The numbers landed like a hammer. Memory now accounts for 50% of global semiconductor revenue. Let that sink in for a moment. For decades, this sector hovered between 20% and 30% of industry income, a cyclical workhorse that funded the more glamorous logic chip divisions. Now, it has overtaken everything. This is not a gentle trend line. It is a structural rupture, driven by an insatiable appetite for AI compute and the high-bandwidth memory that feeds it.
I have watched this industry from the inside for nearly three decades, and I can tell you that when a commodity segment suddenly commands half the revenue pie, it is not merely a market shift. It is a power transfer. The question we should be asking is not whether AI has changed the chip industry. It has. The real question is whether we are prepared for the fragility that comes with this new dominance.
The Context: A Quiet Coup in the Chip World
To understand the magnitude of this shift, we need to look at the mechanics. The AI boom has created a voracious demand for HBM, or High Bandwidth Memory. A single NVIDIA H100 GPU requires 80GB of HBM3. The next-generation B200 doubles that to 192GB of HBM3E. This is not an incremental increase. An AI training chip consumes eight to ten times more memory than a traditional server. The result is that memory manufacturers—Samsung, SK Hynix, and Micron—have been thrust into the center of the technological universe.
This is a profound reversal. For years, memory was the commodity end of the semiconductor business, subject to brutal boom-and-bust cycles. Now, it is the strategic bottleneck. The profit pool has migrated. The top three memory makers are running at over 90% capacity utilization, with HBM lines nearly at 100%. They are not just suppliers anymore. They are the gatekeepers of the AI revolution.
But here is the uncomfortable truth that the celebratory headlines miss. This dominance is built on a foundation of extreme concentration and hidden dependencies. The DRAM market is a triopoly, with Samsung, SK Hynix, and Micron controlling over 95% of global supply. In HBM, the top two players control over 90%. This is not a healthy market structure. It is a single point of failure for the entire global AI infrastructure.
The Core: The Hidden Battlefields of the Memory Wars
Based on my years of auditing supply chains and governance structures, I can tell you that the real competition in memory is no longer about lithography nodes. It has moved to the packaging floor. HBM's core innovation is not the DRAM cell itself, but the TSV, or Through-Silicon Via, stacking technology that allows multiple memory dies to be vertically integrated. This is where the true engineering moat lies.
SK Hynix has taken the lead here, being the first to mass-produce HBM3E. Samsung is scrambling to catch up, and Micron is a distant third. The yield rates tell the story. HBM3E yields are hovering between 60% and 70%, which is low for the industry. Every 10-percentage-point improvement in yield translates to roughly 15-20% more effective capacity. This is the battlefield where fortunes are being made and lost.
Yet, there is a deeper, more troubling dependency that most analysts overlook. HBM does not exist in a vacuum. It must be integrated with the GPU using TSMC's CoWoS advanced packaging technology. This means that the memory giants, despite their market power, are effectively subservient to TSMC's capacity allocation. If TSMC decides to prioritize one customer's HBM packaging over another, the entire memory supply chain shifts. The memory makers are not the true gatekeepers. TSMC is.
This creates a fascinating, and potentially volatile, dynamic. The memory companies are spending over $100 billion annually on new capacity. Samsung is building its P4 fab in Pyeongtaek. SK Hynix is planning a massive cluster in Yongin. Micron is constructing new plants in New York and Hiroshima. But all of this capital expenditure is predicated on the assumption that TSMC will have the CoWoS capacity to package it all. If that assumption fails, we will see a glut of unpackaged HBM and a price collapse.
The Contrarian View: The 50% Figure Is a Peak Signal, Not a New Normal
Here is where I must challenge the prevailing narrative. The market is treating this 50% revenue share as a new paradigm, a permanent shift in the industry's center of gravity. I have seen this movie before. In 2018, during the last memory super-cycle, memory revenue spiked to over 40% of total semiconductor income. It was hailed as a structural change driven by the cloud and mobile. Within eighteen months, prices collapsed by 60%, and the industry entered one of its worst downturns on record.
The current situation has similar fingerprints. The memory makers are engaged in a classic prisoner's dilemma. Each one is rational to expand capacity to capture AI demand. But if all three expand simultaneously, the market will be flooded by 2027 or 2028. The capital expenditure intensity is at historic highs, with capex-to-revenue ratios between 30% and 40%. This is not sustainable.
Moreover, the customer concentration is terrifying. NVIDIA alone accounts for 50-60% of all HBM revenue. This is not diversification. This is a single point of failure. If NVIDIA decides to develop its own memory solutions, or if its AI chip demand slows, the memory industry will face a demand shock of unprecedented proportions. The market is pricing memory stocks as growth companies, but the underlying business model is still deeply cyclical.
There is also the geopolitical elephant in the room. The United States has already restricted advanced logic chips and AI accelerators to China. HBM is the obvious next target. There are already proposals in Washington to include HBM in export controls. China consumes roughly 30% of global memory. If that market is suddenly cut off, the supply-demand balance will shift dramatically. The memory makers are walking a tightrope between Washington and Beijing, and the rope is fraying.
The Takeaway: Building Resilience into the Memory Supply Chain
So, what does this mean for us? It means we need to stop treating memory as a commodity and start treating it as critical infrastructure. The concentration of DRAM and HBM production in South Korea and the United States is a systemic risk. A single fire at a SK Hynix fab, as happened in 2018, could disrupt the global AI supply chain for months.
We need to think about resilience, not just efficiency. This means supporting the localization efforts in the United States, Japan, and even Europe. It means investing in alternative memory technologies like MRAM and ReRAM, which could eventually reduce our dependence on traditional DRAM. And it means being honest about the cyclical nature of this industry, even as we embrace its growth potential.
The 50% revenue share is a remarkable achievement, but it is also a warning. It tells us that we have become dangerously dependent on a handful of companies and a single technology pathway. The AI revolution is real, but the infrastructure supporting it is fragile. Code without compassion is cold, but a supply chain without redundancy is brittle. We need to build systems that can withstand the inevitable shocks, not just ones that maximize quarterly profits.
The memory industry has won the battle for revenue. The war for long-term stability is just beginning. We should not be celebrating. We should be preparing.