Three firms control 90% of global DRAM supply. That is not a market. That is a chokehold.
Over the past seven days, SK Hynix, Samsung, and Micron have collectively announced capital expenditure increases exceeding $75 billion for HBM production through 2026. The AI memory war is real, but the narrative around it is dangerously incomplete. As a due diligence analyst who has spent years mapping crypto token distributions and on-chain concentration, I see the same pattern here: a small set of entities holding the keys to an entire ecosystem's performance. The only difference is the asset class.
Context: The HBM Gold Rush
DRAM—dynamic random-access memory—is the short-term memory of every computing device. For decades, it was a cyclical commodity business driven by PC and smartphone refreshes. Then AI arrived. Training large language models requires massive bandwidth, and the standard solution is High Bandwidth Memory (HBM), a vertically stacked DRAM package that sits inches from the GPU. Every NVIDIA H100 requires six HBM3 modules. The B200 will require eight. Demand has exploded, but supply is limited by a single bottleneck: the three companies that make virtually all of it.
Samsung holds ~41% of the total DRAM market. SK Hynix ~28%. Micron ~21%. The remaining 10% is split among smaller players like Nanya and China's CXMT. In HBM specifically, SK Hynix currently leads with ~50% share, followed by Samsung at ~40%, and Micron at ~10%. This is not a free market; it is a triopoly with structural barriers that make crypto memecoins look like open-source utopias.
The code doesn't lie—but the market does. The 'AI memory war' is a deliberate framing by executives to justify record margins. But the underlying architecture reveals a fragility that any serious analyst must confront.
Core: A Systematic Teardown of the DRAM Oligopoly
1. Technology Barriers: The EUV Wall
To produce cutting-edge DRAM at the 1-alpha or 1-beta nanometer node, a fab requires extreme ultraviolet lithography (EUV) machines. Only one company makes them: ASML. The machines cost over $150 million each and have a delivery lead time of 12-18 months. Samsung, SK Hynix, and Micron collectively hold multi-year contracts for ASML's full EUV output. Any new entrant—including China's CXMT—cannot buy them due to export controls. This is a hardware-enforced monopoly.
During my time auditing smart contracts, I learned that the most effective lock-in is not a vesting schedule; it's a proprietary input that no competitor can replicate. EUV is that input. The three incumbents have turned a supply chain constraint into a competitive moat that will last at least five more years.
2. HBM Packaging: The Hidden Variable
HBM is not just about shrinking transistors. It is about stacking DRAM dies vertically and connecting them with through-silicon vias (TSVs) and microbumps. This packaging process is notoriously difficult. SK Hynix uses a technology called MR-MUF (mass reflow molded underfill). Samsung uses TC-NCF (thermal compression non-conductive film). Both require specialized equipment and years of process tuning. A single misalignment can destroy an entire stack.
I have seen this in crypto projects where a smart contract upgrade fails because the team didn't account for a subtle interaction in a proxy pattern. The same principle applies here: the complexity of HBM packaging creates failure points that only the incumbents have learned to manage at scale.
3. Capital Moloch
The capital expenditure required to build a modern DRAM fab is staggering. A single facility can cost $10-20 billion. Over the next three years, the triopoly will spend an estimated $150 billion combined on capacity expansion—mostly for HBM. This is not a game for venture-backed startups. It is the realm of sovereign-backed giants.
In crypto, we talk about 'mining centralization' due to ASIC dominance. DRAM is worse: the 'mining' of memory requires entire fabs that take years to build and cost more than the GDP of small nations. The incumbents' advantage compounds: each dollar of capex reinforces the barrier to entry.
4. Customer Concentration Risk
Who buys all this HBM? Five customers: NVIDIA, AMD, Intel, Google, and Amazon. NVIDIA alone accounts for roughly 60% of HBM demand. The triopoly is selling to a monopsony (or close to it). This mutual dependency creates a fascinating dynamic: the suppliers have pricing power due to scarcity, but they are also beholden to a single customer's roadmap. If NVIDIA's next GPU architecture fails or shifts to a different memory architecture, the triopoly's investment becomes stranded.
They built on sand; I built on skepticism. The same principle applies to protocols that depend on a single liquidity provider. Diversification is not just good practice; it is survival.
5. Geopolitical Sword of Damocles
The United States, through the CHIPS Act, is actively subsidizing Micron to build factories in New York and Idaho. Japan is funding SK Hynix and Samsung to establish R&D centers and back-end facilities. The motivation is clear: reduce reliance on East Asian production locations and secure memory supply for national security.
But this introduces another vulnerability. The triopoly is now entangled in great-power competition. A sudden escalation in US-China tensions could force Samsung and SK Hynix to choose between the Chinese market (which consumes ~30% of global DRAM) and access to American technology. In 2023, China banned Micron products from critical infrastructure. The retaliation is already happening in slow motion.
Cold logic cuts through the noise of FOMO. The triopoly's stock prices have rallied 50-100% over the past year on AI hype. But the structural risks are rarely priced in. Investors are treating these companies as secular growth stories, when in reality they are playing a high-stakes game of regulatory and technological roulette.
Contrarian: What the Bulls Got Right
Amid all my skepticism, I must acknowledge the bull case. The AI-driven demand for HBM is not a fad. It is a structural shift. The amount of memory per GPU is increasing with each generation. HBM4, expected in 2026, will double bandwidth again. The total addressable market for DRAM could grow from $80 billion today to over $150 billion by 2028, with HBM accounting for more than half.
Furthermore, the triopoly's pricing power is real. Even if NVIDIA and the cloud giants try to diversify, there is no alternative. No startup can design and manufacture a competing HBM product in the next five years. The incumbents can extract monopoly rents during this period, and their gross margins on HBM are estimated at 60-80%—compared to 30-40% for traditional DRAM.
The valuation argument also holds weight. The market still prices these stocks as cyclical commodity plays, with forward P/E ratios around 12-20x. If they were to be re-rated as AI infrastructure plays—akin to NVIDIA's 35x P/E—the upside would be 50-100% from current levels. SK Hynix, the leader in HBM, is already starting to command a premium.
But here is the catch: the bull case assumes the triopoly can maintain its technological edge and navigates geopolitical turbulence without major disruption. History suggests otherwise. In 2017, a fire at a Samsung fab caused a global DRAM shortage. In 2022, US export controls on China disrupted SK Hynix's factory in Wuxi. These are not black swans; they are gray rhinos.
Takeaway: Decentralization Is a Spectrum
The DRAM triopoly is a stark reminder that 'decentralization' in blockchain is not binary. Even if a protocol has 10,000 validators, if those validators run on hardware that is 90% controlled by three manufacturers, the system retains a single point of failure at the infrastructure level.
I spend my days analyzing token distributions and smart contract risks. The triopoly teaches me that the most dangerous centralization is often invisible—embedded in the physical supply chains that underpin the digital world.
The next time you hear a project claim to be 'decentralized,' ask yourself: who makes the memory in your server? Who makes the logic in your phone? The code is law inside the machine, but the machine itself is built by a cartel. Until the industry acknowledges this, we are all renting space in someone else's silicon.
Intermediaries lie. Blocks don't. But the DRAM triopoly is the ultimate intermediary—and it tells no lies because it doesn't have to. It owns the truth at the physical layer.