Baidu's CFO on How It Became a Full-Stack AI Player

Summary

Baidu CFO Henry He described Baidu as a full-stack AI company spanning chips, cloud, models, apps, and robotaxis, but said the most important layer now is cloud because AI demand is shifting from pretraining toward inference and agentic applications. He said roughly 80% of incremental demand is token inference-related, and that Baidu tries to keep token access open internally while relentlessly cutting unit costs. He framed Baidu’s organizational model around younger talent, campus hiring, mentorship, and giving employees more autonomy so they can work like “one-person companies” with AI tools. On the business side, he said Baidu is trying to balance the “impossible triangle” of AI investment, growth, and shareholder returns, while reporting stronger cloud growth, positive operating cash flow, and a mix shift away from search. He also defended Baidu’s approach to safety and alignment as engineering work, and highlighted Apollo Go’s robotaxi scale, outcome-based monetization, and the company’s chip spin-off as part of a broader ecosystem strategy.

Notes

• Baidu is positioned as a full-stack AI company spanning cloud, applications, chips, models, and robotaxis.
• Henry He says the AI stack is moving from infrastructure to applications and from models to agents.
• Cloud is Baidu’s preferred priority layer because it can host Baidu’s model and third-party models.
• About 80% of incremental demand is now token inference-related.
• Inference and application workloads matter more than frontier pretraining for current demand.
• Baidu does not allocate token budgets by title or seniority.
• The company wants to keep token access open and nimble for internal R&D.
• Unit token cost may be falling by half in just a few weeks, according to He.
• Baidu is recruiting younger talent and increasing campus hiring.
• Mentorship and autonomy are core parts of Baidu’s talent strategy.
• Baidu is using internal AI tools to make employees act like one-person teams.
• Baidu’s chip strategy is tied to inference and application workloads, not just pretraining.
• The company’s chip products are meant to reinforce cloud usage through network effects.
• He describes AI investment as an "impossible triangle" of growth, capex, and shareholder returns.
• Baidu says operating profit almost doubled quarter over quarter.
• Cloud revenue grew about 79% year over year.
• Operating cash flow has been positive since Q3 last year.
• He says AI capex is not scaling 2x or 3x, despite the investment push.
• Some AI projects may take 20 to 40 months to generate full cash payback.
• AI safety and alignment are treated as engineering work, not theory.
• Data sanity, post-training, labeling, and SFT are part of Baidu’s safety stack.
• China tech’s advantage is framed as engineering execution and lower-cost robustness work.
• Baidu follows open-source model safety work but also runs its own tests.
• He argues AI in China sits within a broader system involving infrastructure, data regulation, models, and cloud.
• He says cloud-era issues around who owns and accesses data are familiar, not novel.
• Baidu says Ernie 5.1 ranks #1 in a text-format RM arena.
• He says Ernie 5.1 ranks #5 globally in search skill capabilities in the RM arena.
• Apollo Go is running at large scale across 27 cities globally.
• Apollo Go does about 350,000 trips per week, versus Waymo’s about 500,000.
• Baidu search revenue has fallen to about 48% of total revenue this quarter.
• Non-search AI businesses are becoming a larger part of the revenue mix.
• Closed data pockets still exist in China, but cloud and agents should narrow the gap.
• Baidu believes its access to some data pockets is better than peers.
• Before AI, public-car penetration in China was only about 20% to 30%, versus about 90% in the U.S.
• He says AI and cloud infrastructure can help close China’s transportation-service gap.
• Robotaxi could undercut private car ownership if per-mile costs fall below about 60 to 80 cents.
• Today’s robotaxi cost is still around $1 to $2.5 per mile, according to He.
• He expects robotaxi costs to fall over a 5- to 10-year horizon.
• Apollo Go is framed as a car that can earn money while its owner is away.
• Near-term robotaxi competition is still early, with low penetration.
• Robotaxi success depends on technology and operational efficiency.
• Baidu has been working on robotaxis for about 13 years.
• He claims only two cities globally have more than 1,000 robotaxis at scale: San Francisco and one city in China.
• Baidu’s AI success metric is shifting from DAU to DAA, or daily active agents.
• Baidu’s founder Robin is credited with pushing the DAA metric concept.
• Result-based payment is emerging as AI agents prove value.
• Baidu deployed its Famo agent at a major Chinese port for shipment planning and logistics.
• The port example links AI deployment to profit-sharing and measurable efficiency gains.
• AI buying decisions are moving from CTO/CIO buyers to CEO-level buyers.
• Digital employees can expand e-commerce sales across time zones and languages.
• Digital humans can operate 24 hours and answer in multiple languages.
• Baidu has filed a confidential Hong Kong spin-off for chip assets.
• Baidu’s chips are described as highly useful for inference.
• Chips are an ecosystem business, not just hardware.
• A separate public company could make Baidu’s chips feel more neutral and independent to customers.