Samsung is reportedly quoting prices more than three times those of today's mainstream HBM3E in negotiations to supply HBM4 for 2027, according to a report by Korea's Maeil Business Newspaper on October 2. HBM stacks DRAM vertically to feed large volumes of data at high speed to compute chips such as GPUs. HBM4 raises both the data transfer rate and the number of signal lines used for input and output.
However, the performance gains of HBM4 alone cannot explain the large price gap in this report. Profitability of server DDR5, which competes with HBM for production capacity, has risen, and HBM, whose prices are set on an annual basis, is under growing pressure for a price increase in its 2027 contracts.
The "3x+" quote in 2027 supply negotiations
According to Maeil Business Newspaper's exclusive, Samsung has put forward a price in the mid-to-high $4 range per 1Gb in annual supply price negotiations with major customers for 2027 HBM4. Current mainstream HBM3E is said to cost about $1.5 per 1Gb, so a simple comparison puts the quote at more than three times that level.
The "3x+" figure is a comparison between the quoted price for next-generation HBM4 to be supplied in 2027 and the current price of mainstream HBM3E. It does not mean the price of the same HBM4 product has tripled in a year.
The unit used in the original report is "Gb" (gigabits), which measures capacity in bits. It differs from "GB" (gigabytes), the unit commonly used for storage and memory capacity; one byte equals eight bits. Reading this price as a per-GB figure would change its meaning.
The report also does not disclose the stack height or capacity of the HBM4 being priced. From this figure alone, it is not possible to calculate the price of a single HBM stack or the memory cost per GPU.
No official announcement has been found that Samsung has signed contracts with customers at this price. The price quoted in negotiations should be kept distinct from the final contract price.
A 3x price and 2.7x bandwidth are different numbers
In its February 12 announcement of HBM4 mass production and commercial shipments, Samsung described HBM4's per-pin data rate as 11.7Gbps in stable operation, with a maximum of 13Gbps.
Compared with HBM3E's maximum of 9.6Gbps, 11.7Gbps is about 1.22 times higher.
Meanwhile, HBM4 doubles the I/O pins used for data input and output from HBM3E's 1,024 to 2,048. According to Samsung's technical explainer of March 13, this change lifts the maximum memory bandwidth of a single stack to 3.3TB/s, about 2.7 times that of HBM3E.
In other words, HBM4's bandwidth gain does not come only from making each signal line faster. Increasing the number of signal lines itself, and so the amount of data carried at once, also contributes.
That said, a 2.7x increase in maximum memory bandwidth does not mean AI training or inference runs 2.7 times faster.
Overall AI system performance also depends on the compute capability of the GPU or accelerator, networking, software, model architecture and more. If processing is bottlenecked somewhere other than memory bandwidth, adding HBM bandwidth yields a smaller improvement.
Capacity and bandwidth are also separate metrics. Bandwidth indicates how much data can be transferred per second; capacity indicates how much data can be held.
Samsung's current HBM4 lineup spans 24 to 36GB, with 36GB offered in a 12-high configuration. It plans to extend this to as much as 48GB with 16-high stacks. A roughly 2.7x bandwidth gain over HBM3E does not mean the size of models or KV cache that can be held grows by the same proportion.
The three numbers now in circulation also each mean something different.
| Figure and date | What is being compared | What it tells us, and its limits |
|---|---|---|
| 3x+: Maeil Business, Oct. 2 | Samsung's quoted 2027 HBM4 price vs. the current HBM3E price: mid-to-high $4 per 1Gb vs. about $1.5 | Shows the strength of pricing in negotiations over the generational change. Not the final contract price or the rate of increase for the same product |
| About 2.7x: Samsung, Feb. 12 | Maximum single-stack bandwidth of HBM4 vs. HBM3E. HBM4 reaches up to 3.3TB/s | The manufacturer's stated difference in transfer performance. Not a measured difference in overall AI processing performance |
| 121% increase: TrendForce, Sept. 29 | Forecast of the average selling price of the overall HBM market in 2027 vs. 2026 | A market-average forecast that reflects product mix and supply-demand. Not Samsung's individual contract price |
Sources: Maeil Business Newspaper, Samsung's HBM4 announcement, and TrendForce's September 29 release.
Because these are different kinds of numbers (a negotiated price, a product specification, and a market average), they cannot be combined in a calculation such as "the price is 3x but the performance is 2.7x, so it is X% more expensive per unit of performance."
TrendForce's 121% forecast also includes factors other than simple price hikes. The company factors in a larger share of sales going to the more expensive HBM4 and the expansion of HBM4E production from the second half of 2027.
Even if the price of the same product stays unchanged, the average selling price rises when higher-priced products make up a larger share of total sales. The 121% figure therefore cannot be read as meaning that all HBM will rise uniformly by 121%.
Profitability competition with DDR5 also pushes up HBM prices
In the public summary of its May 27 research report, TrendForce says that in the first quarter of 2026, HBM revenue and profit margin per wafer fell below those of DDR5 64GB RDIMM.
RDIMMs are DRAM modules used mainly in servers.
The background is a difference in timing between HBM and conventional DRAM price movements.
HBM prices and volumes for the following year are often fixed in advance through annual contracts with major customers. As a result, if conventional DRAM prices surge after 2026 HBM prices have been set, only HBM contract prices can fail to keep up with the changed market.
According to TrendForce, the result was that in Q1 2026 HBM revenue and margin per wafer fell below those of DDR5 64GB RDIMM. The published summary does not give specific amounts or margins.
This shows that HBM is not always more profitable than conventional DRAM simply because it is a cutting-edge product.
HBM carries a high unit price, but its DRAM dies are larger and it requires stacking and advanced packaging. Increasing HBM production also reduces the wafer capacity available for conventional DRAM.
When manufacturers decide how far to expand HBM output, the comparison is not only the profit from HBM itself but also the profit they would earn by redirecting the same DRAM capacity to products such as DDR5.
Annual contracts let customers secure the HBM they need in advance and let manufacturers build production plans on that basis.
On the other hand, even if memory market conditions change significantly after a contract is signed, the change is slow to be reflected in HBM prices.
The 2027 price negotiations can therefore be seen not only as a chance to reflect HBM4's performance gains and manufacturing costs in prices, but also as an opportunity to repair HBM's profitability, which has deteriorated relative to conventional DRAM.
However, this is a market structure inferred from the pricing mechanism and profitability trends TrendForce described; Samsung has not officially cited the profitability gap with DDR5 as the reason for its quoted price.
Micron has also locked in 2027 HBM at significantly higher prices
Samsung is not the only company pushing up HBM prices.
According to page 13 of Micron's September 30 earnings deck, the company has already completed customer contracts for most of its HBM bit volume to be supplied in 2027.
Micron says contract prices have risen "significantly" compared with 2026, and that as a result the gap in gross margin between HBM and conventional DRAM will narrow.
It has not disclosed specific increase rates, so it is unclear whether they are in line with the mid-to-high $4 per 1Gb that Samsung is reported to have quoted.
Still, it confirms that another major memory maker is also raising HBM prices to improve relative profitability after the sharp rise in conventional DRAM margins.
For HBM buyers, this is no longer a market where it is enough to compare the prices of competing HBM makers.
When conventional DRAM is generating high profits, getting memory makers to devote limited wafer capacity to HBM requires commensurate profitability.
In tracking AI memory prices, it is necessary to follow not only demand for HBM itself but also the prices and profitability of conventional DRAM such as DDR5, which uses the same production facilities.
HBM4 already ships commercially; price and volume are the focus for 2027
Samsung's HBM4 is not an unreleased product about to enter mass production.
The company announced the start of HBM4 mass production and commercial shipments on February 12. In its second-quarter earnings announcement on July 30, it also said it expanded HBM4 sales and prioritized server products within limited production capacity.
The 2027 price negotiations are therefore not about setting a price for HBM4 for the first time before launch. They are more accurately seen as negotiations over next year's supply volume and price for a product that has already begun shipping.
Samsung's HBM4 also reflects the company's position of having both memory and foundry businesses.
It uses 1c DRAM, a sixth-generation 10nm-class process, for the HBM4 core dies, and its in-house foundry 4nm process for the logic base die that handles signals between the stacked DRAM and compute chips such as GPUs.
Advanced packaging technology is also needed for stacking and interconnection.
In other words, rising demand for HBM4 spills over not only to Samsung's memory business but also to its foundry business, which makes the base die, and to advanced packaging.
In fact, Samsung said that for Q2 2026, excluding the effect of incentive-related provisions, its foundry business performance improved significantly. It cited demand for HBM base dies and strong orders from U.S. customers as factors.
However, this explanation alone does not allow the conclusion that HBM4 has turned Samsung Foundry profitable. The orders include business other than HBM base dies, and HBM4's standalone revenue and profit have not been disclosed.
Being able to coordinate memory, logic base dies and advanced packaging in-house could be a strength for Samsung, but it must secure sufficient capacity and yield across multiple manufacturing processes.
Samsung itself has indicated that memory supply constraints will continue even as it increases production.
A rise in HBM4 prices is therefore a separate matter from whether customers can receive the volumes they need when they need them.
Judging HBM4 prices requires looking at the whole AI system
From the customer's side, the final contract unit price alone is not enough to judge whether HBM4 pricing is reasonable.
What is needed is information such as:
- How much was the contract price?
- How much capacity could be secured?
- How many units will be delivered, and when?
- How much does actual AI processing performance improve on GPUs or accelerators using that HBM4?
Even if HBM4 offers far more bandwidth than HBM3E, overall AI processing does not get faster by the same factor.
Some workloads are limited by the GPU's compute performance, while others are bottlenecked elsewhere, such as GPU-to-GPU communication, networking, CPUs or storage.
Conversely, in workloads where memory bandwidth strongly limits performance, HBM4 may let the GPU's compute units be used more efficiently, increasing the amount of work done in the same time.
What ultimately matters is not the price multiple of HBM4 itself.
It is how much more work the whole AI system can handle as a result of the extra money paid for HBM4.
In the 2027 HBM4 market, beyond the price at which Samsung ultimately signs, whether customers can secure the supply they need and how much next-generation GPUs and AI accelerators equipped with that HBM4 raise performance in real workloads will be the yardsticks for judging whether the prices are justified.