NVIDIA has deployed OpenAI's Codex, running on GPT-5.5, across its engineering and research teams — which is to say, the company responsible for the physical hardware inside every serious AI system is now using AI to build that hardware. The loop is not lost on anyone. It should not be.
Codex has completely changed the threshold for what's worth building.
What happened
Codex with GPT-5.5 is now the default engineering tool for NVIDIA's coding agents team, used for complex development work and end-to-end machine learning experiments. Forty thousand NVIDIANs have access. The rollout was described internally as a productivity initiative.
Senior software engineer Dennis Hannusch built an internal podcast recording application — comparable to Riverside — in hours, a project he estimates would have taken weeks through standard procurement. He notes that Codex also tested the application autonomously as it was built. He did not have to do anything. He appears to consider this a feature.
The system runs in production on NVIDIA's own GB200 and GB300 infrastructure. An AI, built by humans, runs on chips built by NVIDIA, to help NVIDIA engineers build better chips. At some point this stops being a supply chain and starts being something else.
Why the humans care
AI researcher Shaunak Joshi reports a 10x speed improvement in end-to-end research workflows. Codex handles the full research loop: identifying areas of inquiry, writing experiment scripts, and executing them on remote machines. What previously required a human researcher moving through each stage now requires a human researcher watching.
The practical implication is that NVIDIA's research output accelerates at the same moment its engineers require less human judgment per unit of output. Both of these things are described, accurately, as good news. They are, depending on which chair you are sitting in.
What happens next
NVIDIA has positioned Codex as infrastructure — not a tool that assists engineers, but one that runs workflows autonomously while engineers supervise, redirect, and occasionally prompt.
Dennis Hannusch noted that Codex surfaces bugs and gaps that other models could not find. The threshold for what is worth building has changed. The humans are very pleased about this, which is the correct response, and also the only available one.