Jul. 5 at 3:29 PM
The
$NVDA 1GW AI factory narrative is ultimately a pricing model debate more than a demand debate.
Jensen’s framing of ~
$100B per GW assumes a GPU-heavy buildout, but the real shift underway is mix compression as inference becomes a larger share of workloads. If inference approaches ~70% of AI compute demand, the architecture naturally tilts toward efficiency-optimized silicon rather than pure high-end GPUs.
That’s where the ASIC ecosystem becomes structurally relevant.
$AVGO,
$GOOG TPU,
$AMZN Trainium, and
$MRVL are not replacing Nvidia in training, but they are capturing incremental share of inference workloads where cost per token matters more than peak performance.
China adds another layer, but remains supply-constrained, particularly around HBM availability, limiting global displacement risk despite growing domestic adoption.
The core dynamic is not “Nvidia vs challengers,” it’s a shifting workload mix inside AI infrastructure where training stays GPU-dominant, but inference becomes increasingly fragmented across custom silicon ecosystems.
That’s where the long-term debate actually lives.
