Fifty percent. That is not a rounding error. That is a historical anomaly. Memory chips now account for half of global semiconductor revenue. The last time memory touched this share, the industry was standing at the top of a supercycle, and three months later prices went over a cliff.
I have spent the last decade reading contracts the way forensic auditors read ledgers. I have traced FTX hot wallets, decompiled DeFi protocols, and profiled ZK proof circuits. Every collapse I have studied followed the same pattern: a narrative that sounds structural, a balance sheet that behaves cyclical, and a hidden constraint that everyone ignored. Memory hitting 50% of semiconductor revenue is a red flag wearing a growth-story costume.
Digital beasts, fragile code: the Axie collapse taught me that a popular asset can die in hours because of a hidden minting function. The AI compute stack is not the Axie sweet spot, but the same law applies. When demand is concentrated at a single point, and supply is controlled by three companies, the system is not strong. It is deferred fragility.
Context: The Memory Stack
Memory is not the flashiest part of a chip. For decades it was the commodity layer of the semiconductor industry: standardized, scale-driven, and brutal in its cyclicality. DRAM and NAND prices moved like the weather. Manufacturers expanded, overbuilt, crashed, and repeated.
AI changed the optics, not the physics. An NVIDIA H100 uses 80 gigabytes of HBM3. A B200 uses 192 gigabytes of HBM3E. AI training servers carry eight to ten times more memory content than traditional servers. That is a real demand shift. It is not a narrative. The question is whether the market is pricing a demand wave or the peak of that wave.
Memory historically represented 20 to 30 percent of global semiconductor revenue. The current 50 percent share is roughly double the historical baseline. The article that triggered this note is accurate: memory now sits at the center of the industry. But the accurate number is not the same as the healthy number.
Core: Reading the Revenue Share
The 50 percent figure deserves a code review. When a number looks too big, I trace its construction.
The Accounting Mirage
Memory revenue is not rising because the world suddenly needs more storage per person. It is rising because HBM carries a premium of three to five times that of DDR5. AI buyers are paying that premium because memory bandwidth is the binding constraint in GPU performance. The revenue share is therefore a premium story, not a volume story.
Samsung, SK Hynix, and Micron control roughly 95 percent of the DRAM market. In HBM, the concentration is even tighter: Samsung and SK Hynix alone control about 90 percent of supply. Micron trails with roughly 10 percent. That is an oligopoly at the exact point of AI bottleneck. The 50 percent revenue share is partially a scarcity rent.
Rents are not earnings. They are timing. When scarcity fades, the rent fades with it.
The Real Bottleneck: CoWoS
Here is the fact I keep returning to. HBM production is not limited by DRAM wafer output. It is limited by advanced packaging. HBM dies are stacked vertically using TSV technology, then integrated with accelerators through TSMC's CoWoS packaging.
TSMC owns the CoWoS bottleneck. The memory makers own the HBM stacks, but they cannot deliver them to NVIDIA, Google, or AMD without TSMC's packaging capacity. This is the ghost in the audit: finding what wasn't there, or rather, what is absent from the memory makers' balance sheets. They report HBM revenue as if they own the full value chain. In practice, the final gate is a Taiwanese foundry that also happens to compete with their customers.
CoWoS capacity has roughly doubled since 2024. It is still insufficient. But TSMC is not expanding CoWoS to make memory makers rich. It is expanding to secure its own bottleneck position. The 50% memory revenue share is an artifact of an intermediary constraint, not a proof of durable margin.
The Yield Curve
From my audit experience, I have learned that theoretical capacity and usable capacity are different animals. HBM3E yields are around 60 to 70 percent at the leading suppliers. HBM4 is expected to start in the 50 to 60 percent range and take six to twelve months to climb. Yield is manufacturing truth. Every yield point above 60 percent is pure incremental margin. Every point below it is hidden risk.
The yield curve matters because it sets the slope of the expansion. Samsung, SK Hynix, and Micron are all adding capacity. Samsung's Pyeongtaek P4, SK Hynix's Yongin cluster, Micron's New York and Hiroshima fabs: combined capex at the three memory makers is already over $100 billion per year. That is not a small number. That is a bet that AI demand remains structurally above trend for another three years.
When I worked on Plonk circuit optimization, I learned that theoretical throughput rarely survives contact with implementation complexity. Memory expansion is the same. The equipment delivery times are twelve to eighteen months for lithography and six to twelve months for HBM-specific tools. The fab ramp takes one to two years. HBM lines require an extra year of packaging qualification. The capacity arriving in 2027 was decided in 2024, based on demand forecasts that are now becoming stale.
The Balance Sheet
The financial story is straightforward. Gross margins at memory makers have recovered from cycle lows: SK Hynix is now in the 40 to 50 percent range, Samsung at 35 to 45 percent, Micron at 30 to 40 percent. The rehabilitation is real. The cash flow is not as healthy as the margins look.
Free cash flow is being destroyed by capex. When a company spends 30 to 40 percent of revenue on capital expenditures, the operating cash flow gets consumed before it reaches equity holders. The accounting stays conservative: R&D is expensed, not capitalized. But the expansion is funded by debt and reduced returns. On an ROIC basis, SK Hynix is the only one clearly above its cost of capital. The other two are financing a growth story with balance sheet risk.
Trust is math, not magic: stripping away the myth, the math says the industry is profitable because HBM scarcity is real. The same math says the current profitability cannot survive the 2027 supply wave.
Customer Concentration
The last problem in the core analysis is the customer. NVIDIA represents roughly 50 to 60 percent of HBM demand. The top five AI customers, including Google, Microsoft, and Amazon, take a huge share of memory output. This is not a diversified revenue base. It is a single-point dependency.
Memory makers have spent years telling investors that demand is structural. Structural demand from one dominant buyer is not structural. It is relational. If NVIDIA changes its packaging strategy, or designs around HBM, or simply reallocates orders between SK Hynix and Samsung, the entire revenue share picture shifts.
I did not see this warning in the article. I saw it in the ledger structure. One concentrated buyer, three concentrated suppliers, and one packaging gatekeeper: that is not a stable system. That is a time bomb with a TSV interface.
Contrarian: Fifty Percent Is Not a New Equilibrium
Here is the contrarian take, and I will state it plainly. A 50% memory revenue share is more likely a peak signal than a new normal.
Memory cycles are not abolished by AI. They are amplified by it. The 2017-2018 supercycle also had a strong demand story: cloud computing, smartphones, and the early data center build-out. Memory share spiked above 40 percent. The market called it structural. Then DRAM prices collapsed, and the share normalized within eighteen months.
Current AI demand is more concentrated and more visible than 2018. That actually makes the cycle risk worse. When demand is concentrated, small changes in AI capex produce violent swings in memory pricing. If one major CSP pauses purchases for two quarters, HBM inventory, which was near zero, will suddenly appear with a lag.
The collective capex race is the classic prisoner's dilemma. Samsung needs to catch SK Hynix in the HBM4 generation. SK Hynix needs to defend its lead. Micron needs to prove it is not the third wheel. Each company has a rational reason to overbuild. The aggregate result is a 2027-2028 supply glut that no single company can prevent.
I have seen this in crypto protocols: every application expands liquidity, then the liquidity fragments, then the market realizes the demand was rented, not owned. Memory capacity is the same rented liquidity. The customers are not loyal. The pricing premium is not permanent.
There is also a geopolitical layer. HBM is becoming a target of export control discussions. If the United States restricts HBM exports to China, the memory industry loses access to a market that consumes roughly 30 percent of global memory. That would be a demand shock layered on top of the supply wave. The article mentioned the possibility of volatility. I would go further: the volatility risk is underpriced.
Takeaway: Count the Cycles
The next six quarters are a race. HBM4 qualification is coming. CoWoS capacity is expanding. Samsung, SK Hynix, and Micron are pouring capital into fabs that will not reach full output until 2027. Every company is making a rational bet that AI demand growth is real. The collective bet is irrational.
When the vault opens itself, the lesson from the leak is not that the vault was secure. It was that nobody bothered to check the combination. The memory market has a reveal date: the first quarter where HBM contract prices stop rising. That will show up in the earnings data before it shows up in the headlines.
I will be watching the ledger. Watch the capital expenditure lines. Watch the CoWoS allocation announcements. Watch Samsung's pricing behavior when HBM4 ramps. If Samsung starts buying HBM4 orders by undercutting SK Hynix, the price war has begun.
Memory is no longer the boring commodity layer of the semiconductor industry. It is the center of the AI narrative. That is why the 50% number is a warning, not a validation. The next bear market will not be manufactured by crypto leverage. It will be built in a memory fab with a perfectly rational expansion plan.