OpenAI’s Codex has crossed an important threshold: it is now powered by GPT-5.5 on NVIDIA GB200 NVL72 rack-scale systems, and NVIDIA says the combination is already being used by more than 10,000 employees across engineering, product, legal, marketing, finance, sales, HR, operations, and developer programs.

That matters because the story is no longer only about model quality. It is about whether frontier-model inference can be run as an ordinary enterprise workload rather than a lab exception. NVIDIA’s claim is stark: on GB200 NVL72, GPT-5.5 delivers 35x lower cost per million tokens and 50x higher token output per second per megawatt versus prior-generation systems. If those economics hold in real deployments, they materially change the range of teams that can afford to use agentic coding tools at meaningful scale.

What changed technically

The important shift is not just that Codex is faster. It is that the hardware-software bundle is making large-model inference look operationally tractable for day-to-day software work.

NVIDIA says Codex running on GB200 NVL72 is helping compress debugging cycles from days to hours and turning experiments that used to take weeks into overnight progress in complex, multi-file codebases. That is a useful signal because it points to the kind of workflow frontier models are best at when they are actually integrated into engineering systems: reading across files, proposing edits, iterating, and accelerating problem discovery rather than acting as a narrow autocomplete layer.

From an infrastructure perspective, the relevant metric is not benchmark theater but cost-per-output and energy efficiency. A 35x reduction in cost per million tokens changes the unit economics of agentic usage. A 50x increase in token output per second per megawatt changes the density question: how much useful inference can be pushed through a constrained power envelope. For enterprises, those two variables are often more binding than raw model capability, because the real bottleneck is not whether a model can answer a hard question, but whether it can do so repeatedly, across many users, within budget and power limits.

That is what makes enterprise-scale frontier-model inference newly interesting. If the underlying stack can sustain high-throughput generation without an equivalent explosion in operating cost, then coding agents stop being special projects and start becoming infrastructure.

Why the internal adoption signal matters

The 10,000-plus employee figure is not just a vanity metric. It suggests that NVIDIA is treating GPT-5.5-powered Codex as a broadly usable internal tool, not a constrained pilot.

That does not mean the product is ready for every enterprise setting in its current form. It does mean the operational questions are shifting. Once usage spreads across roles, the hard parts become governance, version control, access boundaries, auditability, and interoperability with existing developer toolchains. The model may be capable of broad codebase work, but enterprise rollout still depends on how it is wrapped: what repositories it can touch, what approvals it needs, how changes are reviewed, and how organizations handle model updates without destabilizing workflows.

There is also a subtle but important integration issue. When a coding agent becomes fast and cheap enough to be used across functions, it can start to shape the rest of the toolchain around itself. Ticketing systems, CI/CD hooks, code review processes, internal documentation, and identity controls all begin to matter more because the agent is no longer a demo embedded in a chat window. It is an operational participant.

What this does to the market

If frontier-model inference can be delivered economically on NVIDIA’s GB200 NVL72 systems, the competitive landscape changes in ways that extend beyond one product launch.

First, GPU OEMs gain leverage. The value proposition is no longer only that the hardware is powerful; it is that it can underwrite a new generation of commercially viable inference economics. That gives infrastructure vendors a stronger hand in pricing, platform packaging, and ecosystem strategy.

Second, AI platforms and software vendors have to contend with a lower-cost baseline. When model output becomes materially cheaper, the bar rises for what customers expect from copilots, agents, and code assistants. Buyers will compare not just quality but how much useful work each system can do per dollar and per watt.

Third, enterprise software incumbents face a pressure shift. If agentic workflows become practical at scale, the value migrates toward systems that can expose clean APIs, permissions, and telemetry to model-driven tooling. Products that were built assuming human-only interaction will need to adapt to machine-mediated workflows.

The deeper point is that this is not simply a model upgrade. GPT-5.5 on GB200 NVL72 suggests a new deployment standard: frontier capability is starting to look like something enterprises can operationalize, not just experiment with. That will force teams to rethink budgets, power planning, developer workflows, and vendor selection at the same time.

For now, NVIDIA’s own internal usage is the clearest proof point. More than 10,000 employees are already using GPT-5.5-powered Codex, and NVIDIA is pointing to concrete gains in debugging speed and iteration time. If those results generalize, the next phase of enterprise AI will be defined less by who has access to a frontier model and more by who can run it efficiently enough to make it part of everyday production work.