AI memory deep dive: Understanding the industry and who's to benefit

Introduction

My earlier research piece on AI supply chain bottlenecks sparked my interest in the memory category. And once I started pulling at that thread and dive deeper, I found something that surprised me. The High Bandwidth Memory (HBM) market is projected to grow from roughly $35 billion in 2025 to $100 billion by 2028, a ~40% compound annual growth rate.

Some forecasts suggest HBM alone will surpass the entire DRAM market of 2024 by that point. Right now, HBM is sold out through 2026 across all three major suppliers. And this segment is basically an oligopoly between SK Hynix, Micron, and Samsung.

That combination of explosive demand, physically constrained supply, and a concentrated market structure is what made me want to go deeper into this segment.

1 - Why memory became the constraint

For most of the last thirty years, memory was a commodity business. Prices moved in very predictable cycles. Producers built capacity, oversupply hit, prices dropped, stocks came down with them. Rinse and repeat. Samsung, SK Hynix, and Micron between them control over 95% of global DRAM production, which helped stabilize the worst of the volatility, but the cycle itself never disappeared. Investors who lived through 2022 and 2023, when all three were reporting operating losses measured in billions, learned that lesson the hard way.

That cyclical nature is now being tested by something qualitatively different from every prior memory cycle.

The rise of large language models and AI inference at scale has created a specific, technically insoluble demand problem. Modern AI workloads are fundamentally memory bound, not compute bound. GPU performance is bottlenecked by data transfer through memory, with HBM bandwidth being the primary governor of effective throughput.

For context: DRAM is the primary high speed temporary memory used in computers and electronics to store data for active applications. HBM, or High Bandwidth Memory, is a specialized form of DRAM designed for high performance tasks like AI and advanced graphics. Think of it as a skyscraper of memory chips stacked right next to the processor.

What makes this demand curve different from anything I've seen before is who's buying. Hyperscalers are price insensitive (at least right now). They need the memory because without it their GPU clusters don't run at full utilization, and a billion dollar data center idling at 40% is a far more expensive outcome than paying a 30% premium for HBM. Morgan Stanley described this shift explicitly in a November 2025 research note:

"The core driver of memory demand has shifted from price-sensitive traditional customers to price-insensitive AI data centers and cloud service providers."

The numbers: Gartner projected data center systems spending at $653.4 billion for 2026, up 31.7% from $496.2 billion in 2025. Hyperscaler capex has been revised upward every single quarter for two years running. HBM accounted for at least 30% of total DRAM revenue by 2025, and that share is still rising. ABI Research describes AI infrastructure as "triggering a structural reallocation of global memory manufacturing capacity, creating a shortage unlikely to resolve before 2027."

That raises a question I keep coming back to.

2 - Is this still a cycle?

AI isn't like anything the memory industry has encountered before, so it feels illogical to apply the old cyclicality framework to it. But I've learned to be careful with "this time is different" thinking. Getting this question right determines how I look at the entire memory industry and the durability of the investment thesis.

The structural case

The primary argument for structural dynamics rests on a physics constraint. Building a leading edge DRAM facility costs $15 to $20 billion and takes three to five years from investment decision to scaled output.

Micron's Idaho facility, announced in 2024, will produce its first wafer in mid 2027. SK Hynix's Yongin complex won't complete its first fab until May 2027. Samsung follows a similar timeline. So no meaningful new supply arrives before late 2027 at the earliest.

HBM production adds a second layer of constraint. Manufacturing HBM requires approximately three times the wafer capacity compared to conventional DRAM. Every wafer committed to HBM is pulled away from commodity DRAM, tightening both markets simultaneously. DRAM contract pricing surged nearly 172% year over year as of Q3 2025 as a direct result of this dynamic.

KAIST Professor Kim Jeong-ho summarized it well in April 2026:

"As long as AI continues, the memory business will transform into an industry with minimal ups and downs, similar to foundries."

For context, TSMC's transformation from a cyclical foundry into a structurally scarce infrastructure provider happened when leading edge logic manufacturing consolidated into a single viable producer. HBM is experiencing a version of that same consolidation, just with three players instead of one.

Where cyclicality remains

Every prior memory cycle that appeared structural ultimately proved cyclical. The Windows PC supercycle of 1993 to 1996 drove 4x growth in memory content per device and looked permanent. DRAM prices then collapsed 51% in 1996 and another 65% in 1997. The 2016 to 2018 smartphone driven supercycle also looked different. Memory revenue peaked for Micron at $30.4 billion in FY2018 and had halved to $15.5 billion by FY2023.

The mechanism of reversal shares familiar characteristics: supply inertia means all three producers commit capex based on current demand signals and build simultaneously. New capacity arrives into a market where demand has either moderated or been met, and prices collapse under fixed cost leverage. That pattern is clearly visible in Micron's stock price over the past decades. And it's clear as day that we're in the boom phase right now.

How I think about this

I believe the supply demand imbalance will persist into 2027 and possibly 2028 as AI inference demand grows exponentially. But the traditional risk of a bust when a lot of new capacity arrives simultaneously is still very realistic. The key question is whether demand will outrun new supply. If it does, this cycle extends longer than any prior one.

In my view, the core difference this time is that the floor has moved permanently higher. AI inference is not discretionary spend that evaporates in a recession. It's embedded infrastructure. And as prices fall, demand likely increases. It's the Jevons paradox: the falling cost of using a resource causes demand to rise so much that total consumption increases rather than decreases.

But that doesn't eliminate cyclicality entirely. I think the amplitude of peaks and troughs will compress over time, creating a business that's structurally better than the memory industry of the past but still not immune to oversupply dynamics. Something more like a higher floor with a lower ceiling on volatility.

Which brings me to the question that matters most for the investment case.

3 - The runway for HBM

Every new generation of Nvidia's GPU architecture needs more HBM, not less. I believe that is a core part of the memory thesis. What we're seeing is not cyclical demand that peaks and the drops. Demand increases with every product cycle into the foreseeable future.

Written by

Share

Copied