Micron executive: Storage determines AI limits, substantial new capacity will come after 2028

As the AI wave spreads from data centers to the edge and robotics sectors, Micron executives warn that storage has become the most critical bottleneck limiting AI performance. Furthermore, due to the extremely lengthy new production cycles and highly complex manufacturing processes, there won't be a substantial release of new industry-wide capacity until at least after 2028.

On September 15, at the semiconductor-focused forum of the Six Five Summit 2026, Micron’s CEO Senior Advisor, Sumit Sadana, was interviewed by analyst Patrick Moorhead, where he broke down in detail the supply-demand dynamics in the AI era storage industry, capex plans, and future core incremental markets.

In the past year, the entire technology industry’s understanding of storage has undergone a fundamental reshaping. The root of this transformation is that computing power is no longer the sole measurement for AI system capability. In the interview, Sadana pointedly identified the underlying logic of today's AI hardware: “Today, the performance of AI systems is determined first by the performance of the memory subsystem and the memory capacity. These two factors truly define the performance level of AI subsystems.”

He explained that since AI models must be stored in memory and vast amounts of data are transferred back and forth in memory, the bandwidth between the processor and memory has become the key performance bottleneck. Meanwhile, the storage market is facing a structurally entrenched imbalance—back in the early 1990s there were more than 20 DRAM companies, but now only a handful remain; by contrast, there are currently over 20 companies designing processors. This “few-to-many” funnel-shaped supply dynamic makes the memory supply chain extremely fragile in the face of explosive AI demand.

Capex Soaring, Substantial New Supply Not Expected Until 2028

Faced with a huge gap between supply and demand, technological transformations alone can no longer drive bit growth fast enough to meet market appetite, forcing the whole industry into a new heavy asset expansion cycle of “building new cleanrooms and fabs.”

Micron is embarking on an unprecedented capital expenditure plan. Sadana revealed, “Just looking at our capital spending plans: in fiscal 2025, our capex is a little over $13 billion; in fiscal 2026, it will double; and in fiscal 2027, expected to exceed $45 billion. We recently announced in the US that we are increasing our investment from $200 billion to $250 billion and accelerating the timeline.” In addition, Micron is simultaneously advancing about 20 expansion projects in Idaho, New York, as well as in Taiwan, Japan, and Singapore.

However, these efforts are “a distant solution for immediate urgency.” The construction of semiconductor factories is hampered by infrastructure, regulatory approvals, and a serious shortage of skilled technical workers. Sadana gave a clear timeline expectation: “We believe that truly effective new supply will begin to be released around 2028, and that will only be the early stage. The momentum of capacity growth will not truly pick up until a few years later. Therefore, it will take some time for the entire industry to find a new equilibrium.”

Another reason for slow capacity release is the astronomical complexity of memory manufacturing itself. Take the current hot HBM as an example, Sadana said it working at all is “nothing short of a miracle”: “Think about stacking 12 layers of DRAM chips one atop another, with a base chip at the bottom connected to the GPU. The entire system faces severe thermal and power consumption challenges… There are about 2,000 process steps in the fab. From the start of wafer production to final product delivery to customers, the whole cycle takes about five months.”

Goodbye ‘Spot Buying and Selling’: Long-term Contracts and Deep Customization Reshape Business Models

Extremely tight production capacity is forcing a transformation in the storage industry's business model; the era of spot market trading and commodity JEDEC-standard chips is evolving.

“We now sign multi-year contracts, committing to supply our customers, and customers commit to giving us demand forecasts. This is very different from the old one-year agreements—customers used to buy as needed and we shipped if we had available stock.” Sadana stated that this so-called “strategic customer agreement” model yields a very high return on investment, providing foundational security for years of fab capex.

The even more profound change is on the R&D front. In order to differentiate on metrics such as power consumption and model run speed in AI systems, customers are integrating memory design into their 5- to 7-year product roadmaps. Sadana noted, “This model transforms cooperation into something more like an ASIC partnership—the customer has much more autonomy in design, and this relationship is long-term.”

AI Moves to the Edge, Humanoid Robots will Become the Largest Incremental Market

The market is mainly focused on data center AI demand, but Micron believes this is just the beginning. As AI permeates smartphones, PCs (such as with unified memory architecture in Mac mini), and autonomous vehicles, storage demand is set to be everywhere.

For future imagination, Sadana set his sights on the robotics track, seeing it as the next wave of explosive demand following data centers.

“Think about humanoid robots, which will become one of the largest product markets in history.” Sadana, looking ahead to the 2030s market, highlighted the massive memory demand on the edge: “What’s exciting about robotics is that each device must store massive amounts of data—each humanoid robot will have hundreds of GB of DRAM memory and several TB of NAND flash. As a result, they will significantly drive up DRAM and NAND demand… Robots must be able to operate autonomously, not sending data to servers for every action.”

The following is the full interview:

Sumit Sadana 00:00

Today, the performance of AI systems is determined first by the performance of the memory subsystem and memory capacity. These two factors truly define the performance of AI subsystems. Welcome back to the 2026 Summit.

Patrick Moorhead 00:22

The era of AI is upon us. The importance and strategic nature of memory is remarkable. I always thought memory was important, but in light of the AI era, the capabilities it gives us are simply amazing—honestly, if there isn’t enough or the right kind of memory, none of the amazing things we’re seeing around us could happen.

I’m very pleased to announce Micron's presence at this summit. Great to see you, great to have you back. Thanks so much for being here. This is just fantastic. I’ve been in this industry for 35 years—from a memory consumer, to a memory partner now to analyzing the memory market and all adjacent fields. The progress we’ve made is amazing. When we talked last year, I actually categorized memory as a strategic technology—yes, I’ll take credit for that. To be fair, you deserve more praise because you act, I just comment. But nowadays, this view is even more important, and more people are talking about it. In this age of AI, what has changed over the past year to make the industry put so much more emphasis on memory?

Sumit Sadana 01:44

Why has AI actually accelerated this shift? That’s a great question. A year has passed but so much has changed in the world in that year that it feels much longer. From a customer perspective, the transformation has been immense. Today, we are facing a severe shortage of memory across all sectors. Even though we are doing our absolute best to add supply, we still don’t know when supply will be able to meet demand, because demand just keeps growing. The signals from customers in all areas keep increasing year after year.

There are several reasons why customers’ views of memory have changed. One is the huge gap between supply and demand; another is that, considering the demands of AI, system performance requirements are also rising. When you consider what really drives the best system performance, the processor is no longer the only factor. Today, the memory performance, bandwidth between the processor and the memory, and the capacity of memory itself all are crucial. Because AI models need to be stored in memory, enormous amounts of data travel between memories, making bandwidth between the processor and memory a critical performance bottleneck. Therefore, customers are increasingly aware that they need to replan their product roadmaps with different memory strategies.

Sumit Sadana 03:44

Now, what matters is how to design memory for a competitive edge, how to make every customer system stand out. If memory were still the off-the-shelf commodity memory everyone used years ago, JEDEC-standard, it would be difficult to achieve differentiation. Therefore, we now work with customers to incorporate memory design into their multi-year product roadmaps. They’re now focused on how to differentiate, how to change the competitive landscape. This means a different way of thinking about memory, working with companies like Micron to design a new kind of functionality that can be leveraged and applied.

Sumit Sadana 04:31

For example, our collaboration with Nvidia on low-power DRAM (LPDRAM) for data centers is a good example. We were the first to introduce this technology to data centers, have long been the exclusive supplier in this area, and now the technology is being adopted by other companies as well. More customers are realizing the benefits of low-power DRAM—it brings higher density, smaller form factors, greater performance, and significantly reduced power consumption, all of which are crucial for data centers. This is just one of many examples.

Patrick Moorhead 05:07

Looking ahead five years, these technologies are beginning to flourish. When I try to explain why people are so interested in memory, I recall a class I took in college—in the late 1980s—where the basic principle was: if you could keep operations within memory, you’d get better performance. Sure, today things are a bit different. If you have more memory, you get better results, and the closer the memory is to what you’re working on, the better, because as soon as you leave memory, things slow down. In the AI age, this is a huge challenge.

It’s interesting. I’ve gone through nine memory cycles in my career. I worked at an OEM, at a chip company, and now at an analyst firm for 15 years. Memory has always seemed to be the most cyclical—of course, there are other industries with similar cycles, but it’s often viewed as a commodity. You already touched on this in your first answer but…

Sumit Sadana 06:17

Why is AI fundamentally changing this perception? That’s a great question. If you look back at the early 1990s, there were over 20 companies doing DRAM, and the DRAM industry went through tons of consolidation. At the same time, there have only been a few processor companies, but now, processor companies surpass 20. If you sit down and count every processor design firm, the number comes quickly. But there are only a few DRAM companies left.

So, when you think about AI design, and AI subsystem design, it’s not just about data centers—although of course, in data centers you distinguish between training workloads and inference workloads, and even inference is subdividing into many types. A small portion can use SRAM, but a big part still relies heavily on DRAM. We’re designing for as much DRAM capacity as possible in these systems. This trend extends to autonomous vehicles, industrial systems, and look at smartphones and PCs—the unified memory architecture in Mac mini has sparked an "Open Claw" movement. And all this is, in the end, tied to AI.

Sumit Sadana 07:50

Today, system performance is determined first by the performance of the memory subsystem and the capacity of memory. These two factors really define the performance of AI subsystems. So, when memory becomes the central design point for the system, you have to think about memory differently. The accompanying core issue is: how do you quickly deliver differentiated products? I think the memory business going forward will be fundamentally different from the past, and this ties closely to this idea…

Sumit Sadana 08:36

The industry as a whole has finally reached a point where new capacity has to be built, which is surprising because AI-driven demand growth is so strong. Technical transitions alone cannot meet this demand—although in previous years, bit growth from tech transitions was enough. The industry now needs to sharply increase wafer output, which means new cleanrooms must be built. When your existing fab’s cleanroom is full—which most companies face now—you have to simultaneously expand new sites, and that’s a lengthy process.

It’s a very time-consuming process and constrained by many practical factors, like the difference in build speed around the world and the efficiency of regulatory approvals. How do you build all the required infrastructure—power, water, treatment plants—on a greenfield site? The supply of chemicals and everything needed for a fab—all takes considerable time, and this is just the reality for the whole industry.

Sumit Sadana 10:01

So, despite our best efforts, and what we see from others in the industry, we’re still not able to meet demand. Supply will take a long time to catch up, and that’s another important factor influencing customer behavior.

Sumit Sadana 10:27 This is what we’ve been talking about with strategic customer agreements. Those change our business model and operations fundamentally. Now we sign multi-year deals, commit to supply, and customers commit demand forecasts. This is very different from the old one-year contracts—where customers bought what they wanted, when they wanted, and we supplied if we had it. Now it’s a committed supply agreement with a very high ROI, which means we can secure years of capex that’s required.

Of course, this doesn’t even include agent AI, the next stage of AI’s evolution. After that, we’ll head into physical AI, which will be the next huge wave of demand. If you look at these demand waves, they stack on top of and reinforce each other, not replace one another. This represents a massive challenge of how to move so much physical supplying capability online.

Patrick Moorhead 11:39

During this time, to meet all this demand, I think your long-term agreement program perfectly shows the strategic nature of memory is rising. Of course, it’s easy to say “we need these contracts to lock in future capacity”—which is important—but it’s also about collaborative planning.

From a technical perspective, I’m truly impressed, and glad you mentioned edge AI. What’s most interesting to me, whether it’s Open Claw or new client compute designs—for instance, memory tightly coupled with CPU and GPU to boost bandwidth—the architecture in client computers has not changed much for a long time. Next, I want to talk about hyperscale data centers, which we touched on a bit before—a lot of the dynamic is here, especially in capex, and application and model innovation mostly concentrates in this field. But I’m curious about autonomous vehicles, PCs, robots, the intelligent edge, etc…

Sumit Sadana 13:00

Is demand for memory changing these scenarios? Yes, great question. Of course, it all starts with data centers. But when we think about the future of AI, it won’t be confined to the data center. This intelligence wave will move to the edge, and finally, intelligence will be everywhere, embedded in all devices. So it covers consumer electronics, autos, and even the second industrial revolution—where all this intelligence will be embedded globally across enterprises.

Sumit Sadana 13:43

Your personal devices—whether it's your computer at work or home, your smartphone, even some new devices that certain companies are considering—these might look nothing like a smartphone or a PC, adopting completely different designs, because when you build a native AI device, you can imagine totally new approaches versus traditional computers and smartphones. This again returns to how to cut power consumption as much as possible, for instance for devices running on batteries. How to maximize performance, how to run models that offer value to consumers yet are small enough to run on the device without calling back to the cloud—that’s the new frontier for AI.

This exploration will continue. Models that are small enough to run on computers or smartphones will keep getting better over time. As this happens, the benefits for data privacy and confidentiality are huge. Consumers care a lot, and companies that give consumers confidence that everything they interact with on a device is not sent to the cloud will create entirely new applications and huge value.

Sumit Sadana 15:30

Consider autonomous driving, which is advancing quickly thanks to AI progress; next-generation frontier technology will be robotics. Think about humanoid robots, which will become one of the largest product markets in history. Of course, it will take some time, but we’re getting closer to that singularity—eventually, you’ll talk to humanoid robots just like with people, and their cognition will be indistinguishable. Over time, their physical abilities will get ever more advanced, which is revolutionary.

Initially, they’ll be used in environments like factory automation, doing highly specific tasks with limited flexibility. The hardest environment is the home, as it's highly unstructured. But I believe that into the 2030s, robotics will drive massive growth. What's most exciting about robotics is the huge storage required—each humanoid robot will have hundreds of GB of DRAM and several TB of NAND flash. They’ll massively drive new DRAM and NAND demand. Many people fail to grasp this: robots must run autonomously, they can’t send data to the server every time. It’s absolutely true.

Patrick Moorhead 17:25

As you were explaining, I suddenly realized Micron is enabling a leap to gigawatt scale. Just thinking about inventing this technology is mind-blowing. I think people should remember this when they think of Micron.

As we enter the age of AI, strategic vision matters more than ever—be it leveraging cutting-edge devices yet to be invented, or soon-to-be-launched gigawatt-level data centers that will be AI and foundation model-ready.

I’ve heard about this joint design or invention concept, and have spoken with some of your partners about ideas they are trying. I touched on architecture changes earlier—these used to be standardized through JEDEC or similar, but what we’re talking about now is truly deep joint design. Can you talk about how your relationship with partners is changing?

Sumit Sadana 18:45

Of course. Our partners want to know how to gain an advantage in their own markets, and beat their competition, too. The memory-processor subsystem is so crucial to the AI subsystem, because the interplay between memory and processor essentially determines many key variables for AI systems—it sets system power consumption, which types of models can run, their sizes and their speeds.

When you consider all the progress at various sizes and types in the LLM field, you see that to achieve differentiation in hardware design cycles—both for memory and processors—all segments must collaborate closely. It's not a plug-and-play, JEDEC-compliant device that you just certify after the fact. While some segments fit that, increasingly customers need to think: how do we differentiate? I have to create special features off-the-shelf products don’t provide. So, we engage with memory companies on roadmaps—long discussions spanning five to seven years of R&D. We’re involved in the roadmaps of many customers, hearing fascinating proposals, some feasible now, others will need more time and breakthroughs to realize.

Sumit Sadana 20:28

We’re very excited about this diversity of ideas. As you said, these cover such a range of systems—some ultra-low power using only batteries, others for gigawatt-level data centers. We already see innovation in several segments and among leading customers, and will keep witnessing that and working hard to support it.

In some thrilling projects, we’re constantly challenging ourselves, delivering the innovative designs customers dream of. Customers can’t work this closely with many companies, so there are usually only one or two such partners; others take years to catch up. The advantage of this model is that you form very tight partnerships with a few companies, and frequently become the sole source supplier, or one of two for some time. This model turns the relationship into something more like the ASIC model—the customer enjoys far more autonomy and the tie is long-term. That’s why it’s a special, ASIC-style partnership, not the standard memory buyer-seller relationship of a few years ago. Another huge change I see is that this model is now scalable to large volumes.

Patrick Moorhead 22:27

Exactly. This means the volume is so large, and honestly, the capital so significant that both sides invest; and when you start the project, differentiation is key. Everyone is seeking it, whether for performance efficiency or total cost of ownership (TCO), it’s about standing out.

I want to look at the future from an investment perspective. You mentioned capacity investments. It’s interesting, some people expect memory fabs to be built as quickly as writing code. But listen, I once worked at a chip company with a fab, and even today, it’s an extremely difficult, capital-intensive undertaking, where you’ll see any real results only in three or four years. Can you talk about what investments you’re making to truly cement Micron’s place in next-generation tech—be it AI or some other technology or branch of AI? These investments take a long time, and you made that very clear.

Sumit Sadana 23:57

You’re absolutely right. By the way, if anyone knows how to build a DRAM fab fast, we’d love to know, because it would make our lives much easier. Our customers definitely want products as quickly as possible, and we’re doing everything we can.

We’re pushing as hard as we can, on all our investments. Just our capex plan: fiscal 2025, a little over $13 billion; fiscal 2026, it doubles; fiscal 2027, over $45 billion expected. It keeps climbing. We recently announced in the US we’re raising investment from $200 billion to $250 billion and accelerating the timeline—with $50.2 billion planned by the end of next year.

These investments cover all areas. We’re investing in Idaho Fab 1 (ID1), bringing the first production lines online; middle of next year, ID2 follows, and initial wafer lines are expected in late 2028. The Taiwan fab we acquired will also start wafer output in 2027, and we’re expanding that fab. We’re also expanding in Japan and Singapore—so globally, we have about 20 projects of varying sizes underway, driving both front-end and back-end capacity.

Sumit Sadana 26:06

But it all takes a long time. Every new fabs takes this long. In New York, we’ve planned a cluster of four fabs, and the first is expected to come online in 2030. We broke ground in January, and already poured the concrete milestone ahead of schedule.

Sumit Sadana 26:22

There are many projects all over the world, but they all need permits and full supporting infrastructure. There’s a critical shortage of construction talent—think about all the simultaneous building projects in the US and around the world.

Data centers are being built everywhere, they need power, so power plants are also being built; those centers need semiconductors, so both front-end and back-end fabs are being built. There is a dire shortage of skilled technicians—the number of workers is far from enough for these complex builds.

Sumit Sadana 27:15

So, we are investing in communities and nurturing talent to ensure talent reserves—not just for our own projects, but for the whole ecosystem in the long run. We’re making a range of investments, from STEM development to community colleges, to partnering with the towns and communities we operate in. It’s a long-term endeavor for us.

We expect the facilities to keep expanding over time, as these fabs cluster—you can’t just build one fab. You need scale, you need to break through the cost curve as soon as possible, so you have to keep building; which means over the next decade and beyond, expansion will continue. These are long-term projects for us. We believe substantial new supply will only start to be released around 2028, just the initial stage; the real growth momentum comes in the following years. So, it will take some time for the whole sector to reach a new equilibrium, though we can’t predict exactly when that will happen.

Patrick Moorhead 28:54

That’s a great example and the investment data is concrete. However, one thing I think some people misunderstand is the complexity of memory. We all know logic chips are hard. Let's talk about memory’s complexity—memory is among the most complex semiconductor technologies today.

Sumit Sadana 29:19

We always say logic is cutting-edge, but so is memory. It uses EUV lithography, those huge machines. In the fab, there are about 2,000 process steps. From the start of wafer fabrication to final product shipping to the customer takes about five months. It’s incredible—many just can’t imagine it. The wafer production cycle is at least three and a half to close to four months, and then assembly, packaging, and testing another month or even a month and a half sometimes.

Sumit Sadana 30:04

For more complex products, like HBM, it's even more obvious. People call it a miracle for a reason; it’s amazing it even works. Think about stacking 12 DRAM layers, with a base die at the bottom connected to a GPU. The whole system faces severe thermal and power challenges—how to operate at such high bandwidth, move data back and forth between processors and chips while cooling efficiently? Even the packaging process is as complex as a cleanroom; it’s totally different from the flip-chip packaging in logic used for years. From any angle—packaging complexity or front-end process advances like EUV lithography—every DRAM node is getting harder to scale.

Sumit Sadana 31:18

It’s just as true for NAND. We stack it layer by layer—200, 300, 400—this complexity is astronomical. And these NAND products can now store data and do their job—a single SSD can now store 245 TB. To fit such massive capacity in such a tiny package, it’s an everyday miracle performed by engineers and manufacturing teams. The complexity is unbelievable, and over the coming years we’ll keep pushing, scaling geometries tighter and tighter and making this technology widespread—and that will be another miracle.

Patrick Moorhead 32:23

It all takes vast time, effort and technical innovation. The complexity amazes me, and that you can do this across such a wide power envelope—that’s astonishing. I don't know how many firms in the world can do that.

It’s funny—people always say “I’ll never be satisfied”—I think that's because they always want more, more amazing things. Then, people start talking about price and everything. But I think this is an investment—it’s only natural that Micron and others in the memory industry keep investing, because if not, the innovation cycle stops. I look forward to seeing what Micron achieves in the future, not only in fabs but with technology itself. I believe robotics will offer 10x growth; it could be a huge growth market—that’s an extra lever our capacity models haven’t even accounted for yet. Of course, we need more evidence to know when it will really take off.

Thank you so much for this talk. Hard to believe it’s only been a year since our last chat—and look at how much has changed in just a year.