Krux
May 14, 2026
NVIDIA Runs ML Experiments 10x Faster With Codex
Published: May 14, 2026 at 12:13 AM
Updated: May 14, 2026 at 12:13 AM
100-word summary
NVIDIA's machine-learning researchers are now completing end-to-end experiments in one-tenth the time using OpenAI's Codex with GPT-5.5, according to a new case study published by OpenAI. Over 10,000 NVIDIA employees have access to the tool, which runs in secure sandboxes with zero data retention and read-only production access. One internal app went from rough prototype to production-ready in hours instead of days. The system runs on NVIDIA's own GB200 hardware, delivering 50x higher token output per watt than prior chips. Translation: the company building the pick-axes is also mining with them, and the results suggest AI coding assistants can handle real research work, not just boilerplate.
What happened
NVIDIA's machine-learning researchers are now completing end-to-end experiments in one-tenth the time using OpenAI's Codex with GPT-5.5, according to a new case study published by OpenAI. Over 10,000 NVIDIA employees have access to the tool, which runs in secure sandboxes with zero data retention and read-only production access. One internal app went from rough prototype to production-ready in hours instead of days. The system runs on NVIDIA's own GB200 hardware, delivering 50x higher token output per watt than prior chips.
Why it matters
Translation: the company building the pick-axes is also mining with them, and the results suggest AI coding assistants can handle real research work, not just boilerplate.