Everyone Is Ordering More Than They Need

DigiTimes reported August 4, citing industry insiders, that 2027 DRAM and HBM capacity is effectively sold out across Samsung, SK Hynix and Micron. Buyers are reportedly receiving only sixty to seventy percent of the volumes they request, and only sixty to seventy percent of estimated 2027 demand can be met. Roughly seventy percent of DRAM output is being allocated to AI server and HBM applications, leaving thirty percent for everything else. Contracts are being written as three-to-five-year long-term agreements on an advance-payment deposit model. NAND is reported sold out for 2027 at Samsung, Micron and SanDisk, with Kioxia and SK Hynix expected to confirm by the end of August. ADATA chairman Simon Chen has publicly confirmed the shortage. Separately, TrendForce reported August 4 that Nvidia has begun evaluating lower memory configurations for its Rubin Ultra accelerator, including HBM4e 8-Hi and HBM4 12-Hi, which would reduce on-package memory from 288GB to approximately 192GB. Nvidia has already halved SOCAMM capacity on Vera Rubin Superchip modules over LPDDR5X availability, and cloud providers and server OEMs have been reducing RDIMM capacities since the first half of 2026.

Take the reported allocation at face value. A buyer asks for one hundred units and receives sixty-five. What does that buyer do next quarter? They ask for one hundred and fifty, because they have learned that requests get cut by a third. So does the buyer beside them, and the one beside that. The order book now shows demand that nobody actually intends to consume, and every supplier is planning capacity against a number that includes it. This is not a hypothesis about human nature. It is the documented failure mode of allocation-constrained component markets, and it has preceded the end of every major memory cycle. Demand did not collapse in 2001 or in 2008. What happened was that suppliers discovered a meaningful share of the backlog was insurance rather than intention, and the correction arrived through the order book rather than through the economy. The sold-out headline is the evidence for it, not against it. Allocation is the mechanism that manufactures phantom demand. You cannot have double-ordering without shortage, and you rarely have a shortage this acute without double-ordering. Two things happening simultaneously make this cycle's version worse. The largest buyer is engineering its way toward buying less. If Nvidia moves Rubin Ultra from twelve-high to eight-high stacks, that is a third less high-bandwidth memory per accelerator, at the customer absorbing the majority of premium DRAM allocation. Combine reduced consumption per unit at the marginal buyer with inflated order books across everyone else, and the gap between planned capacity and real demand widens from both directions at once. And efficiency is the tail nobody models. DeepSeek's V4 Flash arrived at roughly ninety-nine percent below the output cost of the frontier alternative. If an architectural shift materially reduces memory required per unit of inference, the 2027 book was written against an assumption that no longer holds. The deposits deserve scrutiny too, because they are being cited as proof the demand is committed. Three-to-five-year agreements with advance payment do give suppliers cash before the wafers exist, and that is genuinely new. But a deposit is a fraction of contract value. If spot prices fall below contract prices by more than the deposit, walking away is the rational choice, and buyers have made it before. What deposits actually purchase is timing — a quarter or two of stickiness while the glut develops. None of this requires the current shortage to be fake. It is not. SK Hynix has committed $38.1 billion to two new fabs, and the capacity does not arrive until 2029. Samsung says no meaningful supply increase before 2028. Memory prices have risen five to sevenfold since the start of 2025, and memory has gone from ten percent of an iPhone's bill of materials to more than a third. A shortage this severe is exactly what generates the behaviour described above. Both things are true. The shortage is real, and the order book overstates it. The top does not require the scarcity to be false. It only requires the backlog to be smaller than it looks.