China's search giant quietly assembled the entire AI stack while everyone else was still arguing about which layer to own.
The Summary
- Baidu now controls every layer of its AI operation: custom chips, the Ernie model, cloud infrastructure, and Apollo Go robotaxis
- CFO Henry He revealed how the company thinks about token economics, AI safety alignment, and using AI to reshape its own hiring and org structure
- The original search business now serves as integration layer for a vertically-integrated AI play that started in the late 1990s
- This is what Web4 verticalization looks like when a company commits to building everything in-house
The Signal
Most AI companies are scrambling to patch together partnerships. Baidu built the whole thing. The Chinese tech firm now manufactures its own chips, trains its own foundation model called Ernie, runs its own cloud system, and deploys its own autonomous vehicles through Apollo Go. That's not a strategy. That's control of the entire value chain from silicon to self-driving.
CFO Henry He's comments on the Odd Lots podcast reveal something more interesting than the stack itself: how Baidu uses AI internally to evolve its organizational structure and attract talent. When your CFO is talking about maximizing token spend like it's a line item in the budget, you're not experimenting with AI. You're running on it.
"The company uses AI to evolve its own organizational structure and attract new talent."
The robotaxi piece matters more than it looks. Apollo Go isn't just another autonomous vehicle experiment. It's a real-world testbed for everything Baidu builds:
- Custom inference chips handling real-time decision making
- Ernie models processing sensor data and traffic patterns
- Cloud infrastructure coordinating fleet operations at scale
This is vertical integration with a forcing function. Every layer of the stack has to work because actual cars are moving actual people.
The shift from search company to AI infrastructure provider happened quietly. Baidu was founded in the late 1990s as China's answer to Google. Now search is just one application sitting on top of a foundation model, which sits on top of custom hardware, which powers everything from cloud services to robotaxis. The business model didn't pivot. It telescoped outward.
He also discussed how Chinese tech firms approach safety and alignment, and how Baidu thinks about the global robotaxi competition. The details matter because China's AI development is following a different path than the West. Less venture capital chaos, more state-backed infrastructure plays. Less "move fast and break things," more "own the stack and control the outcome."
The Implication
The full-stack play is back. Not because integration is inherently better, but because AI infrastructure is too critical to depend on partners who might pivot, price gouge, or shut you out. Baidu's bet is that owning chips through cloud through models through applications gives them speed and control that loosely coupled competitors can't match.
Watch for Western AI companies to follow this path. The agent economy needs reliable infrastructure. When your AI agents are booking flights, managing portfolios, and controlling physical systems, you can't afford downtime because your GPU supplier changed terms. Vertical integration isn't about being big. It's about reducing points of failure when the stakes are high.