Picture this: You’re sitting in a conference room in April 2026, watching Doug Clinton from Intelligent Alpha explain his AI investment thesis on CNBC’s Squawk Box. He’s not talking about some obscure startup with a flashy demo. He’s talking about Nvidia and Google. The two companies everyone already knows are winning.
And that’s precisely the point.
As someone who spends my days analyzing agent architectures and intelligence systems, I find Clinton’s assessment fascinating—not because it’s contrarian, but because it isn’t. From a technical perspective, the “safest bets” in AI aren’t the ones promising magical breakthroughs. They’re the ones that already control the infrastructure layer.
The Infrastructure Advantage
Let’s talk about what “safe” actually means in the context of AI investments. Clinton points to strong revenue growth and rising AI demand as his core reasoning. But beneath those financial metrics lies a more fundamental truth about how AI systems actually work.
Nvidia doesn’t just make chips that happen to be good for AI. They’ve built an entire ecosystem—CUDA, cuDNN, TensorRT—that makes it extraordinarily difficult for competitors to displace them. When you’re training a large language model or running inference at scale, you’re not just buying hardware. You’re buying into a software stack that your engineers already know, that your codebase already depends on, and that every major framework optimizes for first.
This is what economists call “switching costs,” but in AI, it’s more like switching impossibility. Try migrating a production system from CUDA to anything else. I’ll wait.
Google’s Quiet Dominance
Google’s position is different but equally entrenched. They’re not selling picks and shovels—they’re using them to mine their own gold. TPUs power their internal infrastructure. DeepMind’s research feeds directly into products. And crucially, they have something most AI companies desperately need: data at scale, collected over decades, with user permission baked into terms of service that billions have already accepted.
When I evaluate agent systems, I look at three things: compute access, training data quality, and deployment reach. Google has all three. They can train models, they can serve them to users, and they can iterate based on real-world feedback faster than almost anyone else.
What “Safe” Really Means
Clinton’s framing of these as the “safest” bets reveals something important about where we are in the AI cycle. We’re past the phase where every startup with a transformer model could claim to be the next big thing. We’re entering the phase where infrastructure matters more than ideas.
From an agent architecture perspective, this makes sense. The most sophisticated AI agents aren’t standalone systems—they’re compositions of multiple models, retrieval systems, and execution environments. Building these requires access to compute, pre-trained models, and production-grade serving infrastructure. Nvidia provides the compute. Google provides everything else, including the models themselves through their various APIs and services.
The Technical Moat
What makes these companies “safe” isn’t just their current market position. It’s the technical depth of their moats. Nvidia’s advantage compounds with every new architecture that gets optimized for CUDA. Google’s advantage compounds with every new service that integrates Gemini or uses their infrastructure.
Compare this to other AI companies that are essentially reselling access to foundation models with a thin wrapper on top. Their moat is a product feature, not a platform. Their advantage is temporary, not structural.
The Researcher’s Perspective
As someone who works on agent systems, I see this playing out in real time. When we prototype new architectures, we default to Nvidia GPUs because the tooling is mature. When we need to deploy at scale, we evaluate Google Cloud alongside AWS and Azure—but Google’s AI-specific offerings often win on performance and ease of integration.
This isn’t about brand loyalty. It’s about technical reality. The companies that control the infrastructure layer have a structural advantage that’s difficult to overcome through better algorithms alone.
Clinton’s assessment isn’t bold or surprising. But sometimes the obvious answer is obvious for good reasons. In a space full of hype and speculation, betting on the companies that actually build and control the infrastructure isn’t exciting. It’s just smart.
đź•’ Published: