NVIDIA Unveils Rubin CPX GPU for Million-Token AI Workloads

September 9, 2025

Details

  • NVIDIA introduced the Rubin CPX GPU on September 10, 2025, designed specifically for AI models requiring million-token context windows.
  • This is described as a “new class of GPU,” featuring architecture shifts surpassing NVIDIA's existing Blackwell and Hopper lines.
  • NVIDIA claims enterprises could generate up to $5 billion in token-based revenue for every $100 million invested in Rubin CPX hardware and deployment.
  • Million-token inference powers applications like entire movie subtitling, comprehensive code-base analysis, and reviewing full contracts end-to-end without data chunking.
  • The GPU targets cloud providers and datacenters seeking to capitalize on large-context language models, offering lower latency and higher throughput than current GPUs.

Impact

With Rubin CPX, NVIDIA challenges competitors like AMD’s MI300X and Google’s TPU v5e by enabling far larger context windows. The revenue-driven hardware pitch recasts GPU buying as a profit-generating investment, potentially shifting procurement strategies industry-wide. Emphasizing memory and context length over raw compute marks a strategic pivot that could reshape hardware design and enterprise AI adoption in the coming years.

Read more