Who NVIDIA buys from and sells to
Designs the GPUs, networking and rack systems at the centre of AI data-centre spending; owns no fabs.
Part of AI Capex spending
Buys from 6
Sells to 6
- Foxconnassembles GB-series racks for hyperscalerstwo direct customers 22% and 14% of revenue (FY2026 10-K)
- MicrosoftMSFTAzure accelerator capacity
- MetaMETAtraining clusters for its own models
- AmazonAMZNAWS instances, alongside its own silicon
- AlphabetGOOGLCloud capacity, beside its own TPU
- OracleORCLOCI capacity, bought largely on debt
Where the chain could break
- Every leading-edge die and CoWoS package comes from TSMC sites in Taiwan
- Two direct customers were 36% of FY2026 revenue; top hyperscalers buy about half
- US export licences decide whether it can sell data-centre GPUs into China
Competes with
- AMDAMDthe only other merchant GPU at scale
- BroadcomAVGOcustom accelerators that displace merchant parts
- AlphabetGOOGLTPUs rented out on Google Cloud
Mapped by hand from filings and company reporting, as of October 2026. It shows the relationships that matter most, not every contract: most are never disclosed, and agreements change.