NVIDIA’s agent stack raises hardware and governance flags

NVIDIA is no longer just selling the engines for artificial intelligence. At its March 2026 GTC event, it rolled out an “agent” stack that spans models, developer tools, a runtime called OpenShell, and optimization software, with Adobe, Salesforce, SAP, ServiceNow, Siemens, Cisco, and others already lined up to use it. (nvidianews.nvidia.com) An artificial intelligence agent is software that does a chain of jobs instead of answering one prompt. NVIDIA’s pitch is that one vendor can now supply the model, the toolchain, the runtime that executes actions, and the infrastructure layer underneath it. (developer.nvidia.com) (docs.nvidia.com) The new runtime matters because agents are riskier than chatbots. NVIDIA says OpenShell sits between the agent and the company’s systems, deciding what the agent can see, what commands it can run, and where its inference is sent, while using sandboxing and kernel-level isolation to keep it off the host machine. (docs.nvidia.com) (developer.nvidia.com) Guardrails are the rulebook for those agents. NVIDIA’s own NeMo Guardrails product filters inputs and outputs with configurable policies, which means the company is not only supplying chips and software but also a growing share of the policy machinery enterprises may rely on in production. (developer.nvidia.com) (docs.nvidia.com) That is where the governance question starts. If a company builds on NVIDIA models, NVIDIA runtime controls, NVIDIA guardrails, and NVIDIA enterprise deployment tools, the practical answer to “who sets the rules” can drift from the customer’s security team toward the defaults, update cycles, and interfaces of one vendor’s stack. (docs.nvidia.com 1) (docs.nvidia.com 2) The hardware question sits underneath all of this. Modern NVIDIA artificial intelligence systems depend heavily on advanced packaging called Chip on Wafer on Substrate, or CoWoS, which Taiwan Semiconductor Manufacturing Company describes as the packaging platform for high-performance computing chips with large interposers and multiple high-bandwidth memory stacks. (tsmc.com) (3dfabric.tsmc.com) That packaging step has been one of the tightest bottlenecks in artificial intelligence hardware. Industry reports in 2025 and 2026 said NVIDIA had locked up more than half of global CoWoS capacity, with estimates around 60% for 2026, which is why critics worry that every new NVIDIA-led software layer can reinforce demand for the same already-constrained supply chain. (astutegroup.com) (financialcontent.com) NVIDIA is also tying in cuOpt, its optimization engine for routing and scheduling problems with millions of variables and constraints. That makes the stack broader than a chatbot toolkit, because the same platform can plan delivery routes, allocate resources, and run agents that act on those plans. (docs.nvidia.com) (nvidia.com) The attraction for big software companies is obvious. NVIDIA says its open AI-Q blueprint can mix frontier and open models and cut some query costs in half, while partners like Adobe, Atlassian, Box, Cadence, CrowdStrike, Red Hat, Salesforce, Siemens, and Synopsys get a ready-made path to ship agent features faster. (nvidianews.nvidia.com) The risk is that enterprises may think they are buying a toolkit and end up standardizing on a full operating system for agentic work. Once the same vendor supplies the compute, the runtime, the optimization layer, the guardrails, and many of the reference models, switching costs stop looking like software migration and start looking like rebuilding a factory. (developer.nvidia.com) (docs.nvidia.com) So the argument around NVIDIA’s agent stack is not whether the pieces are real. The argument is whether a company that already dominates the scarce hardware layer should also become the default place where enterprises define what their agents are allowed to do. (nvidianews.nvidia.com) (docs.nvidia.com)