Nvidia has officially announced that ‘Feynman’, the successor to its next-generation Artificial Intelligence accelerator ‘Rubin’, will be equipped with custom High-Bandwidth Memory .

This has led to projections that the HBM market will now enter a full-scale foundry (semiconductor contract manufacturing) battle.

Jensen Huang, CEO of Nvidia, briefly mentioned the next-next-generation AI accelerator ‘Feynman’ during the opening keynote of GTC 2026 on March 16 (local time).

As the name ‘Feynman’ was first introduced at GTC 2025 last year, it was anticipated that detailed specifications would be revealed during this year’s keynote. However, there were no specific mentions.

However, presentation materials indicated that Feynman will be equipped with custom HBM. While it was stated last year that ‘next-generation HBM’ would be included, this year’s announcement clarified it as custom HBM.

Custom HBM is a product with performance and power efficiency optimized to meet customer demands. Its key feature, unlike general HBM, is the inclusion of logic (computation) functions on the base die that controls data.

The base die, which serves as the brain of HBM, connects and communicates with various processors such as GPUs. Samsung Electronics and SK hynix are manufacturing the base die for their 6th-generation product, HBM4, using advanced foundry processes rather than conventional DRAM processes.

This approach offers the advantage of not only improving power efficiency and performance but also providing customized products tailored to the needs of each client. With global big tech clients like Nvidia and Google demanding custom base dies optimized for their AI processors, it is projected that foundry capabilities will become increasingly crucial in HBM manufacturing.

Designing a custom base die that a client desires requires memory engineers and foundry engineers to collaborate from the very beginning. It necessitates the establishment of a complex process that considers everything from the design stage to mass production and packaging.

Samsung Electronics expects its strength as an ‘Integrated Device Manufacturer (IDM)’, which covers all business areas from memory and foundry to logic design and advanced packaging, to contribute to expanding its position in the HBM market. The company’s strategy is to enhance the quality of its next-generation HBM by maximizing synergy between its memory and foundry businesses.

In the case of HBM4, the base die was manufactured using Samsung Electronics’ own 4-nanometer foundry process. As a result, it passed quality tests in a single attempt without any redesign requests from Nvidia, becoming the first among the three memory companies to begin shipments. For HBM5, development is underway applying a 2-nanometer process to the base die.

While Samsung Electronics has focused on mobilizing its total internal capabilities, SK hynix has chosen to collaborate with Taiwanese foundry TSMC. By partnering with the world’s No. 1 foundry, TSMC, SK hynix can expect to secure TSMC’s clients or see an expansion in HBM orders.

The company has also moved to secure its own talent. In January, the inclusion of base die design personnel in its recruitment for experienced professionals drew attention. As custom HBM is expected to become a key battleground, SK hynix’s hiring move was interpreted as a strategic step in preparation for this.

Meanwhile, when CEO Huang was asked on this day about ‘Feynman’, the next-next-generation successor to the AI accelerator ‘Rubin’, he replied, “If I answer all of that question, there will be no reason for you to come to GTC next year. I’ll save it for next year.” This suggests that Feynman, scheduled for release in 2028, is expected to be in the spotlight at next year’s GTC.

Source:Business Korea