"STORE" IS A NEGLECTED BOTTLENECK FOR AI, SUPPLY IS TIGHT OR CONTINUES BEYOND 2026

Original by Li Jia

Original source:See you on Wall Street

"AI COMPETITION IS NOT JUST A MATH CONTEST, BUT A STORAGE COMPETITION. @AMBASSADOR: @AMBASSADOR: @AMBASSADOR: #GAZA #GAZA。

On 5 June, the podcast, A Bit Personal, received a rare in-depth interview at home. In addition to the usual industry insights, this personalized conversation allowed him to speak on his own initiative about growth experiences, family influence and career choices。

AI IS STILL AT AN EXTREMELY EARLY STAGE, AND THIS IS ONE OF THE CORE JUDGMENTS OF SANJ。

In his view, AI needed more than just more calculus, as the big model, Agent AI, and reasoning applications evolved。

Longer context windows, larger model sizes and growing Token consumption are contributing to the continued rise in storage demand。

THE ESSENCE OF AI IS DATA, WHICH CANNOT BE STORED, SO STORAGE WILL BE ONE OF THE MOST IMPORTANT INFRASTRUCTURE IN THE PROCESS OF AI CAPACITY ENHANCEMENT。

At the same time, the supply side was not adequately prepared. Sanjay noted that the storage industry was currently facing supply constraints that were not short-term supply-demand mismatches, but rather structural supply constraints. Advanced storage products require more rounding, while the construction of new rounding plants often takes three to four years, with subsequent production being equally lengthy。

More importantly, as the technology nodes advance, the output of storage capacity per round is declining。He concluded that supply constraints in the industry were expected to continue beyond 2026。

In explaining why storage technology has been underestimated for a long time, Sanjay put it simply:"There is often a misunderstanding of memory, and there is no understanding of how hard it is to create it."FROM PHYSICS, CHEMISTRY TO MATERIAL SCIENCE, TO MASS PRODUCTION, TO ENSURE THAT EACH OF THE TRILLION BITS IS ACTED CORRECTLY, THE UNDERLYING TECHNOLOGY IS EXTREMELY DIFFICULT. IN HIS VIEW, AI WAS ALSO A STORAGE COMPETITION, WHICH THE MARKET HAD LONG IGNORED。

In a longer-term perspective, Sanj believes that the bottom logic of corporate and individual success has not changed. Whether it's promoting a $200 billion investment plan or leading Light through the storage industry cycleThe key words he has repeatedly emphasized are resilience, discipline and permanenceI don't know. Investment must be based on data and fundamentals, and leaders need to look at industrial trends and understand technological details in depth。

As he had learned from his father, success required both persistence and the ability to seize opportunities at critical times。

THE CORE POINT OF THE INTERVIEW WITH GEO SANGER MEHROTRA WAS AS FOLLOWS:

STORAGE IS AN UNDERVALUED BOTTOM BOTTLENECK OF AI, WHOSE MANUFACTURING DIFFICULTY AND STRATEGIC VALUE ARE WELL BEYOND MARKET RECOGNITION。It's going from "calculations competition" to "storage competition." The expansion of models, the lengthening of context windows, and the proliferation of Token consumption have made AI more dependent not only on greater calculus but also on greater "rememberability". There is not enough storage capacity and bandwidth, and no strong computing power can be released。

The structural constraints on the supply side determine that storage shortages are not short-term fluctuations but long-term。The advanced storage product consumes more crystals, while it takes three to four years to build a new mill, with equally long capacity to climb the slope. At the same time, the advancement of technical nodes has led to a decrease in output per crystal circle. Supply and demand are staggered and supply stress continues at least beyond 2026。

The difficulty of creating memory is always underestimated, but it is precisely the deepest moat in the industry。From physics, chemistry to material science, from design to mass production, it is ensured that hundreds of trillion bits are in every error and that its engineering is extremely complex. Storage of chips is as difficult to produce as any semiconductor area, or even more difficult in many ways。

Success comes from resilience, discipline and long-termism, rather than short-term judgment。Whether it is driving $200 billion in investment or moving through cyclical fluctuations in the storage industry, leaders need to see both industry trends and technical details. Just as fathers do not give up after having been denied visas three times, success requires both the persistence of the endurance and the ability to seize opportunities at critical times。

STORAGE IS BECOMING AI BACKBONE

When it comes to the current historical position of the storage industry, Sanjay goes on to say, "I've been in the industry for over 45 years. This is the most exciting moment I've ever had in the whole industry."

HE FURTHER EXPLAINS THE STRATEGIC SIGNIFICANCE OF STORAGE FOR AI:

"NO SEMICONDUCTOR, NO AI. AND STORAGE IS THE BACKBONE OF AI, AND IT IS THE KEY FOUNDATION THAT SUSTAINS ITS EVOLUTION."

IN HIS VIEW, THE ROLE OF STORAGE IS NO LONGER JUST A PART OF THE DEVICE, BUT A DIRECT CARRYING OF "SMART" ITSELF: "TODAY, STORAGE IS NOT JUST LETTING THE DEVICE RUN, IT IS SUPPORTING THE "SMART" ITSELF IN AI, HELPING ARTIFICIAL INTELLIGENCE TO BECOME SMARTER."

With the expansion of models, the outbreak of reasoning needs, and the rapid rise of intelligent AI (Agent AI), the logic of growth of storage demand seems clear to Sanjay: "Ai will only grow with the growing number of models, as the demand for reasoning continues to grow, as AI moves from training to reasoning, and from data centres to the margins, the demand for storage will only increase — it will require greater capacity, higher performance and lower effort

He also referred specifically to the dependence of token economics on storage: "When you look at token economics, it is equally highly dependent on storage. As the use of token increases, the context window becomes longer, the need for KV caches increases, and the models themselves growAI NEEDS NOT ONLY THE ABILITY TO CALCULATE, BUT ALSO THE ABILITY TO REMEMBER."

Supply stress will continue beyond 2026

Sanji has made a clear judgement on the supply and demand issues of greatest concern to the market:Supply stress across the industry will continue beyond 2026 and for a considerable period of time。

He explained the structural constraints on the supply side: "It will take a long time to build a round mill. It usually takes three to four years from groundbreaking to the first crystal round output. It will then continue to climb the slope and gradually increase production”

more crucially, the rise in technical difficulty is compressing the output efficiency of the unit round: "the productivity gains of each generation of new technologies, that is, the bit increments that each round can bring, are diminishing

Soneji revealed that Misang had predicted this trend as early as 2021。

AT THAT TIME, HIGH BANDWIDTH STORAGE (HBM) ACCOUNTED FOR LESS THAN 1 PER CENT OF THE ENTIRE STORAGE INDUSTRY, BUT THEY HAVE SEEN THAT FUTURE GENERATIONS OF HBMS WILL CONSUME LARGE AMOUNTS OF SILICON CHIPS AND HAVE A MAJOR IMPACT ON SUPPLY PATTERNS:"SO AS EARLY AS 2021, WE SAID THE INDUSTRY NEEDED TO BUILD A NEW ROUND FACTORY FROM SCRATCH. IT'S JUST THAT NO ONE REALLY PREDICTED THAT AI WOULD EXPLODE SO QUICKLY

SANJAY DID NOT DIRECTLY EXCLUDE THE MARKET'S FEAR OF “RE-OVERSUPPLY,” BUT STRESSED THAT AI WAS STILL AT AN EARLY STAGE AND LONG-TERM STRUCTURAL GROWTH AT THE DEMAND END WAS THE BASIS OF ITS CONFIDENCE: “FROM THE DEMAND SIDE, THIS IS STILL AT A VERY EARLY STAGE. WE THINK THERE'S STILL A LONG WAY TO GO AFTER AI."

Bottom logic of $200 billion investments: discipline

The US-based US-based investment of $200 billion in the construction of storage manufacturing systems has been one of the most interesting capital decisions in the semiconductor industry in recent years. On the bottom line of this decision, Soneji has repeatedly emphasized the word "discipline":

"In no case is investment blindly made, it must be disciplined and based on data. You have to understand technology, you have to understand applications, you have to understand where they go. And you will also work closely with your clients to understand where they are going in the future, and what role Mi-gwang plays in it.”

He further explained the discipline at the enforcement level: “Today, we are investing in the construction of new round mills from scratch. The first step is to build the plant and infrastructure. When these plants are completed, we will remain disciplined in installing equipment and generating real capacity — continuously assessing demand projections, assessing how much growth technological progress can bring and assessing how demand will change”

When asked if there had been any self-doubt, Sanjie answered simply:

"We have no self-doubt. We absolutely believe in storage opportunities, and today that is very clear. Of course, in our operations it is always important to maintain adaptability and agility”