Print desk·2026-08-28·.md
NVIDIA sells the rack. CoreWeave rents the hour.
Same GPU story. Opposite books. One is silicon. The other is the neo-cloud that bills the hour.
NVIDIA is ~88% data center. CoreWeave is ~88% GPU cloud. The customer of one is often the other.
The mix
Hover a slice
- Data center88%—
- Gaming8%—
- Auto / OEM / other4%12%
- GPU cloud—88%
- NVIDIACoreWeave
Data-center GPUs. Gaming is the old book. Auto and OEM are the rest.
Pays: Hyperscalers, neo-clouds, and anyone training or serving models.
Builds: Silicon, systems, CUDA, and the people who keep the stack sold.
GPU hours. Reserved clusters. The customer is a lab or a cloud that ran out.
Pays: AI labs, funds, and anyone who cannot wait on a hyperscaler.
Builds: Infra, facilities, and the people who keep the GPUs billed.
The seat
A systems person who wants to ship the chip should sit NVIDIA. A person who wants to keep GPUs billed — power, cluster, reservation — should sit CoreWeave. Do not treat 'AI infra' as one market. The FAE who keeps the socket is not the person who keeps the cage cold.