OpenAI's new GPT-5.5 powers Codex on NVIDIA infrastructure, and NVIDIA is already using it

As AI agents transform development workflows, OpenAI's GPT-5.5 is running on NVIDIA's GB200 NVL72 rack-scale system through Codex. NVIDIA revealed that more than 10,000 employees are already using this Codex.

The GB200 NVL72 is presented as delivering a 35x reduction in cost per 1 million tokens and a 50x improvement in tokens per second per megawatt compared to the previous generation. As a result, debugging that used to take days has been cut down to hours, and running experiments on complex multi-file codebases overnight has become feasible.

Enterprise deployment was designed around security and auditability.

• Codex supports remote SSH connections to approved cloud VMs, so it doesn't expose actual company data externally.
• NVIDIA IT deployed per-employee cloud VMs to provide agent-only sandboxes.
• It uses a zero-data retention policy, read-only permissions, and the CLI together with Skills.

This rollout is also an extension of over a decade of collaboration between NVIDIA and OpenAI. Since Jensen Huang delivered the first DGX-1 to OpenAI in 2016, the two companies have worked together on the gpt-oss day-zero partnership, the deployment of 10 gigawatts worth of NVIDIA systems, and the joint operation of the first GB200 NVL72 100,000-GPU cluster. In the end, GPT-5.5 is presented as the result of this infrastructure running at full capacity.