Three days' worth of DRAM on hand at the two companies that make most of the world's memory. That is the number KB Securities put on Samsung and SK Hynix in the third quarter — inventories below ten days, a condition research director Kim Dong-won described as "sellable-inventory exhaustion," not a demand recovery but something closer to a run on the shelves.
Last week this column framed the HBM4 crunch as a scarcity story and China's $7 billion catch-up bet as its counterweight. This week the scarcity has a second face: the profit pool has moved, and it has moved toward the memory makers.
The evidence for the inversion is no longer circumstantial. KeyBanc's John Vinh cut Nvidia's 2026 Rubin production plan from roughly 2 million units to about 1.5 million — not on design slips, not on CoWoS, but on HBM4 certification and supply delays at SK Hynix and Micron. When a memory qualification schedule moves an accelerator forecast by half a million units, the constraint has migrated. Nvidia can design whatever it wants. It cannot ship what Samsung and SK Hynix cannot stack.
When a memory qualification schedule moves an accelerator forecast by half a million units, the constraint has migrated.
The mechanism is wafer intensity. HBM4 consumes roughly three times the wafer capacity of standard DRAM, per KB Securities, and that multiplier is the whole story compressed into one ratio. Every wafer allocated to HBM4 is three wafers not allocated to the DDR5 that fills servers, PCs and industrial equipment. Samsung has reportedly routed 50% of its 4nm capacity to HBM4; the three major makers are steering more than 70% of new capacity toward HBM and server DRAM. Consumer memory is being squeezed as a byproduct of the AI buildout, not as a demand story of its own.
The pricing power that follows is extraordinary by the standards of a commodity once nicknamed the rice of the electronics industry. Samsung's Q3 DRAM operating margin approached 80%, with NAND at 70–75% — against TSMC's roughly 60% gross margin and Nvidia's roughly 75%. Samsung is quoting HBM4 at about $4 per Gb for 2027 supply, triple HBM3E's roughly $1.50. Bernstein sees HBM4 climbing toward $53 per GB in 2027; UBS expects the overall HBM average price to rise about 79% from 2026. TrendForce puts the 2027 HBM ASP increase at 121%.
This is the margin transfer, and it is real. The company that used to be the cyclical whipping boy of the semiconductor industry now earns margins that exceed the foundry and the accelerator vendor it supplies. Goldman Sachs now expects tightness across DRAM, NAND and HBM through 2028, and argues the market still hasn't priced the durability of the cycle — memory stocks trade at mid-single-digit P/E ratios on 2027 earnings even after the run.
But the transfer is bounded, and the bound is the same 3x that creates it. Memory margin is purchased with accelerator output. Every dollar Samsung extracts per gigabit is a gigabit that does not ship inside a Rubin rack, and at some point the arithmetic becomes self-limiting: the memory makers' windfall is capped by the physical number of accelerators the industry can build, which is exactly what their own wafer allocation is throttling. KB Securities expects DRAM and NAND bit demand to exceed supply by more than 10 percentage points in 2027. That gap is not a growth curve. It is a ceiling.
You can see the ceiling in the spot market. DDR5 availability is described as extremely scarce, with buyers raising bids that suppliers won't match; DDR4 8G has been drifting sideways for weeks because there is nothing left to trade. Consumer DRAM lead times have stretched from 4–6 weeks to 8–12; enterprise SSDs run 16–20 weeks; HBM4 makers are reportedly not accepting new customer orders at all. Micron's Sanjay Mehrotra said on the October 1 call that most 2027 HBM supply is already contracted, with no inflection point for a return to balance in sight. Samsung says the same about 2027 bookings.
That is where the shortage stops being a procurement problem and becomes a policy problem. Morgan Stanley estimates Nvidia has locked about 37.3% of 2027 global HBM supply, with long-term commitments surging from $119 billion last quarter to $279 billion — roughly $92 billion due in the remainder of fiscal 2027, $87 billion in 2028, $88 billion in 2029. One customer has pre-committed to more than a third of a global strategic input, three years forward, at a price that has nearly doubled from HBM3E's $17–18 per GB to $31–32.
Historically, that configuration triggers two responses. The first is export-control hardening: when a scarce input is locked up by a single buyer, governments treat the lock-up itself as the strategic fact, not the underlying technology. The second is antitrust attention, because long-term agreements that pre-empt supply are functionally allocation decisions made by a private party rather than a market. Neither has arrived yet. Both now have a number attached.
The China thread from 09/29 closes here as a structural gap rather than a 2026 answer. CXMT qualification for HBM3E does not relieve 2026–2027 tightness — the capacity, the stacking yields and the customer certifications are not there, and SMIC's advanced-node expansion of roughly 40,000 wafers per month after 2025's 50,000 is a rounding error against the demand curve. The catch-up bet is real. It is also a 2028-and-beyond story, which means the lock-up runs unopposed through the window that matters.
One more tension worth naming. SK Hynix, the revenue leader in HBM, is reportedly struggling with hybrid bonding and has not begun sample production, while Samsung has already delivered hybrid-bonded samples. The market leader's technical lead is narrower than its share suggests, and the qualification delays KeyBanc flagged are consistent with that. The windfall is concentrated in two Korean balance sheets, but it is not evenly distributed between them — and the customer absorbing the cost knows it.
Goldman's read is that this cycle is different: higher demand visibility, constrained supply growth, more powerful long-term agreements, more efficient capex planning. That is a plausible description of a durable upcycle. It is also a plausible description of a cartel-adjacent equilibrium, and the difference will be decided by how the $279 billion in locked commitments is treated by the regulators who have not yet looked.