vendor
AMD
NVIDIA
at the branch
claim converts
stays marketing
Power-constrained AI buildout makes efficiency a contested competitive axis
data-center power envelope
capex gated by power + cooling
kW in
AI accelerator rack
power to tokens
accelerator dies
co-packaged optics
liquid cooling loop
perf / watt
power walls make efficiency the lever
AMD claim
38x node-level eff.
20x-by-2030 rack-scale
2020 base, contested
NVIDIA claim
Rubin: 10x token cut
5x photonics, CPO
liquid-cooled, claimed
lower $/token
at rack scale
TCO + design wins
share gain
cooling · CPO · co-design
e.g. OpenAI MI450
stays marketing
no share gain
claim doesn't convert
no design wins