Samsung's $34 Billion HBM Packaging Bet Lands in a Crowded 2028

Samsung is putting ₩46 trillion — roughly $34 billion — into high-bandwidth memory packaging at its Onyang complex in Asan, South Korea, and the facility won't start operations until June 2028. That's the number that matters for anyone tracking the memory cycle: the money is being committed during today's shortage, but the supply arrives 18 months or more from now, alongside capacity from SK hynix and Micron.

The program covers five advanced packaging lines. Of the total, ₩6 trillion (13%) goes to facility construction and ₩40 trillion (87%) to equipment — a split that tells you where the money in modern HBM actually goes. Stacking, bonding, inspection and testing tools dominate the bill, and the equipment spend will be phased across the full five-line program rather than the newly announced building alone.

What does the breakdown mean for suppliers?

For semiconductor equipment makers, the allocation is the story. Advanced HBM packaging requires wafer thinning, die stacking, bonding, inspection and testing, and every additional stacked die makes acceptable yields harder to hold. KLA is positioned through process control, inspection and metrology; Applied Materials and Lam Research supply front-end DRAM equipment; specialized packaging toolmakers benefit from rising stack and bonding requirements.

Equipment suppliers also carry a different timing risk than memory makers. They recognize revenue while facilities are being built and equipped — before the resulting chips hit the market. So suppliers can win this cycle even if memory pricing sours once the new capacity switches on. The flip side: equipment demand is cyclical too, and excess memory capacity eventually translates into delayed orders.

Why won't Onyang fix the current shortage?

Samsung's new facility is back-end packaging, not front-end DRAM wafer fabrication. HBM dies come out of DRAM fabs; packaging plants assemble them into stacked products. Adding packaging capacity doesn't add wafer supply. Finished HBM needs qualified dies, packaging equipment and acceptable yields — constrain any one and output stalls regardless of capacity elsewhere.

The physical footprint: roughly 31,000 square meters of cleanroom across three floors, expanding the Onyang campus from 270,000 to 420,000 square meters — a 55.6% increase. Operations begin June 2028, a year earlier than originally planned. But "beginning of operations" isn't full-volume production; meaningful supply depends on equipment installation and customer qualification schedules after that date.

Who else is building?

Samsung isn't alone. SK hynix announced in August 2026 roughly ₩54.3 trillion for two projects: ₩35.2 trillion for its Yongin Y2 fab and ₩19.1 trillion for Cheongju M17. The M17 cleanroom opens December 2028, Y2 in June 2029. Micron is adding DRAM capacity in Idaho and New York under multiyear programs — though these aren't dedicated HBM fabs, and their HBM contribution depends on product allocation and packaging availability.

The convergence matters because memory supply has historically hinged on the collective capex decisions of a handful of producers. Simultaneous expansion is how gluts get built.

Is demand growth fast enough to absorb it?

The Information Network estimates global HBM revenue grew from $21.1 billion in 2024 to $38.7 billion in 2025 — about 83.4% — with projections of $50.5 billion in 2026 and $63 billion in 2027. That's nearly tripling in three years, driven by Nvidia and other AI accelerator makers pushing memory bandwidth requirements up.

But the growth rate is moderating: 83.4% in 2025, a projected 30.5% in 2026 and 24.8% in 2027. In absolute terms the market still adds about $12.5 billion in 2027 versus $11.8 billion in 2026, so demand isn't declining. The question is whether it absorbs the capacity landing in 2028–2029. Remember that HBM revenue mixes volumes and pricing — added supply can pressure ASPs and dent revenue growth even as stack shipments rise.

One wrinkle supports current pricing: HBM eats wafer capacity and its complexity reduces effective bit output, keeping conventional DRAM tight even amid aggressive capex. That support weakens if HBM production gets more efficient, packaging bottlenecks ease, or AI infrastructure spending slows.

What's the risk window?

The pivotal period is 2028–2029, when Samsung's Onyang, SK hynix's M17 and Y2, and Micron's expansions hit operational milestones. Added capacity doesn't automatically mean oversupply — manufacturers can delay installs, cut utilization, shift product mix or postpone phases. The real danger is capital committed before demand is clear: once fabs and tools are in place, companies face pressure to run them and recover investment. High fixed costs mean lower utilization raises per-unit costs just as falling prices squeeze margins.

For Micron investors, watch the triangle of capex, revenue growth and free cash flow. Samsung and SK hynix juggle the same math plus technology leadership and customer qualifications. Samsung's announcement is both an opportunity and a warning — and the industry's next turning point comes not in today's shortage, but when the capacity built to end it finally switches on.