The money is moving from foundation models to the industries where AI can actually kill you if it gets it wrong.
The Summary
- Lux Capital's Deena Shakir says the AI investing thesis is shifting from model-building to deployment in health care and robotics, following recent Anthropic and OpenAI disclosures that highlighted security and real-world implementation challenges
- The new bet: vertical applications built on trust and partnerships, not just raw capability improvements
- Translation: foundation models are becoming infrastructure. The returns are now in the application layer, especially in high-stakes domains.
The Signal
Lux Capital doesn't chase hype. They funded Anduril when everyone else was funding meal kits. So when partner Deena Shakir says AI is now an operating system and the next boom is vertical, that's a signal the venture capital thesis has fundamentally changed.
The timing matters. This comes right after both Anthropic and OpenAI made disclosures about deployment challenges and security considerations. Not capability benchmarks. Not parameter counts. Real-world friction points. That shift from "look how smart our model is" to "here's how we're making it safe enough to actually use" is the tell.
"AI is now an operating system" means the foundation layer is solved enough that returns move up the stack.
Health care and robotics aren't random picks. They're the two domains where AI encounters the hardest constraints: regulatory approval, physical consequences, and zero tolerance for hallucinations. A chatbot that makes up a citation is embarrassing. A surgical robot that hallucinates a blood vessel location kills someone. These aren't just harder markets to enter. They're markets where trust becomes the actual product.
The venture math is changing:
- Foundation models require billion-dollar training runs with unclear monetization paths
- Vertical AI companies can use existing models, focus capital on domain expertise and regulatory moats
- Health care AI alone is projected to hit $188 billion by 2030, and none of it goes to the model providers
This is the application layer gold rush. The picks and shovels phase already happened. Now we're in the phase where someone builds the actual mine, hires actual miners, and sells actual gold. Health care and robotics are where AI stops being a demo and starts being a business with revenue, margins, and defensibility.
The Implication
If you're building in AI, the question isn't "how do I make a better model." It's "what regulated, high-stakes industry can I enter where trust and domain knowledge create a real moat." Health care, robotics, energy, infrastructure. Anywhere the cost of being wrong is high and the incumbents move slow.
For investors, this is the rotation. Foundation models are becoming table stakes infrastructure. The asymmetric returns are in companies that combine AI with regulatory expertise, physical systems integration, and customer relationships that took decades to build. Watch where the serious money goes in Q3 and Q4 2026. It won't be another LLM lab.