Jensen Huang just told investors the quiet part out loud: the next trillion-dollar companies won't be buying Nvidia chips — they'll be built on top of them.
The Summary
- Nvidia is positioning itself as infrastructure for the entire AI stack, not just a chip vendor to hyperscalers — selling compute as the foundation for startups building agents, robotics, and autonomous systems
- Sarah Guo of Conviction VC, one of the leading AI-native investors, confirms this shift: the action is moving from infrastructure companies buying GPUs to application companies building on Nvidia compute
- The picks-and-shovels metaphor masks a bigger play: Nvidia is creating platform lock-in at the compute layer, not the application layer
The Signal
For years, the Nvidia narrative was simple: they sell chips to Microsoft, Google, Meta. Those companies build the models. Everyone else uses the APIs. That vertical is now a side business.
The new playbook: startups access Nvidia compute directly through cloud providers or specialized inference platforms, build agents or robotic systems, and scale without ever touching model training. They're renting the mine, not buying pickaxes. Nvidia gets recurring revenue from compute cycles instead of one-time hardware sales. It's AWS for intelligence.
"The next wave isn't companies that train models. It's companies that deploy agents at scale on someone else's compute."
Sarah Guo's thesis at Conviction has been betting on exactly this layer: application companies that assume intelligence is abundant and cheap, and build products that couldn't exist without it. Her portfolio spans:
- Agent frameworks that orchestrate multi-step reasoning
- Robotics companies training humanoid systems in simulation before deploying to hardware
- Autonomous vehicle startups running inference at the edge with cloud backup
What makes this different from previous platform shifts: the moat isn't the application. It's the compute dependency. An agent startup can't easily switch from Nvidia CUDA to AMD or custom silicon without rewriting core inference pipelines. A robotics company training on Nvidia simulation environments can't port that work to another stack without starting over.
The Big Tech buyers are still there, still spending billions. But they're no longer the growth story. The growth story is ten thousand startups each spending $50K to $5M per year on compute. That's a different revenue profile. More distributed. Stickier. Harder to displace.
This also explains why Nvidia keeps releasing new software layers: CUDA, TensorRT, Omniverse for robotics simulation, NIM microservices for inference. Each one makes the compute harder to leave. The chip is just the entry point.
The Implication
If you're building an AI-native company, your compute strategy is now a core strategic question, not an ops detail. Choosing Nvidia means access to the best tooling and the deepest talent pool. It also means you're renting your competitive advantage from a supplier who sees you as a customer, not a partner.
Watch for the counter-move: open inference standards, AMD's ROCm maturation, or startups offering compute-agnostic agent frameworks. The lock-in is real, but it's early enough that alternatives could still fracture the market. For now, though, Nvidia isn't just selling picks and shovels. They own the trail to the goldfield.