This is the part of the AI boom I don’t trust: the crowds, the stadium energy, the feeling that one company’s conference is where the future gets “decided.” When an industry starts packing arenas, it’s usually a sign of momentum—and a sign that people are ready to believe almost anything.
Nvidia’s GTC 2026 pulled in about 40,000 attendees in San Jose, spread across the convention center and a nearby arena. That’s not a niche developer meetup anymore. That’s a full-on pilgrimage. The headline message, based on what’s been shared publicly, is that Nvidia is pushing hard on “physical AI” and “agentic AI,” with Jensen Huang using the keynote to underline accelerated computing and big, transformative AI as the center of gravity.
On paper, it all sounds exciting. Smarter systems. More capable tools. AI that doesn’t just answer questions, but can take action. AI that can connect to the physical world, not just live in chat windows. If you’re building products, or running an IT team, or trying to keep up with what’s coming next, I get why you’d want to be in that room.
But I think we should admit what a 40,000-person conference really signals: power is concentrating. Not just “AI is happening.” This is “AI is happening through a narrow funnel.” Nvidia isn’t merely selling hardware or hosting talks. It’s setting the mood, the language, and the default roadmap for a lot of companies that don’t want to miss the wave.
That’s where my discomfort starts.
When “agentic AI” is the theme, the promise is obvious: less busywork, more automation, systems that can coordinate tasks and make decisions faster than humans can. Imagine a small business owner who finally gets an assistant that can handle invoices, respond to customers, and reorder supplies without constant babysitting. Imagine a hospital admin team that can schedule staff, manage inventory, and reduce paperwork. If this works, it’s real relief for people who are drowning.
But the risk is also obvious: once you tell people the system can act on its own, they will let it. Not carefully. Not slowly. They’ll do it because they’re tired, understaffed, and under pressure to cut costs. And then you’re one mistaken action away from a mess you can’t easily unwind.
Say you’re running a mid-sized company and you give an “agent” permission to negotiate with vendors, approve refunds, or change ad budgets. At first it saves time. Then it makes one confident mistake at scale. Not evil. Not dramatic. Just wrong. Who pays for that? The person who clicked “enable,” not the company that sold the dream.
Physical AI raises the stakes even more. The second AI leaves the screen and starts shaping real-world outcomes—machines, movement, logistics, anything tied to bodies and buildings—the tolerance for error drops fast. A bad chatbot answer is annoying. A bad physical decision can be dangerous or expensive. If the pitch is “we can bring AI into the physical world,” my first reaction isn’t awe. It’s: are we actually ready for the liability, the safety standards, and the boring but necessary controls?
And here’s the thing: conferences don’t reward caution. They reward momentum.
A keynote is not where you hear about the messy edge cases. A panel is not where you hear about the internal failures that made a team pull the plug. The whole machine runs on excitement. That’s not Nvidia’s fault, exactly. It’s the incentive of the moment. People buy tickets and fly in because they want to be told they’re early, not because they want a lecture on what could go wrong.
Still, Nvidia is choosing to be the center of this. And the center has responsibility.
When the focus is “accelerated computing,” it’s basically a bet that faster and bigger is the path forward. That might be true for some breakthroughs. But it also locks in a worldview: progress equals more compute, more infrastructure, more dependency on whoever can supply it. If you’re a startup, that can feel like a golden road and a toll road at the same time. You can build incredible things—if you can afford the toll. If you can’t, you’re not just behind; you’re invisible.
Meanwhile, big companies will lean into this because it fits how they operate. They like vendors that can promise end-to-end solutions and a clear path to “innovation.” The result could be a market where a few players get to define what “modern AI” even means. That’s good for standardization, sure. It’s also how ecosystems become locked, where alternatives struggle not because they’re worse, but because they don’t have the gravity.
To be fair, there’s another way to read this 40,000-person turnout: it’s a sign the industry is finally taking implementation seriously. Live sessions and panels can mean people are trying to move beyond demos and into real systems. Maybe this is the phase where the hype gets forced into contact with reality, and reality wins. That would be healthy. I just don’t think the crowd energy naturally pushes in that direction.
So yes, Nvidia pulling tens of thousands of people into San Jose is impressive. It’s also a warning label. We’re not just choosing tools anymore. We’re choosing who gets to steer the direction of AI, what kinds of AI get funded, and how quickly “autonomous” becomes the default setting in places that can’t afford mistakes.
If you’re building, buying, or regulating any of this, the question isn’t whether Nvidia’s vision is bold—it clearly is—it’s whether we’re comfortable letting one company’s momentum set the pace for how fast agentic and physical AI gets pushed into everyday life.
What would it actually take for you to believe this push toward more autonomous, more physical AI is being rolled out at the right speed, not just the most profitable speed?