GPT-5.5 and Codex Are Inside NVIDIA Already. Here's What That Actually Means.
OpenAI's Codex coding agent has been upgraded to run on GPT-5.5, the company's latest model, and it's serving inference on NVIDIA's GB200 NVL72 rack-scale hardware. That last part matters more than it might seem.
Over 10,000 NVIDIA staff across engineering, legal, finance, HR, sales and pretty much every other function you can think of are already using it. Early feedback from inside the company reportedly includes phrases like "mind-blowing" and "life-changing," which is either genuine enthusiasm or the most successful internal change management campaign in Silicon Valley history. Probably some of both.
The practical results being cited are at least plausible. Debugging cycles that previously dragged on for days are apparently closing in hours. Complex multi-file codebases that would have taken weeks to experiment with are turning around overnight. Teams are generating end-to-end features from natural language prompts with fewer broken iterations than earlier models produced.
None of that is independently verified, obviously. But the infrastructure claims are at least quantifiable. NVIDIA says the GB200 NVL72 delivers 35x lower cost per million tokens and 50x higher token throughput per second per megawatt compared to the previous generation. If those figures hold up at scale, they make frontier-model inference genuinely viable for enterprise deployment rather than just aspirationally so.
Jensen Huang sent a company-wide email urging staff to adopt Codex, signing off with "Let's jump to lightspeed. Welcome to the age of AI." Make of that what you will.
Security First, Apparently
Enterprise AI deployments live or die on whether IT will actually sign off on them. NVIDIA's internal rollout took the sensible approach of provisioning dedicated cloud virtual machines for every employee, giving each Codex agent its own isolated sandbox. Agents connect via SSH to these approved VMs, which means company data doesn't wander off anywhere it shouldn't.
Production system access is read-only, governed by a zero-data retention policy, and controlled through the same command-line interfaces and automation tooling NVIDIA already uses internally. Whether this setup scales cleanly across 10,000-plus users without incident remains to be seen, but the architecture is at least thoughtful.
Ten Years of Mutual Dependence
The GPT-5.5 launch and this Codex rollout are the current output of a partnership that started in 2016, when Huang personally delivered the first DGX-1 supercomputer to OpenAI's San Francisco office. That origin story gets trotted out a lot, but the subsequent decade of collaboration is genuinely substantial.
NVIDIA was a day-zero partner on OpenAI's recent open-weight model release, optimising weights for TensorRT-LLM and supporting ecosystem frameworks including vLLM and Ollama. OpenAI has committed to deploying more than 10 gigawatts of NVIDIA infrastructure for next-generation training and inference. The two companies also co-developed the first GB200 NVL72 100,000-GPU cluster, completing multiple large-scale training runs on it and claiming a new benchmark for reliability at that kind of scale.
GPT-5.5 is the product of that cluster running properly. Which means NVIDIA is both the supplier of the hardware and one of the first major enterprise customers running the model on it. Vertical integration has rarely been this literal.