patrick soheili

Google (TPU v8-v10), AWS, Microsoft, Meta, Alibaba, Tencent, and Baidu are maturing custom accelerators, while AMD, Intel, and Arm are moving into rack-level systems to optimize the full data path; this dual push — hyperscaler custom silicon and merchant rack-scale — creates a sustained demand for open interconnect and packaging technology that Nvidia keeps proprietary.

HBM and advanced memory supply is bottlenecked at Samsung, SK Hynix, and Micron; hyperscaler demand is insatiable and concentrated, while new leading-edge memory fabs require years and billions to build, creating a persistent supply-demand imbalance that favors incumbents.

Only Samsung, SK Hynix, and Micron can produce high-bandwidth memory at advanced nodes; all are at full capacity with no near-term relief because building new fabs takes years and billions, creating a persistent supply constraint that underpins pricing power for memory incumbents.

Every major cloud provider (Google, AWS, Microsoft, Meta, Chinese giants) is investing heavily in proprietary AI accelerators (TPU, Trainium, Maia, MTIA, etc.) to reduce Nvidia dependency and tailor compute to their specific inference/training needs; this fragmentation creates demand for open, standards-based interconnect technology.

HBM and advanced memory production is concentrated in Samsung, SK Hynix, and Micron; building new leading-edge fabs takes years and billions, while AI-driven demand is insatiable, creating a multi-year structural undersupply.

Major cloud providers (Google, AWS, Microsoft, Meta, Chinese giants) are developing custom AI accelerators to reduce Nvidia dependency, but require high-performance interconnect IP (like Eliyan's) that Nvidia keeps proprietary, creating a merchant market for chiplet/rack-scale connectivity.