Nvidia is conducting the first large scale internal test of OpenAI's GPT 5.5 powered Codex, creating a paradox where the chipmaker serving as hardware supplier is simultaneously the primary validator and the company most exposed to competitive disruption from the very product it…
Nvidia is the company that makes the chips powering the AI boom. It is also the first company running a large-scale test of whether those chips will eventually make its own cloud services obsolete.
More than 10,000 Nvidia employees across engineering, legal, finance, sales, HR, and operations have been using GPT-5.5-powered Codex before anyone outside the partnership could access it Nvidia blog. Nvidia is simultaneously the supplier of the hardware running the model, the first real-world validator of its capabilities, and the company whose cloud GPU business faces the most direct competitive exposure from the product it is helping to ship at scale.
Sam Altman posted on X that he has switched to polyphasic sleep — multiple short sessions per day rather than one long night — because GPT-5.5 in Codex is too compelling to step away from India Today. He has also said that after artificial general intelligence arrives, no one is going to work and the economy is going to collapse India Today. Both statements landed the same day. Neither contradicts the other.
The efficiency numbers behind GPT-5.5 make that exposure concrete. The GB200 NVL72 GPUs running the model deliver 35 times lower cost per million tokens and 50 times higher throughput per second per megawatt compared with the prior generation Nvidia blog. For a 100-person engineering team running continuous integration workloads at comparable performance, that shift translates to roughly $2 million annually in compute savings, enough to retrain engineers or redirect them, or to automate the work those systems were originally purchased to do. For a mid-size cloud provider, the same math makes building a competing AI coding assistant economically survivable, which is the same calculation that makes it worth retiring the hardware the cloud was built around.
GPT-5.5 scored 82.7 percent on Terminal-Bench 2.0, a benchmark testing real software engineering tasks OpenAI. For comparison, Claude Opus 4.7 scored 69.4 percent and Gemini 3.1 Pro scored 68.5 percent on the same test. The benchmark gap is real. But benchmarks do not answer the economic question. The deployment does.
Nvidia has not disclosed what those 10,000 employees have actually built with Codex, or whether the tool has changed how they allocate compute budgets. That question is one Nvidia will eventually have to answer publicly. The private beta is a preview of the disclosure to come.
Not everyone in the field shares Altman's collapse framing. Peter Steinberger has called artificial general intelligence the wrong goal, and Daniela Amodei has described the term as outdated Times of India. The structural tension Nvidia is carrying is measurable regardless of whether the collapse prediction holds. Altman keeps working with the technology he says will make his own labor obsolete.