SK hynix holds firm on 90% yield threshold for hybrid bonding despite HBM stacking race

As the semiconductor industry weighs the adoption of hybrid bonding technology needed to manufacture next-generation high-bandwidth memory chips, SK hynix has underscored its position that a yield rate — the share of defect-free chips — of at least 90 percent must be secured before the technology can be introduced.

Competition to stack HBM dies ever higher has intensified focus on when hybrid bonding will be adopted, but the company's stance signals it will not rush the transition until heat management and mass-production stability are sufficiently confirmed.

At the "Future of AI Computing" conference recently hosted by SEMI, the global semiconductor equipment and materials industry association, in San Jose, Kim Ji-hoon, senior director at SK hynix America, took part in a panel discussion and outlined the company's next-generation HBM strategy, according to industry sources Saturday.

Kim said that stacking HBM beyond 12 layers creates bottlenecks in heat dissipation, yield and packaging, and warned that these physical constraints cannot be ignored.

HBM is built by stacking multiple DRAM dies vertically. While a higher stack increases data capacity and processing speed, it also raises packaging complexity, making yield management and thermal control more difficult.

As HBM stacks have grown from eight layers to 12, 16 and 20, the industry has come to view hybrid bonding as the key technology for the next generation.

SK hynix disclosed at Hot Chips 2026, a high-performance semiconductor and computing architecture conference held in Silicon Valley on Aug. 23 (local time), that it is developing hybrid bonding to enable HBM stacks of 20 layers or more.

Current production relies on thermocompression bonding, or TC bonding, which uses heat and pressure to join dies. TC bonding connects DRAM layers electrically through small solder joints called micro-bumps placed between each die, a method that adds to the overall thickness of the package and increases power consumption during data transfer.

Hybrid bonding, by contrast, joins copper directly to copper without micro-bumps, reducing the total thickness of increasingly tall HBM stacks while enabling higher performance — making it an essential technology as stack counts rise.

Hybrid bonding had been expected to arrive as early as the sixth-generation HBM4, but yield challenges tied to the technology's complexity have pushed back that timeline. The industry now expects hybrid bonding to debut with the seventh-generation HBM4E 16-layer product or the eighth-generation HBM5 20-layer stack.

At the panel discussion, Kim stressed SK hynix's cautious and rigorous approach, saying that "a yield of 90 to 98 percent" and "a lifespan of three to five years" are non-negotiable standards before hybrid bonding can be put into actual use.

The remarks signal that the company will not rush to adopt hybrid bonding by lowering those benchmarks simply to keep pace in the HBM stacking race.

The stance is reinforced by signals from Nvidia, SK hynix's largest HBM customer, which is considering equipping its next-generation AI accelerators with lower-specification HBM.

According to market research firm TrendForce, Nvidia had planned to use HBM4E 12-layer chips in its next-generation AI accelerator Rubin Ultra, set to launch in the second half of next year, but is now weighing options including sixth-generation HBM4 in 8- and 12-layer configurations and HBM4E in an 8-layer configuration.

The shift is believed to stem from Nvidia's assessment that it cannot secure enough HBM4E 12-layer supply to meet demand amid a DRAM shortage. TrendForce said that if Nvidia ultimately decides to downgrade its HBM specifications, it will likely do so by reducing the number of stacked DRAM layers.

Semiconductor equipment makers that developed and unveiled hybrid bonding tools early are waiting for Samsung Electronics, SK hynix and Micron to adopt the equipment and begin mass-producing next-generation HBM.

Mukund Srinivasan, vice president at Applied Materials, the world's largest semiconductor equipment company, addressed the delayed adoption of hybrid bonding at a press briefing in Seoul last month. "When it will be commercialized is a matter of timing — that is for the customer to answer," he said. "As has always been the case at every inflection point with new technology, there is a learning curve that must be experienced. The issues will be overcome."