Goldman Sachs Call: HBM Price Forecast Raised to 2.5x Current Levels
0xBroomberg
A Goldman Sachs memory-market expert call has raised the 2027 HBM price outlook from 'doubling' to roughly 2.5× current levels. Aggressive capacity expansion by the big three is being largely offset by HBM's heavier wafer consumption, leaving the timing of this upcycle's cool-down still uncertain.
Why did price expectations jump so suddenly?
The direct trigger: spot prices re-accelerated in late June, gaining over 15% in the first three weeks of July.
Markets had expected a slower Q3 price ramp — TrendForce's upper-bound forecast was about 18%. After the spot rally, Q3 contract-price estimates were revised up across the board, and Q4 shifted from flat to another ~20% increase.
This means → the "second-half cool-down" script has broken down. Experts are calling this streak an "unprecedented super-cycle."
How high could HBM prices go by 2027?
The prior consensus was that 2027 HBM prices would "more than double." The latest range has been lifted to +130%–150%, with aggressive calls at 170% and outliers at 200%.
The number closest to consensus now is ~+150%, implying a price roughly 2.5× today's level.
In plain terms = HBM contracts are mostly negotiated in Q4 and Q1, but DRAM contract prices kept rising afterward. Earlier forecasts only priced in H1 gains and assumed Q3 weakness — that assumption is now broken.
Capacity is surging — why isn't actual supply keeping up?
The big three's combined wafer capacity may grow 13%–18% this year, far above the historical average of under 5%: Samsung from 780k to ~880k wafers/month, SK hynix from 545k to ~645k, Micron from ~340k to 400k.
But wafer-capacity growth ≠ sellable-bit growth. HBM carries a die penalty and a trade ratio — switching production to HBM consumes more wafers per unit, and the higher HBM's share, the more conventional DRAM capacity gets crowded out.
This means → a 13%–18% wafer increase translates to only 10%–14% bit growth. Including density gains from process migration, the DRAM bit-growth consensus is about 16% this year and 17% next year — still below the historical average of roughly 20%.
How are long-term agreements changing the game?
LTAs — long-term agreements — now typically run two to five years. About two-thirds of server-DRAM contracts are already LTAs, and server DRAM accounts for 50%–60% of total DRAM.
That puts LTA penetration across all DRAM contracts at an estimated 30%–40%, and still rising.
Typical terms include ~30% prepayment (some reaching 40%), take-or-pay clauses, and capacity lock-ins, driven mainly by hyperscale cloud buyers. This reflects a market where downstream customers would rather pay upfront to lock supply than ride an upcycle unhedged.
How real is Samsung's HBM4 comeback?
Samsung has begun volume production and shipment of HBM4, achieving a pin speed of 11.7 Gbps — some samples hit 13 Gbps. Nvidia raised its HBM4 threshold from 11.6 to 11.7 Gbps, and Samsung is best positioned to meet it consistently.
The three vendors are diverging sharply: SK hynix uses 1b DRAM core dies + TSMC 12 nm logic; Micron uses 1b DRAM + its own process to cut costs; Samsung takes the most aggressive path — an advanced in-house node for the base die and the latest 1c node for core dies.
In plain terms = Samsung's HBM3E qualification delays dragged on for over eighteen months, pushing its share below 20%. New management shifted strategy from "maximize margin" to "win back share first" — higher costs bought performance headroom, which ironically became Samsung's ticket back into Nvidia's supply chain.
Could Chinese suppliers break this cycle?
CXMT's share by volume is about 16% — close to Micron — but only about 8% by revenue, because its products still sell below global-peer prices. YMTC's share has risen to roughly 13%, on par with Micron and Kioxia.
The near-term gap is less about volume and more about yield, product quality, and reliability; high-end customer qualification still takes time.
This means → the real threat may materialize in two to three years — a window that coincides with when global memory supply and demand are expected to re-balance (2028 at the earliest, 2030 at the latest). At that point, incremental Chinese supply could amplify the risk of industry overcapacity.
Content is for reference only, not financial advice.