Key points
- Cantor and Fubon Research each estimate SK Hynix's HBM4 at about $31 to $32 per gigabyte for Nvidia, near double HBM3E's $17 to $18.
- SK Hynix has not confirmed a price. The figures are analyst estimates built from Nvidia's parts cost, not a disclosed contract.
- Samsung is negotiating to price its HBM4 at the same level SK Hynix gets from Nvidia, a step up from where its HBM3E has sold.
Separate estimates from Cantor and Fubon Research put Nvidia's (NVDA) price for SK Hynix's (SKHY) newest memory in nearly the same range: roughly twice the cost of the previous generation.
At an industry conference this month, Cantor, a US investment bank, estimated that SK Hynix's HBM4 costs about $32 per gigabyte. Fubon Research put the chips destined for Nvidia at $31 to $32 per gigabyte. By comparison, the current-generation HBM3E sells for about $17 to $18 per gigabyte. On those estimates, HBM4 is close to twice as expensive.
SK Hynix has not confirmed any of this. The company says only that it prices memory according to customer requirements. Both estimates were built by adding up the components in an Nvidia system; they did not come from a contract seen by either firm. The precise figure is best treated as an informed estimate, not an established price.
Why HBM4 costs more to make
There are real reasons HBM4 is more expensive. It moves data over far more connections than HBM3E. The logic chip at its base is also larger. Both changes make the part harder to build, and they lower the share that come out of the factory working. Memory prices are rising across the whole market this year too, which pulls HBM up with everything else.
Samsung wants the same terms
Samsung Electronics is competing for the same business. Korean reports say Samsung is negotiating to contract its HBM4 at the same unit price SK Hynix gets from Nvidia, a sharp step up from its HBM3E. Samsung spent years trailing in this segment of the memory market, so winning the same price would be one indication that it has closed the gap.
HBM is the reason both companies have had the year they have. It is the memory that sits next to Nvidia's chips in AI servers, and demand for it has run ahead of supply. If the newest version really does cost about double the last one, that flows straight into both companies' profits in the second half. I wrote about the record cash they are already returning to shareholders here, and about the foreign buying that pushed both stocks to a record this week here.
The remaining question is whether SK Hynix can hold its lead at the highest end of the market. According to Korean reports, a decision on supplying Nvidia with the fastest HBM4 will come near the end of the year, after the chips pass Nvidia's testing.
Sources
- 아이뉴스24, 캔터가 파악한 SK하이닉스 HBM4 가격, on the Cantor estimate from the FMS 2026 conference.
- 머니투데이, 엔비디아 HBM 원가 부담과 삼성·SK하이닉스 HBM4 단가, on the Fubon Research per-gigabyte figures.
- 딜사이트, 삼성전자, SK하이닉스와 HBM4 가격 맞춘다, on Samsung negotiating to sign its HBM4 at the same unit price as SK Hynix.
This is general market commentary and not investment advice.
