Samsung's 52% HBM CAGR Is Really a 2026 Problem
A set of Goldman Sachs HBM forecast tables has been circulating widely, and two numbers are doing most of the work: the HBM market doubles in 2027 (+108%), and Samsung grows HBM shipments at a 52% CAGR from 2026 to 2028.
Read the same tables vertically and a condition appears. Total DRAM bit supply grows 17-18% a year for all three vendors. Growing HBM at three times that rate means taking capacity from somewhere. This post is about where.
KEY TAKEAWAYS
1. Samsung's HBM bit supply actually fell 7% in 2025, from 604mn GB to 564mn GB. The forecast then has it at 1,503mn GB in 2026 — a 2.7x step-up in a single year. The 52% CAGR hides where the difficulty sits.
2. The 2027 doubling is not a volume story. It decomposes into +45% bits and +44% price per GB. Half of it is pricing.
3. Supply is quoted in bits, but fabs are constrained in wafers. Assuming HBM burns 3x the wafer per GB, HBM takes roughly 34% of Samsung's DRAM wafers by 2028 versus 14.7% of its bits.
The hard year is 2026, not 2028
A 52% CAGR suggests three evenly paced years. The table does not show that.
Samsung's HBM bit supply went from 604mn GB in 2024 to 564mn GB in 2025 — down, not up. Over the same period SK hynix went from 794mn GB to 1,550mn GB. Samsung ended 2025 at roughly a third of SK hynix's volume.
The forecast then puts Samsung at 1,503mn GB in 2026. That is 2.7x in twelve months. Growth after that decelerates to 66% in 2027 and 40% in 2028. In other words, essentially all of the execution risk in that 52% CAGR is concentrated in the first year.
Samsung's HBM bit supply fell in 2025 and has to nearly triple in 2026.There is a basis for the step-up. Samsung announced HBM4 mass production shipment on February 12, 2026 — described in its own release as an industry first — and in its Q2 2026 results said it was expanding HBM4 supply and had shipped the first HBM4E samples. SK hynix said on its Q2 call that HBM4 mass production began in the second quarter, with yields close to those of HBM3E.
But starting mass production and hitting a volume target are different events. Yields are not disclosed. What is observable from the outside is the language in each quarter's results and the pace of capex execution.
Half of the 2027 doubling is price
HBM TAM growth is quoted as +67% in 2026, +108% in 2027 and +45% in 2028. Those figures decompose cleanly into bit growth times per-GB price growth, and the arithmetic checks out in all three years: 1.76 x 0.94 for 2026, 1.45 x 1.44 for 2027, 1.42 x 1.02 for 2028.
One distinction matters here. Price per GB is not price per stack. As layer count and capacity per stack rise, the price of a stack can go up while the price per GB comes down. The -6% pencilled in for 2026 should not be read as HBM getting cheaper.
Which makes the +44% in 2027 the more interesting number. Per-GB pricing rising while capacity per stack keeps growing implies a genuine shortage year. The forecast then flattens to +2% in 2028, which reads as that shortage clearing.
Bits are the wrong unit
Every supply line in these tables is denominated in bits. Fabs are not constrained in bits. They are constrained in wafers.
HBM consumes far more wafer per GB than commodity DRAM: core dies are thinned and stacked, TSVs are drilled through them, and as layer count rises a single bad die can scrap the whole stack. Published estimates put the penalty at roughly 3-4x per GB. Take the conservative end and rerun the supply table.
| Vendor | HBM share of bits | HBM share of wafers (est.) |
|---|---|---|
| Samsung 2026E | 8.6% | 22.0% |
| Samsung 2028E | 14.7% | 34.0% |
| SK hynix 2026E | 14.6% | 33.9% |
| SK hynix 2028E | 20.1% | 42.9% |
Push the same assumption one step further. To grow bit supply 17% a year while shifting mix toward HBM at this pace, wafer starts have to grow 22-23% a year — about 1.5x over two years. That holds for both Korean vendors.
That is consistent with what the companies are saying in their Korean disclosures. On its Q2 2026 call SK hynix guided to capex in the high-40 trillion won range for the year, citing a pulled-forward M15X ramp and the Yongin fab 1 cleanroom opening in early 2027.
The corollary is the part that reaches consumer prices: if HBM takes a third of the wafers, commodity DRAM gets what is left. HBM and DDR5 contract pricing are not independent variables.
CXMT: 207% growth, 38% self-sufficiency
CXMT posts the highest HBM growth rate of the three at a 207% CAGR, but from a much smaller base — 125mn GB in 2026 to 1,179mn GB in 2028. Chinese HBM demand over the same window goes from 547mn GB to 3,086mn GB.
Expressed as self-sufficiency that is 23% in 2026, 56% in 2027, and back down to 38% in 2028. The ratio peaks and then declines because demand outruns supply. On total DRAM, CXMT still covers only about half of Chinese demand by 2028.
What I actually watch
| Signal | Why it matters |
|---|---|
| Samsung's quarterly HBM language | 2.7x for the full year needs a visible second-half acceleration |
| Wafer starts, not bit growth | Cleanroom opening dates and equipment orders are the observable version |
| DDR5 contract pricing | Direct read on how much wafer HBM is absorbing |
| ASIC qualification news | ASIC HBM demand compounds at 87% vs 41% for GPU, and is 39% of the market by 2028 |
Value chain read-through
| Segment | If this forecast holds |
|---|---|
| Memory makers | Wafer availability, not HBM bit targets, sets the ceiling on results |
| Front-end equipment | 22-23% annual wafer growth implies sustained order flow through 2028 |
| Bonding and back-end | Rising layer counts raise both process difficulty and tool intensity |
| Commodity DRAM | Supply stays tight as HBM crowds out standard product |
Risks to this view
- These are one bank's estimates. I worked from the exhibit tables that circulated publicly and could not read the report text, so the assumptions behind them are unverified. Forecasts get revised every quarter.
- The wafer conversion rests entirely on the 3x assumption. At 4x the conclusion gets larger; at 2x it gets smaller.
- The +44% price year assumes shortage. If capacity arrives faster, or AI capex moderates, that is the first number to break.
- If Samsung misses the 2026 step-up, the shortfall can show up as a tighter market or as share shifting to competitors. It does not resolve in one direction.
The habit worth breaking
The riskiest way to read a forecast table is to look only at the CAGR. A 52% CAGR spread evenly across three years and a 52% CAGR front-loaded into year one are entirely different propositions. And supply plans quoted in bits conceal the constraint that actually binds, which is wafers.
Sources: Goldman Sachs Global Investment Research, Exhibits 37 and 50 (from publicly circulated exhibit images); Samsung Newsroom HBM4 mass production release, February 12, 2026; Samsung and SK hynix Q2 2026 results and earnings call; wafer-per-GB penalty from published industry estimates of 3-4x. Everything here is from public sources.
Comments
Post a Comment