日本語

NVIDIA Just Made the Agent Stack Two Tiers Deep. Here's the CTO Infrastructure Test for Your Next Platform Renewal

NVIDIA Open Agent Platform launch with 17 enterprise adopters at GTC 2026

The enterprise agent platform shortlist you've been building just got more complicated. Not because a fifth vendor joined the race, but because NVIDIA may not be a platform vendor at all. It may be the layer the other four sit on.

At GTC 2026, NVIDIA announced the nvidia open agent platform alongside 17 publicly named enterprise adopters. That list includes Salesforce, SAP, ServiceNow, Adobe, Cisco, Atlassian, Red Hat, and Palantir. When that many strategic vendors align on the same runtime in a single announcement, the question stops being "which platform should I pick?" and starts being "what exactly am I picking when I pick a platform?"

For CTOs who have been tracking the enterprise agent platform race across Microsoft Agent 365, Anthropic's self-hosted sandboxes, and Google's Antigravity 2.0, this announcement reframes the whole evaluation. Those pieces were about platforms. This one is about the infrastructure layer that may sit under all of them.

Why 17 Strategic Vendors Picking the Same Runtime Is the Story

According to NVIDIA's GTC 2026 announcement, NVIDIA launched three open components that form the core of the platform: Nemotron (open models tuned for agentic reasoning), AI-Q (a blueprint for building enterprise-knowledge agents that perceive, reason, and act), and OpenShell (an open-source runtime that enforces policy-based controls for security, network, and privacy). The optimization skill library is called cuOpt.

But the component list is not the story. The story is the 17 adopters who announced alongside it.

Adobe is using the platform for hybrid long-running creative agents. Salesforce embedded it in Agentforce. Amdocs built what it calls a Cognitive Core on top of NVIDIA AI-Q and Nemotron, using it to monitor customer interactions and billing data. These aren't pilot programs or exploratory proofs of concept. They're production-grade architectural choices by organizations that ship enterprise software to millions of users.

When Salesforce builds Agentforce on top of NVIDIA's runtime, that runtime becomes part of the Agentforce decision for every Salesforce customer. The same logic applies to SAP, ServiceNow, and Cisco. You may not be evaluating NVIDIA directly. But your vendors already did.

Key Facts

  • NVIDIA's Open Agent Platform launched at GTC 2026 with three open components: Nemotron (models), AI-Q (blueprint), OpenShell (runtime) (NVIDIA, March 2026)
  • 17 enterprise launch adopters publicly named, including Adobe, Salesforce, SAP, ServiceNow, Atlassian, Palantir, Cisco, and Red Hat (NVIDIA Newsroom)
  • Bain frames the launch as "AI becomes the operating layer," with NVIDIA positioning beneath application-layer platforms rather than competing with them (Bain & Co. analysis, GTC 2026)

Futurum's analysis framed this clearly: NVIDIA is staking a claim on autonomous-agent infrastructure, not the application layer. The Bain & Co. GTC 2026 analysis described the moment as "AI becomes the operating layer." That framing is precise. NVIDIA isn't competing with Microsoft or Salesforce for the agent platform budget line. It's competing for the infrastructure budget line. And a significant chunk of the enterprise software market just told you which side they came down on.

Nemotron, AI-Q, and OpenShell: What CTOs Actually Need to Understand About the Three Open Components

The two-tier agent stack: NVIDIA infrastructure under Microsoft, Google, OpenAI, and Anthropic platforms

Most CTO briefings on NVIDIA's platform will lead with Nemotron because it's the model, and models are easy to demo. Don't let the demo determine your evaluation. The three components do different jobs, and the one that matters most for infrastructure decisions is OpenShell.

Nemotron is a family of open models tuned specifically for agentic reasoning. The distinction from general-purpose large language models matters. An agentic model needs to decide when to call a tool, how to decompose a task into sub-steps, and when to stop and ask for human input. Nemotron is trained on that problem specifically. For CTOs building agents on top of vendor platforms like Salesforce or SAP, Nemotron may never be visible. But for teams building custom agents or evaluating whether vendor platforms are using capable underlying models, it's a relevant data point.

AI-Q is a blueprint rather than a product. It defines an architecture for enterprise-knowledge agents that can perceive (ingest structured and unstructured enterprise data), reason (apply agentic logic to that data), and act (trigger downstream workflows or surface outputs). The blueprint pattern matters because it gives the 17 adopters a shared reference architecture. When Amdocs builds its Cognitive Core and Salesforce builds Agentforce on the same AI-Q blueprint, the resulting agents share assumptions about data access, state management, and action scope. That interoperability is a quiet structural advantage.

OpenShell is where the infrastructure argument lives. It's an open-source runtime that enforces policy-based controls for security, network boundaries, and privacy. For a CTO, this is the component that answers the question: "If I build on NVIDIA's infrastructure, who controls the guardrails?" The answer OpenShell gives is: you do, because it's open-source and you can audit, modify, and enforce it. That's a different answer than you get from a closed proprietary runtime baked into a vendor's cloud platform.

The distinction matters most at enterprise scale, where compliance teams need to inspect what any given runtime will and won't let an agent do.

The Two-Tier Question: Platform Versus Infrastructure

The standard CTO evaluation framework for agent platforms treats the decision as a horizontal comparison: Microsoft Agent 365 versus Google Gemini Enterprise versus OpenAI Workspace Agents versus Anthropic Claude Managed Agents. Pick one, or pick a primary with a backup. That framework was reasonable through early 2026.

NVIDIA's GTC announcement makes that framework obsolete.

The right mental model now is vertical, not horizontal. There's a platform tier and an infrastructure tier. The platform tier is where Microsoft, Google, OpenAI, and Anthropic compete. The infrastructure tier is where NVIDIA is placing its bet. When Salesforce embeds NVIDIA's runtime into Agentforce, the two tiers are stacked. Salesforce is your platform decision. NVIDIA is (potentially) your infrastructure decision. They're not alternatives. They're layers.

This matters because the procurement process, the evaluation criteria, the team responsible, and the renewal timeline are all different for a platform decision versus an infrastructure decision.

A platform decision is an application-layer decision. You're choosing a user-facing surface, an integration layer, a control plane, and a distribution mechanism. The buyer is typically the VP of engineering, the head of IT, or the CTO directly. The evaluation criteria are user experience, integration coverage, cost per seat, and vendor support.

An infrastructure decision is a lower-level architectural choice. You're choosing what the agents run on: the model family, the runtime guardrails, the blueprint assumptions baked into the agent's reasoning architecture. The buyer is an architect or a platform team lead. The evaluation criteria are openness, auditability, composability, and long-term dependency risk.

Mixing these two decisions in a single procurement process produces bad outcomes. You end up optimizing on user experience metrics for a decision that should be evaluated on runtime security. Or you spend engineering capacity auditing the model layer of a product you'll never directly configure.

Re-Architecting Procurement: Why the Platform Decision and Infrastructure Decision Should Be Separate

The practical implication of the two-tier model is a change in how procurement is organized. Right now, most enterprise technology procurement treats AI agents as a single budget line in a single evaluation process. One RFP, one vendor shortlist, one buying committee.

That structure made sense when the stack had one tier. It doesn't make sense now.

A two-tier procurement process looks different. The platform evaluation (which vendor's surface, control plane, and integrations serve your business units best) happens at the application layer. The infrastructure evaluation (what runtime is your platform built on, what are its guardrail assumptions, how auditable is it, and do you accept that dependency) happens separately, with different people asking different questions.

In practice, many organizations will discover their infrastructure decision has already been made for them. If you've committed to Salesforce's Agentforce, and Salesforce has committed to NVIDIA's runtime, your infrastructure decision is NVIDIA by default. That's not necessarily a problem. But it should be a conscious choice documented in your architecture log, not an invisible assumption buried in a vendor's product roadmap.

The teams thinking about AI pattern governance and buy-versus-build decisions by AI pattern will recognize the dynamic: decisions made at the platform layer carry architectural assumptions that propagate down through the stack. The NVIDIA announcement makes those assumptions visible for the first time.

The CTO Infrastructure Test (5 Questions to Run Before Next Renewal)

Before your next platform contract renewal, or before you sign a new one, run this test on the infrastructure layer. These questions are designed to be vendor-neutral. Apply them to any platform that uses an agent runtime underneath.

1. What runtime is this platform built on, and is that runtime open-source? If the runtime is proprietary and closed, you can't audit what it will and won't allow an agent to do. For regulated industries or high-security environments, this is a disqualifying constraint. OpenShell's open-source stance is a deliberate answer to this question.

2. Who controls the model layer, and can you swap it? Platforms locked to a single model family create dependency risk. If the vendor changes model pricing, deprecates a model version, or changes fine-tuning policies, your agents change with it. An infrastructure layer like Nemotron, which is open and separable, gives you the ability to pin a model version or switch providers without rebuilding the application layer.

3. Can your compliance team inspect the guardrail logic? Policy-based guardrails are only credible if they're reviewable. "Trust us, the guardrails work" is not an acceptable answer for a system acting on customer data or billing records. Demand documentation of what the runtime will block, what it will log, and what it will allow.

4. If this infrastructure vendor changes its pricing or terms, what is your switching cost? The openness of the components matters here. An open-source runtime has a switching cost. A proprietary runtime has a much higher one. Price your infrastructure dependency risk accordingly.

5. Is your platform vendor's roadmap locked to a single infrastructure provider, or do they support multiple? A platform vendor committed to a single infrastructure provider transfers that dependency to you. A platform vendor that supports multiple infrastructure options (or that exposes an abstraction layer over the runtime) preserves your optionality. This is a question to put directly to your platform vendor in the next QBR.

What to Do This Week

The NVIDIA announcement doesn't require immediate action. But it does require a mental model update and one concrete item added to your next architecture review.

Update your mental model. Stop thinking of the agent platform decision as a horizontal comparison between four vendors. Start thinking of it as two separate decisions stacked vertically: platform layer (Microsoft, Google, OpenAI, Anthropic) and infrastructure layer (increasingly NVIDIA, but potentially others). The Gartner AI coding agents realignment coverage is useful context for how the stack is evolving.

Audit your current platform commitments for hidden infrastructure assumptions. If you've already chosen a platform, ask your vendor which runtime underlies their agent execution environment. If the answer is NVIDIA's stack, you've already made an infrastructure choice. Document it and evaluate whether you'd have made the same choice consciously.

Add the five-question infrastructure test to your next platform renewal review. Platform contracts typically have 12-to-24-month renewal cycles. The infrastructure test above can be completed in a half-day workshop with your architecture team before the next renewal conversation. Start with the renewal coming up soonest.

And if you haven't yet read the prior pieces in this series, the Microsoft Agent 365 control plane analysis and the Google Antigravity 2.0 platform comparison are the natural complements to this one. Those pieces evaluated the platform tier. This one evaluated the infrastructure tier underneath it.

Frequently Asked Questions

What is NVIDIA's Open Agent Platform?

NVIDIA's Open Agent Platform is a set of three open components launched at GTC 2026: Nemotron (open models tuned for agentic reasoning), AI-Q (a blueprint architecture for enterprise-knowledge agents), and OpenShell (an open-source runtime that enforces policy-based security, network, and privacy guardrails). It's not an end-user application. It's the infrastructure layer that enterprise software vendors embed into their own agent platforms. At launch, 17 enterprise vendors including Salesforce, SAP, ServiceNow, Adobe, and Cisco publicly announced adoption.

How is NVIDIA different from Microsoft Agent 365, Google Gemini Enterprise, and OpenAI Workspace Agents?

Those three are application-layer platforms. They give enterprise users a surface, a control plane, and a set of integrations for deploying and governing agents inside their organizations. NVIDIA is an infrastructure-layer provider. It supplies the runtime, model, and blueprint that those platforms (and the software they run on) may be built on top of. When Salesforce uses NVIDIA's runtime inside Agentforce, NVIDIA and Salesforce are not competitors. They're two tiers of the same stack.

Should a CTO evaluate NVIDIA separately from their agent platform selection?

Yes, but the evaluation looks different. A platform evaluation focuses on user experience, integrations, control plane depth, and vendor support. An infrastructure evaluation focuses on openness, auditability of guardrails, model portability, and dependency risk. In many cases, the infrastructure choice will be made implicitly through the platform choice. The job of the CTO is to make that implicit choice explicit and to run the five-question infrastructure test before signing any platform contract.

Source: NVIDIA Open Agent Platform announcement. Coverage from VentureBeat, Futurum, and Bain & Co..