The Full-Stack Pivot: Why Baidu is Betting on Inference Over Scale

In this conversation, Baidu CFO Henry He outlines the shift from model-first AI to agent-first execution. The core idea is that competitive advantage in AI has moved away from massive pre-training, which is becoming a commodity, toward the application layer and inference efficiency. The consequence of this shift is a change in how companies measure success: moving from token spend to daily active agents (DAA). This conversation is for leaders and investors who want to understand why vertical integration is a necessary architecture for capturing economic value as AI moves from a tool to an autonomous worker.

The Shift from Model-First to Execution-First

Most organizations treat AI as an R&D cost center, a tool that requires a budget. Henry He argues that this perspective is flawed because it ignores the systemic nature of AI. When AI is viewed as a tool, it remains a cost to be managed. When it is viewed as an agent, it becomes a revenue-generating asset that can operate autonomously.

The insight here is that the cloud is the must-win layer of the stack because it serves as the platform for inference. As He notes, 80% of incremental demand for computing power is now driven by inference, not pre-training.

Right now, the pre-training is important but 80% of the incremental demand today on a token are inference related. I think this part of the full picture is what I want to emphasize.

-- Henry He

This creates a competitive advantage for those who can lower the unit cost of inference. While competitors fixate on the leaderboard performance of their models, Baidu is focusing on the completion of real-world tasks. The system responds to this by shifting incentives: engineers are no longer rewarded for token maxing, but for the efficiency of the agents they deploy.

The Hidden Cost of General AI

Conventional wisdom suggests that AI labs should be general-purpose. He suggests the opposite. By vertically integrating, designing their own chips, models, and cloud infrastructure, Baidu is positioning itself to capture the economic value of the data flywheel.

This integration creates a feedback loop that is difficult to replicate: the robotaxi data informs the foundation model, which improves the inference capabilities of the cloud, which then powers the digital employees used by e-commerce merchants. This is a system designed for durability, not just scale.

The agent is smart enough to think about the planning, the task and completing the task. And obviously in the way of interacting with human beings it actually become more smarter and in a way that working with more efficient planning of that.

-- Henry He

The implication is that the moat is not the model itself, which is increasingly commoditized, but the ability to embed the agent into high-value, repetitive enterprise workflows where the client is willing to share a percentage of the efficiency gains.

The 18-Month Payoff of Vertical Integration

He describes the financial management of this transition as an impossible triangle: driving high growth, maintaining ambitious AI investment, and returning capital to shareholders. The resolution to this triangle is not to spend more, but to spend more responsibly by focusing on the full lifecycle of an AI project.

The payoff is delayed, often requiring 20 to 40 months to see full cash returns. Most organizations fail here because they lack the patience to manage the pacing of these investments. By treating the AI stack as a unified system rather than a collection of disparate tools, Baidu is building a structure that can survive the price wars and hardware bottlenecks that define the current AI infrastructure race.

Key Action Items

• Shift from Tool to Agent Metrics: Stop measuring AI success by headcount or token spend. Over the next quarter, begin tracking Daily Active Agents (DAA) to identify which internal processes are actually being automated rather than just supported.
• Audit for Inference Efficiency: If your AI strategy is focused solely on model performance, you are misallocating capital. In the next 6 to 12 months, prioritize infrastructure that lowers the unit cost of inference, as this will be the primary driver of margin.
• Adopt Top-Down AI Integration: Stop selling AI as a technical tool to the CTO. As He notes, the real value is captured when the CEO is involved. Over the next 12 to 18 months, reframe AI initiatives as efficiency gains that share in the profit of the business unit they support.
• Prioritize Vertical Integration: Evaluate where your organization is reliant on external infrastructure. Over the next 18 to 24 months, identify one critical layer of your stack, whether it is data curation, model deployment, or specific hardware needs, where you can build proprietary capabilities to reduce dependency.
• Embrace One-Person Teams: Leverage internal AI agents to empower individuals to act as one-person companies. This requires a cultural shift toward autonomy and trust, which pays off in long-term talent retention and operational speed.