The ledger remembers what the market forgets. In 2018, memory accounted for roughly 40% of global semiconductor revenue. That was the peak of the last supercycle. The correction that followed erased $200 billion in market value within twelve months. Now, memory has crossed 50%. The question is not whether AI is reshaping the chip industry. It is whether we are looking at a new equilibrium or the same cyclical peak wearing different clothes.
Let me be precise about what the data shows. AI demand for HBM and DDR5 has pushed memory to half of all semiconductor revenue. Historically, that figure sat between 20% and 30%. The shift is not incremental. It is structural. But structure and cycle are not mutually exclusive. The industry can be experiencing a genuine paradigm shift while simultaneously approaching a cyclical apex. Both things can be true.
The HBM Bottleneck Is Not Where You Think
Most analysis focuses on DRAM wafer capacity. That is a mistake. The binding constraint in HBM production is not the memory cell. It is the TSV and CoWoS packaging layer. TSV—silicon through-via—is the vertical interconnect that stacks memory dies. CoWoS is the wafer-level packaging that integrates HBM with logic chips. Both are advanced packaging technologies. Both are controlled by a handful of players.
TSV capability sits with Samsung, SK Hynix, and Micron. CoWoS capacity sits with TSMC. This creates a structural dependency that most market participants underestimate. Memory manufacturers can build all the DRAM wafers they want. Without TSMC's CoWoS allocation, HBM cannot reach the GPU. The packaging bottleneck is the real gatekeeper.
Based on my experience stress-testing DeFi liquidity in 2020, I recognize this pattern. The market focuses on the visible constraint—wafer starts, utilization rates, bit growth. The invisible constraint—packaging capacity, equipment lead times, test infrastructure—is where the actual risk lives. In crypto, we called it the oracle problem. Here, it is the packaging problem.
The Prisoner's Dilemma of Memory Expansion
Samsung, SK Hynix, and Micron are collectively spending over $100 billion annually on capacity expansion. Each company is rationalizing this as a response to AI demand. Each company is also aware that collective over-expansion leads to the 2027-2028 supply glut. This is a textbook prisoner's dilemma.
The historical precedent is clear. In 2017-2018, all three manufacturers expanded aggressively to capture AI and server demand. The result was a 60% price collapse in DRAM over 2019. The current cycle has a similar setup, with one critical difference: HBM's technical complexity creates a higher barrier to entry. Not every manufacturer can produce HBM3E at scale. The yield rates tell the story. HBM3E yields sit at 60-70% for the top three players. HBM4 initial yields are expected at 50-60%. That is a significant constraint on effective supply.
But yields improve. Every percentage point of yield improvement adds 15-20% effective capacity. The question is timing. If yields improve faster than AI demand grows, the supply-demand balance shifts. The market is pricing HBM scarcity through 2026. The risk is that scarcity normalizes by mid-2027.
The NVIDIA Concentration Risk
NVIDIA accounts for 50-60% of HBM demand. That is a concentration risk that should concern every memory investor. The ledger remembers what the market forgets: when a single customer dominates a supplier's revenue, the supplier's pricing power is an illusion.
NVIDIA is not a passive buyer. It is the most powerful semiconductor company in history. It has the balance sheet to vertically integrate. It has the technical talent to develop in-house memory solutions. The probability of NVIDIA moving HBM production in-house within five years is low—the capital intensity is prohibitive. But the probability of NVIDIA aggressively negotiating HBM prices is 100%. That is not speculation. That is the natural behavior of a monopsony buyer.
Memory manufacturers are diversifying. Google's TPU, AMD's MI series, and custom ASICs from hyperscalers provide alternative demand. But NVIDIA remains the anchor customer. The customer concentration risk is not a tail risk. It is a structural feature of the current market.
The Geopolitical Overlay
Memory has largely escaped the US-China export control regime. That may be changing. US lawmakers have proposed restricting HBM exports to China. If implemented, this would reshape the global memory market. China accounts for approximately 30% of global memory consumption. Removing that demand would create a supply surplus in the rest of the world.
The counter-argument is that memory is too strategically important to weaponize. Japan and South Korea are allied nations. Their memory industries are deeply integrated with US technology. Restricting HBM exports would hurt allied companies more than China. This is a rational argument. Geopolitics, however, is not always rational.
The friend-shoring trend is already visible. Micron is expanding in the US and Japan. Samsung is building in Texas. SK Hynix is investing in Indiana. This regionalization of memory production will increase costs. It will also create a more resilient supply chain. The trade-off is real.
The Valuation Question
Memory stocks trade at 15-20x forward earnings. That is historically elevated. The traditional memory cycle supported 5-10x multiples. The market is paying a premium for AI-driven growth. The question is whether that premium is justified.
SK Hynix trades at 10-15x earnings. That is the cheapest of the three majors. The market is discounting its NVIDIA concentration risk. Micron trades at 15-20x. Samsung's semiconductor business is valued within a conglomerate discount. The relative valuations suggest the market is pricing different risk profiles.
My assessment: the market has partially priced the AI growth story but has not priced the 2027-2028 supply risk. That is the asymmetry. The upside is visible. The downside is underappreciated.
The Contrarian View: This Is Not 2018
The bears will point to history. Memory at 50% of semiconductor revenue is a peak signal. The last time we saw this, the correction was brutal. The bears are not wrong about the pattern. They are wrong about the context.
This cycle is different in three ways. First, AI demand is not a capex cycle. It is a structural shift in computing architecture. AI inference will require memory bandwidth for the foreseeable future. Second, HBM's technical complexity creates a higher barrier to supply. The packaging bottleneck is real. Third, the memory industry has consolidated. Three players control 95% of DRAM. They have learned from past cycles. They are more disciplined about capacity additions.
We do not build on hype; we build on consensus. The consensus is that AI demand is real. The consensus is also that memory is the bottleneck. The question is whether the consensus is early or late.
The Takeaway
Memory at 50% of semiconductor revenue is a signal. It tells us that AI is not a software story. It is a hardware story. The hardware story runs through memory. That is the structural thesis.
The cyclical thesis is equally clear. Capacity expansion is accelerating. Yields are improving. The supply-demand balance will normalize. The question is when, not if.
For investors, the positioning is straightforward. The structural thesis supports long-term exposure. The cyclical thesis supports tactical risk management. The two are not in conflict. They are in sequence.
The ledger remembers what the market forgets. The market is pricing AI growth. It is not pricing the cycle. That is the opportunity. And that is the risk.