美光CEO访谈：「存储」是AI被忽视的瓶颈，供给紧张或延续至2026年后

Original author: Li Jia

Original source: Wall Street Sights

"The AI race is not just a computing power race, but also a storage race." This is the assessment given by Micron Technology CEO Sanjay Mehrotra.

On the June 5th podcast program "A Bit Personal," Sanjay gave a rare in-depth interview, recorded at his home. Beyond the usual industry insights, this personal conversation led him to proactively discuss his upbringing, family influences, and career choices.

One of Sanjay's core judgments is that AI is still in its very early stages.

In his view, as large models, Agent AI, and reasoning applications continue to evolve, AI will need not only greater computing power but also enhanced "memory capabilities."

Longer context windows, larger model sizes, and ever-increasing token consumption are all driving continuous growth in storage demand.

The essence of AI is data, and data cannot exist without storage. Therefore, storage will become one of the most critical infrastructures in the process of improving AI capabilities.

At the same time, the supply side is not fully prepared. Sanjay pointed out that the current challenge facing the storage industry is not a short-term supply-demand mismatch but a structural supply constraint. Advanced storage products consume more silicon wafers, and building new fabs typically takes three to four years, with subsequent capacity ramp-up also being a lengthy process.

More importantly, as technology nodes advance, the increase in storage capacity output per wafer is decreasing. He predicts that the industry's supply tightness is expected to continue well beyond 2026.

When explaining why storage technology has been underestimated for so long, Sanjay stated bluntly: "People often misunderstand memory. They don't know how difficult it is to manufacture memory." From physics and chemistry to materials science, and then to ensuring the correct behavior of every single bit among trillions in mass production, the underlying technological difficulty is extremely high. He believes that the AI race is also a storage race, a fact that has been overlooked by the market for a long time.

From a longer-term perspective, Sanjay believes the fundamental logic for the success of companies and individuals hasn't changed. Whether it's driving a $200 billion investment plan or leading Micron through the storage industry cycles, the key words he repeatedly emphasizes are resilience, discipline, and long-termism. Investments must be based on data and fundamentals. Leaders need to see both industry trends and deeply understand technical details.

As he learned from his father, success requires both the resilience to persevere and the ability to seize opportunities at critical moments.

Key insights from the interview with Micron Technology CEO Sanjay Mehrotra are as follows:

Storage is the underestimated bottleneck for AI; its manufacturing difficulty and strategic value far exceed current market perception. AI is extending from a "computing power race" to a "storage race." Increasing model sizes, longer context windows, and surging token consumption mean AI depends not only on stronger computing power but even more on enhanced "memory capabilities." Without sufficient storage capacity and bandwidth, no amount of computing power can be unleashed.

The structural constraints on the supply side determine that the storage shortage is not a short-term fluctuation but a long-term condition. Advanced storage products consume more wafers, while building new fabs takes three to four years, and the capacity ramp-up is equally slow. Meanwhile, technological node advancements lead to diminishing output growth per wafer. Given this supply-demand mismatch, supply tightness will last at least until after 2026.

People always underestimate the difficulty of manufacturing memory, but this is precisely the industry's deepest moat. From physics, chemistry, and materials science to design and mass production ensuring trillions of bits function without error, the engineering complexity is immense. The manufacturing difficulty of memory chips is no less than any other semiconductor field, and in many aspects, it is even harder.

Success comes from resilience, discipline, and long-termism, not from short-term trend-chasing. Whether it's driving a $200 billion investment or navigating the cyclical fluctuations of the storage industry, leaders need both to see industry trends and delve into technical details. Just as his father persisted despite having his visa denied three times, success requires both the tenacity to endure and the ability to seize opportunities at crucial moments.

Storage is Becoming the Backbone of AI

Discussing the current historical position of the storage industry, Sanjay stated: "I've been in this industry for over 45 years. This is the most exciting time for the entire industry that I have ever experienced."

He further elaborated on the strategic significance of storage for AI:

"Without semiconductors, there is no AI. And storage is the backbone of AI, the key foundation supporting AI's continuous evolution."

In his view, storage’s role is no longer just a component within a device; it directly carries "intelligence" itself: "Today, storage isn't just about making devices run; it's about supporting the 'intelligence' within AI, helping artificial intelligence become smarter."

With increasing model sizes, exploding reasoning demand, and the rapid rise of Agent AI, the growth logic for storage demand is very clear to Sanjay: "As models get bigger, and as reasoning demand continues to grow, AI moves from training to inference, from data centers to the edge. The demand for storage will only increase – it needs greater capacity, higher performance, and lower power consumption."

He specifically mentioned tokenomics' dependence on storage: "When you look at tokenomics, it also heavily relies on storage. As token usage grows, context windows become longer, KV cache demand increases, and models themselves get larger. AI needs not just the ability to compute, but also the ability to 'remember'."

Supply Tightness to Continue Beyond 2026

Regarding the supply-demand issue that concerns the market most, Sanjay gave a clear judgment: The entire industry's supply tightness will continue beyond 2026 and will last for a considerable period.

He explained the structural constraints on the supply side: "Building a fab takes a long time. From breaking ground to the first wafer output, it typically takes three to four years. Then you continue the ramp-up, gradually increasing the production volume."

More critically, the increasing technological difficulty is compressing the output efficiency per wafer: "The production efficiency gains from each new technology generation – meaning the bit increase per wafer – are diminishing."

Sanjay revealed that Micron had anticipated this trend around 2021.

At that time, High Bandwidth Memory (HBM) accounted for less than 1% of the entire storage industry, but they already saw that future HBM generations would consume a large amount of silicon and significantly impact the supply landscape: "So back in 2021, we said the industry needs new fabs built from the ground up. It's just that no one really predicted AI would explode at such a rapid pace."

Regarding market fears of a new glut once supply catches up, Sanjay didn't explicitly rule it out, but he emphasized that AI is still in its early stages, and the long-term structural growth on the demand side is the basis for his confidence: "From the demand side, everything is still in a very, very early stage. We believe AI still has a very long way to go."

The Underlying Logic of the $200 Billion Investment: Discipline

Micron's announcement of a $200 billion investment to build a memory manufacturing ecosystem in the US is one of the most attention-grabbing capital decisions in the semiconductor industry in recent years. For the underlying logic of this decision, Sanjay repeatedly emphasized the word "discipline":

"Investment is absolutely not made blindly; it must be disciplined and based on data. You have to understand the technology, understand the applications, and understand where these applications are headed. You also have to work closely with customers, understand where they are going in the future, and what role Micron plays."

He further explained the discipline in execution: "Today, we are investing in building a batch of new fabs from the ground up. The first step is to build the shells and infrastructure. Once these shells are built, we will still maintain discipline when installing equipment and creating actual capacity – continuously evaluating demand forecasts, assessing how much growth technological advancements can bring, and analyzing how product demand will change."

When asked if he ever had self-doubt, Sanjay's response was direct:

"We don't have self-doubt. We absolutely believe in the storage opportunity. Today, this is very clear. Of course, in our business, what's always important is maintaining adaptability and agility."