Samsung has set HBM4 supply prices at over three times the current HBM3E rate, a move that will be locked in through annual contract negotiations with major AI customers.
This round of HBM4 pricing marks the first time in years that memory makers have proactively raised prices by a factor of three in front of AI customers.
According to a report by Korean financial media on October 2, Samsung Electronics has set its supply quote for next year's HBM4 in the mid-to-high range of $4 per gigabit (Gb), more than three times the current flagship HBM3E price of approximately $1.50 per Gb, and has already initiated annual contract negotiations with key customers. We previously detailed the HBM annual bargaining mechanism in our October 1 article, where the three major manufacturers negotiate once a year and lock prices for the contract period; thus, once signed, the supply price is fixed for the entire year, making every number on the table critical to actual financial outcomes.
$4
HBM4 price per Gb
$1.50
HBM3E current price per Gb
2048-bit
I/O count doubled
These three figures represent Samsung's supply quote per gigabit for 2027 HBM4, the current reference price per gigabit for HBM3E, and the doubled I/O count of HBM4 compared to HBM3E—specifically, the interface expanding from 1024-bit to 2048-bit, which serves as the starting point for the cost re-evaluation of this generation of products.
First, consider Samsung's own reported figures. In February 2026, Samsung announced that the industry's first commercial HBM4 had entered mass production and was being delivered to customers such as NVIDIA: the DRAM core uses the sixth-generation 10nm-class (1c) process, while the logic base die utilizes a 4nm logic process—marking the first time in HBM history that the base die has fully transitioned to an advanced logic node. The transfer speed is 11.7Gbps, approximately 46% higher than the JEDEC industry standard of 8Gbps, with a maximum single-stack bandwidth of 3.3TB/s, roughly 2.7 times that of HBM3E; 12-layer stacks provide 24GB to 36GB of capacity, while 16-layer stacks are planned to reach 48GB. Official figures indicate that power efficiency has improved by approximately 40% and thermal performance by approximately 30% compared to HBM3E.
A widely circulated claim needs correction. Many reports assert that HBM4 is synonymous with hybrid bonding, but Samsung's official stance is that it is simultaneously developing two process routes: TC-NCF (Thermal Compression Non-Conductive Film) and Hybrid Bonding (HCB). Samsung Vice President Kim Dae-woo clarified at a technical forum that MR-MUF is more efficient for stacks of up to eight layers, while TC-NCF offers advantages for 12 layers and above. Consequently, TC-NCF remains the primary production method for HBM4, with hybrid bonding serving as an option for stacks exceeding 16 layers. The bulk of the cost increase lies less in the bonding method and more in three concrete areas: the significantly higher wafer unit cost of the 4nm logic base compared to DRAM processes, the linear increase in TSV vias and bonding alignment counts with layer count, and the yield ramp-up in the first year of the new process. Supply chain reports suggest high-end stack yields have reached 80%, but this figure comes from channel information and has not been officially confirmed.
Industry interpretations of this triple price quote are split in two directions.
The first interpretation views this as an opening negotiation price, expecting customers to push prices back down to near incremental costs, as they have in previous years. The rationale is strong: SK hynix completed HBM4 development and built its mass production system in 2025, and Micron began mass production of 36GB 12-layer HBM4 in Q1 2026. With all three suppliers ramping up capacity, customers always hold the option of a second supplier, allowing them to cut orders if quotes are unreasonable.
The second interpretation sees this as a transfer of pricing power. Samsung’s own capacity outlook states that by 2027, HBM will occupy nearly 30% of its DRAM capacity. This squeeze on general-purpose DRAM is one of the primary drivers behind this year’s across-the-board memory price increases. Customers currently have no substitutes and no room to wait. In the next-generation platform designs of GPU manufacturers, HBM4 is already written into the interconnect specifications; switching memory vendors would require a platform redesign. Micron also issued a shortage warning for 2027 memory supply in early October. None of the three major manufacturers are slowing down their shift of capacity toward HBM, meaning the buyer’s cost of waiting will only increase.
I lean more toward the second interpretation, but with a discount: the triple price is an opening bid, not the final transaction price. My assessment is that the actual 2027 HBM4 transaction price will likely fall in the range of $2.50 to $3.00 per Gb—significantly above incremental costs, but not reaching the mid-$4 range. This judgment is falsifiable: if long-term contract prices for NVIDIA or AMD exceeding $3.50 per Gb are revealed before the end of 2026, it indicates seller pricing power is stronger than estimated, and this discount was misplaced. Conversely, if prices are pushed back below $2.00, this round is a typical opening-bid tactic, and Samsung’s high posture is merely a negotiation strategy.
There is also an easily overlooked timing gap. When Samsung announced the start of mass production, Executive Vice President Sang Joon Hwang simultaneously outlined a roadmap: HBM4E samples are scheduled for the second half of 2026, with customized HBM arriving in 2027. The shift toward customer-specific customization means that next year’s discussions will extend beyond price to include specifications. Whoever signs the customization terms first will secure a strategic advantage in 2027 platform designs. The $4 price tag, in a sense, also serves as a benchmark for negotiating leverage in the era of customization.
See you in the comments. If you are a procurement officer at a GPU manufacturer holding a second-source option, at what price would you sign off on this $4 quote?