OpenAI’s New GPT-5.5 Powers Codex on NVIDIA Infrastructure — and NVIDIA Is Already Putting It to Operate
OpenAI’s latest frontier model, GPT-5.5, is now powering Codex, the company’s agentic coding application, on NVIDIA’s GB200 NVL72 rack-scale systems. The deployment marks a significant step in the evolution of AI-assisted software development, enabling enterprises to run complex coding tasks with improved efficiency and security. NVIDIA confirmed that over 10,000 of its employees across engineering, product, legal, marketing, finance, sales, HR, operations, and developer programs are already using the GPT-5.5-powered Codex tool to achieve measurable gains in workflow speed and reliability.
The integration leverages NVIDIA’s GB200 NVL72 infrastructure, which delivers 35x lower cost per million tokens and 50x higher token output per second per megawatt compared to prior-generation systems. These performance gains make frontier-model inference viable at enterprise scale, according to NVIDIA’s internal benchmarks. Engineers report that debugging cycles once spanning days now close in hours, and experimentation that previously took weeks is progressing overnight in complex, multi-file codebases. Teams are shipping end-to-end features from natural-language prompts with stronger reliability and fewer wasted cycles than earlier models.
Beyond internal use, NVIDIA emphasized that the collaboration with OpenAI reflects more than a decade of full-stack partnership. The relationship began in 2016 when Jensen Huang, NVIDIA’s founder and CEO, hand-delivered the first NVIDIA DGX-1 AI supercomputer to OpenAI’s San Francisco headquarters. Since then, the two companies have co-optimized hardware and software across the AI stack, including work on OpenAI’s gpt-oss open-weight model, where NVIDIA optimized model weights for TensorRT-LLM and supported frameworks like vLLM and Ollama.
As part of the alliance, OpenAI has committed to deploying more than 10 gigawatts of NVIDIA systems for its next-generation AI infrastructure — a buildout that will place millions of NVIDIA GPUs at the foundation of its model training and inference operations for years ahead. The companies are also early silicon and codesign partners, with OpenAI providing feedback that informs NVIDIA’s hardware roadmap in exchange for early access to new architectures. This collaboration produced the joint bring-up of the first GB200 NVL72 100,000-GPU cluster, which completed multiple large-scale training runs and set a new benchmark for system-level reliability at frontier scale.
To ensure secure deployment within enterprise environments, NVIDIA IT implemented cloud virtual machines (VMs) for every employee running Codex. Each agent operates in a dedicated sandbox with access to real company data through approved remote Secure Shell (SSH) connections, preventing external exposure. A zero-data retention policy governs the system, and agents interact with production systems using read-only permissions via command-line interfaces and Skills — the same agentic toolkit NVIDIA uses for internal automation workflows. Users control the agent through a familiar interface, maintaining auditability without sacrificing performance.
Jensen Huang encouraged company-wide adoption in a recent internal email, stating: “Let’s jump to lightspeed. Welcome to the age of AI.” The message underscores NVIDIA’s strategy of not only accelerating AI agent use internally but also helping partners build the world’s best, lowest-cost, and most power-efficient models for broader enterprise adoption.
As AI agents move beyond developer workflows into broader knowledge work — processing information, solving complex problems, generating ideas, and driving innovation — tools like GPT-5.5-powered Codex represent a shift toward AI-augmented enterprise productivity. With verified performance gains and security controls in place, the NVIDIA-OpenAI collaboration illustrates how infrastructure optimization and model advancement can combine to deliver tangible results at scale.