Deploying GPT-5.5 Powered Codex on NVIDIA GB200 NVL72: A Practical Guide for Enterprise AI Agents

By ✦ min read

Overview

AI agents are transforming developer workflows, and the next frontier is knowledge work—processing information, solving complex problems, and driving innovation. OpenAI's Codex, now powered by the cutting-edge GPT-5.5 model on NVIDIA GB200 NVL72 rack-scale systems, enables this transformation. With over 10,000 NVIDIA employees across engineering, product, legal, marketing, finance, sales, HR, operations, and developer programs already using GPT-5.5-powered Codex, the results are measurable: debugging cycles that once took days now close in hours, and experimentation that required weeks turns into overnight progress. This guide provides a detailed, technical walkthrough for deploying GPT-5.5-powered Codex on NVIDIA infrastructure, covering everything from prerequisites to common pitfalls.

Prerequisites

Before beginning, ensure you have the following:

• NVIDIA GB200 NVL72 rack-scale system (or access to equivalent hardware) capable of delivering 35x lower cost per million tokens and 50x higher token output per second per megawatt compared with prior-generation systems.
• OpenAI API credentials with access to GPT-5.5 (the frontier model powering Codex).
• Enterprise network setup with Secure Shell (SSH) connectivity to approved cloud virtual machines.
• Cloud virtual machines (VMs) provisioned for each user, acting as dedicated sandboxes.
• Zero-data retention policy configured in your environment.
• Read-only permissions for production systems accessed through command-line interfaces and agentic toolkits (Skills).

Step-by-Step Instructions

Step 1: Provision Cloud Virtual Machines with NVIDIA GB200 NVL72

To ensure each agent has its dedicated computer, provision cloud VMs on the NVIDIA GB200 NVL72 system. Use the following example command to spin up a VM with optimal GPU allocation:

nvidia-smi invoke --create-vm --gpu-type=A100 --gpu-count=8 --memory=512GB --storage=2TB

Assign each VM to a specific employee for accountability and auditability. Ensure the VMs are in the same network segment as the intended production systems.

Step 2: Configure Remote SSH Connections

Codex relies on SSH for secure remote access. Set up SSH keys and configure the ~/.ssh/config file to point to the provisioned VMs:

Host codex-agent-vm
    HostName 192.168.1.100
    User agent-user
    IdentityFile ~/.ssh/codex_key
    Port 22

Enable agent forwarding if needed, but ensure the connection remains isolated to approved VMs only.

Step 3: Deploy Codex with GPT-5.5

Install the Codex application on each VM. Use the NVIDIA container toolkit to pull the latest Codex image with GPT-5.5 support:

docker pull nvidia/codex:gpt5.5-latest
nvidia-docker run -d --name codex-agent --gpus all -p 8080:8080 nvidia/codex:gpt5.5-latest

Configure the environment variables for API keys and model parameters:

export OPENAI_API_KEY='your-api-key'
export MODEL='gpt-5.5'
export MAX_TOKENS=4096

Step 4: Apply Zero-Data Retention Policy

To comply with enterprise security, enforce a zero-data retention policy. Modify the Codex configuration file (typically /etc/codex/config.yaml) to disable logging and caching:

logging:
  enabled: false
cache:
  type: none
retention:
  policy: zero

Restart the Codex service for changes to take effect.

Step 5: Set Read-Only Permissions for Production Access

Agents access production systems via command-line interfaces and Skills—the agentic toolkit NVIDIA uses for automation. Ensure user accounts used by Codex have read-only permissions. Use the following to verify:

sudo -u codex-user ssh production-server 'echo "test" > /tmp/test.txt'  # Should fail with permission denied

If it succeeds, adjust permissions using sudoers or SSH command restrictions.

Step 6: Run and Monitor Agent Workflows

Start an example agent session using natural-language prompts for code debugging:

codex --query "Debug the following multi-file codebase: /path/to/project/src/ --focus on error handling"

Monitor performance using NVIDIA’s nvidia-smi and codex-specific metrics. Track token cost and throughput:

nvidia-smi --query-gpu=timestamp,utilization.gpu,memory.used --format=csv -l 5

Step 7: Scale Across Teams

To replicate for all employees, as done at NVIDIA (over 10,000 users), create a central management dashboard. Use Kubernetes to orchestrate multiple Codex agents across VMs:

kubectl apply -f codex-deployment.yaml --namespace codex

Common Mistakes

• Neglecting the zero-data retention policy: Forgetting to disable logging can expose sensitive data. Always verify after deployment.
• Using shared VMs without isolation: Each agent needs its own dedicated computer; sharing VMs reduces performance and security.
• Incorrect SSH configuration: Misconfigured SSH keys lead to connection failures. Test with ssh -v before relying on Codex.
• Overlooking token cost optimization: GB200 NVL72 provides 35x lower cost per million tokens, but still monitor usage to avoid unexpected bills.
• Granting write permissions: Read-only access is critical—any write capability can lead to data corruption or unintended changes in production.
• Not testing with small prompts first: Jumping straight to large, complex tasks wastes tokens. Start with simple queries to validate setup.

Summary

Deploying GPT-5.5 powered Codex on NVIDIA GB200 NVL72 enables enterprise-scale AI agents that dramatically reduce debugging and experimentation time. By provisioning dedicated cloud VMs, configuring SSH securely, enforcing zero-data retention, and setting read-only permissions, you replicate the setup that over 10,000 NVIDIA employees use daily. Avoid common mistakes like ignoring retention policies or sharing VMs. As Jensen Huang urged, “Let’s jump to lightspeed. Welcome to the age of AI.”