SK Hynix announced on 18 April that it has begun mass production of its sixth-generation high-bandwidth memory (HBM4), becoming the first memory chipmaker to ship the new standard commercially. The Seoul-based company said initial shipments from its M15X fabrication plant in Cheongju will reach Nvidia, Broadcom, and a third undisclosed US hyperscaler customer in May.
HBM4 delivers 2 terabytes per second of memory bandwidth per stack — roughly 60% higher than the HBM3E generation currently dominant in AI accelerators — and operates at 20% lower power. Each 12-high HBM4 stack contains 36 gigabytes of memory, up from 24 gigabytes in HBM3E. The chips are manufactured on SK Hynix's 1b-node DRAM process with through-silicon via packaging produced at the company's new Icheon facility.
"Mass production of HBM4 is the most consequential technology milestone for our company this decade," Kwak Noh-Jung, chief executive of SK Hynix, said at a press conference in Seoul. "We have committed full-year 2026 HBM4 volume to customers, and our 2027 capacity is more than 80% pre-booked."
SK Hynix did not disclose wafer output figures or unit pricing. Industry analysts at TrendForce estimated the company's HBM4 monthly output at 35,000 to 45,000 units during the production ramp, rising to 90,000 units per month by Q4 2026. Per-stack pricing is estimated at US$620 to US$780, a premium of roughly 70% over HBM3E pricing.
The announcement places SK Hynix approximately two quarters ahead of rival Samsung Electronics, which is still qualifying HBM4 samples with Nvidia. Micron Technology, the third major HBM supplier, told investors on its March quarterly call that it expects initial HBM4 shipments in Q3 2026. SK Hynix held 53% of the global HBM market in 2025, according to Omdia data, with Samsung at 38% and Micron at 9%.
Nvidia is expected to be the largest initial buyer of HBM4. The company's next-generation "Rubin" GPU architecture, scheduled for launch in the second half of 2026, uses eight HBM4 stacks per package for a total of 288GB of on-package memory — a 50% increase over the Blackwell Ultra generation currently shipping. Nvidia confirmed HBM4 adoption in March but has not disclosed supplier allocations.
Broadcom, which designs custom AI accelerators for Google, Meta, and several other hyperscalers, disclosed in its 6 April earnings call that it had secured HBM4 supply commitments for its next-generation Tomahawk Ultra and custom XPU chips. The company did not name SK Hynix specifically but confirmed the chips would enter volume production in "late summer."
SK Hynix's announcement comes amid continued demand constraint in the HBM market. The company's February presentation to investors indicated that all of its 2026 HBM capacity, across HBM3E and HBM4, was sold out before the end of Q1. Samsung made similar statements at its Q4 2025 earnings call.
Total capital expenditure for SK Hynix in 2026 is projected at 28 trillion Korean won (roughly US$20.2 billion), up from 19 trillion won in 2025, with the majority allocated to HBM capacity expansion. The company announced in February that it would build a new M16 fab dedicated to HBM4 and HBM4E production, with an estimated completion date of late 2027.
Shares of SK Hynix closed up 6.4% in Seoul trading on 18 April at 252,500 won, a record high. The company's market capitalisation now stands at approximately 184 trillion won (US$133 billion). Samsung Electronics shares fell 1.8% on the same session.
Korea's Ministry of Trade, Industry and Energy said in a statement that the commercial launch of HBM4 production strengthens Korea's position as the leading memory manufacturing hub. The ministry confirmed that HBM-related exports accounted for US$31 billion of Korea's Q1 2026 semiconductor export total of US$58 billion.
SK Hynix will release detailed first-quarter earnings on 24 April.