May. 18 at 6:22 PM
Morgan Stanley: We estimate hyperscaler capex to build a 1GW datacenter w/ current-gen
$NVDA GPUs (Blackwell) is up to ~2x the cost of current-gen custom
$AVGO ASICs (
$GOOGL TPU,
$AMZN Tranimum)...but compute power efficiency matters, & this is where Nvidia shines...w/ compute performance/watt 2x-8x ahead of custom ASICs"
---- Note: Custom ASICs vastly beat NVIDIA on compute performance-per-watt for mature, high-volume, predictable workloads (which is why Google runs Gemini primarily on TPUs). NVIDIA wins b/c they are the fast, flexible, ultra-powerful "default" that can run any model seamlessly out of the box.
Not an apples-to-apple comparison: Custom ASICs (TPUs, Trainium) are designed on longer development cycles.NVIDIA operates on an aggressive, annual roadmap (Hopper-Blackwell->Vera Rubin). This chart effectively benchmarks NVIDIA's future roadmaps against current or lagging ASIC generations, skewing the real-world efficiency picture