OpenAI's Latest Frontier Model Now Runs on NVIDIA GB200 NVL72, Slashing Debugging from Days to Hours

OpenAI has introduced GPT-5.5, its most advanced frontier model yet, powering the Codex agentic coding application. The model operates on NVIDIA's GB200 NVL72 rack-scale systems, delivering a 35x reduction in cost per million tokens and a 50x increase in token output per second per megawatt compared to prior-generation hardware. Over 10,000 NVIDIA employees across engineering, product, legal, marketing, finance, sales, HR, and operations are already using this new system, describing the results as 'mind-blowing' and 'life-changing'.

NVIDIA CEO Jensen Huang sent a company-wide email urging adoption: 'Let's jump to lightspeed. Welcome to the age of AI.'

Measurable Gains Across the Enterprise

NVIDIA engineers have had access to GPT-5.5 through Codex for several weeks. The impact is dramatic. Debugging cycles that once spanned days now close in hours. Complex, multi-file code experiments that required weeks of effort can now be completed overnight. Teams are shipping end-to-end features directly from natural-language prompts with significantly improved reliability and fewer wasted cycles.

Background: A Decade-Long Partnership

The GPT-5.5 deployment marks the latest milestone in a partnership that began in 2016, when Jensen Huang personally delivered the first NVIDIA DGX-1 to OpenAI. Over ten years, the collaboration has evolved from hardware support to full-stack AI co-development. NVIDIA now works with every frontier model company, not only to accelerate internal AI agent use but also to help partners build the most cost-efficient and powerful models for the broader market.

What This Means: Enterprise AI Agents at Scale

This rollout demonstrates that frontier-model inference is now economically viable at enterprise scale. Key enablers include the GB200 NVL72's ability to handle massive token loads with low latency and power consumption. For security, NVIDIA IT deployed dedicated cloud virtual machines (VMs) for each employee, ensuring agents run in isolated sandboxes with full auditability. Codex supports remote SSH connections to approved VMs, processing real company data without exposure. A zero-data retention policy governs the system, and agents access production systems with read-only permissions through command-line interfaces and the internal Skills toolkit.

How It Works

• Secure Remote Access: Employees use Codex via a familiar user interface that connects to a cloud VM through SSH.
• Read-Only Permissions: Agents cannot modify production data, mitigating risk.
• Skills Toolkit: The same automation workflows used across NVIDIA are available to each agent.

Impact on Developer Workflows

The gains extend beyond coding. Knowledge work — processing information, solving problems, and generating new ideas — is being transformed. As one NVIDIA engineer put it, 'Experimentation that used to take weeks now happens overnight.' The combination of GPT-5.5's reasoning capabilities and the GB200 NVL72's economics is enabling a new class of AI agents that can operate autonomously on complex, multi-file codebases.

Stay tuned for further updates as NVIDIA and OpenAI continue to push the boundaries of what AI agents can achieve in enterprise environments.