Broadcom is serious about AI chips – next year it will become one of the main consumers of HBM3E from SK hynix

Nvidia and SK hynix are now closely linked in the field of advanced HBM3E memory, since the former is the latter’s dominant client in this area, and this state of affairs allows the Korean manufacturer to lead in a dynamically developing market segment. Rumor has it that Broadcom will also be a major recipient of SK hynix’s HBM3E next year.

Image Source: SK Hynix

As The Elec notes, Sk hynix will begin supplying modern versions of HBM for Broadcom’s needs in the second half of next year. The initiator of this relationship was Broadcom. It is generally accepted that it is developing specialized accelerators for artificial intelligence systems for the needs of Apple or OpenAI. Unlike Nvidia’s solutions, they will rely on the use of specialized neural units, and therefore will be of limited use.

If initially SK hynix expected to be able to produce from 140 to 150 thousand silicon wafers with HBM chips in a standard size of 300 mm by next year, then taking into account the needs of Broadcom, it will be forced to increase production volumes to 160 or 170 thousand silicon wafers. Based on the results of the fourth quarter, the company expects that HBM will generate 40% of SK hynix’s total revenue in the DRAM segment. As Broadcom joins the ranks of clients, this share may increase.

Broadcom earlier this month said it had developed a new chip packaging technology that allows for the integration of HBM-type memory chips. Even if the latter is not useful to third-party clients like OpenAI and Apple, Broadcom can launch the production of server processors with such memory, similar to Fujitsu solutions.