet who can actually do that. Smith, this is great. It's funny, I have to note, people are talking a lot about, oh, I can't get enough of this. I look at that as they're not getting enough awesomeness. And they want more awesome. And then there's discussion about, oh my gosh, the pricing and stuff like that. But it's like, I look at this as investment. Micron and other folks in the memory industry, it's okay for them to make investments. Because quite frankly, if we don't make more investments, the cycle of innovation is going to stop. And I'm excited to see what Micron can do in the future with its investments, not only in the foundries and the fabs, but also in the technology itself. I do believe that robotics could be a 10x, right? It could be the big growth market after, again, not after we built out the data center or the data center edge, but it will be another adder that I don't think we've even contemplated. in our capacity models, but I know we need to see a little bit more evidence of it, of when it's going to hit before it hits. Smith, thank you so much for this conversation. I can't believe we did this a year ago, and the conversation, the situation, where we were and where we are today. I've had a lot of fun, and I know the audience has too.

Sumit Sadana: 

Thank you so much, Pat. I really appreciate the dialogue and great time to have that discussion last year and what a difference a year makes. So thank you.

Patrick Moorhead: 

I know, it was amazing. So I want to thank all of you for tuning in here for the Six Five Summit 2026 AI Unleashed. It is definitely unleashed. Check out all of our content about Micron as memory as well. We've covered it a lot. And check out all the analysis on Daniel's and my website as well. Take care. Hit that subscribe button. Bye-bye.

‍