HBM Price Outlook 2027: Reading the Goldman Memory Call

The short answer #

Memory manufacturers are spending more, but that does not mean buyers will immediately see more qualified DRAM and HBM. A circulated summary of a Goldman Sachs memory-expert call held on July 28 argues that conventional DRAM pricing could continue rising through the second half of 2026 and that HBM contracts may face a larger reset in 2027.

The most aggressive number in the summary is an estimated 87% year-over-year increase in Samsung Electronics' 2027 HBM average selling price, compared with a cited sell-side consensus of 52%. LimChip has not found a publicly accessible Goldman transcript or report that independently confirms those figures. They should be treated as a reported forecast—not as an announced supplier price or a market-wide increase that applies to every HBM product.

The direction of the argument is supported by public market evidence. TrendForce forecast conventional DRAM contract prices to rise 13–18% quarter over quarter in 3Q26. It also reported that 2027 HBM4 contract negotiations were already under way and that suppliers had stronger pricing power after conventional DRAM economics moved sharply higher.

For component buyers, the useful conclusion is narrower than “all memory will double.” Near-term supply remains tight, but the price path will differ by product generation, supplier, qualification status, contract structure and delivery window.

Why announced capacity does not immediately reduce prices #

The apparent contradiction—more investment but higher prices—comes from confusing installed capacity with deliverable output.

New tools must first complete process conversion and yield ramp. The resulting dies still need packaging, testing and, for server or accelerator use, customer qualification. HBM adds stacking and advanced-packaging steps, so its output depends on more than DRAM wafer starts alone.

Product mix matters as much as total capacity. A supplier can expand production resources while allocating a greater share to HBM and high-density server products. HBM uses larger dies and more manufacturing resources per delivered bit than conventional DRAM. That leaves less incremental output for ordinary server, PC, mobile and embedded products even when headline investment is rising.

This is why a fab announcement is not a reliable short-term buying signal. Procurement teams need to know which product generation will run, when yields stabilize, whether packaging capacity is available and when the exact device is qualified—not only how many tools were installed.

Conventional DRAM may move before HBM contracts catch up #

Conventional DRAM and HBM do not always reprice on the same schedule. Conventional products can move through quarterly contracts or shorter negotiations, while HBM supply is often tied to accelerator programs and longer contract cycles.

TrendForce reported that annual HBM pricing had not fully reflected the rapid rise in the broader DRAM market. Its analysis found that the per-wafer revenue and profitability of HBM fell below a 64GB DDR5 RDIMM in the first quarter of 2026. That does not mean HBM became simple or inexpensive to manufacture. It means existing HBM pricing had lagged the value of capacity used for conventional server DRAM.

That gap gives suppliers a reason to seek higher prices when 2027 HBM4 agreements are negotiated. The shift to HBM4 can also change die size, base-die design, stack configuration, packaging cost and customer qualification, making a single percentage difficult to apply across Samsung, SK hynix and Micron.

Other analyst assumptions illustrate the uncertainty. UBS was reported in May to expect a 30% year-over-year increase in HBM3E and HBM4 average selling prices in 2027—well below the 87% figure in the circulated Goldman-call summary. These estimates may use different product mixes, customers and comparison periods. They are scenarios, not interchangeable measurements.

Long-term agreements change availability, not just price #

The Goldman-call summary also says more than half of server DRAM is covered by long-term agreements. Because commercial contracts are confidential, LimChip could not verify that precise share from a public primary source.

The underlying mechanism is still important. A long-term agreement can reserve volume through committed purchases, prepayment, take-or-pay terms, cancellation penalties or periodic pricing formulas. This gives the buyer greater supply visibility, but it does not necessarily provide a fixed or capped price.

When more output is committed before production, less inventory remains available for short-notice transactions. The residual market becomes more sensitive to a qualification delay, packaging bottleneck or accelerator launch. A buyer without direct allocation may therefore face a different price and delivery situation from a hyperscale customer covered by a multi-year agreement.

For ordinary OEM and EMS purchasing, the phrase “server DRAM is under LTA” is not enough. Buyers should confirm whether the quoted material is committed factory allocation, distributor stock or independent-market inventory, and whether the price is fixed, formula-based or valid only for a short window.

What would weaken the bullish price case? #

The price outlook is not one-way. Several developments could ease the market:

• conventional DRAM yield or bit output improves faster than expected;
• new packaging capacity removes an HBM production bottleneck;
• additional suppliers complete customer qualification;
• AI server deployments or accelerator launches are delayed;
• buyers reduce memory content or postpone systems because total platform cost becomes too high;
• suppliers redirect more capacity toward conventional DRAM.

Chinese memory capacity can also add supply, particularly in mainstream and mature products. Its effect should be judged by qualified shipments, yield and exact product coverage rather than wafer-start announcements. Additional commodity supply may influence some DRAM prices before it becomes an approved substitute for high-end server DRAM or HBM.

Hybrid copper bonding is relevant to future high-stack HBM, but it is not an immediate supply cure. New interconnect processes must pass their own yield, reliability and customer-qualification ramps. For a 2026–2027 procurement decision, proven output matters more than a technology roadmap.

What buyers should check now #

The correct response is not to stockpile every memory product. Rank requirements by production impact and redesign difficulty. Secure the least replaceable approved ordering codes first, and keep negotiable or easily qualified products under normal review.

LimChip's memory IC catalogue can be used to research published DRAM, DDR, LPDDR, eMMC and other memory part numbers. Current price, lead time, date code and lot condition still need to be confirmed for the required quantity; a market forecast cannot answer those MPN-level questions.

Bottom line #

The reported Goldman view is plausible in direction but aggressive in magnitude. Conventional DRAM may remain firm in the second half of 2026 because new investment takes time to become qualified output and HBM continues to consume a large share of advanced resources. HBM could then reprice higher in 2027 as HBM4 agreements reflect tighter capacity and more complex manufacturing.

That does not establish that all HBM prices will rise 87%, or that every DRAM part will follow the same curve. Buyers should treat the forecast as an early planning signal and watch qualified output, contract terms and exact product availability for confirmation.

Sources #

• TrendForce 3Q26 memory pricing forecast: conventional DRAM contract prices expected to rise 13–18%
• TrendForce: tight DRAM supply strengthens supplier pricing power in 2027 HBM negotiations
• TrendForce HBM market outlook: capacity crowd-out and 2027 HBM4 contract negotiations
• Micron fiscal Q3 2026 results and HBM4/HBM4E product update
• UBS 2027 HBM shipment and ASP assumptions reported by Investing.com

The Goldman percentages and server-DRAM LTA share discussed above come from a circulated summary of the July 28 expert call. No public primary transcript was available for verification at publication. Forecasts are time-sensitive and may change; this article is component-supply analysis, not investment advice.