06-reference/research

hbm vs commodity dram wafer allocation displacement

2026-11-09·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)

Neither fab has disclosed its HBM split, but the displacement ratio is robust: a wafer moved to HBM4 loses ~70-80% of its commodity bit potential whichever ratio you believe

The question

"What share of first-phase wafers at SK hynix Yongin Y1 and Samsung P5 is allocated to HBM vs commodity DRAM, and how many DDR5 bits does each HBM4/4E wafer displace?"

This is the derivative of tripwire #3 in the memory capital-cycle framework. Tripwire #3 fires on greenfield wafer capacity coming online, but a wafer start is not a commodity bit. Converting the 2027-28 online dates into commodity bit supply needs two multipliers: the per-fab HBM share, and the HBM-to-DDR5 wafer-per-bit trade ratio. Sub-question (a) is a disclosure question. Sub-question (b) is an arithmetic question.

What we already know (from the vault)

What the web says

Convergences and contradictions

The displacement arithmetic (Ray calculation, auditable)

The honest way to compute this is to anchor on the one company-guided quantity (the bits-for-bits ratio r) rather than on die areas nobody publishes.

Step 1 - bits per commodity DDR5 wafer (the reference unit). Assumptions, all Ray-chosen and all contestable: 300mm wafer (70,686 mm² gross), 16Gb monolithic DDR5 die on a 1b/1c-class node, die area 35 mm², 0.93 packing-and-edge-exclusion factor, 90% die yield.

Step 2 - bits per HBM4 core-die wafer. HBM bits per wafer = 3.4 TB / r. With r = 3.5 for HBM4 (Micron's ~3 for HBM3E, "higher" for HBM4, capped by the base-die offset above): ~1.0 TB per wafer. With r = 4.25 for HBM4E (32Gb core dies, 16-Hi, thinner dies, lower stack yield): ~0.8 TB per wafer.

Step 3 - the displacement answer. Converting one 300mm wafer from DDR5 to HBM4 core dies:

Sensitivity, and the one genuinely robust finding here. The whole calculation rides on r, but the two outputs behave very differently. Across r = 3.0 to 5.0, HBM bits delivered swings from 1.13 to 0.68 TB per wafer (a 66% spread), while net commodity bits removed only moves from 2.25 to 2.70 TB (a 20% spread). A wafer moved to HBM loses 67-80% of its commodity bit potential across the entire plausible range of r. For tripwire purposes that is the number to carry, because it survives being wrong about r. Any claim about how much HBM you get does not.

Two further levers. Bits per commodity wafer drift up 15-20% a year with node shrink and the 24Gb monolithic die, so the absolute TB figures re-base annually while the ratios hold. And the base-die question above is unresolved in a direction that matters: if Micron's 3x already nets out the foundry base die, r for HBM4 in DRAM-wafer terms could be nearer 3.2 than 3.5.

Step 4 - mix translation (company-guided input, Ray calculation). On a fixed wafer base, commodity bits scale with (1 - HBM wafer share). Taking Kim Taewoo's 20% to "nearly 30%": 0.70 / 0.80 = commodity bit supply from the same wafers falls ~12.5% in 2027 on mix alone. Industry wafer capacity has to grow ~12.5% next year just to hold commodity bits flat before any node-shrink gain.

Step 5 - what Y1's first phase is actually worth in commodity bits. Y1 phase 1 at +60k wafers/month (secondary-source cadence), expressed in commodity-wafer-equivalents:

Synthesis for RDCO

The answer to (a) is that it is not disclosed, and the correct move is to stop asking for a percentage and start carrying a range. Neither SK hynix nor Samsung has published an HBM-versus-commodity allocation for Yongin Y1 or Pyeongtaek P5, and the free analyst layer does not price it per-fab either. The usable bounds are company-level and industry-level: SK hynix at ~30% of DRAM wafers on HBM today with ~40% a 2027 analyst guess, and HBM at 20% of industry DRAM wafer capacity rising to nearly 30% in 2027 on a Samsung executive's own guidance. Y1's first cleanroom is reported as HBM4/HBM4E-leaning, which argues the top of that band, but the only attribution strong enough to cite says "designated for DRAM" and nothing more. That bounds Y1 phase 1 somewhere between 17k and 45k commodity-wafer-equivalents per month, and leaves P5 outside the window entirely.

The answer to (b) is more useful than it first looks, because the robust half of it is the half the tripwire needs. Anchored on Micron's bits-for-bits ratio, one wafer converted to HBM4 gives up roughly 3.4 TB of DDR5 bits and returns roughly 1.0 TB of HBM, so the net removal is about 2.4 TB per wafer, around 71% of the wafer's commodity potential. Run r from 3.0 to 5.0 and the HBM bits delivered move by two-thirds while the commodity bits removed move by less than a fifth. So a framework that only needs to know how much commodity supply a wafer stops producing is on firm ground; a framework that needs to know how much HBM arrives is not. That asymmetry should shape which claims we are willing to make.

For the memory-cycle tripwire list, three concrete changes follow. First, tripwire #3's primary leg counts wafers and should count commodity-bit-equivalent wafers, because an HBM-dedicated greenfield phase delivers roughly a quarter of the commodity relief its wafer count implies. "2+ vendors bring greenfield wafer capacity online" can be satisfied in 2027 by capacity that moves commodity bits almost not at all, which is exactly the trap the brownfield Step A already sprang in 2026. Second, tripwire #3's alternate leg should be rescored as moving away from trigger: the Samsung 250k wafers/month milestone the vault recorded for end-2026 now reads ~180k in 2026 with 250k pushed to 2027, which is a slip of about a year and a 2026 growth rate of ~6% against the ~47% on file. Third, the mix shift itself is a quietly large supply fact that the list does not instrument at all: HBM going 20% to ~30% of industry wafer capacity subtracts ~12.5% from commodity bit supply at constant wafers, which is the same order of magnitude as a year of capacity growth and cuts the other way.

The net effect of all of this is to push the commodity-bit-relief date later rather than earlier, consistent with the vault's existing re-specification of #3 to "first wafer out at 2 of 3 makers, no earlier than Q3-Q4 2027, material bits around Q1-Q2 2028." Nothing here is a position recommendation and nothing here changes a position. It is a measurement change: the tripwire should be read in bits, and the one number worth hard-coding is that an HBM-allocated wafer is worth roughly 0.2 to 0.33 of a commodity wafer in bit terms, about 0.29 at the HBM4 central case.

Why this is in the vault

It closes open follow-up #1 of [[2026-09-15-dram-greenfield-fab-online-dates]], which the memory-cycle framework named as the missing multiplier that prevents greenfield wafer-start dates from being converted into commodity bit supply, and it supplies the unit conversion that tripwire #3 in [[2026-07-07-dram-hbm-phase2-phase3-early-signals]] needs in order to be scored in bits rather than raw wafers. It also delivers a fresh reading on that tripwire's alternate leg, which the vault had flagged as untracked since December 2025.

Open follow-ups

Related

Sources

Vault:

Web:

Correction logged

[[2026-07-07-dram-hbm-phase2-phase3-early-signals]] is fine, but the memory-cycle thesis stated: "Every HBM wafer also crowds out ~3 commodity DRAM wafers (Micron's stated conversion ratio)." The sentence sat in BOTH live theses -- [[2026-05-18-memory-cycle-v1.1]] (line 32) and [[2026-05-18-nvidia-supply-chain-v1]] (line 36), both at status: framing-pending-founder-deploy-gate -- not only in [[2026-05-17-memory-cycle-v1]], which is superseded-by-v1.1 and audit-trail only. That is not what Micron said and it is not dimensionally possible — one wafer cannot crowd out three wafers. Micron's statement is bits-for-bits: the same bit count needs ~3x the wafer supply in HBM. The correct restatement is "every HBM bit crowds out ~3 commodity bits," equivalently "an HBM wafer yields ~1/3 the bits a commodity wafer would." The error overstates HBM's supply drag by roughly 3x if read literally. CORRECTED 2026-11-09 in both live theses plus the two downstream repetitions ([[2026-07-03-cxmt-dram-capacity-memory-oligopoly-risk]], [[2026-05-15-jane-street-dwarkesh-tour-ai-datacenter]]); the superseded v1 was deliberately left unedited as an audit record.