Baidu's CFO on How It Became a Full-Stack AI Player | Odd Lots — yt-summary

o5QiW-jvAag.md

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

Channel: Bloomberg Podcasts · Duration: 49:11 · Watch: youtube.com/...

In a live recording at Bloomberg Invest Asia in Hong Kong, Odd Lots hosts Joe Weisenthal and Tracy Alloway interview Henry He, CFO of Baidu, about the company's full-stack AI strategy spanning chips, cloud, foundation models (Ernie), and robotaxis (Apollo Go). He argues Baidu is the closest structural analog to Google — own chips, own cloud, own model, own physical-AI applications — and reveals the company has filed to spin off its chip business in Hong Kong, ships robotaxis competitively with Waymo, and is shifting its key metric from daily active users to "daily active agents."

---

Cold Open & Introduction 00:00

The episode opens with a teaser from the main interview — Joe asking Henry whether employees at Baidu get individualized token budgets — before the hosts introduce themselves and the context. Joe and Tracy explain they were in Hong Kong for the Bloomberg Investor conference (and an Odd Lots trivia night), and that the trip gave them a chance to talk to senior executives at Chinese tech companies during a moment of intensifying US–China AI competition.

• 00:00 — Cold open: Joe: asks whether two Baidu employees in the same department would get identical token budgets, or whether the company measures who extracts more value per token.
• 01:44 — Henry: says the technology evolves too fast to lock in policies by job title; Baidu wants to stay "open and nimble" on token allocation.
• 02:07 — Joe frames the moment: Chinese internet giants concentrated on China, American ones had the rest of the world, and now they're heading into the first real head-to-head battle in AI and self-driving.

Full Stack AI Strategy Discussion 03:16

Joe asks Henry to identify which layer of Baidu's stack — chips, cloud, model, application — is the must-win. Henry frames the backdrop as a shift "from infrastructure to applications and from auto to agents," making it hard to isolate a single layer since ROI depends on all of them working together. If forced to choose, he picks cloud: it hosts Ernie (Baidu's own model), can host third-party models, and connects to their custom chip for inference — and 80% of incremental token demand today is inference-related, not pre-training.

• 03:39 — Joe: asks which layer of the full stack (cloud, application, chips, model) is a must-win for Baidu.
• 04:12 — Henry: says AI is shifting from infrastructure to applications, from auto to agents; can't isolate one layer because ROI requires all of them.
• 04:43 — Henry: if forced to pick one, chooses cloud — it hosts Ernie and other models, and their chip connects to it for inference; 80% of incremental demand today is inference-related.

Token Budget and Productivity Measurement 05:23

Joe introduces what he calls the new standard financial-journalist question — "what's your token budget?" — and asks how Baidu measures productivity from token spend. Henry splits the measurement into two buckets: internal R&D (reaching higher technology standards, AGI-level model performance) and external client delivery (how many real tasks get completed). He stresses that the frameworks now linking foundation models to real-world task completion — not just chat — are what make token spend meaningful, and that "completion" is becoming more important than raw consumption.

• 05:23 — Joe: jokes that "what's your token budget" will become the standard question for financial journalists, like asking about headcount.
• 05:45 — Henry: splits measurement into two buckets — internal R&D (reaching higher tech standards, AGI) and external client delivery (real tasks completed).
• 06:16 — Henry: frameworks that link foundation models to real-world tasks matter more than raw token consumption; "completion part of the tokens is more important today."

Individual Token Allocation Philosophy 07:22

Joe presses on whether individual employees get differentiated token budgets. Henry resists the premise of rigid allocation — he doesn't want to define token limits by job title or seniority — and instead argues for treating speed, cost, and output efficiency as a package. He notes that token unit costs are dropping so fast (halving in weeks) that any policy would be stale before it's implemented, and that the younger talent Baidu is recruiting naturally uses tokens sensibly without being told.

• 07:22 — Joe: presses — would Tracy get a bigger token budget than Joe if she's more productive with AI?
• 08:16 — Henry: rejects defining token limits by title or seniority; wants allocation to be nimble and tied to speed, cost, and output efficiency as a package.
• 08:47 — Henry: token unit cost is dropping so fast — "maybe half in like a few weeks" — that rigid policies would be obsolete before implementation.
• 09:18 — Henry: younger recruits are "getting smarter than people expected" and don't waste tokens; AI is "not only a tool, it's a mindset."
• 03:09 — Tracy introduces the guest: Henry He, CFO of Baidu, recorded live at Bloomberg Invest Asia — Odd Lots' first live recording anywhere in Asia.

Talent Acquisition and Company Culture 09:53

Tracy asks about the fierce competition for top engineers and Baidu's pitch to talent. Henry says the company is one of the few in China still increasing campus recruiting, is restructuring mentor relationships to nurture younger talent, and is giving employees more autonomy — including encouraging "one person company" teams that use AI agents for internal tasks. He highlights an internal open-call tool (nicknamed "doodle" in Chinese) and says the full-stack breadth at Baidu is itself a recruiting differentiator.

• 09:53 — Tracy: asks what Baidu's pitch is to top engineers when US firms treat them "like sports stars" with million-dollar offers.
• 10:56 — Henry: Baidu is among the few Chinese companies still increasing campus recruiting and investing in younger talent and mentor relationships.
• 11:27 — Henry: promotes the "one person company" concept — employees using AI agents as a team; an internal tool nicknamed "doodle" enhances efficiency.
• 12:30 — Henry: the full-stack nature of Baidu — chips, cloud, model, applications — is a unique value proposition for talent in the China tech space.

Custom Silicon Strategy 12:46

Joe observes that American companies are "becoming more Chinese" by vertically integrating into custom silicon, and asks Henry to explain Baidu's rationale. Henry says the computing-power mix has shifted from pre-training (which dominated last year) toward inference and combined tasks. Baidu's chip (Kunlun) is targeted at the inference-and-application segment — not competing with super-scale pre-training silicon — and Henry sees positive network effects between their chip, cloud platform, and applications, with supply and demand well-matched in that defined market.

• 13:04 — Joe: notes US companies are now doing the vertical integration (custom silicon) that Chinese firms have long practiced.
• 13:36 — Henry: computing demand has shifted from pre-training last year to inference and combined tasks now.
• 14:06 — Henry: Baidu's chip focuses on inference and application — a well-defined market with clear boundaries, not super-scale pre-training — and sees positive network effects with their cloud and applications.

Capital Allocation and the "Impossible Triangle" 14:50

Tracy asks how Baidu balances heavy AI capex against returning capital to shareholders. Henry calls this the "Impossible Triangle" — driving growth, maintaining investment density, and keeping ROI conscious — and says the most recent quarter partially resolved it: operating profit nearly doubled quarter-over-quarter, cloud revenue grew 79% year-over-year (nearly double the market growth rate), and operating cash flow turned positive about three quarters ago. Crucially, capex did not double or triple to achieve those results. He stresses the need to look at the full cash-cycle — every dollar spent today takes 20–40 months to earn back, depending on category.

• 15:11 — Tracy: asks how Baidu balances capital allocation across a mature search business and AI infrastructure while returning capital to shareholders.
• 15:41 — Henry: calls it the "Impossible Triangle"; recent quarter: operating profit nearly doubled QoQ, cloud revenue grew 79% YoY, operating cash flow turned positive ~3 quarters ago.
• 16:12 — Henry: capex did not double or triple to achieve those results — the goal is driving growth while keeping AI investment density reasonable.
• 16:44 — Henry: every dollar spent takes 20–40 months to earn full cash back, depending on category; must look at the entire lifecycle and pacing of investment.
• 17:14 — Henry: wants to invest "in a more responsible manner to shareholders" without diminishing ambitious AI investment — "keep the right density."

AI Safety and Alignment Discussion 17:51

Joe asks whether Baidu invests in AI safety and alignment research, noting the "AI psychosis" among American lab leaders. Henry frames alignment as part of the post-training "harness" — the data flywheel, data quality, labeling, and SFT pipeline — and argues China's engineering ecosystem in this area is a comparative advantage: deep talent pools, lower cost, and an established data-labeling and alignment industry dating back to the mobile internet era. He downplays the radical-quantum-leap framing, saying it's "not so radical" and more about applied engineering capability.

• 17:51 — Joe: asks whether Baidu invests in AI safety/alignment and shares the same concerns as American AI lab leaders.
• 18:46 — Henry: frames alignment as part of the "harness" — data flywheel, data quality, labeling, post-training — and says China's engineering ecosystem in this area is a comparative advantage.
• 19:49 — Henry: calls it "not so radical quantum leap" — the talent pool, lower cost, and efficiency of China's data and alignment industry are the key differentiators.
• 20:15 — Joe: presses on US model reports showing models detect testing 4% of the time and change behavior; asks if Baidu does similar research.
• 20:54 — Henry: Baidu is part of the open source community, follows newest research, does own testing; views the community as "collegial," not adversarial.

Government Relations and Policy Environment 21:23

Tracy contrasts US AI companies' self-regulation with China's more hands-on government approach. Henry says China's advantage isn't just engineers but framework designers, and that the country's infrastructure has been upgraded to a world-leading level over 10–20 years. He argues AI regulation isn't a new concept for Chinese policymakers — data ownership and access rules were already built up during the cloud era — and that AI is a new tool layered on existing compliant platforms. He describes the environment as "transparent and open."

• 21:23 — Tracy: contrasts US AI self-regulation with China's government being more hands-on and treating AI as national strategy.
• 22:30 — Henry: China's advantage is not just engineers but framework designers; infrastructure has been upgraded to world-leading level over 10–20 years.
• 23:04 — Henry: data regulation rules were already robust from the cloud era — AI is a new tool on existing compliant platforms, not a new concept for policymakers.
• 23:34 — Henry: describes the policy environment as "transparent and open," with support from policymakers, industry, and academia.

OpenClaw Discussion 24:20

Joe tries to bait Henry with gossip about Dario Amodei — formerly a Baidu employee and now the most hawkish US AI CEO on chip exports to China. Henry deflects neatly, pivoting to a story about "Peter," the founder of the OpenClaw community, who reached out to Baidu asking them to contribute search capabilities to the skill marketplace. Baidu's engineers linked up and joined the marketplace. Henry notes that Ernie 5.1 is now ranked #1 globally in text format and #5 in search-skill capabilities on the global arena. Joe notes the OpenClaw creator "has it out" for Dario, giving the audience "a little drama."

• 24:40 — Joe: asks for office gossip about Dario Amodei, formerly at Baidu and now the most hawkish US AI CEO on chip exports to China.
• 25:12 — Henry: deflects and pivots to a story about "Peter," founder of the OpenClaw community, who posted on Instagram asking to work with Baidu.
• 25:43 — Henry: Peter noticed search is an important capability for the skill marketplace — foundation models lack real-time information, and Baidu (like Google) is a search powerhouse.
• 26:17 — Henry: Baidu's engineers linked up with Peter and joined the skill marketplace; Ernie 5.1 is ranked #1 globally in text format and #5 in search skills on the arena.
• 27:17 — Joe: notes the OpenClaw creator originally called it "OpenClawed," got sued by Anthropic, and "is not the biggest fan of Dario's approach" — giving the audience "a little drama."

Data Advantage and Comparison with Google 27:57

Tracy raises a point from AI analyst Grace Shao that China lacks a data edge because much of its data is unstructured. Henry makes the case that Baidu is the closest structural analog to Google — own TPU-equivalent chip (Kunlun), own cloud, own foundation model (Ernie), own physical-AI application (Apollo Go robotaxi), plus iQIYI for long-form multimodal content. He shares a striking number: Apollo Go delivers ~350,000 trips per quarter across 27 cities, only 20–25% fewer than Waymo's ~500,000 weekly trips globally. Search revenue has dropped below 50% of total (from 80–90%), replaced by growing AI application and software revenue. He admits candidly that different "camps" in China don't share data openly, but argues AI is pushing everything to public cloud, which narrows the gap — China's public cloud penetration is 20–30% vs 90% in the US.

• 27:57 — Tracy: cites AI analyst Grace Shao's view that China lacks a data edge because much collected data is unstructured and hard to harness for model training.
• 28:57 — Henry: lays out the Google parallel — Google has TPU + cloud + model; Baidu has Kunlun chip + cloud + Ernie model + Apollo Go robotaxi + iQIYI for multimodal content.
• 29:29 — Henry: Waymo delivers ~500,000 trips/week globally; Apollo Go delivers ~350,000 trips/quarter across 27 cities — only 20–25% fewer.
• 30:33 — Henry: search revenue has dropped below 50% of total (from 80–90%); AI applications and software are now the growth powerhouse.
• 31:38 — Henry: admits candidly that different "camps" in China don't open up data enough to share — "which is reality."
• 32:09 — Henry: China's public cloud penetration is 20–30% vs 90% in the US, but AI is pushing everything to cloud and narrowing the gap; Baidu's full-stack structure follows the proven Google pattern.

Robotaxi Business Deep Dive 32:53

Joe is excited about the prospect of direct Apollo-vs-Waymo competition, with London as the first head-to-head market. Henry argues robotaxi will fundamentally change car ownership: in the US, the all-in cost of owning a car (insurance, gas, parking) averages $0.60–0.80 per mile — the tipping point between owning and renting. Current robotaxi costs are $1–2.50 per mile but falling fast. Once they cross the $0.60–0.80 threshold, the economics flip: EVs don't need gas, don't need parking (the car drives itself away), and can earn money for the owner while they're busy. Apollo Go shipped cars in London last quarter, partners with Uber, Lyft, and Grab, and operates 24/7 — expanding the midnight market human drivers don't serve. Only two cities globally have over 1,000 robotaxis: San Francisco (Waymo) and one city in China (Apollo Go). Success factors are technology plus operational efficiency, and Baidu has been working on this for 13 years.

• 33:09 — Joe: excited about direct Apollo-vs-Waymo competition, with London as the first head-to-head market between US and Chinese internet giants.
• 34:13 — Joe: asks what dimension will determine the winner — application quality, vehicle production volume, or something else.
• 34:43 — Henry: reveals he's a licensed race car driver; says the key word is "change the car ownership" model.
• 35:16 — Henry: US all-in car ownership cost is $0.60–0.80/mile (insurance, gas, parking) — the tipping point between owning and renting; current robotaxi cost is $1–2.50/mile but falling fast.
• 35:49 — Henry: once robotaxi hits $0.60–0.80/mile, ownership flips — EVs need no gas, no parking, and the car can earn money for the owner while they're in a meeting.
• 36:54 — Henry: Apollo Go shipped cars in London last quarter; partners with Uber, Lyft, and Grab — calling a car from those apps may get you a Baidu car.
• 37:25 — Henry: robotaxis work 24 hours, expanding the midnight market human drivers don't serve; penetration is still very low with huge TAM.
• 37:59 — Henry: only two cities globally have 1,000+ robotaxis — San Francisco (Waymo) and one city in China (Apollo Go); partnerships with Uber/Lyft are "collegial" because demand exceeds supply.

Daily Active Agents (DAA) Metric 38:31

Tracy asks how "daily active agents" — a metric Baidu's founder Robin Li has championed — actually turns into revenue, since it's less obvious than search ad monetization. Henry says DAA measures not just how many people use agents but how difficult the tasks they complete are. He gives a concrete example: an agent deployed at one of China's biggest ports for shipment and logistics planning, saving idle-time costs and improving revenue, with the port sharing profit back to Baidu. The sales process has shifted too — where AI used to be sold to CTOs and CIOs as a cost-center tool, it's now pitched directly to CEOs who have top-down budget authority and see AI as a revenue driver, not a cost.

• 38:31 — Tracy: asks how daily active agents (DAA) — the metric CEO Robin Li has championed — actually turns into revenue, unlike search ads.
• 39:33 — Henry: DAA measures not just usage volume but task difficulty — agents can plan and complete tasks, not just serve as tools.
• 40:37 — Henry: concrete example — an agent deployed at a major Chinese port for shipment and logistics planning, saving idle-time costs; the port shares profit with Baidu.
• 41:42 — Henry: sales process has shifted from CTO/CIO conversations (cost center) to direct CEO meetings (top-down revenue initiative with budget).
• 42:15 — Henry: four cycles — agent reuse across clients lowers cost basis, clients see value and profits, willingness and ability to pay both increase; "very different with the traditional IT."

AI Revenue Models and Digital Employees 42:54

Joe asks whether revenue-sharing or per-task pricing is the model Baidu sees across AI applications. Henry introduces the "digital employees" product line: AI-powered digital humans that do live-stream e-commerce sales, working across time zones and languages, handling Q&A with fluent foundation-model knowledge. Baidu monetizes by charging for result improvement. Tracy jokes about her own "Tao Bao psychosis" — buying three couches for a 500-square-foot apartment — and Henry suggests she'll need an AI agent to help sell them.

• 42:54 — Joe: asks whether per-task or revenue-sharing pricing (e.g., per successful insurance claim resolved) is the model Baidu sees across AI applications.
• 43:49 — Henry: introduces "digital employees" — AI digital humans that do live-stream e-commerce sales 24/7 across time zones and languages.
• 45:22 — Henry: Baidu monetizes digital employees by charging for result improvement; the Q&A fluency comes from the foundation model.
• 45:52 — Tracy: jokes about her "Tao Bao psychosis" — ended up with three couches in a 500 sq ft apartment; Henry: suggests she'll need an agent to help sell them.

Chip Business Spinoff Announcement 45:54

Joe asks Henry to "break some news" on the chip spinoff. Henry reveals that Baidu has filed a confidential filing in Hong Kong to spin off its chip assets, and the process is on track. He notes — surprisingly — that his own background is in chip design (bachelor's and master's), and that he "became CFO by accident." After the spinoff, customers will view the chip products as more neutral and independent, which should expand testing and adoption. He stresses that chips are an ecosystem play — hardware, software developers, suppliers — and a separate listing will help build that ecosystem. The conversation closes with sign-offs and plugs for the Odd Lots newsletter and Discord.

• 46:23 — Joe: "let's break some news" — asks about the chip business spinoff at this first live Odd Lots recording in Asia.
• 46:54 — Henry: reveals Baidu has filed a confidential filing in Hong Kong to spin off its chip assets; process is on track.
• 47:25 — Henry: after spinoff, customers will view chips as more neutral and independent; it's an ecosystem play — hardware plus software developers plus suppliers.
• 47:54 — Henry: reveals his own training was in chip design (bachelor's and master's) — "I became CFO by accident."
• 47:59 — Joe: recalls that Baidu figured out a way for developers to easily port their stack to Baidu's chip stack; thanks Henry for coming on.
• 48:31 — Sign-offs: plugs for Odd Lots newsletter (bloomberg.com/oddlots), Discord (discord.gg/oddlots), and producer handles.