Company maps
NV
silicon dossier
NVIDIA
NVIDIA remains the default toll road for frontier AI until networking, software, and supply-chain integration become modular enough to unbundle.
Confidence
78%
The company sells more than chips: it sells a working scale-up system, a software moat, and a supply-chain coordination layer.
Hypotheses
NVLink-scale systems keep NVIDIA advantaged even as individual accelerator competition improves.
The main threat is not one better chip; it is a credible full-stack alternative with supply and software maturity.
Open questions
Do hyperscaler custom chips cap NVIDIA margins or merely segment the market?
Does inference shift value from GPUs toward memory, networking, or specialized accelerators?
What would change our mind
A hyperscaler proves a materially cheaper full-stack alternative at frontier scale.
Inference demand shifts decisively to architectures where NVIDIA software matters less.
Recent evidence
Frontier labs and hyperscalers continue building around NVIDIA scale-up domains.
Memory bandwidth, networking, and rack-level power are becoming as strategic as raw FLOPS.
Linked themes
scale-up networkingHBMinference economicsdata centers