Nvidia Deploys OpenAI Codex To 10,000 Employees

NVIDIA has provisioned OpenAI's agentic coding app Codex, powered by GPT-5.5, to over 10,000 NVIDIAN employees across engineering, product, legal, marketing, finance, sales, HR, operations, and developer programs. The service is hosted on GB200 NVL72 rack-scale systems and, according to NVIDIA, yields substantial cost and efficiency improvements. Jensen Huang framed the rollout as a shift into the "age of AI," and Sam Altman confirmed the expanded internal access.

GPT-5.5 is presented as a step up in agentic capability, matching prior per-token latency while performing more complex, multi-step tasks. NVIDIA reports the GB200 NVL72 deployment achieves 35x lower cost per million tokens and 50x higher token output per second per megawatt relative to previous systems, making frontier-model inference more viable at enterprise scale. Early internal outcomes include faster debugging, shorter experiment cycles, and end-to-end feature delivery from natural language prompts. Key technical claims and operational changes:

• •Codex now operates as an agentic, multi-tool application able to plan, use external tools, and carry tasks to completion using GPT-5.5.
• •Serving on GB200 NVL72 emphasizes power efficiency and throughput improvements for low-latency inference workloads.
• •OpenAI reports token efficiency gains, meaning fewer tokens required per Codex task, which lowers operational cost and extends usable quotas for token-limited products.

This is a significant enterprise validation of two linked trends: the maturation of agentic models for knowledge work, and infrastructure-driven reductions in inference cost. For practitioners, the deployment demonstrates that advanced agentic models are crossing from R&D proofs to company-wide productivity tools. GPT-5.5 claims stronger multi-step reasoning, tool use, and efficiency compared with prior frontier models, which matters for teams building internal developer platforms and CI/CD-integrated assistants. NVIDIA's emphasis on energy and throughput gains underscores an accelerating feedback loop: more efficient hardware lowers the marginal cost of running larger, more capable models, which in turn drives more aggressive internal adoption and tighter hardware-model co-design.

The combination of model capability and optimized serving means organizations can consider broader, lower-risk rollouts of agentic assistants beyond a small developer pilot. Token efficiency and power improvements change cost modeling for internal tooling, making continuous-coding assistants, automated test generation, and multi-file refactors operationally affordable for larger user bases. The reported productivity effects, if reproducible, can compress development timelines across many teams and shift how companies staff and structure engineering work.

Validate claims in real-world pilots, especially around reliability, hallucination modes, and guardrails for legal, finance, and HR workflows. Monitor how OpenAI exposes GPT-5.5 on the API and what additional safeguards or rate-control mechanisms appear for enterprise customers. Also watch competitors and platform partners to see whether similar efficiency gains and enterprise rollouts accelerate industrywide adoption.

The NVIDIA-OpenAI internal deployment is an early, material example of agentic models moving to broad enterprise use. It highlights the importance of hardware-model co-optimization and forces practitioners to reassess operational costs, safety controls, and the scope of automatable knowledge work.