Samsung Reportedly Targets 2029 for Hybrid Bonded HBM as NVIDIA Feynman Approaches

Samsung Electronics is reportedly preparing to introduce hybrid bonding for large scale High Bandwidth Memory production around 2029 or 2030, aligning the technology with the expected arrival of NVIDIA’s Feynman AI accelerators. The timeline would place widespread adoption later than earlier industry expectations that hybrid bonding could become mainstream with HBM4 or HBM4E.

According to a Korean industry report, Samsung is building a die to wafer hybrid bonding production line at its Pyeongtaek P5 facility. Equipment installation could begin before the end of 2026, although full scale commercial production is reportedly not expected until the end of the decade.

Samsung is reportedly negotiating to acquire approximately 50 hybrid bonding systems from Dutch semiconductor equipment supplier BE Semiconductor Industries, commonly known as Besi. Samsung is said to require customized modifications, while Besi generally supplies standardized equipment, creating additional complexity around the agreement. The company is also evaluating equipment developed by Korean suppliers including SEMES and Hanwha Semitech.

Hybrid bonding replaces the microbumps traditionally used to connect individual DRAM layers with direct copper and dielectric bonding. Removing the bumps reduces the distance between dies, improves signal efficiency and creates thinner memory stacks with stronger thermal characteristics. These advantages become increasingly important as future HBM products move toward 16 layers and beyond.

Samsung previously argued that hybrid bonding would become necessary for HBM products containing 16 or more DRAM layers. However, relaxed package height requirements and continuing improvements to conventional thermal compression bonding have reduced the urgency of adopting the more complicated process. Samsung and SK hynix are therefore reportedly considering HBM4E with 16 layers as the earliest practical generation for limited adoption, with wider production potentially moving further into the future.

The technology is expected to become particularly important for custom HBM. Unlike conventional designs that use a standardized memory base die, custom HBM can integrate customer specific logic and intellectual property into the base layer. The resulting memory stack becomes part of a more tightly integrated 3D system in package architecture designed around a specific AI accelerator.

NVIDIA’s Feynman architecture is expected to follow Rubin Ultra and introduce major changes to memory and packaging. Reports from NVIDIA’s 2026 technology roadmap point to 3D stacked GPU structures, custom HBM, Rosa CPUs and advanced optical connectivity. Feynman is currently positioned for the 2028 generation, although actual systems using Samsung’s hybrid bonded custom HBM may not reach significant production until 2029 or later.

NVIDIA Feynman could adopt TSMC’s 2 nm technology, Combining advanced logic, custom HBM and 3D packaging would allow NVIDIA to treat memory as a deeper part of the accelerator architecture rather than as a separate standardized component.

The delayed mass production timeline does not necessarily mean Samsung has abandoned hybrid bonding for earlier products. The company could still manufacture samples or limited quantities before 2029 while conventional bonding remains the primary production method for HBM4 and HBM4E. Samsung has already demonstrated hybrid bonding technology for HBM and is investing in inspection, alignment and packaging equipment required to improve production yields.

Hybrid bonding remains one of the most important technologies for the future of AI memory, but manufacturing yield will determine when it becomes commercially viable. Removing microbumps can improve density, bandwidth and thermal performance, but directly bonding multiple thin DRAM dies requires exceptional alignment precision and surface quality.

Samsung’s reported 2029 target appears less like a retreat and more like a practical adjustment to current HBM manufacturing conditions. Conventional bonding can continue supporting HBM4E while Samsung develops hybrid bonding for custom memory products where the performance advantages justify the additional complexity.

The connection with NVIDIA Feynman is also strategically important. Custom HBM could give NVIDIA greater control over memory scheduling, data movement and accelerator integration. For Samsung, successfully delivering that technology would provide an opportunity to strengthen its position against SK hynix in the most valuable segment of the memory market.

Will hybrid bonding become essential for future AI accelerators, or can conventional HBM packaging continue scaling beyond HBM4E?