Coming out of this year’s Nvidia GPU Technology Conference, we were shocked at how far the conversation had progressed. A year ago, the headlines were that AI had graduated from pilots to real-world enterprise deployments. It’s now a gamble. Leaders in this wave are not just implementing AI into their businesses, they are reimagining their operating models around AI.
Even a few months ago, companies were looking for ways to make AI work reliably at scale. Today’s question is even more difficult. How do you manage autonomous agents operating across thousands of business units? How do you maintain your security posture when AI systems are making decisions on their own? How do you ensure that your system design and data infrastructure can actually support the speed these platforms need to operate?
A few themes that really matter to corporate leaders stood out.
- The cost of using AI is falling rapidly, the pace of model development is accelerating, and every business needs a strategy for tackling AI.
- Agentic AI has moved from concept to platform, making governance a central challenge for companies looking to scale it.
- Security, system design, and data infrastructure must evolve in tandem with AI adoption, rather than as separate, slow-moving workstreams.
- The calculus of building and buying AI is becoming more nuanced. A model-agnostic platform is a permanent investment.
- Physical AI is embedded in real-world implementation timelines in automotive, robotics, and healthcare.
The economics of AI are changing rapidly
The cost of using AI is decreasing significantly. Nvidia’s Vera Rubin platform, the successor to Blackwell currently in production, delivers up to 10x more inference throughput per watt and 10x the cost per token. At the same time, the pace of development of frontier models continues to accelerate. Models are becoming more capable with shorter release cycles, and the tools you build on top of them are dramatically more accessible.
For most companies, this is not about training their own models. This is about the ever-expanding scope of what is economically achievable with AI. Use cases that weren’t planned six months ago, such as real-time decision-making, customer interactions, and operational automation, are worth a fresh look. Nvidia has raised its demand forecast for AI chips to $1 trillion by 2027. Its growth is overwhelmingly driven by inference workloads rather than training.
But cheaper inference has a second-order effect that is often overlooked. As the cost per AI action falls, AI usage across businesses will explode. Governance, data quality, and security need to be in place.
Agentic AI now has a platform
At last year’s GTC, production-grade agent AI was just beginning to gain traction. This year, Nvidia announced an agent toolkit based on NemoClaw, an enterprise reference design built on OpenClaw, an open source agent framework that exploded in popularity in early 2026. NemoClaw overlays the flexibility of OpenClaw with runtime sandboxing, privacy routing, and network guardrails to give enterprises a controlled operating environment for their agents without locking them into a closed ecosystem.
A recurring pattern in our client work is that companies struggle to scale up agent pilots. This is the more difficult problem of running agents across hundreds or thousands of locations, business units, or customer touchpoints. Companies often have the technical capabilities but lack a reliable platform with strong safety guardrails that can run at machine computing speeds without posing unacceptable risks. Companies that are making measurable progress are thinking carefully about escalation paths, auditability, and what happens when an agent makes an unexpected call.
Agent AI meets reinforcement learning
One of the technically important topics at GTC, and one that has a direct impact on how companies design agent systems, is reinforcement learning applied to multi-turn agent workflows. Rather than chaining prompt-response pairs, the leading team trains agents through trajectories (sequences of actions in which each step builds context from previous outputs), and agents receive turn-by-turn reward signals based on accuracy and performance.
As a result, agents can self-improve across multi-step tasks of writing code, evaluating it, iterating on it, and converging on a better solution without human intervention at each stage. This evolution is a real step beyond the “call a model and get a response” paradigm that most enterprise deployments still rely on.
A key enabler and the dividing line between theory and operation is measurability. As with other reinforcement learning problems, convergence depends on the ability to define and track reward signals. For enterprise agent systems, this means having a clear, automated way to measure agent impact, output quality, and accuracy across multiple dimensions simultaneously. Organizations that cannot measure what their agents are doing cannot improve them or catch the silent setbacks where gains in one axis come at sacrifices in another. Teams that can build evaluation into their workflows from the beginning and operationalize these feedback loops effectively will move forward.
AI-assisted development accelerates with risk
AI is now writing critical parts of production code. Developer co-pilots, code generation agents, and AI-assisted testing tools have gone from novelty to default workflow in most major engineering organizations. The GTC session on Context Engineering for AI Code Reviews drew a packed house. As AI takes over more of the codebase, teams want to know how the best engineering organizations manage quality.
That speed comes with a trade-off. AI-generated code is dramatically accelerating development, allowing teams to ship faster and tackle projects they couldn’t even consider a year ago. However, the number and type of errors are also increasing in parallel. Code written with AI can introduce subtle bugs, security vulnerabilities, and architectural choices that seem fine on their own, but compound at scale. Debugging AI-generated logic requires different strengths than debugging code written by humans.
For company leaders, the transition to AI-assisted development is not just an engineering issue, but a product quality and business risk issue. Organizations that do it well treat AI-generated code with the same rigor they apply to output from competent but new team members: a quick but reviewable way. What’s wrong is banking on speed improvements without considering the new forms of technical debt that are accumulating underneath.
Agent-based AI requires a new security posture
As agents gain the ability to autonomously access systems, move data, invoke external tools, and make decisions, vulnerabilities change in ways not built into traditional security architectures.
NemoClaw’s security design, with its runtime sandbox, privacy routing, and network guardrails, directly recognizes that agent AI is exposed to new categories of risk. Agents that can write code, make API calls, and access corporate data are powerful. They also bear significant liability if improperly managed.
It is instructive to look back at the history of the early development of the Web. Mainstream commerce didn’t catch on until trust layers (SSL, browser sandboxing, standardized authentication) made it sufficiently secure for businesses. Agentic AI is at a similar inflection point. Although functionality has been proven, the trust infrastructure is still being built. Platforms that properly leverage this layer will unlock the next wave of enterprise value.
Early companies are treating agent-based AI security as its own discipline, separate from application security and data privacy. What does this agent have access to? What can it do without human approval? How do you audit a set of automated decisions? How do you defend against adversarial manipulation of agent input? Most organizations haven’t yet built the power to handle this.
Data governance is the difference between winning and losing
At GTC, we positioned structured enterprise data (SQL, Spark, data warehouses) as the foundation for trusted AI, along with GPU-accelerated libraries designed to bring the data layer up to the speeds required by AI systems. Agent performance is limited by the speed at which agents can access and query organizational data.
This “AI-enabled” data foundation is where many companies’ AI strategies fall short. While companies invest in models, agents, and infrastructure, they treat data governance as a slow, parallel effort. In the world of agents, that gap becomes a structural shortcoming. The trustworthiness of an agent is determined by the data it operates on. And that data must be classified, managed, authorized, and queryable at machine speed.
Bet on the platform, not the model
In addition to NemoClaw, which is built on the open source OpenClaw framework, Nvidia now offers a family of open models across six domains and has launched the Nemotron Coalition, a collaboration with Mistral AI, Perplexity, LangChain, and others to co-develop the next generation of open frontier models. This year’s GTC celebrated the 20th anniversary of Compute Unified Device Architecture (CUDA), Nvidia’s platform that first enabled graphics processors to be used for general-purpose computing. The strategic logic has not changed. An open ecosystem accelerates hardware adoption. What’s new is that this approach extends far beyond computing to models, agents, and orchestration.
For business leaders, the build-versus-buy question is worth carefully revisiting, especially as the economics of AI change rapidly. Fine-tuned and custom models still make sense for certain use cases, languages, and domains. However, the pace of frontier model development continues to accelerate, and custom models built today can quickly become less efficient and effective than tomorrow’s cutting-edge models. This is a long-term cost that is often underestimated in the short term.
More durable investments are infrastructure, governance, security, and auditability that are built to be model agnostic, so you can swap models as frontiers move without rewriting everything around you.
Physical AI is entering the implementation schedule
You couldn’t miss the physics AI at GTC this year. There were more than 100 robots on display and partnership announcements indicating that the technology is no longer in the demo stage. Nvidia’s robotaxi platform is currently partnering with major automakers and will have Uber in 28 cities by 2028. The world’s largest industrial robotics companies are building on this platform. It was also described that the healthcare industry is experiencing a “ChatGPT moment” for physical AI, with surgical robotics partnerships with several leading medical technology companies.
Key enablers: Simulation pipelines and synthetic data generation enable real-world data collection to be augmented with compute-generated training data, compressing iterative cycles for physical AI development. The Disney Olaf robot that walked on stage (trained through reinforcement learning in a GPU-accelerated simulation) was showmanship, but at its core it was remarkable. That means physical AI systems can learn complex behaviors in hours instead of months. For manufacturing, logistics, and healthcare companies, the planning horizon for physical AI is narrowing.
conclusion
From agent platforms to speculative economics to physical AI, the major announcements at GTC 2026 all pointed to a world where AI is the infrastructure on which business runs, not something bolted onto the side. The organizations best suited to capture value from this change are not those with the most sophisticated models or the largest computing budgets. They are now investing in the data foundation, governance framework, and security posture that makes everything else work, building a platform that is flexible enough to evolve as fast as the technology.
