NVIDIA and OpenAI Unveil GPT-5.6 Running on GB200 NVL72 Rack-Scale Systems for Enterprise AI

For anyone who has used ChatGPT at work to draft a document, debug some code, or pull together a report, the latest announcements from NVIDIA and OpenAI matter in a practical way. The two companies are deepening a partnership that shapes how fast, how affordable, and how capable those tools are — and the next generation of that infrastructure is now being put to work.

NVIDIA posted on social media this week GPT-5.6, OpenAI’s newest frontier model family, is trained and run on NVIDIA’s GB200 NVL72 rack-scale systems. The post also references GB300 NVL72 systems, though that hardware has not yet been independently confirmed in official technical documentation and should be treated as unverified at this stage. The announcement positions NVIDIA’s own workforce as early adopters, with teams across the company already using ChatGPT Work — OpenAI’s work-focused product environment — powered by these models.

What Is GPT-5.6 and Why Does It Matter?

GPT-5.6 is not a single model. OpenAI has structured it as a family of three tiers, each aimed at different use cases and budgets. Sol is the flagship — OpenAI describes it as its most capable option and its strongest cybersecurity model to date. Terra sits in the middle, balancing performance with lower running costs. Luna is the fastest and most affordable of the three, suited to high-volume, cost-sensitive tasks.

All three are described by OpenAI as being built for coding, enterprise work, scientific research, cybersecurity, and what the company calls agentic workflows — meaning AI that can plan and carry out multi-step tasks with limited human intervention. OpenAI has also classified all three variants as “High capability” in cybersecurity and biological and chemical risk domains under its internal Preparedness Framework. That is not a statutory regulatory label, but it does signal that OpenAI itself regards these models as powerful enough to warrant careful oversight.

GPT-5.6 is available across ChatGPT, Codex, and the OpenAI API, with a global rollout planned shortly after the announcement.

NVIDIA’s Role: The Hardware Behind the Headlines

The GB200 NVL72 is NVIDIA’s current rack-scale AI system, designed to handle the enormous computational demands of running large language models at enterprise scale. NVIDIA claims the GB200 NVL72 delivers around 35 times lower cost per million tokens and about 50 times higher token output per second per megawatt compared with previous-generation systems. Those are vendor-reported figures rather than independently audited results, but they give a sense of the efficiency leap NVIDIA is claiming.

At the same time, the partnership between NVIDIA and OpenAI goes back to 2016. NVIDIA describes it as a “full-stack” collaboration — covering GPUs, rack-scale systems, networking, and the software tooling used to train and deploy frontier models. It is, in other words, not just a hardware supply arrangement but a deep technical integration between two of the most influential companies in AI.

10,000 NVIDIA Staff Already Using Codex

One of the more striking details in NVIDIA’s announcement is the internal adoption figure. According to NVIDIA’s own data, over 10,000 of its employees — across engineering, legal, marketing, finance, HR, sales, and operations — are already using GPT-5.5-powered Codex to automate workflows, speed up debugging, and improve how quickly features are delivered.

GPT-5.5 is the model that currently powers Codex in production. GPT-5.6 builds on it, and NVIDIA teams are described as early adopters of ChatGPT Work running on the newer model for operational automation and faster insight generation. The specific cost savings from GPT-5.6 in production have not yet been backed by formal published metrics — those qualitative benefits are consistent with NVIDIA’s earlier claims about GPT-5.5, but independent verification for the newer model is still limited.

Jensen Huang, NVIDIA’s chief executive, has previously described the company’s AI infrastructure ambitions in broad terms. On the OpenAI side, Sam Altman, OpenAI’s chief executive, said of the partnership: “OpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity, and NVIDIA’s infrastructure is central to how we build and deploy the models that make that possible.”

Caution Around the Claims

It is worth being clear about what is confirmed and what is not. The GB200 NVL72 as OpenAI infrastructure is verified. The GB300 NVL72 reference in the tweet is not yet confirmed in official technical documentation. Detailed specifications for GPT-5.6 — including full benchmark results, context window sizes, and exact hardware configurations — remain limited in publicly available documentation. Some information circulating online emerged through early previews and backend logs before full official documentation was published, so a degree of caution is sensible.

Critics have also raised broader questions. Some commentators warn that tighter integration between a handful of US-based firms — NVIDIA controlling the hardware, OpenAI controlling the frontier models — concentrates significant power in a small number of hands and reduces transparency. Others point out that claims about “automation” and “cost reduction” in enterprise AI often lack independent verification and can understate risks around job displacement, especially for junior developers, data analysts, and back-office staff whose work overlaps with what these models can now do.

UK regulators are watching closely. The Information Commissioner’s Office has been clear that organisations using cloud-based AI services must comply with data protection law, and the Department for Science, Innovation and Technology has encouraged responsible AI adoption in line with the principles established at the AI Safety Summit.

What This Means for Kent Residents

For businesses and developers here in Kent using ChatGPT, Codex, or the OpenAI API — whether that is a software start-up in Canterbury, a law firm in Maidstone, or a sole trader using AI to draft correspondence — improvements in the NVIDIA and OpenAI infrastructure stack can translate into faster, cheaper, and more capable tools over time. The University of Kent and other local institutions involved in computer science or cybersecurity research will also be watching GPT-5.6’s capabilities closely, especially Sol’s emphasis on defensive security applications. As with any powerful AI system, UK residents will want to see independent verification of the performance claims and clear answers from regulators on how these tools are governed when used in public services or sectors handling sensitive personal data.