We came away from this year’s Nvidia GPU Technology Conference struck by how far the conversation has moved. A year ago, the headline was that AI had graduated from pilots to real enterprise deployment. That’s now table stakes. Leaders of this wave aren’t just deploying AI in their businesses—they’re rebuilding their operating models around it.

Even a few months ago, companies were asking how to get AI to work reliably at scale. Today’s questions are harder: How do we govern autonomous agents operating across thousands of business units? How do we keep our security posture intact when AI systems are making decisions on their own? How do we make sure our system design and data infrastructure can actually support the speed at which these platforms need to operate?

Several themes stood out that really matter for enterprise leaders:

• The cost of using AI is dropping fast, the pace of model development is accelerating, and every business needs a strategy for engaging with it.
• Agentic AI has moved from concept to platform, and governance is the central challenge for enterprises trying to scale it.
• Security, system design, and data infrastructure need to evolve in lockstep with AI adoption—not as separate, slower-moving workstreams.
• The build vs. buy calculus for AI is getting more nuanced; model-agnostic platforms are durable investments.
• Physical AI is crossing into real deployment timelines in automotive, robotics, and healthcare.

The economics of AI are shifting fast

The cost of using AI is falling sharply. Nvidia’s Vera Rubin platform, the successor to Blackwell now in production, delivers up to 10 times higher inference throughput per watt and one-tenth the cost per token. At the same time, the pace of frontier model development keeps accelerating: Models are getting more capable on shorter release cycles, and the tools for building on top of them are becoming dramatically more accessible.

For most enterprises, this isn’t a story about training your own models. It’s a story about the increasingly wide range of what’s now economically viable to do with AI. Use cases in real-time decision making, customer interaction, and operational automation that didn’t pencil out six months ago deserve a fresh look. Nvidia raised its AI chip demand outlook to $1 trillion through 2027—and the growth is driven overwhelmingly by inference workloads, not training.

But cheaper inference has a second-order effect that’s easy to miss: As the cost per AI action drops, the volume of AI usage across your business explodes. Your governance, data quality, and security need to be ready.

Agentic AI has a platform now

At GTC last year, production-level agentic AI was just starting to gain traction. This year, Nvidia announced its Agent Toolkit, anchored by NemoClaw—an enterprise reference design built on top of OpenClaw, the open-source agentic framework that exploded in popularity in early 2026. NemoClaw layers runtime sandboxing, privacy routing, and network guardrails on top of OpenClaw’s flexibility, giving enterprises a controlled operating environment for agents without locking them into a closed ecosystem.

A recurring pattern in our client work is companies struggling to scale up their agent pilots. It’s a much tougher problem, running agents across hundreds or thousands of locations, business units, or customer touchpoints. Companies often have the technical capability but lack a reliable platform with strong safety guardrails that can execute at the speed of machine compute without introducing unacceptable risk. The companies making tangible progress have thought carefully about escalation paths, auditability, and what happens when an agent makes a call you didn’t expect.

Agentic AI meets reinforcement learning

One of the more technically significant themes at GTC, and one with immediate implications for how enterprises design agentic systems: reinforcement learning applied to multi-turn agent workflows. Rather than chaining together prompt-response pairs, leading teams are training agents through trajectories—sequences of actions where each step builds context from prior outputs, and the agent receives per-turn reward signals based on correctness and performance.

The result is agents that can self-improve across multistep tasks: writing code, evaluating it, iterating, and converging on better solutions without human intervention at each stage. This evolution is a real step beyond the “call the model, get a response” paradigm that most enterprise deployments still rely on.

The critical enabler, and the part that separates theory from production, is measurability. As with any reinforcement learning problem, convergence depends on your ability to define and track reward signals. For enterprise agentic systems, that means having clear, automated ways to measure agent impact, output quality, and correctness across multiple dimensions simultaneously. Organizations that can’t measure what their agents are doing can’t improve them, nor catch the silent regressions where gains on one axis come at the cost of another. Teams that build evaluation into the workflow from the start and can operate these feedback loops effectively will pull ahead.

AI-assisted development is accelerating along with the risks

AI is now writing a meaningful share of production code. Developer copilots, code generation agents, and AI-assisted testing tools have moved from novelty to default workflow across most leading engineering organizations. GTC sessions on context engineering for AI code review drew packed rooms. Teams are eager to learn how the best engineering organizations are managing quality as AI takes on more of the codebase.

That velocity comes with a trade-off. AI-generated code is dramatically accelerating development—teams are shipping faster and tackling projects they wouldn’t have scoped a year ago. But the number and types of errors are growing in parallel. AI-written code can introduce subtle bugs, security vulnerabilities, and architectural choices that look fine in isolation but compound at scale. Debugging AI-generated logic requires a different muscle than debugging human-written code.

For enterprise leaders, the shift to AI-assisted development is a product quality and business risk issue, not just an engineering one. The organizations getting it right are treating AI-generated code with the same rigor they would apply to output from a capable but new team member: fast but needing review. The ones getting it wrong are banking the velocity gains without accounting for the new forms of technical debt accumulating underneath.

Agentic AI demands a new security posture

As agents gain the ability to autonomously access systems, move data, call external tools, and make decisions, the vulnerabilities change in ways traditional security architectures weren’t built for.

NemoClaw’s security design, with its runtime sandboxing, privacy routing, and network guardrails, is a direct acknowledgment that agentic AI creates a new category of exposure. An agent that can write code, execute API calls, and access enterprise data is powerful. It’s also, if improperly governed, a serious liability.

A historical look at the early rollout of the web is instructive. Mainstream commerce didn’t take off until the trust layer—SSL, browser sandboxing, standardized authentication—made it safe enough for enterprises. Agentic AI is at a similar inflection point: The capabilities are proven, but the trust infrastructure is still being built. The platforms that get this layer right will unlock the next wave of enterprise value.

We’re seeing early movers treat agentic AI security as its own discipline, separate from application security or data privacy. What can this agent access? What can it do without human approval? How do you audit a chain of automated decisions? How do you defend against adversarial manipulation of agent inputs? Most organizations haven’t built the muscle for this yet.

Data governance is the make-or-break factor

Structured enterprise data—SQL, Spark, data warehouses—was positioned at GTC as the foundation of trustworthy AI, with GPU-accelerated libraries designed to bring the data layer up to the speed AI systems demand. The performance of any agent is bounded by the speed at which it can access and query organizational data.

This “AI-ready” data foundation is where many enterprise AI strategies are falling short. Companies invest in models, agents, and infrastructure while treating data governance as a parallel, slower-moving effort. In an agentic world, that gap becomes a structural disadvantage. Your agents are only as trustworthy as the data they operate on—and that data needs to be classified, governed, permissioned, and available for querying at machine speed.

Bet on the platform, not the model

In addition to NemoClaw, built on the open-source OpenClaw framework, Nvidia now offers open model families across six domains and launched the Nemotron Coalition, a collaboration with Mistral AI, Perplexity, LangChain, and others to codevelop the next generation of open frontier models. This year’s GTC marked the 20th anniversary of the Compute Unified Device Architecture (CUDA), Nvidia’s platform that first empowered graphic processors to be used for general-purpose computing. The strategic logic hasn’t changed: Open ecosystems drive hardware adoption. What’s new is that this approach now extends well beyond compute into models, agents, and orchestration.

For enterprise leaders, the build vs. buy question is worth revisiting carefully, especially with the economics of AI shifting as fast as they are. Fine-tuning and custom models still make sense for certain use cases, languages, and domains. But the pace of frontier model development continues to accelerate, and a custom model built today may quickly become less efficient and effective than tomorrow’s state-of-the-art. That’s a long-term cost that’s easy to underestimate in the short term.

The more durable investments are in infrastructure, governance, security, and auditability—and built to be model-agnostic, so you can swap models as the frontier moves without rewriting everything around them.

Physical AI is entering deployment timelines

Physical AI was impossible to miss at GTC this year, with more than 100 robots on display and partnership announcements that confirmed it’s no longer a demo-stage technology. Nvidia’s robotaxi platform now has commitments from major automakers and a planned Uber deployment across 28 cities by 2028. The world’s largest industrial robotics companies are building on the platform. And healthcare was described as going through its “ChatGPT moment” for physical AI, with surgical robotics partnerships across several major medtech players.

A key enabler: Simulation pipelines and synthetic data generation are making it possible to augment real-world data collection with compute-generated training data, compressing the iteration cycle for physical AI development. The Disney Olaf robot that walked onstage—trained via reinforcement learning in GPU-accelerated simulation—was showmanship, but the underlying point was noteworthy: Physical AI systems can learn complex behaviors in hours rather than months. For companies in manufacturing, logistics, and healthcare, the planning window on physical AI is tightening.

The bottom line

Every major announcement at GTC 2026, from agentic platforms to inference economics to physical AI, pointed to a world where AI is the infrastructure your business runs on, not something you bolt onto the side of it. The organizations best positioned to capture value from this shift aren’t those with the most sophisticated models or the biggest compute budgets. They’re the ones investing now in the data foundations, governance frameworks, and security postures that make everything else work—and building their platforms to be flexible enough to evolve as fast as the technology does.