Overview of Baidu's CFO on How It Became a Full-Stack AI Player
This Bloomberg Odd Lots interview with Baidu CFO Henry He centers on how Baidu is positioning itself as a full-stack AI company across chips, cloud, foundation models, and applications. He argues that the AI market is shifting from pre-training toward inference and task completion, making cloud the most important layer for Baidu today. The conversation also covers capital allocation, talent, AI safety, China’s policy environment, robotaxis, and how Baidu plans to monetize AI through agents and “digital employees.”
Baidu’s Full-Stack AI Strategy
Henry He describes Baidu as an integrated AI stack with multiple linked layers:
- Custom chips for AI infrastructure
- Cloud for deployment and inference
- Ernie, Baidu’s foundation model
- Applications, including robotaxis and enterprise AI products
His key point: the AI value chain is moving from infrastructure to applications and from models to agents. In that environment, he says the cloud is the “must-win” layer because it hosts Baidu’s own models and can also support other models.
Why cloud matters most now
- Pre-training still matters, but He says inference now accounts for the majority of incremental token demand.
- Baidu’s cloud is positioned as the platform where models are deployed and real tasks are executed.
- The company wants to profit from the transition from chat to actionable AI agents.
Tokens, Productivity, and the New AI Workflow
A running theme in the interview is the idea of a “token budget” and how companies measure AI productivity.
He’s view:
- Internal AI use should be judged by better model performance and better task completion.
- The more important metric today is not how many tokens are consumed, but how many real-world tasks are completed.
- Baidu does not believe in rigid token quotas by title or seniority; it prefers a more flexible, open approach.
He also says:
- Token costs are falling rapidly.
- Younger employees are using AI effectively and often won’t waste resources.
- AI is becoming a mindset, not just a tool.
Talent, Organization, and Internal AI Adoption
To attract top talent, Baidu emphasizes:
- Autonomy
- Trust
- Real responsibility
- Full-stack exposure across model, cloud, chip, and application layers
He says Baidu is unusually open to:
- Campus recruiting
- Mentorship from senior staff
- Giving employees more freedom to act like “one-person teams” using AI tools
He also mentions internal productivity tools that help workers operate more independently and efficiently.
Capital Allocation: The “Impossible Triangle”
He frames Baidu’s financial challenge as an “impossible triangle”:
- Grow the AI business
- Maintain disciplined capex
- Return capital to shareholders
Recent results he highlighted:
- Operating profit nearly doubled quarter over quarter
- Cloud revenue grew about 79% year over year
- Operating cash flow turned positive starting in Q3 of the prior year
His message is that Baidu wants to keep investing in AI, but in a more responsible, ROI-focused way:
- Spend must be paced carefully
- Returns may take 20–40 months depending on the project
- Capex should stay aligned with cash generation and lifecycle economics
Custom Silicon and the Chip Spin-Off
He says Baidu’s chip strategy is tied to its broader AI stack, especially inference and applications rather than massive pre-training.
Main rationale for custom silicon:
- Better fit for Baidu’s cloud and model workloads
- Improved performance for inference-heavy use cases
- Stronger ecosystem control and positive network effects
He also confirms that Baidu has filed for a spin-off of its chip assets in Hong Kong, which he presents as a way to unlock value and make the chip business more independent and market-facing.
AI Safety, Alignment, and Data Quality
He treats AI safety more as an engineering discipline than a philosophical one.
Baidu reportedly invests in:
- Data sanity
- Post-training
- Alignment checks
- Robustness testing
- Labeling and quality control
His view is that China’s tech ecosystem already has strong experience with:
- Engineering rigor
- Cost-efficient data pipelines
- Cloud governance and data access controls
He downplays the idea of “AI psychosis” or existential panic, saying the focus is on building reliable systems and continuously improving them.
China’s Policy and Data Environment
He avoids commenting directly on public policy, but suggests China’s AI environment is:
- Supportive
- Transparent
- Open to industry and academic input
- Built on a decade-plus of infrastructure and cloud development
He argues that China’s cloud and data governance environment is already mature enough to support large-scale AI deployment, and that more workloads will move to public cloud over time.
Robotaxis, Search, and the Consumer AI Race
A major part of the discussion focuses on Apollo Go, Baidu’s robotaxi service.
He argues robotaxis could change transportation economics by making it cheaper than car ownership in some markets:
- He estimates current robotaxi costs are still too high to replace ownership broadly
- But as scale grows, the cost per mile could fall into the range that makes it economically compelling
- Robotaxis may also create a “physical AI agent” that can work 24/7 and even generate revenue for owners
He says Baidu is expanding internationally and already has partnerships with:
- Uber
- Lyft
- Grab
He also notes that Baidu’s search business is still important, but its share of revenue is falling as AI applications grow.
Agents, Digital Employees, and Monetization
He sees daily active agents (DAA) as the AI-era equivalent of DAU in mobile internet.
How Baidu expects to make money:
- Task-based pricing
- Sharing in cost savings or productivity gains
- Revenue-linked compensation in enterprise settings
Examples he gives:
- AI agents managing port logistics
- “Digital employees” helping e-commerce merchants
- AI systems supporting live commerce across time zones and languages
His core point: as agents become better at completing work, customers are more willing to pay for outcomes rather than software access alone.
Key Takeaways
- Baidu wants to be a full-stack AI company, not just a model provider.
- Cloud is the most important strategic layer right now because inference and deployment are where demand is shifting.
- Baidu is pursuing AI with a focus on ROI, efficiency, and real task completion.
- The company believes robotaxis and digital agents can become major commercial businesses.
- Custom chips, data quality, and integration across the stack are central to Baidu’s long-term strategy.
- The interview presents Baidu as one of China’s most mature and ambitious AI platforms, with a model that increasingly resembles Google’s integrated stack.