🧠 Food for thought

Implications, context, and why it matters.

Why a tech giant trusts an AI agent with its internal work

• Nvidia treats Codex as an AI agent that can carry out tasks, not a chatbot that only replies to prompts, said CEO Jensen Huang 1.
• Nvidia says running GPT-5.5-powered Codex on GB200 NVL72 rack-scale systems cuts cost per million tokens by 35x and lifts token output per second per megawatt by 50x versus earlier systems 2.
• For security, Nvidia says it uses zero-data retention and gives staff cloud virtual machines (VMs) 2. The Codex app can connect by remote Secure Shell (SSH) to approved cloud VMs, while agents get read-only access to production systems 2.
• This setup lets Codex handle jobs that run for more than 24 hours. It relies on “compaction,” a feature that compresses context so work stays within memory limits 3.

Beyond Nvidia, Codex is moving into business software

• Nvidia’s rollout serves as a test case for a broader push that moves Codex past its earlier role as a GitHub Copilot-style coding assistant for GitHub users 1.
• OpenAI is repeating the approach with other large companies. Canada-based technology services company CGI says tens of thousands of its engineers, experts and consultants use OpenAI tools including Codex 4. The two companies are also expanding the ways they sell services together 4.
• OpenAI and Nvidia also work together on chip design. Nvidia says OpenAI helps test and shape future chips, while OpenAI gets early access to new system designs 2.
• That work extends a letter of intent, a nonbinding preliminary agreement, for OpenAI to deploy at least 10 gigawatts of Nvidia systems for its next-generation AI infrastructure 5.