Nvidia rolls GPT‑5.5 into teams

Nvidia said more than 10,000 employees are now using OpenAI’s Codex, powered by GPT‑5.5, across engineering, legal, marketing, finance, sales, human resources and operations. (blogs.nvidia.com) Codex is OpenAI’s coding agent, a tool that can write, inspect and debug software instead of only answering chat prompts. Nvidia said staff use it to ship features, analyze codebases and cut debugging cycles from days to hours. (blogs.nvidia.com) (techspot.com) Nvidia said the deployment runs on its GB200 NVL72 Blackwell systems, which deliver 35x lower cost per million tokens and 50x higher token output per second per megawatt than prior GPT‑4o-class infrastructure. (blogs.nvidia.com) (techspot.com) The company built the rollout around isolation rather than broad system access. Nvidia said each employee gets a dedicated cloud virtual machine, while Codex agents connect over approved Secure Shell sessions and keep read-only access to production systems. (blogs.nvidia.com) (techinasia.com) Nvidia also said the deployment uses zero-data retention, meaning prompts and outputs are not kept for model training or long-term storage in the standard way. That setup is meant to let employees work with internal company data without exposing it outside approved environments. (blogs.nvidia.com) The rollout extends a relationship between Nvidia and OpenAI that Nvidia dates to 2016, when Jensen Huang delivered the first DGX‑1 system to OpenAI’s San Francisco office. Nvidia said the two companies have since worked together across training systems, inference hardware and software. (blogs.nvidia.com) OpenAI and Nvidia are presenting the deployment as a model for agentic artificial intelligence inside large companies: wide employee access, but narrow permissions, auditable actions and separate execution environments. Nvidia said employees control the cloud-based agent through a familiar internal interface rather than direct unrestricted system access. (blogs.nvidia.com) The pitch is that coding agents become easier to approve when the tool can act, but only inside a sandbox. Nvidia’s example is less about a chatbot for questions than a locked-down worker for specific tasks. (blogs.nvidia.com)