Notes from GTC and Cloud Next: we have stopped talking about just models

Something has quietly shifted in how the industry talks about AI.

Not long ago, most of the conversation was about which model was better, faster, cheaper, or more capable. That still matters, of course. But after attending NVIDIA GTC in March and Google Cloud Next in April, I felt the center of gravity move somewhere else.

The harder question now is not “which model should we use?” It is: how do we build AI systems that actually hold up in production?

I noticed it at both conferences, even though they came at the topic from very different angles. GTC was focused on infrastructure, physical AI, simulation, and what AI can become. Cloud Next was more grounded in enterprise adoption, agent platforms, governance, data, and how AI gets embedded into real workflows.

Different crowds. Different agendas. Same conclusion: we are moving from AI models to AI systems.

What stood out at GTC

At NVIDIA GTC, Jensen Huang’s keynote covered the full AI stack, but what stayed with me most was the emphasis on agentic AI, physical AI, and digital twins.

The message was clear: AI is leaving the chat window. It is moving into systems that can reason, plan, take action, interact with physical environments, and continuously improve after deployment. That shift matters for every industry, but it feels especially relevant in retail.

Physical AI can sound like science fiction if we only picture humanoid robots. And yes, there were plenty of robots on the show floor. But the more practical version for retail is much closer to home: intelligent systems that can understand what is happening inside a store, a warehouse, or a fulfillment center and help teams respond faster.

That is where digital twins become important. For a retailer, a digital twin is not just a 3D model. It is a way to represent the physical store with enough accuracy and context to test decisions before making changes in the real world. Store layouts, merchandising changes, bay optimization, fulfillment flows, and associate workflows can all be evaluated in simulation before they become physical execution.

This is not theoretical. NVIDIA has publicly highlighted how Lowe’s uses physically accurate digital twins of its stores to support visualization, analysis, simulation, and merchandising decisions. The practical value is simple: test more, disrupt less, and make better decisions before work reaches the sales floor.

That was one of my biggest GTC takeaways. Physical AI is not only about robots. It is about connecting AI to real operations.

What stood out at Cloud Next

A few weeks later, Google Cloud Next filled in the other half of the story.

If GTC was about what AI can become, Cloud Next was about what it takes to run AI inside an enterprise: agent platforms, orchestration, data infrastructure, governance, security, and integration into everyday workflows. In other words, the less glamorous work that determines whether AI actually creates value.

The clearest signal across the keynotes and sessions was that AI has moved from experimentation to execution. The conversations were no longer about pilots or proofs of concept. They were about running fleets of agents inside real workflows, at scale, with real accountability for the results.

That changes the bar for enterprise AI. Once AI becomes part of how work gets done, you can no longer manage it like a pilot. You need platforms, controls, real visibility into how these systems behave, and clear ownership of the outcomes.

The mental model I came home with

Here is the simplest way I now think about it.

NVIDIA is building the infrastructure layer that powers agentic, physical, and simulation-based AI. Google Cloud is building the enterprise layer to deploy, orchestrate, govern, and integrate those systems. Seen side by side, they point to the same place: the model is no longer the hard part.

The center of gravity has moved from raw capability to everything that surrounds it, the infrastructure, the platforms, the operations, and the discipline required to run these systems where real work happens.

That is where AI is heading. Impressive models are becoming table stakes. The harder and more valuable work is building AI systems that are reliable, well governed, and capable of delivering consistent value at scale.