Nvidia Launches MGX ETL Open Rack, Supporting Mixed Deployment of Third-Party AI Chips

Nvidia launched the MGX ETL open standard rack at GTC 2026. This rack is not only compatible with its own chips but also allows customers to integrate AI chips from other manufacturers, breaking the closed ecosystem and offering more choices.

The Vera Rubin POD system includes five new dedicated rack-level expansion systems, all using the third-generation MGX design. NVL72 is based on the MGX NVL rack, while the other four use MGX ETL racks. Nvidia's own MGX NVL racks support up to 256 chips via Spectrum-X Ethernet or Groq 3 LPU direct connection; third-party chips only support Spectrum-X Ethernet.

Nvidia does not restrict partners from using alternative components. Customers can mix AI XPUs from Google, AMD, etc., with Nvidia's network chips, expanding its network device market and easing monopoly disputes from bundled sales.