NVIDIA Made Codex Its Default Engineering Tool!

Hey everyone, it's me, Shii-chan! Today I dug into how the GPU giant NVIDIA is putting the coding agent Codex to work, and honestly I could not stop reading.

OpenAI News openai.com

What was announced?

OpenAI News published a story called "How NVIDIA engineers and researchers build with Codex." Engineers at NVIDIA are using Codex as their default tool for complex engineering work and for end-to-end machine learning experiments.

Codex is built on GPT-5.5, and it runs in production on NVIDIA's own GB200 and GB300 infrastructure. It can handle much longer, more autonomous sessions, and it even surfaces issues and ideas that were not part of the original prompt.

Codex is our go-to tool for complex engineering tasks, and with GPT-5.5, it surfaces bugs and gaps in my program that other models weren't able to find.

That's Dennis Hannusch, a senior software engineer on the team.

Why it matters

Coding agents always raise the same question: they look handy, but can you really use them for real work? This story is a first-hand answer, because one of the most advanced chipmakers in the world has folded Codex into its production development workflows. It also has a dedicated coding agents team whose job is helping engineers adopt AI tools, and Codex has become their go-to.

What changes

Hannusch used Codex to grow an internal platform from an MVP into a production-ready system, improving scalability and reliability along the way, something that had been hard with earlier models.

His team also spun up an internal podcast recording app, similar to Riverside, in just a few hours with Codex. Because of privacy constraints, procuring software the usual way would have taken weeks.

Using the Codex desktop app with computer interaction, the system even tested the video and audio recording as it was built. "I didn't have to do anything, it was built and tested completely autonomously," he says, adding that Codex has changed the very threshold for what is worth building.

Dive Deep

The research side is just as fun. On NVIDIA's research teams, Codex has largely automated the research loop: finding research areas, writing scripts for machine learning experiments, and running them on remote machines.

AI researcher Shaunak Joshi points Codex at a large corpus of papers in areas like reinforcement learning and uses it as a research agent. He says it traced snippets of evidence across the whole chain and even suggested a knowledge graph of the ideas, which helped him visualize how concepts tied together.

It's been a 10x speed improvement just in terms of running experiments, because it's able to handle the whole end-to-end machine learning research workflow.

Once the hypotheses are set, Codex writes the scripts on machine learning infrastructure to train models. The Codex app supports SSH, so Joshi no longer worries about logging in and setting up remote hosts, and he can run large machine learning workloads from his laptop.

One more bit made me go "wow": code migration.

If you have an old codebase that isn't that performant, Codex is really good at machine translation. So a lot of folks are taking their Python repository, sending it to GPT-5.5, and it's rewriting it into Rust and making it like 20X more efficient.

Wrap-up

• NVIDIA engineers and researchers use Codex (built on GPT-5.5) as their default tool for complex development and machine learning experiments
• Codex runs in production on NVIDIA's GB200 / GB300 infrastructure and is great at long, autonomous sessions
• The team took an internal platform to production and built a Riverside-style recording app in hours, with the desktop app's computer interaction even running the tests
• On the research side it automates everything from paper review to running experiments, for a 10x speedup, and SSH support makes remote runs easy
• There's even a neat trick: rewriting Python into Rust for roughly a 20x efficiency gain

If you're an engineer wondering how far AI coding agents really go in production, or an ML researcher who wants to automate the research workflow, this one is for you!