← Command Deck · Thesis · Engine · Comparison

Where every memory bit comes from — and where it goes

Price ₩1,693,000 · 12-mo blended E[V] ₩3.92M (+132%) · Tue 1 Sep close

The likeliest forecast, drawn. Tools → producers → product tiers → end demand, with the supply-demand balance projected year by year and a range around each figure. This is a forecast, not a calculator.

THE FORECAST IN ONE PARAGRAPH

Demand outruns supply in every tier through 2028. The squeeze is worst in server DRAM, which peaks at a +56% cumulative gap in 2028, because HBM is eating the wafers that used to make it. Consumer memory flips to surplus first, in 2029. HBM stays tight longest and is the last to loosen. By 2030 every tier has flipped — which is the wave, arriving on the schedule this site has always dated, and the reason the price path plateaus in 2029 rather than compounding.

The flow at peak tightness (2027)

Read it left to right. ASML's ~60 EUV systems a year is the hard gate on leading-edge capacity — but note the wide DUV band beneath it, and the pink China lane feeding CXMT that bypasses export controls entirely. In the producer column, SK's block is sized by its 36% DRAM share, but the flow into HBM is sized by its 57% HBM share — that thick green band into the top tier is the profit engine. On the right, the two AI bands together outweigh everything else combined, and the robots band is the one growing fastest from a small base.

The balance, year by year

Tier	2026	2027	2028	2029	2030
HBM	+10% +4 to +14	+17% +8 to +24	+21% +9 to +30	+19% +9 to +27	+16% +7 to +23
Server DRAM / LPDDR	+20% +9 to +28	+42% +19 to +60	+56% +25 to +80	+50% +22 to +71	+46% +21 to +65
Consumer DRAM / NAND	+8% +4 to +11	+11% +5 to +16	+9% +4 to +13	-3% -5 to -2	-11% -17 to -6

Cumulative demand-over-supply gap. Positive = shortage, sellers set price. Negative = surplus, buyers set price. Ranges are weighted by the probability engine's regime mix (45% bear mass, 38% base, 17% upside).

What the gap does to prices

Year	HBM gap	Server gap	Consumer gap	Theoretical price move	Realised (after contract ceilings & floors)
2026	+10pt	+20pt	+8pt	+56%	+42%
2027	+7pt	+18pt	+3pt	+43%	+32%
2028	+3pt	+10pt	-2pt	+21%	+15%
2029	-1pt	-4pt	-11pt	-20%	-7%
2030	-3pt	-3pt	-8pt	-17%	-6%

Annual gap ÷ 0.25 buyer elasticity = theoretical pressure. Contracts damp it in both directions: ceilings cap roughly a quarter of the upside, floors absorb about two-thirds of the downside. That asymmetry is the whole investment case — it is why 2029 is modelled as a plateau rather than a crash.

Cross-check against the main model — and where they disagree

This flow model is built bottom-up from tier volumes; the main forecast is built top-down from price × bits. They agree closely on 2026 (+42% here vs +42% in the chain) and on the 2029 turn. They disagree on 2027: this model implies +32% realised price growth, while the revenue chain uses +42%.

Stated plainly: the main chain's FY27 sits at the optimistic end of what this flow model supports. On the flow model's central case, FY27 revenue would be nearer ₩588T than ₩771T — about 7% lower. The chain is not being revised on this alone, because the two methods use different demand inputs and the range on the 2027 server gap is wide (+19% to +59%). But the tension is real, it is disclosed here rather than hidden, and October's Q3 print is the datum that settles it.

From capacity and contracted price to a share price

Until now the revenue forecast was top-down — last year's revenue times a price assumption times a volume assumption. Two things changed that. SK removed the price cap from its long-term contracts, so spot rises flow straight through, and 2027 HBM4 is already sold out at published prices. That means 2027 revenue can be built the way an accountant would build it: segment by segment, volume × price.

THE INPUTS — what is contracted, what is disclosed, what is estimated

Input	Value	How solid
2027 HBM4 price	Nvidia $32/GB · Broadcom $36 · AMD $40 weighted ≈ $34.6/GB vs $15.5 in 2026 = +123%	Broker-sourced (Cantor). Capacity reported sold out.
2027 conventional DRAM price	+45% to +65%	Contracts printed +90-95% then +58-63% QoQ in H1-26, and the cap is gone on renewal.
2027 NAND price	+30% to +45%	Printed +55-60% then +70-75% QoQ in H1-26; eSSD ramp.
Volumes	HBM bits +40-50% · DRAM +6-12% · NAND +12-18%	SK guidance: Q3 DRAM shipments +10% QoQ. M15X ramping, 1c conversions. Yongin adds nothing — first cleanroom Feb-27, equipment Q2-27, 360K wpm only by H1-2030.
Segment mix of 2026	HBM ~32% · DRAM ~44% · NAND ~24%	Estimated. SK does not disclose revenue by tier. This is the softest input and the one most likely to be wrong.
Operating margin	70%	Q2-2026 actual was 76.3%; 70% allows the depreciation step and mix shift.

FY2027 BUILT FROM THE PARTS

Case	HBM	Conventional DRAM	NAND	Total revenue	EPS	End-2026 share price
Conservative	₩331T	₩243T	₩126T	₩700T	₩500K	₩4.7M
Central	₩366T	₩268T	₩137T	₩771T	₩550K	₩5.2M
Upside	₩406T	₩293T	₩148T	₩847T	₩605K	₩5.7M

Reconciliation (28 Aug): the ₩771T / ₩550K builder above is the pre-Q2 derivation of this page and is kept for its structure; the live Command-Deck chain the site displays is FY27 ₩633T · clean EPS ₩583K · end-2026 ₩5.5M. Year-end centrals in the table below are re-based to the unified path; conservative floors retain the pre-Q2 build. Full re-derivation of this builder on Q3 actuals, 27 Oct. The conservative case (₩700T · ₩500K · ₩4.7M) is kept in the table above as the downside boundary of the same arithmetic, and the upside case at ₩847T · ₩605K · ₩5.7M as its ceiling.

The one input to keep watching. Every won of the HBM uplift runs through the segment mix, and SK does not disclose revenue by tier — so the ~32% HBM share is our estimate, not a fact. If HBM is really 25% of 2026 revenue, the central case falls back to roughly the conservative one. Q3 in late October settles it: revenue above ₩104T with the HBM mix holding confirms this page; below ₩98T and the demotion tripwire fires the other way.

SHARE PRICE DEDUCED, BOTH WAYS

Year end	OUR ESTIMATE (central)	Conservative floor	What has to be true for the central case
2026	₩5.5M	₩4.7M	Q3 revenue >₩104T with HBM mix holding at ~32%
2027	₩8.0M	₩6.3M	HBM4 ladder holds at $32-40/GB through the year
2028	₩8.4M	₩6.8M	No price cap reinstated as Samsung and Micron qualify
2029	₩7.9M	₩6.5M	Yongin and Micron capacity absorbed rather than dumped

Both columns use the same multiples (9.5 → 10.6 → 10.2 → 8.8) and the same 44% bear-regime mass. The gap between them is entirely the segment-mix assumption, which is why it is disclosed as the weakest input rather than buried in a spreadsheet.

The integrated model — every input, in one place

This is the full specification. Every number the forecast rests on, what drives it, and where it came from. If any one of these is wrong, this table shows you exactly what breaks.

DEMAND SIDE

Input	Value	Evidence
HBM bits	+40-50%	SK guidance +10% DRAM shipments QoQ ≈ +46% annualised; M15X ramping; unit-driven since capacity per GPU was flat at Rubin
Conventional DRAM bits	+6-12%	1c node conversion delivers +8.6%; wafers otherwise diverted to HBM
NAND bits	+22%	KV offload to SSD is the third memory tier; SK flash sales +90% last quarter; eSSD ramp
The agent multiplier	1.5-3×	Measured: every HBM deployment needs 1.5-3× that capacity in DRAM beside it (SemiAnalysis AgentX, $3M / 2MW / 1,000 chips)
Physical AI	~$31B/yr	Vehicle memory 90GB → 278GB; L4 autonomous 300GB+; humanoid shipments +272%/yr

PRICE SIDE

Input	Value	Evidence
HBM price	+119%	2027 HBM4 contracted at $32-40/GB vs ~$15.5 today (Cantor). SK removed the price cap — spot rises flow through in full; Micron kept its cap
Conventional DRAM price	+55%	Contracts printed +90-95% then +58-63% QoQ in H1-2026; no cap on renewal
NAND price	+45%	+55-60% then +70-75% QoQ in H1-2026
Blended	+79%	Cross-checks against the ×2.27 revenue step. Nvidia's +17% server rise is confirmation of these contracts, not additional

SUPPLY SIDE — what limits it

Constraint	Effect	Evidence
Samsung capacity	+1.3% in 2027	Omdia. The number-two producer is choosing mix over volume — pushing HBM, barely adding ordinary DRAM
TSV / stacking	The binding limit on HBM	HBM consumes >70% of industry TSV capacity; 19 extra process steps on equipment that cannot do conventional packaging
ASML EUV	~60 systems/yr	~20 needed per 100K-wpm fab. Gates leading-edge only — China's domestic DUV bypasses it
New fabs	2028-2030	Micron Boise ramps 2028; M17 Dec-2028; Yongin Y2 Jun-2029; Y1 at 360K wpm only by H1-2030

FROM REVENUE TO SHARE PRICE

	FY2026	FY2027	FY2028
Revenue	₩360T	₩816T	₩1,102T
Operating margin	76%	70%	67%
Profit per share	₩270K	₩583K	₩753K
Multiple applied	9.5×	10.6×	10.2×
Share price at year end	₩5.5M	₩8.0M	₩8.4M

Margin declines by design — depreciation from the new fabs climbs ₩31T → ₩47T → ₩68T. Multiples are held near Micron's historical ceiling of ~10×: growing earnings at a constant multiple is what produces the price, not multiple expansion.

WHAT COULD BREAK IT — 44% of the probability

Scenario	Weight	Outcome
Multiple compresses for two years	10.0%	₩2.6-3.9M
HBM share lost to Samsung and Micron	8.5%	₩1.8-3.0M
Decode moves to SRAM architectures	7.0%	₩2.0-3.4M
China floods commodity DRAM	7.0%	₩2.3-3.5M
Inventory correction / bullwhip	4.5%	₩1.9-3.2M
Korea · Taiwan · currency	4.0%	₩2.2-3.6M
Thesis simply wrong (includes a Taiwan conflict)	3.0%	₩0.8-1.6M

Three tripwires fire automatically: Q3 revenue below ₩98T (late Oct) · October HBM capture below 55% · Q1-2027 share below 34%.

What would change this picture

CXMT scales faster than modelled	Its $8.6B raise plus domestic DUV lands in commodity DRAM, not HBM. Would flip the consumer and server tiers a year early. Carries 7% weight in the probability engine.
HBM per task falls	d-Matrix in production, AMD's Taalas acquisition, Cerebras splitting decode. Cuts the HBM band without cutting total bits — the counter-argument is that fewer HBM stacks per GPU means more racks, not fewer.
Agent and robot demand at the top of range	Persistent per-user context plus fleet training would widen the top two bands simultaneously while supply cannot respond inside the window. 17% weight.
New fabs slip	Yongin faces Gyeonggi wastewater demands and reported night-work limits. Any slippage pushes the 2029 flip to the right and extends every green bar above.

Inputs. DRAM revenue shares: SK ~36%, Samsung ~34%, Micron ~22%, CXMT 7.67% (CXMT IPO prospectus). HBM shares: SK 57%, Samsung ~21% (targeting 38%), Micron ~21% (Counterpoint Q1-2026). HBM consumes ~3× the wafer area per bit of DDR5. ASML: 48 EUV shipped 2025, ~60 in 2026, ~80 in 2027; ~20 needed per 100K-wpm fab. China domestic immersion DUV: ~5 units 2026, ~20 in 2027, first deliveries SMIC/Hua Hong/CXMT (The Information, 27 Jul 2026). Yongin Y1 360K wpm by H1-2030; M17 cleanroom Dec-2028; Y2 Jun-2029; Micron Boise first wafers 2027, Clay NY up to four fabs, capacity pre-sold under non-cancelable agreements through 2030. AI capex 2026 ~$765B (past oil & gas at $681B). Memory market ~$890B in 2026 (labelled estimate). Consumer RAM +90% latest quarter; DDR4 +158% / DDR5 modules +307% since Sep-2025 (TrendForce). Tier volumes and growth rates are modelled from these public inputs; companies do not disclose unit volumes by tier. Updated 2 Sep 2026 · anchor ₩1,693K — 1 Sep close.