Maps

Competitor NVIDIA vs Custom Silicon

The defining debate in AI compute, and the reason the fund owns both NVIDIA and Marvell Technology.

The Two Camps#

Merchant GPU (NVIDIA): general-purpose, CUDA-locked, best-in-class for training + flexible inference; the default. Moat = software ecosystem + annual cadence + systems integration.

Custom ASIC (hyperscaler in-house): Google TPU, Amazon Trainium/Inferentia, Microsoft Maia, Meta Platforms MTIA, purpose-built, cheaper-per-workload, reduce NVIDIA dependence. Mostly designed with Marvell Technology or Broadcom, fabbed by TSMC.


Where Each Wins#

| Workload | Likely winner | |, -|, -| | Frontier training | NVIDIA (flexibility + ecosystem) | | Internal, stable, high-volume inference | Custom ASIC (cost) | | Enterprise/neocloud/sovereign | NVIDIA (don't have chip teams) | | Hyperscaler internal workloads | Increasingly custom |

The realistic future is both coexist: NVIDIA keeps the majority of merchant value; custom silicon takes a growing slice of inference.


The Fund's Elegant Hedge#

PCA SOF doesn't have to call the winner:

The fund stays long the AI cluster build-out regardless of compute architecture. This is the structural logic of holding Marvell alongside NVIDIA. → Marvell Technology, AI Supply Chain Map.


Linked Notes#

Back to PCA SOF