Why Only Three Companies Make the Memory That Runs AI

Three companies make the memory that every AI chip on Earth depends on. They're sold out years in advance, their margins beat NVIDIA's, and in May 2026 all three bought a stake in their own biggest customer. The supplier became the owner.This episode goes inside one component: HBM, the high-bandwidth memory bolted next to every AI accelerator — and the hardest chip on Earth to make. Two questions the headlines skip: why can only three firms make it, and what happens when the supplier owns the customer.The answer to the first is arithmetic. An HBM chip is twelve to sixteen wafer-thin memory dies stacked vertically, and the package only works if every die works — so yield multiplies instead of averaging. At a 95% per-die yield, a sixteen-high stack collapses to about 44%. A new entrant doesn't need to be good; it needs to be near-perfect — and NVIDIA reportedly relaxed its own HBM4 spec to fit what the three best memory makers on Earth can actually yield.Three firms making essentially all of it, sold out years forward, is an oligopoly — and SK hynix posted a 72% operating margin in early 2026, an all-time high that outshines the company buying its chips. Memory is now 63% of an AI chip's component cost, and roughly 90% of all HBM is made in one country: South Korea.The move that anchors it all: on May 28, 2026, Micron, Samsung, and SK hynix all joined Anthropic's $65 billion round at a $965 billion valuation — the first time the entire memory oligopoly took equity in a single AI lab.Which leaves the real question: a durable moat, or the top of a violent cycle? Memory has always been brutally boom-bust, the bears cluster on a 2027 inflection, and all three crossed a trillion dollars in market value within 21 days of each other. The honest read holds both — a real moat, with a real expiration date.Deep Dive goes beneath the headline on the stories that matter — AI, finance, security, and the systems running underneath.RELATED EPISODESThe AI Chip War: Why Everyone's Watching the Wrong Fight — named memory as the binding AI bottleneck; this goes inside it.NVIDIA Just Forecast $91 Billion Without China — memory eating the AI chip's bill of materials, now 63%.How LLM Inference Actually Works — the memory wall, and the wafer-scale SRAM bet that routes around it.Why Wall Street Is Betting Billions on Nuclear Power for AI — the same 'AI needs X' thesis priced into trillion-dollar firms.CHAPTERS00:00 Three firms, sold out, and they bought their customer01:04 Why only three: the yield math05:14 Why China can't just become the fourth06:21 An oligopoly that's sold out for years08:11 Margins that outshine NVIDIA10:18 Memory is 63% of the chip — who pays for it13:21 The Korea chokehold15:59 The supplier becomes the owner17:58 A durable moat, or the top of a cycle?20:34 Can anyone route around HBM?SOURCESHBM stack-yield is multiplicative — at 95% per die, a 16-high stack ≈ 44%; one bad die kills it (nomadsemi; SemiEngineering)Three firms ≈ 97% of HBM, ~90% in Korea; NVIDIA reportedly relaxed its HBM4 spec to fit yields (AI Frontiers; TrendForce)Samsung HBM3E ~18 months late; dismantled its HBM team in 2019 (Tom's Hardware; SamMobile)SK hynix Q1 2026: 72% operating margin, sold out ~3 years; Micron 75% gross margin (CNBC; Yahoo Finance)HBM = 63% of AI-chip component cost; DDR5 kit ~$80–100 → $350–600; Dec 2024 BIS HBM rule (Epoch AI; BIS/CSIS)Anthropic $65B round at $965B — all three memory makers join, May 28 2026; Intel customer-equity counter-precedent (TechCrunch; Bloomberg)———This episode is for educational and informational purposes only and does not constitute financial, investment, or trading advice, nor a recommendation to buy or sell any security. Deep Dive is not a registered investment adviser. All investing involves risk. Consult a licensed financial professional before making investment decisions.