Samsung has locked up 70% of its high-bandwidth memory (HBM) capacity through 2031, according to a Yahoo Finance report, extending a multi-year memory shortage into the next decade.

The 70% figure refers to HBM output Samsung has committed under long-term supply agreements. With AI accelerators consuming the bulk of global HBM demand, locking that volume removes most of Samsung's near-term HBM from the spot market and concentrates remaining allocation among a small number of buyers.

What HBM is and who supplies it

High-bandwidth memory stacks multiple DRAM die vertically and connects them through through-silicon vias, delivering wider buses and higher bandwidth than standard DDR memory. The format now ships primarily inside AI accelerators from Nvidia, AMD, and custom silicon programs at hyperscalers.

Three vendors dominate global HBM production:

• Samsung
• SK Hynix
• Micron

Samsung's 70% commitment covers its share of that three-supplier market. SK Hynix and Micron have also been expanding HBM capacity and signing multi-year agreements of their own.

What 70% through 2031 changes

Memory makers historically sold most DRAM through short-term contracts tied to spot pricing. Moving 70% of HBM into multi-year deals changes how the market clears.

Three effects follow:

• Pricing power shifts to the supplier. Long contracts insulate revenue from spot crashes but remove the price-discovery mechanism DRAM buyers historically relied on.
• Allocation becomes rationed. Buyers without standing agreements compete for whatever HBM remains on the open market.
• Lead times extend. Multi-year deals imply Samsung is forecasting sustained HBM demand and aligning fab capacity around those volumes.

The 2031 horizon exceeds the typical two-year DRAM contract cycle by a wide margin. The deal effectively signals Samsung expects AI-driven memory demand to outlast any near-term capex cycle.

Why the shortage persists

HBM consumes more wafer area per bit than DDR5 because of the die-stacking process, TSV manufacturing steps, and lower yields on advanced nodes. Each new HBM generation also requires qualification by the accelerator maker before it ships at volume, which extends the gap between product announcement and mass shipment.

Even with all three suppliers running at high utilization, HBM supply has lagged demand from AI accelerator shipments. The result has been multi-cycle HBM price inflation and persistent allocation pressure for smaller customers.

What the 2031 timeline signals

Locking 70% of HBM output through 2031 indicates Samsung expects demand to remain elevated for at least seven years. For accelerator designers, datacenter operators, and OEMs, the implication is straightforward: HBM supply will not loosen meaningfully before the early 2030s, and memory will continue to account for a significant share of the AI server bill of materials.

What to track next

• New HBM allocations opened by capacity additions at SK Hynix or Micron
• HBM4 qualification results and which supplier wins next-generation accelerator sockets
• Long-term contract pricing trends versus historical contract-versus-spot spreads
• Hyperscaler capex commentary indicating whether accelerator build rates are slowing or accelerating