How artificial intelligence is transforming memory from a supporting component of computing into one of its most critical technologies — and placing Micron at the center of a profound industrial transformation

For most of the history of computing, memory lived in the shadow of the processor.

The processor was the star. Faster CPUs defined successive generations of personal computers. GPUs later transformed graphics and, eventually, artificial intelligence. Memory was essential, of course, but largely invisible: a component whose job was to store information and deliver it when the processor needed it.

Artificial intelligence is changing that relationship.

The extraordinary computational requirements of modern AI systems are creating a problem that raw processing power alone cannot solve. The world’s most advanced processors can perform enormous numbers of calculations, but those calculations are useful only if equally enormous quantities of data can reach the processors quickly enough.

This is the emerging memory wall.

And it is one of the reasons why Micron Technology, a company founded nearly half a century ago in Boise, Idaho, suddenly finds itself operating at the center of one of the most consequential technological transformations of our time.

Micron does not build the GPUs that have become synonymous with artificial intelligence. It builds something those GPUs increasingly cannot work effectively without: the advanced memory that surrounds them.

The distinction matters.

As artificial intelligence moves from experimental models toward infrastructure deployed across data centers, businesses, automobiles, robotics and personal devices, memory is becoming not simply larger in quantity but more sophisticated, faster, more energy-efficient and more tightly integrated with computation.

The AI revolution, in other words, is becoming a memory revolution as well.

---

COMPANY PROFILE

Company: Micron Technology, Inc.
Founded: 1978
Headquarters: Boise, Idaho, United States
Industry: Semiconductors
Core business: Memory and storage
Principal technologies: DRAM, NAND, NOR, High Bandwidth Memory (HBM), SSDs
Chairman, President & CEO: Sanjay Mehrotra
Stock exchange: Nasdaq
Ticker: MU
Global presence: Manufacturing, research and commercial operations across multiple continents
Website: micron.com

---

FROM A BOISE BASEMENT TO THE AGE OF AI

Micron’s origins could hardly be further removed from today’s trillion-dollar artificial-intelligence infrastructure race.

The company began in 1978 with four people working in the basement of a dental office in Boise. Its first contract involved designing a 64K memory chip. By 1980 it was building its first fabrication plant; in 1981 it shipped its first 64K DRAM product.

Over the following decades, Micron evolved alongside the computer industry itself.

Personal computers required memory. Servers required more. Smartphones dramatically expanded both memory and storage requirements. Cloud computing created vast data centers. Automobiles became computers on wheels.

Through all of these transitions, memory remained indispensable.

Micron consequently developed into one of the world’s major memory manufacturers, producing the technologies that allow digital systems to retain, access and move information. Today its portfolio encompasses DRAM, NAND and NOR memory as well as sophisticated storage products.

But artificial intelligence represents something different from previous computing cycles.

It does not simply require more memory.

It increasingly requires better memory.

And that distinction may prove fundamental to the industry’s future.

THE MEMORY WALL

Consider what an advanced AI accelerator actually does.

Modern processors can execute calculations at extraordinary speed. But an AI model may contain hundreds of billions of parameters, and increasingly complex systems must continuously retrieve, manipulate and update immense quantities of data.

A processor waiting for data is wasted computational power.

The challenge therefore becomes not only how quickly a processor can calculate, but how quickly information can be delivered to it.

This imbalance between rapidly increasing computational capability and the ability to move data efficiently is generally described as the memory wall.

Artificial intelligence has made the problem considerably more important.

Large-language models, multimodal AI, reasoning systems and increasingly autonomous agents require enormous amounts of information to move repeatedly between memory and processors. As AI systems evolve from answering individual questions toward performing extended sequences of tasks, the amount of data that must remain rapidly accessible can increase dramatically.

Memory bandwidth therefore becomes part of computational performance itself.

The fastest engine in the world is of limited use if fuel cannot reach it quickly enough.

This is where High Bandwidth Memory enters the story.

HBM: MEMORY BECOMES A HIGH-PERFORMANCE TECHNOLOGY

Traditional memory modules sit some distance from the processor and communicate through relatively constrained pathways.

High Bandwidth Memory takes a radically different approach.

Multiple memory dies are vertically stacked and connected using extremely dense electrical pathways. The resulting package can sit close to an AI accelerator and move enormous quantities of information while consuming less energy per transferred bit.

HBM is extraordinarily difficult to manufacture compared with conventional memory.

That complexity is precisely what makes it valuable.

Micron’s latest HBM4 technology illustrates how quickly the field is advancing. Its 12-high HBM4 product uses a 2,048-pin interface and operates above 11 Gb/s per pin, delivering more than 2.8 terabytes per second of bandwidth from a single stack. According to Micron, this represents more than twice the bandwidth of its previous HBM3E generation while improving power efficiency by more than 20 percent.

In early 2026, Micron began volume shipments of its 36GB HBM4 product designed for NVIDIA’s Vera Rubin generation of AI systems. It has also produced samples of a 48GB, 16-high HBM4 stack, increasing capacity per HBM placement by one third.

These numbers are technical, but their significance is straightforward.

AI systems need enormous amounts of data immediately available to enormously powerful processors. Every new generation of accelerator increases the pressure on the memory subsystem.

The processor and the memory are therefore becoming progressively more interdependent.

Memory is no longer merely where computation keeps its data.

It is increasingly part of what determines how fast computation can happen.

AI IS CHANGING THE ECONOMICS OF MEMORY

For decades, memory manufacturing was notorious for its cyclicality.

The mechanism was relatively simple.

Demand increased. Memory prices rose. Manufacturers expanded capacity. New capacity arrived. Supply exceeded demand. Prices collapsed. Manufacturers reduced investment. Eventually demand caught up, beginning another cycle.

It was one of the semiconductor industry’s most persistent patterns.

Artificial intelligence does not necessarily abolish that cycle. Semiconductor manufacturing remains capital-intensive, demand can fluctuate, and no technological boom expands forever.

But AI may be altering some of the forces behind it.

First, the quantity of memory required by advanced computing systems is increasing.

Second, the performance requirements of that memory are becoming much more demanding.

Third, HBM consumes substantial manufacturing resources and involves additional processing and packaging complexity.

And fourth, creating new leading-edge semiconductor capacity takes years rather than months.

This means that an increase in AI demand cannot simply be answered by opening a factory and producing more advanced memory the following quarter.

The physical world imposes constraints on the digital one.

Industry research increasingly reflects this tension. TrendForce described the rise of agentic AI in 2026 as producing a structural expansion in memory demand, driven in part by the transition from individual inference requests toward continuous iterative AI processes.

Its May 2026 forecast raised the estimated value of the global memory market to approximately $889 billion for 2026 and more than $1.28 trillion for 2027.

Forecasts of that magnitude should always be treated cautiously. Technology markets can change rapidly, capacity eventually responds to shortages, and extreme pricing conditions rarely persist indefinitely.

But the direction of travel is difficult to ignore.

Memory has become one of the central physical resources required to scale artificial intelligence.

THE SHIFT FROM TRAINING TO INFERENCE

There is another reason the development deserves attention.

The first stage of generative AI was dominated by training: constructing enormous models using enormous computational clusters.

The next stage increasingly involves inference — actually using those models, potentially billions of times, across businesses and society.

And inference may eventually represent a far larger computational activity than training.

An AI model might be trained periodically. Once deployed, however, it can answer questions, analyze documents, write software, process images, operate machines or coordinate autonomous agents continuously.

The rise of reasoning models intensifies this phenomenon because the system may perform numerous computational steps before generating an answer.

Agentic AI potentially pushes it further still. Instead of receiving a question and producing an answer, an autonomous system may execute dozens or hundreds of interconnected actions while retaining and processing contextual information throughout the process.

This is why memory requirements cannot be understood simply by counting the number of AI models being trained.

The larger question is how frequently intelligence will eventually be used.

If AI becomes embedded throughout economic activity, the infrastructure supporting inference may become enormous.

Memory sits inside that infrastructure.

THE NUMBERS ARE ALREADY CHANGING

The transformation is not merely theoretical.

Micron’s fiscal third quarter of 2026, ending May 28, produced results on a scale dramatically different from the previous year.

Revenue reached $41.46 billion, compared with $23.86 billion in the preceding quarter and $9.30 billion in the equivalent quarter one year earlier.

GAAP gross margin reached 84.6 percent.

Operating cash flow was $25.39 billion.

The company’s Cloud Memory Business Unit generated $13.77 billion in quarterly revenue, while its Core Data Center Business Unit generated another $11.52 billion.

For the following fiscal quarter, Micron guided to approximately $50 billion in revenue, with an expected GAAP gross margin of approximately 86 percent.

Numbers of this magnitude require context rather than extrapolation.

Memory pricing has been exceptionally favorable, supply conditions are tight, and semiconductor economics can change significantly when additional capacity eventually enters the market.

Nevertheless, the scale and speed of the change provide tangible evidence of the forces reshaping the industry.

What was once largely a component business has become strategically important to the infrastructure race surrounding artificial intelligence.

MORE THAN HBM

It would be a mistake, however, to reduce Micron’s AI story entirely to HBM.

AI infrastructure requires a hierarchy of memory and storage technologies.

HBM supplies extraordinarily fast access close to accelerators. Server systems require large quantities of conventional DRAM. Massive datasets need enterprise storage. AI inference at the edge creates additional memory requirements in computers, smartphones, automobiles and industrial equipment.

Micron participates across much of this hierarchy.

During fiscal Q3 2026, the company reported that its PCIe Gen6 high-performance SSD based on its G9 NAND technology had entered high-volume production. It also began shipping a 245-terabyte QLC SSD, an illustration of the extraordinary storage density now being developed for data-intensive infrastructure.

Micron is simultaneously advancing 256GB DDR5 server modules, SOCAMM2 memory, next-generation mobile memory and automotive products.

This diversification matters because AI is unlikely to remain confined to hyperscale data centers.

The longer-term trajectory points toward intelligence being distributed throughout the technological environment.

Servers will need memory.

AI PCs will need memory.

Smartphones running increasingly sophisticated local models will need memory.

Robots will need memory.

Autonomous vehicles will need memory.

Industrial systems will need memory.

The proliferation of artificial intelligence therefore has the potential to increase both the quantity and technological sophistication of memory across numerous categories of computing.

A RARE INDUSTRIAL STRUCTURE

The global advanced-memory industry is unusually concentrated.

At the leading edge of DRAM and HBM, the competitive landscape revolves primarily around three companies: South Korea’s Samsung Electronics and SK hynix, and America’s Micron Technology.

That concentration reflects the extraordinary difficulty of entering the industry.

Leading-edge memory manufacturing requires enormous fabrication facilities, advanced semiconductor equipment, sophisticated process technology, extensive intellectual property, packaging expertise and years of manufacturing experience.

Capital alone is insufficient.

Micron passed 60,000 lifetime patents in 2025, an indication of the accumulated technological knowledge required to compete at this level.

This does not mean Micron enjoys an uncontested position.

Quite the opposite.

SK hynix has established formidable leadership in HBM. Samsung possesses immense semiconductor resources and has been aggressively advancing its HBM4 technology. TrendForce’s 2026 research indicates intense competition among all three suppliers as next-generation AI platforms come to market.

That competition is one of the most important elements to watch.

HBM leadership is not permanent. Each generation creates another technological race involving bandwidth, thermal characteristics, energy efficiency, capacity, manufacturing yield and customer qualification.

Micron must continue executing.

But the fact that only a handful of companies can realistically participate at the leading edge is itself significant.

AI may be creating enormous demand for a technology whose supply cannot easily be replicated.

BUILDING CAPACITY FOR THE NEXT DECADE

Micron’s response to this environment is physical and enormous.

In July 2026 the company increased its planned investment in U.S. fabrication and technology to more than $250 billion through 2035.

Its long-term objective is to manufacture approximately 40 percent of its DRAM in the United States.

In Boise, Micron is developing two leading-edge high-volume fabrication plants alongside its existing research operations. Output from the first new Idaho fab is scheduled to begin in 2027.

In Clay, New York, the company has begun developing what could ultimately become a four-fab manufacturing complex. Potential investment there could reach approximately $100 billion over more than two decades.

Micron is also expanding and modernizing manufacturing in Virginia and investing in the broader semiconductor supply chain.

The scale is difficult to visualize.

A modern semiconductor fab is not simply a factory. It is one of the most complex manufacturing environments humans have ever constructed: immense cleanrooms, extreme control of contamination, sophisticated chemical and material systems and some of the most advanced industrial equipment on Earth.

These facilities also take considerable time to build and qualify.

That creates an interesting asymmetry.

Software demand can explode almost instantaneously.

Physical semiconductor capacity cannot.

AI may live in the cloud, but the cloud ultimately lives inside buildings filled with machines.

THE GEOPOLITICS OF MEMORY

There is also a geopolitical dimension.

Advanced semiconductors have become strategic infrastructure.

Governments increasingly view domestic semiconductor production not merely as an industrial policy objective but as an issue involving national security, technological sovereignty and supply-chain resilience.

Micron occupies an unusual position in this landscape: it is the principal U.S.-headquartered producer competing at the leading edge of global memory manufacturing.

The United States has consequently supported domestic semiconductor expansion through the CHIPS Act and associated incentives. Micron has secured up to approximately $6.4 billion in direct CHIPS funding supporting projects in Idaho and New York as well as expansion in Virginia.

The objective extends beyond economics.

A world increasingly dependent on artificial intelligence will also become increasingly dependent on the physical components required to create it.

Memory is one of those components.

POWER BECOMES PART OF THE EQUATION

The AI infrastructure race has another constraint: electricity.

High-performance computing systems consume enormous amounts of energy, and memory itself contributes meaningfully to system power consumption.

This makes energy efficiency more than an environmental characteristic.

It becomes an engineering requirement.

A memory architecture capable of moving substantially more data while consuming less energy per bit can improve the economics and practical feasibility of an entire AI system.

This helps explain why Micron emphasizes power efficiency alongside raw bandwidth in its HBM development.

As AI clusters grow from thousands to tens of thousands and eventually potentially hundreds of thousands of accelerators, small improvements multiplied across entire facilities become significant.

The future of memory therefore involves at least three simultaneous races:

capacity, bandwidth and energy efficiency.

Winning on only one may not be enough.

WHAT COULD CHALLENGE THE STORY?

Technological transformations rarely proceed in straight lines.

Several forces could make the future less favorable than today’s extraordinary conditions suggest.

The first is capacity.

High prices and exceptional profitability encourage manufacturers to invest. Micron, Samsung and SK hynix are all expanding advanced-memory capabilities. Eventually new supply will enter the market.

The second is competition.

Micron competes against two exceptionally capable semiconductor manufacturers. SK hynix has been particularly strong in HBM, while Samsung has the resources and technological breadth to contest leadership aggressively.

The third is AI infrastructure spending itself.

Today’s demand assumes continued enormous investment by hyperscalers and other technology companies. If the economic returns from AI infrastructure disappoint, spending could slow.

The fourth is technological change.

Computing architectures evolve. Engineers constantly seek ways to reduce memory requirements, improve caching, compress models and redesign systems. A breakthrough that materially reduces memory intensity could alter long-term demand assumptions.

The fifth is the industry’s own history.

Memory has repeatedly taught observers that shortages eventually become surpluses. It would be premature to declare cyclicality dead merely because current conditions are exceptional.

And finally there is execution.

Building leading-edge fabs, improving manufacturing yields, developing successive generations of HBM and qualifying them with demanding customers are extraordinarily difficult tasks. Massive capital commitments create massive operational responsibilities.

These are not peripheral considerations.

They are central to understanding what happens next.

THE MORE INTERESTING QUESTION

The future of Micron therefore depends on something larger than whether today’s memory shortage persists.

It almost certainly will not persist forever in exactly its present form.

The more interesting question is whether the nature of memory itself has changed.

If AI requires progressively larger quantities of increasingly sophisticated memory; if bandwidth becomes a persistent constraint on computational performance; if inference spreads intelligence through data centers, computers, automobiles, robots and industrial systems; and if producing the required memory remains technologically and financially difficult, then memory occupies a different strategic position from the one it held during earlier computing eras.

That possibility is what makes Micron particularly interesting today.

Not because memory has ceased to be cyclical.

Not because competition has disappeared.

And certainly not because technological forecasts are guaranteed.

But because artificial intelligence is forcing the computing industry to reconsider where performance actually comes from.

For decades, the answer was dominated by processors.

The AI era is demonstrating that computation is a system.

Processors must calculate.

Networks must connect.

Electricity must power.

Cooling systems must remove heat.

And memory must deliver enormous quantities of information precisely when the processors need it.

Failure at any one of these points becomes a constraint on everything else.

THE QUIET TECHNOLOGY BEHIND INTELLIGENCE

There is something almost paradoxical about Micron’s position.

Artificial intelligence is among the most visible technologies ever created. People interact directly with AI systems, watch them generate images, write software, translate languages and reason through increasingly complicated problems.

Memory remains almost invisible.

Yet behind every remarkable AI output lies an immense physical choreography of processors, networks, storage and memory moving information at extraordinary speed.

Micron operates inside that invisible layer.

The company that began with four people designing a 64K memory chip in a Boise basement is now producing memory capable of transferring terabytes of information every second and committing hundreds of billions of dollars to the infrastructure required for the next generation.

That journey says something important about technological revolutions.

The companies that define them are not always those producing the most visible products.

Sometimes they build the indispensable technology underneath.

And as artificial intelligence moves from an extraordinary new capability toward a fundamental layer of the global economy, memory may prove to be one of the places where the physical limits — and possibilities — of that revolution are ultimately determined.

Micron has spent nearly half a century learning how to build it.

Foto di The Six