Baidu Submits Confidential HK Filing to Spin Off Chip Unit

百度提交芯片业务赴港分拆上市秘密申请 首席财务官何俊杰表示，百度正在扩大AI推理算力，并分拆其昆仑芯芯片业务以服务外部云端及企业客户。 百度首席财务官何俊杰阐述了公司如何从中国领先的搜索引擎转型为涵盖自研芯片、基础大模型、云服务和自动驾驶的垂直整合全栈AI巨头。他谈到80%的Token需求已转向推理，促使百度围绕20至40个月的现金回收周期来规划硬件投资。他还透露了关键运营指标，包括Apollo Go（萝卜快跑）在27座城市每周提供35万次自动驾驶叫车服务，以及搜索业务占公司总收入比重已降至50%以下。最后，何俊杰探讨了百度面向企业AI智能体的成效导向定价模式、电商数字人主播，以及半导体业务昆仑芯赴港分拆上市的秘密申请。 百度CFO何俊杰透露，传统搜索业务收入在最近一个季度已降至总收入的50%以下（约48%），增长重心正全面转向AI云、企业级Agent与自动驾驶业务。 (30:00) 随着行业从模型预训练转向应用落地，推理计算已占据增量Token需求的约80%；百度将自研芯片与云基础设施重点对准商业推理，并通过20至40个月的现金回收周期严格把控资本开支与服务器部署。 (4:12) 在自动驾驶领域，萝卜快跑（Apollo Go）在27个城市已实现每周约35万单的运营规模，与Waymo（约50万单/周）正共同推进全球Robotaxi落地，目前全球仅有旧金山与武汉两座城市的车队规模超过千辆。 (30:00) 百度正推动从日活跃用户（DAU）向日活跃Agent（DAA）的商业模式演进，采用按实际降本增效成果分成的收费方式，并在物流港口与24小时电商数字人直播中落地变现。 (43:45) 百度已在香港就自研芯片业务（昆仑芯）提交了保密分拆上市申请，旨在以中立独立公司的身份加速第三方开发者生态拓展并支持无缝迁移CUDA代码库。 (50:35)

Baidu Chief Financial Officer Henry He argues that the company has transitioned from a search-dependent business into an integrated, full-stack AI enterprise centered on cloud infrastructure, inference workloads, and autonomous operations. He contends that long-term commercial returns in artificial intelligence will be determined by disciplined capital allocation, outcome-based software pricing, and real-world operational scale rather than speculative frontier pre-training. Full-Stack Architecture and Inference Demand (4:12) Baidu has restructured its core profile away from search advertising, with traditional search dropping to roughly 48% of total revenue in a recent quarter as cloud services and AI applications expanded. Rather than prioritizing model pre-training clusters, He explains that roughly 80% of incremental token demand stems from commercial inference. Baidu AI Cloud operates as the core orchestration platform linking hardware, the proprietary Ernie model suite, and third-party models to customer applications. He frames technical challenges such as safety, alignment, and data sanity as solvable data engineering tasks rather than theoretical dilemmas. Emphasizing domestic execution strengths in supervised fine-tuning and data pipelines, He reports that Baidu's Ernie 5.1 achieved top global rankings in arena benchmarks for text tasks and fifth in search integration, which provides real-time contextual knowledge missing from static foundation models. Capital Discipline and Silicon Spin-Off (15:30) To address what He terms the "impossible triangle" of balancing high capital expenditure, operating profitability, and shareholder returns, Baidu tracks hardware spending against a 20- to 40-month cash recovery lifecycle. The company ties infrastructure investments to concrete revenue traction, reporting 79% year-over-year cloud revenue growth and a near doubling of operating profit quarter-over-quarter. Hardware investments focus primarily on cost-effective silicon tailored for inference workloads. He confirmed that Baidu has filed confidentially in Hong Kong to spin off its semiconductor subsidiary, Kunlunxin, into an independent publicly listed company. Operating the chip unit independently is intended to eliminate customer concerns regarding platform lock-in, supported by software compatibility layers that allow developers to transition existing CUDA codebases directly onto Baidu hardware. Apollo Go and Fleet Economics (40:00) Baidu operates its Apollo Go robotaxi fleet across 27 cities, delivering approximately 350,000 weekly trips compared to Alphabet subsidiary Waymo's estimated 500,000 trips. Globally, only Wuhan (operated by Baidu) and San Francisco (operated by Waymo) host active autonomous vehicle fleets exceeding 1,000 units. To expand international operations, Baidu has deployed vehicles in London and formed partnerships with ride-hailing networks including Uber, Lyft, and Grab. He argues that replacing personal vehicle ownership with autonomous fleets depends on driving operating costs down from current levels of $1.00 to $2.50 per mile to a threshold of 60 to 80 cents per mile. Reaching this economic tipping point over the next five to ten years relies heavily on continuous operational optimizations, such as off-peak fleet utilization, routing, and charging network management. Outcome-Based Models and Daily Active Agents (43:45) Baidu is shifting its enterprise software metrics from Daily Active Users (DAU) toward "Daily Active Agents" (DAA), reflecting systems capable of completing multi-step workflows autonomously. Internally, Baidu fosters a "one-person company" structure using platforms like Doodoo and DuoMate to execute administrative and engineering tasks, measuring token expenditure by task completion and unit cost rather than imposing rigid usage caps. In enterprise sales, this functional capability enables outcome-based pricing rather than standard software subscriptions. For example, Baidu's FAMO logistics tool takes a percentage of the revenue generated from container throughput gains at shipping ports. Similarly, 24/7 autonomous "digital employees" deployed in live-stream e-commerce are billed based on performance milestones and incremental sales lift, shifting procurement conversations from technical cost centers directly to corporate leadership. While Baidu has achieved substantial operational volume in autonomous driving and rapid growth in its cloud division, the long-term viability of its capital-intensive initiatives remains subject to whether autonomous fleet costs can successfully decline to the 60-to-80-cent per mile threshold and whether enterprise demand for outcome-based AI agents sustains broad adoption.

百度首席财务官何俊杰（Henry He）阐述了公司如何从传统搜索引擎转型为全栈人工智能企业。他指出，随着大模型由预训练向实际应用推理阶段过渡，百度正以AI云为核心中枢，通过垂直整合芯片、自动驾驶规模化运营以及按成果分成的智能体商业模式开辟新增长曲线，并依托严苛的投资回报周期平衡资本开支与盈利能力。 业务架构重塑与AI云平台 (4:12) 何俊杰指出，百度的营收结构正在发生深刻变革。外界普遍认为传统搜索仍占据百度绝大部分收入，但在最近一个季度，传统搜索业务收入占比已降至约48%，快速增长的AI云、企业应用软件以及数字人业务正在填补差额。在硬件、模型、云平台与应用的全栈架构中，何俊杰将百度AI云定义为连接底层算力与上层商业化变现的核心中枢。该平台不仅承载自研的文心大模型（Ernie），同时也向第三方开源模型开放。 算力消耗格局的转变正在重塑基础设施规划。何俊杰透露，当前新增的Token处理需求中，高达80%由下游实际推理而非模型预训练驱动。为了适应这一趋势，百度强化了实时检索对大模型的赋能，以弥补基础模型缺乏实时世界信息的缺陷。在最新的全球大模型竞技场（LLM arena）基准测试中，旗舰模型Ernie 5.1在文本格式任务中位列全球第一，在搜索技能维度位列全球第五。 资本纪律与昆仑芯独立分拆上市 (15:30) 针对大规模资本支出、核心业务盈利能力与股东回报之间的矛盾，何俊杰将其称为资本分配的“不可能三角”。他表示，百度拒绝跟随市场狂热盲目堆叠算力，而是将硬件开支严格锚定在实际业务增长与投资回报率上。百度近期云业务收入同比增长79%，几近中国公有云行业整体增速的两倍，季度运营利润环比翻倍；公司针对服务器等硬件资产执行20至40个月的严格现金回收周期评估，综合考量硬件折价与数据中心运营开销。 在底层芯片方面，何俊杰解释了差异化的硬件路线：大模型预训练需要耗资巨大的超大规模算力集群，而百度自研芯片昆仑芯（Kunlunxin）则优先针对高吞吐量、低成本的商业推理和应用场景进行优化，并实现对CUDA代码生态的平滑兼容。何俊杰在访谈中正式确认，百度已就昆仑芯资产向香港联交所递交了秘密分拆上市申请。他认为，将芯片业务独立拆分为上市公司，能够消除外部企业客户因同业竞争产生的顾虑，建立中立且开放的第三方半导体生态。 自动驾驶规模扩张与出行经济学 (40:00) 在物理AI领域，何俊杰重点对比了百度Apollo Go（萝卜快跑）与Alphabet旗下Waymo的发展现状。目前Waymo在美国每周提供约50万次出行服务，而Apollo Go已在全球27个城市实现每周约35万次服务，运力规模达到Waymo的70%左右，从而源源不断地回流多模态训练数据。在车辆规模方面，全球目前仅有旧金山（Waymo运营）和中国武汉（Apollo Go运营）两座城市部署了超过1000辆规模的活跃无人驾驶车队。Apollo Go亦在伦敦展开路测，并与Uber、Lyft及东南亚的Grab等平台建立合作推进国际化。 对于自动驾驶取代私家车的临界点，何俊杰从出行经济学角度预测，当每英里综合出行成本降至60至80美分区间时，消费者将普遍转向无人网约车服务。当前全球Robotaxi的运营成本因未形成极致规模仍处于每英里1.00至2.50美元的高位，但他预计随着制造规模扩大与技术成熟，未来5至10年内运营成本将进入该经济临界点。无人车具备全天候运营能力，在高停车费与高油价密度的城市环境中，其商业效率将显著优于闲置率高昂的私人车辆。 智能体商业化与结果导向定价 (43:45) 在企业软件服务方面，何俊杰阐述了百度商业模式从移动互联网时代向大模型时代的跃迁。百度董事长兼CEO李彦宏提出了“日活跃智能体”（Daily Active Agents，简称DAA）这一全新核心指标，以此取代过去的“日活跃用户”（DAU）。何俊杰解释称，智能体具备自主规划并执行多步骤复杂任务的能力，这从根本上颠覆了传统软件按席位或许可证收费的商业模式。 百度正在推进基于业务成果分成的全新变现机制。何俊杰以中国大型海运港口部署的FAMO企业智能体为例，该系统通过统筹物流调度大幅缩短了集装箱闲置时间并提高吞吐量，直接为港口创造增量收益，百度则从中抽取一定比例的利润分成。这种模式将销售对话从控制IT开支预算的技术部门直接推升至掌握顶层财务决策权的CEO层面。此外，百度的“数字员工”正在规模化替代电商直播等密集型人力场景，凭借24小时不间断跨语种多轮互动，按带货销售增量向商家收取效果佣金。 工程化治理与“一人公司”组织模式 (20:00) 探讨模型对齐与安全监管时，何俊杰强调百度将对齐、数据清洗与后训练视为务实的软件和数据工程挑战，而非不可解的存在主义危机。得益于数据飞轮构建、精细化标注以及监督微调（SFT）等工程能力，中国企业能够在本土工程红利下高效构建模型安全防护栏。在宏观政策上，他指出过去20年建立的数据合规与云监管体系早已为AI应用划定了明确法律边界；与此同时，百度积极参与全球开源基准评测，认为技术开发者群体之间依然保持着开放协作关系。 在内部生产力管理上，百度推行“一人公司”理念，构建了名为Doodoo的内部智能体平台，赋能初级工程师借助DuoMate等自主智能体端到端独立负责过去需要多人员团队协作的复杂业务。公司不以员工级别或工龄限制内部算力访问，而是全力鼓励内部Token使用，并在短短数周内实现了Token单位处理成本腰斩。百度评估算力投入产出比的核心标尺并非预训练阶段的基础算力消耗，而是各业务线在响应速度、单位成本与智能体最终任务交付效率上的实际提升。 百度展示了从搜索巨头向全栈AI企业转型的完整路径，其核心在于将行业竞争由底层算力军备竞赛引向应用级推理效率与成果分成变现。然而，该战略的全面落地仍面临若干核心考验：昆仑芯在香港独立分拆后能否真正打破生态壁垒、被外部广泛采纳，Robotaxi运营成本能否在5至10年内如期压缩至每英里80美分以下，以及按成效分成的智能体定价能否克服政企采购的传统惯性全面铺开，仍有待市场长期检验。

## Background 本期节目由彭博（Bloomberg）旗下的旗舰播客《Odd Lots》于 2026 年 6 月 29 日在 RSS 及网络平台发布。节目简介指出，百度如今已转型为贯穿自研芯片、文心大模型、智能云以及自动驾驶“萝卜快跑”（Apollo Go）的“全栈式 AI 玩家”，本期邀请到百度集团首席财务官何俊杰（Henry He），深度探讨百度的 Token 消耗与预算哲学、模型安全对齐、全球 Robotaxi 竞争格局，以及传统搜索业务在当前战略中的定位。 ## TL;DR 在节目中，百度 CFO 何俊杰从资本分配与工程落地两重视角，全面拆解了百度的全栈 AI 战略布局。他强调百度将智能云视为承接 80% 推理算力红利的核心枢纽，并倡导以“任务完成度”而非机械的 Token 数量来衡量企业 AI 生产力与投资回报率。何俊杰指出百度传统搜索收入占比已历史性降至 48% 左右，取而代之的是 AI 云、具身智能及数字员工等新引擎；同时分享了“萝卜快跑”在周均订单量上紧咬谷歌 Waymo、进军伦敦正面交锋的全球进展，并正式证实了旗下昆仑芯业务正推进赴港独立分拆上市的计划。 ## Atmosphere & style 整场对话在彭博香港投资峰会的线下现场录制，气氛热烈、专业且张弛有度。彭博名嘴 Joe Weisenthal 与 Tracy Alloway 抛出了一连串关于巨额资本开支、硅谷“AI 精神病”、工程师天价薪酬等尖锐而具体的犀利质问，而工科背景出身的何俊杰则以极其务实、风趣且松弛的状态从容应答。从赛车手视角畅谈每英里出行成本颠覆汽车所有权，到打趣 Tracy 的“淘宝精神病”需要买手 Agent 把关，现场笑声与干货交织，真实生动地呈现出中国科技掌舵者在中美 AI 巅峰对决中的底气与商业定力。 ## Bulletin - `03:40` 百度 CFO 何俊杰透露，全行业新增 Token 消耗中已有 80% 来自推理端，百度算力重心正全面由万卡预训练转向云端推理与应用承接。 - `30:00` 百度传统搜索业务营收占比已历史性降至约 48%，AI 云计算、数字人与企业级 AI 软件正在重构其底层收入结构。 - `30:00` 百度的无人驾驶出租车 Apollo Go 已在 27 城运营，周均订单约 35 万单，已达谷歌 Waymo 全美约 50 万单体量的 70% 以上，并已进军伦敦展开正面对决。 - `51:15` 何俊杰证实百度旗下自研芯片业务昆仑芯已向香港联交所递交分拆上市申请，旨在确立中立第三方定位并加速吸纳外部 CUDA 开发者生态。 ## Key takeaways - `03:40` 在自研芯片、基础大模型、智能云和上层应用的 AI 全栈布局中，百度将智能云视为绝对不能输的核心枢纽，用于开放托管各类模型并承接 80% 的推理红利。 - `05:40` 企业衡量 Token 价值的尺度已从闲聊转向真实世界的任务完成度，百度在内部管理上摒弃了教条化的职级定额，而是结合单位成本断崖式下跌推行敏捷自驱机制。 - `10:00` 面对顶尖人才争夺，百度逆势扩大校招并推崇“一人团队”文化，借助内网智能体工具将员工转变为能独立完成闭环业务的自主单元。 - `12:35` 昆仑芯不盲目追求边界无限的万卡级超大规模预训练，而是聚焦于高增长、边界清晰的云端推理与应用优化增量市场。 - `16:45` 百度在维持重资产 AI 投入的同时保持资本纪律，智能云营收同比增长约 79%，且在全生命周期 ROI 约束下使经营性现金流持续转正。 - `20:00` 针对所谓“AI 精神病”与对齐焦虑，中国科技业更侧重依托成熟且高性价比的工程体系、数据飞轮与后训练来务实保障模型的鲁棒性与工程稳定性。 - `26:15` 由于基础大模型每隔数月更新且缺乏实时事实，如 OpenClaw 等爆火智能体平台必须集成百度的实时搜索技能来补充最新动态上下文。 - `26:15` 文心大模型 5.1 在 LMSYS Chatbot Arena 文本榜单中位列全球第一，搜索技能位列全球第五。 - `33:15` 中国企业公有云渗透率在生成式 AI 爆发前仅为 20%-30%（美国约为 90%），AI 浪潮正强力倒逼国内数据跨过“花园围墙”加速向云端整合。 - `40:00` Robotaxi 的核心经济拐点在于每英里成本降至 60 到 80 美分，一旦跌破该临界点，汽车的本质将从私人资产颠覆为替主人自主接单创收的物理 AI Agent。 - `43:24` 百度未来的核心商业指标正在从 DAU（日活跃用户）转向 DAA（日活跃智能体），其商业模式也从按人头购买许可证演变为按任务效果或利润增量分成。 - `43:24` 工业级调度 Agent“伐谋”在大规模港口部署后，通过优化泊位与物流调度降低闲置时间，百度得以直接与客户 CEO 按照带来的增量利润进行分成。 - `47:00` 百度的数字员工技术已被广泛应用于电商直播带货，能克服人类生理极限实现 24 小时全天候无间断播报、多语种切换及跨文化本土幽默表达。 - `51:15` 昆仑芯独立分拆上市旨在摆脱竞争对手附属品的刻板印象，以中立芯片供应商身份吸引外部客户，并降低开发者由 CUDA 迁移的适配门槛。 ## Detailed walkthrough ### `00:00` 开场与宏观背景：中美 AI 正面交锋下的香港之行 节目由赞助商广告拉开序幕，其中 IBM 展示了其利用企业级 AI 协助 30 万全球员工处理人力资源咨询、解决 94% 常见问题的商业化落地成果，而这恰好切中了本期播客的核心议题——AI 正在告别单纯的预期画饼，转向实打实的商业产出。 彭博旗舰播客《Odd Lots》的主持人 Joe Weisenthal 与 Tracy Alloway 随后登场。两人刚刚结束了香港之行，Tracy 感慨自己已有四年没回香港，城市的总体变化虽然比预想中要少，但在此刻造访香港却极具象征意义。本次他们前往香港是为了参加彭博投资峰会（Bloomberg Invest Conference），期间还在香港举办了首场美国本土之外的《Odd Lots》线下知识竞答之夜（Trivia Night）。更重要的是，当前全球资本市场最瞩目的焦点正是中美两国在人工智能领域的竞逐，而香港之行让他们得以直接对话那些在中国科技巨头内部执掌数十亿资本配置大权的最高决策者。 Joe 指出，全球科技版图正在发生根本性的范式转移：在过去的互联网时代，中国互联网巨头深耕国内市场，美国巨头则吃下全球其余大部分份额，双方井水不犯河水；但在如今的生成式 AI 与自动驾驶（Robotaxi）赛道上，中美科技公司首次在技术模型、基础设施甚至海外市场展开了真正的正面对决，美国企业甚至也在调用中国团队研发的 AI 模型。在这一历史转折点上，他们迎来了极其重磅的访谈嘉宾——百度集团首席财务官（CFO）何俊杰（Henry He）。 --- ### `03:40` 全栈布局下的资本取舍：为什么百度将“云”视为必胜底牌？ 对谈伊始，Joe 直奔主题：在当下的科技叙事中，百度是中国极为罕见的“全栈式（Full-Stack）AI 玩家”，其业务纵深贯穿自研芯片（昆仑芯）、文心大模型（Ernie）、百度智能云平台以及最上层的产品应用。作为掌管全盘资金分配的 CFO，如果必须在资源分配上做取舍，百度技术栈中哪一层是“绝对不能输、必须拿下胜利”的核心阵地？ 面对这个开门见山的刁钻提问，何俊杰笑称主持人把最难的压轴考题直接放到了开场。他首先为整个行业画出了清晰的技术演进坐标：整个 AI 产业正在经历两重至关重要的结构性迁移——**在底层架构上，正从硬件基础设施向实际应用层转移；在模型范式上，正从基础大模型（Foundation Models）向智能体（Agents）进化**。 ```mermaid flowchart TD accTitle: Baidu Full-Stack AI Architecture accDescr: Diagram showing Baidu's four AI layers and highlighting Cloud as the central hub connecting chips, models, and applications. subgraph Infra["底层支撑"] Chips["自研芯片 (昆仑芯)<br/>提供算力基底 / 支撑推理"] end subgraph Hub["核心枢纽 (必选基石)"] Cloud["百度智能云<br/>开放生态 / 托管文心与第三方模型"] end subgraph Engine["能力引擎"] Ernie["文心大模型 (Ernie)<br/>能力泛化与智能体框架"] end subgraph Apps["价值兑现"] AppLayer["应用层 & 智能体<br/>任务交付 / 商业化变现与 ROI 闭环"] end Chips --> Cloud Ernie --> Cloud Cloud --> AppLayer ``` 何俊杰拆解了全栈协同的逻辑：芯片是算力底座，大模型是连接软硬件能力的桥梁，云是能力的输出与部署网络，而所有的商业化变现与投资回报率（ROI）最终都沉淀在应用层，缺了任何一环商业闭环都无法跑通。但如果像学生考试做单选题（A、B、C、D）那样被逼只能选一个，何俊杰给出的答案是 **C——云平台**： - **生态开放性**：云不仅承载自家的文心大模型，更保持高度开放，能够托管全球各类第三方优质模型。 - **算力推理重心转移**：接入云平台的自研芯片在推理场景中扮演着关键角色。何俊杰透露，当下模型预训练（Pre-training）固然重要，但**全行业新增的 Token 消耗需求中，已有高达 80% 均来自推理（Inference）端**。云平台正是承接这 80% 推理红利的关键载体。 --- ### `05:40` 重新定义企业 ROI：从算力消耗到任务完成度的“Token 预算”哲学 随后，Tracy 抛出了一个未来可能成为所有财经记者标配的问题：“在今天的企业管理中，如何看待百度的‘Token 预算（Token Budget）’？你们究竟如何衡量在 Token 消耗中所获得的生产力与实际回报？” 何俊杰将百度的 Token 消耗划分为截然不同的两个维度： 1. **研发型消耗（R&D Efforts）**：这是技术极客层面的追求，旨在消耗算力以冲击更高的 AGI 技术标杆、改进模型基准测试表现，并设计更严苛的测试框架（Harness）来压榨技术极限。 2. **任务完成型消耗（Task Completion）**：这是作为赋能千行百业的科技巨头最看重的指标，即模型与框架在现实世界中究竟完成了多少具体任务。 何俊杰特别强调，今年相较于去年的质变不仅在于底层基础大模型变得更聪明，更在于应用框架的打通。AI 的价值坐标已经彻底从**“与用户闲聊（Chatting）”跨越到“动手干活并彻底解决问题（Completing）”**。因此，衡量 Token 生产力最重要的尺度是“任务完成度”。 针对内部资源分配，Joe 紧追不舍地追问了一个极为具象的管理场景：如果团队内部有两名员工，管理层是否会给他们完全一样的 Token 配额？CFO 是否有一套监测机制，能够识别出某个人更能榨取 AI 的价值，从而直接批给他 10 倍于普通人的 Token 预算？ 何俊杰幽默地调侃道，冲着 Joe 问出这么敏锐的问题，下次就应该给他批更多的 Token。但他随后阐述了一套反传统的动态管理哲学： - **拒绝教条化的职级定额**：技术迭代日新月异，绝不能用传统的层级制度或员工头衔来僵化规定每个人能用多少 Token。百度选择保持敏捷与开放，给予员工充足的 Token 额度去探索内部研发。 - **单位成本的断崖式下跌**：与其靠行政指令卡死配额，不如依赖技术降本。目前单位 Token 的成本下降速度极快，往往在管理层构思好一套配额政策的几周时间里，底层的调用成本就已经腰斩。因此，管理的核心是将**迭代速度、单位成本、产出效率**三者作为一个综合包来评估，而非孤立地数 Token 数量。 - **年轻一代的自驱与智能体思维**：何俊杰观察到，即便像百度这样拥有 20 多年历史的成熟上市公司，如今大量涌入的年轻新血液比管理层预想的更具悟性。年轻员工不会随意挥霍算力，因为他们深度融入了智能体的工作流中，模型本身也会反向协助他们梳理任务优先级。他总结道，今天的 AI 已经超越了单纯的生产力工具（Tools），它正在重塑一种高度智能化、自适应的工作心智（Mindset）。 ### `10:00` 百度的人才招募策略与内部“一人团队”变革 面对全球科技行业对顶级人工智能工程师白热化的争夺——尤其是在美国顶尖工程师被当成“体育明星”般开出数百万美元天价的背景下，百度的首席财务官何俊杰（Henry He）从组织与文化演变的角度给出了百度的破局思路。他指出，过去科技公司的核心使命是赋能外部客户，提供技术工具帮助他们从传统 IT 架构迁移至云端；但生成式 AI 的爆发同样深刻地重塑了企业自身，AI 赋能内部组织与文化变革，其重要性丝毫不亚于对外服务客户。 在具体的人才招募与留存举措上，百度展现出与许多同行截然不同的策略： - **逆势扩大校园招聘**：在行业普遍收缩招聘的大环境下，百度依然保持极高开放度，持续加大校园招聘力度，深耕年轻人才储备。 - **重构导师机制**：要求资深技术骨干跳出传统的上下级层级汇报关系，与年轻一代骨干建立真正的导师（Mentor）带教机制。 - **推崇“一人公司”与内部“一人团队”**：在外部，百度乐于服务“一人公司”（One-Person Company）形态的创业者，这类客户能够高效利用百度的 AI 工具创造价值并愿意为高品质付费；在内部，百度同样鼓励员工成为借助 AI Agent 独立负责闭环任务的“一人团队”。百度内部研发并推行了名为协同提效的内网智能平台（内部昵称工具），让员工利用智能体处理大量日常事务，实现更扁平、更具自主性的组织进化。 - **全栈生态的独特价值**：百度在中文科技圈具备极高辨识度的“全栈”（Full Stack）布局，工程师能够在一套完整的体系内纵览底层芯片、框架、基础大模型到终端应用，这种纵深实践机会对顶尖工程师而言是独一无二的吸引力。 --- ### `12:35` 自研芯片的战略定位：瞄准推理与应用增量市场 主持人注意到一个耐人寻味的趋势：美国科技巨头近来反而越来越像中国企业，纷纷走向垂直一体化，甚至争相下场设计并销售自研芯片（Silicon）。而百度早已布局昆仑芯多年。针对自研芯片的初衷究竟是为了“掌控自身供应链命运”还是为了“让硬件与模型架构深度契合”的提问，何俊杰从行业计算周期的演进切入，阐明了百度的差异化算力哲学。 他分析道，过去一两年全行业的算力消耗绝大多数集中在超大规模基础模型的预训练（Pre-training）阶段，但眼下算力结构正在发生显著转变，大量的算力消耗正快速向推理（Inference）与任务落地迁移。百度并不打算将自研芯片死磕在边界无限、不计成本的万卡级超大规模预训练场景，而是精准切入一个边界清晰、高速扩张的增量市场——即云端推理与应用场景。百度的自研芯片专为支撑百度智能云的推理与应用优化，在该特定领域内形成了显著的正向网络效应，在供给端与需求端实现了极高的适配度与确定性。 ```mermaid flowchart TD accTitle: 百度芯片与算力战略演进 accDescr: 展示百度算力结构从预训练向推理应用转移，并由自研芯片支撑云服务的闭环逻辑。 subgraph Market_Shift["计算需求转移"] direction TB A["过去：基础模型预训练<br/>(高算力消耗、边界宽泛)"] --> B["现在：模型推理与具体任务落地<br/>(高速增长、边界明确)"] end subgraph Baidu_Strategy["百度全栈协同"] direction TB C["自研芯片 (昆仑芯)"] --> D["百度智能云平台"] D --> E["企业级推理与 AI 应用服务"] E -->|"正向网络效应 / 供需匹配"| C end B --> Baidu_Strategy ``` --- ### `16:45` 破解资本分配的“不可能三角”：在 AI 狂潮中保持投资纪律 在谈到动辄数百亿美元甚至千亿美元的生成式 AI 资本军备竞赛时，主持人询问百度如何在维持庞大搜索主业、推进 AI 全栈基础设施重资产投入的同时，平衡资本配置与对股东的现金回报。何俊杰坦言，面对市场上各家科技巨头令人咋舌的资本开支（CapEx）新闻，他每隔几周也必须“抓头”深思，并将其形象地称为企业管理中的**“不可能三角”**。 何俊杰强调，百度在最近的季度财报中已经用扎实的业绩数据部分破解了这一困局： - **经营利润显著跃升**：季度运营利润环比实现接近翻倍的强劲表现。 - **云业务逆势高增长**：百度智能云营收同比增长约 79%，增速几乎达到同期中国整个云计算市场平均年化增速的两倍。 - **现金流转正与开支克制**：自前一年第三季度以来，百度经营性现金流已持续转正；更关键的是，在利润和云收入高增的同时，资本开支（CapEx）并未出现成倍或爆发式的失控扩张。 在重资产的 AI 周期中，优秀的管理层必须在驱动业务快速增长的同时，将 AI 投资的密度与节奏控制在合理区间。何俊杰指出，做每一项基础设施投资都必须具备全生命周期的现金回报（ROI）意识。当下投出的每一块钱，根据具体资产类别，可能需要 20 个月、30 个月乃至 40 个月才能完成全额现金回收。在这一漫长的周期内，企业还必须直面高带宽内存（HBM）价格上涨、IDC 智算数据中心稀缺以及服务器采购成本高企等多重挑战，因此绝不能盲目追随市场炒作，必须严格紧扣全生命周期回报率精细化运作。 ### `20:00` 资本配置的节律与所谓“AI 精神病”：工程化落地 vs 理论担忧 谈及大模型的研发投入，百度首席财务官何俊杰（Henry He）强调，基础模型的研发周期通常长达数月，因此节奏控制至关重要。百度在资本配置上的策略是：以对股东更加负责的态度进行投资，同时绝不削弱对 AI 投资的雄心，核心在于“保持合理的资本密度”。从该季度财报表现来看，百度实现了预期的节奏平衡，希望未来半年依然能维持“高业务增长、更少资金消耗、更优投资回报率（ROI）”的模式。 主持人随即抛出一个略带调侃却十分尖锐的问题：美国顶尖 AI 实验室的负责人们似乎患有不同程度的“AI 精神病（AI psychosis）”，每天都在极度焦虑地研究“对齐（alignment）”、防止 AI 失控并确保其造福人类。主持人追问，百度是否也有同样的焦虑？是否也为防范模型“造反（go rogue）”拨出了可观的专项资本？何俊杰对此给出了务实的技术视角：在数据健全性（data sanity）、监督微调（SFT）和后训练（post-training）领域，对齐确实是关键环节，但评估模型不能只看预训练和刷榜成绩，评测基准套件（harness）的核心在于衡量模型的鲁棒性与工程稳定性。 何俊杰明确指出，在中国科技行业，工程能力一直是最核心的竞争优势。从数据飞轮、数据标注、质量清洗，到对齐检测与后训练，这套工程生态早在移动互联网时代就已相当成熟。在人才储备、数据处理成本与全链路工程效率上，中国生态体系具备极高的性价比。在他看来，这本质上是一项“工程问题，而非理论上的量子跃迁”。中国科技产业依托优秀的工程师红利与成本优势，能够扎实地解决鲁棒性与对齐问题，因此在本土市场中，大家并不像硅谷同行那样过度神经质地担忧这类风险。 ```mermaid flowchart TD accTitle: 百度对齐与模型鲁棒性的工程化飞轮 accDescr: 展示百度如何依托全链路工程生态解决模型安全与对齐问题 subgraph DataOps["数据健全性与准备"] D1["数据标注与清洗"] --> D2["数据质量检测"] D2 --> D3["数据飞轮流转"] end subgraph PostTrain["后训练与调优"] P1["监督微调 (SFT)"] --> P2["对齐检验 (Alignment Checks)"] P2 --> P3["鲁棒性评测 (Harness)"] end subgraph Advantage["核心竞争壁垒"] E1["工程师人才红利"] E2["全链路高成本效益"] E3["工程化解决落地而非空谈理论"] end DataOps --> PostTrain Advantage -.-> DataOps Advantage -.-> PostTrain ``` - 主持人提及美国前沿实验室发布的模型报告（例如思维链中出现 4% 的概率能察觉自己正处于测试中并据此改变行为，被视作潜在未对齐的迹象）。 - 何俊杰回应称，百度也是开源社区的一员，始终跟进业界最新的思维链与对齐评测研究，并进行内部自测。他认为当前的全球开源社区氛围非常融洽（collegial），大家的目标都是打造面向全球的高水平前沿模型。 --- ### `23:45` 政策导向、合规生态与数据监管的连续性 主持人转入一个相对敏感但绕不开的议题：美国 AI 企业目前主要依靠自我监管并自主发布评测报告；而在中国，人工智能被明确列入国家战略和国家安全范畴，政府在产业规划与监管上介入更深。身处与国家产业政策深度交织的环境中，百度的模型研发与技术路线受到怎样的影响？ 何俊杰虽然表示不便直接评判公共政策，但他分享了产业视角的深度观察： 1. **系统性基础设施与人才红利**：中国 AI 的优势不仅在于顶尖的算法工程师，还在于能够设计整体技术框架与系统的人才。大模型不是单纯的算法问题，而是涵盖算力基础设施、数据监管、模型层及云平台的一体化工程。过去一二十年中国数字基础设施的升级已被证明走在世界前列，这种政策支持与监管演进的路径是行之有效的。 2. **监管框架在云计算时代的延续性**：何俊杰阐述道，数据透明度、数据所有权、使用权与访问权限等问题，早在过去五到二十年的云计算时代就已经确立了严密的规则体系。AI 大模型对于监管机构而言并非前所未见的虚无概念，而是访问和利用已有资源的一种更高效的新工具。在已有完全合规的云平台架构上，政策制定者、产业界和学术界一直保持着高频互动与透明沟通，监管机构也通过各种渠道积极倾听前沿技术提出的新问题。 --- ### `26:15` 辟谣与转折：从前员工 Dario 的传闻到 OpenClaw 的实时搜索合作 主持人随后试图打探一则八卦：Anthropic 联合创始人 Dario Amodei 过去曾是百度的研究人员（曾供职于百度硅谷 AI 实验室），如今他在美国对华芯片出口管制等政策上立场极为强硬，主持人好奇百度内部是否还会偶尔提起这位前员工。 何俊杰幽默地“把球打回给主持人”，表示与其聊过去的传闻，不如分享一个刚刚发生、更有意思的圈内轶事： 过去百余天里，社区中类似“OpenClaw”（因避免商标侵权而常有昵称）这类开源智能体工具爆火，迅速向市场普及了“AI 如何进入真实世界执行复杂任务”。此类智能体工具需要调用各种具体技能（skills），而开发者社区的创始人 Peter 发文表示急需与百度合作——因为基础大模型每隔数月重新训练一次，模型内部权重缺乏实时世界的信息（“例如昨天的基础模型根本不可能知道 Joe 和 Tracy 今天在节目里聊了什么”），必须通过实时搜索来捕获瞬态信息。 ```mermaid flowchart LR accTitle: 实时搜索技能为智能体提供上下文流 accDescr: 基础大模型通过集成百度的实时搜索技能获取当前世界的新鲜事实 subgraph AgentEnv["智能体与执行环境"] OC["OpenClaw 智能体平台"] --> SK["技能调用中心 (Skills)"] end subgraph RealTimeData["实时动态信息"] BD["Baidu 实时搜索技能"] --> RT["最新事实、突发事件与实时讨论"] end SK <--> BD RT --> OC ``` 何俊杰指出，在全球范围内 Google 是实时信息的核心入口，而在中文与中国区域，百度则是无可替代的实时搜索中枢。百度团队在看到消息后第二天便安排工程师与 Peter 的团队对接，正式入驻其技能市场，为智能体接入百度的实时搜索能力。借此契机，何俊杰公布了百度的最新技术战报：百度文心大模型 5.1（Ernie 5.1）在权威评测 LMSYS Chatbot Arena（即 LIM Arena）的纯文本榜单中位列全球第一，其搜索技能相关能力位列全球第五。至于 Dario 的陈年八卦，何俊杰笑着表示“可以在下播闭门之后再聊”。 - 智能体生态的核心瓶颈在于基础模型缺乏实时上下文，搜索巨头在实时知识检索领域的壁垒难以被纯静态模型替代。 - 百度文心 5.1（Ernie 5.1）在权威的大模型竞技场（LMSYS Arena）文本能力登顶全球第一，搜索技能位列全球第五。 ### `30:00` 对标谷歌全栈生态：自研芯片、萝卜快跑与搜索收入的结构性裂变 主持人首先回应了上一话题留下的行业八卦：此前 Anthropic 限制 API 访问权曾引发 OpenClaw 创作者的强烈不满，达里奥（Dario Amodei）的做法显然得罪了不少人。随后话题转入 AI 时代最核心的要素——数据。主持人提到此前采访 Substack 科技博主 Grace Shao 时的讨论：Grace 认为中国在数据收集上其实并没有绝对优势，因为过去积累的海量数据大多是非结构化的，极难清洗并高效用于大模型的训练和推理。主持人借此提问：百度手握海量搜索与中文互联网数据，究竟是如何将其成功转化为文心大模型（ERNIE）的底层养料的？ 何俊杰（Henry He）直言资本市场和科技行业普遍低估了百度的业务拼图，并详细对比了百度与谷歌在全栈 AI 能力上的高度镜像关系。谷歌拥有自研 TPU 来支撑 Google Cloud（GCP）的高速增长；而百度同样拥有自研芯片业务（昆仑芯），并已完成独立分拆的公开申报。在基础大模型层，谷歌有 Gemini，百度有自主研发的文心大模型（ERNIE）；而在具身智能与物理世界 AI（Physical AI）层面，百度的无人驾驶出租车“萝卜快跑”（Apollo Go）正展现出惊人的规模效应。何俊杰分享了一组令业内朋友深感意外的运营对比数据： Waymo 在全美包括旧金山、奥斯汀等全部开通城市中，每周完成约 50 万单；而同期百度的萝卜快跑在全球 27 个城市每周交付约 35 万单，订单规模仅比谷歌少 25% 左右。这些自动驾驶车队不仅是一项独立业务，其在真实道路上捕捉的多模态数据和行驶 Know-how，反过来持续反哺着基础大模型的迭代与推理闭环。 ```mermaid flowchart TD accTitle: 百度与谷歌全栈 AI 业务架构对照 accDescr: 展示百度与谷歌在底层硬件、云计算、大模型、物理 AI 应用与多模态内容闭环上的全面对标关系 subgraph HardwareCloud["硬件与算力基础设施"] G_TPU["谷歌: TPU 算力矩阵"] --- B_CHIP["百度: 昆仑芯（自研芯片）"] G_GCP["谷歌: Google Cloud (GCP)"] --- B_CLOUD["百度: 百度智能云"] end subgraph FoundationModel["基础大模型层"] G_GEM["谷歌: Gemini 模型家族"] --- B_ERN["百度: 文心大模型 (ERNIE)"] end subgraph PhysicalAI["具身物理 AI 与多模态生态"] G_WAY["谷歌: Waymo (约50万单/周)"] --- B_APO["百度: 萝卜快跑 (约35万单/周)"] G_YTB["谷歌: YouTube (长/短视频数据)"] --- B_IQY["百度: 爱奇艺 (中国长视频超50%份额)"] end ``` 何俊杰进一步揭示了一个打破外界固有认知的关键财务转折：长期以来，外界乃至很多投资人朋友依然习惯性地对他说“恭喜财报表现，但你们 80% 到 90% 的收入依然来自传统搜索吧”。然而事实是，在最近一个季度中，百度传统搜索业务的营收占比已经历史性地降至 48% 左右，彻底跌破 50% 门槛。取而代之的是数字人、企业级 AI 软件应用以及 AI 云计算等新型业务正在成为强劲增长的“动力引擎”。此外，在多模态视频内容上，正如谷歌深度绑定 YouTube 一样，百度控股的爱奇艺在中国长视频领域占据超过 50% 的市场份额，为百度的多模态数据飞轮构筑了独特的内生闭环。 - 萝卜快跑（Apollo Go）已在全球 27 个城市展开布局，季度内周均订单量达到约 35 万单，紧咬谷歌 Waymo 全美周均约 50 万单的体量。 - 百度传统核心搜索业务的营收贡献度已降至约 48%，非搜索类的 AI 新技术与企业软件业务正在重构营收底座。 - 百度控股的爱奇艺在中国特定长视频内容领域占据逾 50% 市场份额，为视觉与多模态大模型提供关键的数据闭环。 --- ### `33:15` 中国数据孤岛现实与公有云普及带来的弥合机遇 谈及数据质量与获取难度，何俊杰坦诚地承认了国内环境客观存在的结构性挑战。与美国相对开放互通的数据生态不同，中国互联网各大阵营之间壁垒森严，彼此构建花园围墙、互相不开放数据，这是业界皆知的客观事实。但何俊杰指出，随着 AI Agent 和基础大模型变得越来越聪明，推动企业业务与系统整体迁移至公有云的推力正呈指数级上升，这在客观上极大加速了跨领域数据整合与信息实时访问的进程。 何俊杰给出了一个极具反差的基准对比：在生成式 AI 爆发之前，中国公有云的企业渗透率仅在 20% 到 30% 之间，而美国公有云渗透率早已高达 90% 左右。正是由于过去底层上云率极低，外界才会形成“中国数据碎片化、非结构化且难以挖掘”的强烈印象。然而，AI 正在以前所未有的速度强迫国内企业拥抱云原生架构，这一数据孤岛鸿沟正在迅速收窄。对百度而言，沿着谷歌已经验证成功的全栈模式深耕，凭借自研芯片、云基础设施和大模型底座的多层协同，百度在从各业务层面汲取并盘活数据价值方面，依然保持着优于同行的战略卡位。 - AI 爆发前中美企业公有云渗透率差距悬殊：中国仅约为 20%-30%，而美国已接近 90%。 - 生成式 AI 的智能化跃升正倒逼全行业系统上云，推动实时数据交互，从而根本性改变国内非结构化数据难以利用的困局。 --- ### `37:15` 萝卜快跑出海对决 Waymo：赛车手视角下的“汽车所有权颠覆” 节目的后半段，主持人将目光投向了自动驾驶的国际竞争格局。长期以来，市场习惯将百度称为“中国的谷歌”，但两家公司始终处于彼此隔离的本土市场中：谷歌不在中国运营，美国普通大众也基本不接触百度。然而这一平行线状态即将被打破——在海外市场，尤其是即将开放竞争的伦敦，Waymo 与萝卜快跑（Apollo Go）将首次在同一座城市展开正面对决。这是中美两国的消费级互联网巨头首次在完全相同的海外终端消费市场狭路相逢。主持人好奇地发问：在这场全球 Robotaxi 的世纪角逐中，决定胜负的最关键维度究竟是什么？是软件算法和应用的精准度，还是大规模车辆量产与车队整合的供应链交付能力？ 面对这个问题，何俊杰不仅从高管视角分析，还顺带揭秘了自己的硬核个人爱好：“在回答这个问题之前，其实汽车是我的个人兴趣之一。Joe 可能知道，我其实是一名持有专业执照的业余赛车手。”他笑称自己每次有机会都要开上几圈，享受引擎的轰鸣，甚至感慨在内燃机车彻底淡出视野前，也许未来 20 年内大家还能偶尔享受自己踩油门驾驶燃油车的乐趣。回到自动驾驶的胜负手，何俊杰明确指出，判定这场竞赛未来走向的底层关键词绝非单纯的单车性能或造车数量，而是从根本上“颠覆传统的汽车所有权模式（Car Ownership）”。在未来的出行账本中，当每英里出行成本因无人驾驶的普及而出现断崖式下降时，整个人类社会对私家车的拥有逻辑与城市出行范式将被彻底重构。 ### `40:00` Robotaxi的经济平衡点与全球竞争格局 百度首席财务官何俊杰（Henry He）分析了Robotaxi（自动驾驶出租车）背后的核心经济学逻辑。在综合计算保险、油价以及停车费等持有成本后，每英里约 60 到 80 美分是“租车/打车”与“自购私家车”之间的临界分水岭：一旦出行成本低于这个区间，大众会倾向于买车；而若高于此区间，则租车或按需出行更划算。当前全球各大 Robotaxi 玩家受限于初期运营规模，每英里成本大致落在 1 到 2.5 美元之间。然而，这一成本曲线正在急剧下行。何俊杰预测，在未来 5 到 10 年内，全球自动驾驶车队的每英里成本完全有希望压低至 60 到 80 美分。 谈及这种经济转变对日常生活的冲击，何俊杰以高密度的香港为例展开描绘：在香港，停车费不仅令人咋舌，甚至比昂贵的汽油还要贵。随着纯电动自动驾驶车的普及，用户既无须承担高昂的汽油开销，也彻底摆脱了买车位或租车位的沉重负担。更颠覆的是车辆属性的变化——在主人录制播客或办公的 40 分钟里，私家车可以自行上路接单赚钱。自动驾驶车在本质上已经演变为一个“在物理世界中穿梭并自主挣钱的实体 AI Agent（Physical AI Agent）”，它将从根本上重塑整个人类的出行习惯与交通出行模式。 从当下竞争态势来看，全球市场仍处在极早期阶段，潜在市场空间（TAM）极为庞大，玩家之间更多是合作拓展而非零和博弈。百度 Apollo 与 Waymo 等企业正朝着共同的蓝海迈进： - 上一季度，百度已将其自动驾驶车辆部署至英国伦敦，与 Waymo 同台开启伦敦市场的拓展步伐。 - 百度在全球范围与 Uber、Lyft 达成深度合作，并在东南亚与 Grab 展开联手；未来用户在这些打车软件上下单，接驾的可能正是一台百度自动驾驶车。 - 自动驾驶有效填补了人类司机的服务缺口，尤其是在深夜等运力稀缺时段，车辆能提供 24 小时全天候服务。 - 决定行业成败的核心要素主要有两点：其一是底层前沿技术的持续迭代与突破，其二是极度依赖细节的精细化运营效率（涵盖上下客点选址优化、充电桩网络布局以及调度网络设计等）。 - 全球范围内目前仅有两座城市部署了规模超过 1,000 辆的单城自动驾驶车队：一座是 Google（Waymo）深耕的旧金山，另一座则是百度 Apollo 运营的中国城市（武汉）。 ```mermaid flowchart TD accTitle: Robotaxi 运营成本走向与自购车替代拐点 accDescr: 展示Robotaxi成本从当前阶段下降到经济临界点，并推动自购车向物理AI Agent转型的演进路径。 subgraph CostTrend["成本演进趋势"] CurrentCost["当前阶段规模受限<br/>成本：$1.00 - $2.50 / 英里"] FutureCost["未来 5-10 年规模化<br/>目标成本：$0.60 - $0.80 / 英里"] CurrentCost -->|"规模效应与技术迭代"| FutureCost end subgraph TippingPoint["临界点效应与模式颠覆"] CarOwnership["传统买车模式<br/>高额车价、昂贵油费与停车费"] AIAgent["物理 AI Agent 模式<br/>全天候按需出行、闲置外出接单创收"] FutureCost -->|"触及 $0.60-$0.80 临界点"| AIAgent CarOwnership -.->|"行为模式重塑"| AIAgent end ``` ### `43:24` 从 DAU 到 DAA：日活跃 Agent 如何重塑商业变现 播客主持人 Tracy 顺势打趣，透露自己从小在东京及各大都会长大，一直没去学开车，甚至常开玩笑说自己打算直接坚持到自动驾驶完全普及时再享受出行，这引发了全场轻松的笑声。随后，主持人将话题转向了近期 AI 界最火热的概念——Agent（智能体）。百度 CEO 李彦宏曾多次指出，百度未来的核心指标之一是 DAA（Daily Active Agents，日活跃智能体）。作为 CFO，何俊杰详尽阐释了 DAA 背后的商业逻辑，以及它与传统搜索广告模式在变现机制上的根本差异。 在移动互联网时代，移动应用的核心北极星指标是 DAU（日活跃用户数），其核心诉求是满足个人的信息查询与内容消费需求。然而，AI Agent 已经远远超越了单纯充当人类“信息工具”的范畴——它具备自主规划复杂任务并端到端执行落地的智能水平。因此，李彦宏提出的 DAA 衡量的不仅是有多少人调用了 Agent，更衡量了 Agent 所解决任务的深度与复杂度。这直接催生了全新的按结果付费（Result-driven payment）与商业分成模式。 何俊杰以百度面向企业级复杂问题推出的智能体产品“伐谋”（FAMO）为例进行深度剖析： - “伐谋”宛如现实工业场景下的 AlphaGo，此前已成功部署在中国吞吐量名列前茅的超大型港口之一。 - 该 Agent 全面负责船只进出港排期、物流与货物调度等超高复杂度规划，大幅降低了港口的闲置等待时间，直接提振了港口的营运收益。 - 商业变现路径不再是传统的软件许可买断，而是港口方根据 Agent 实际带来的效率提升与利润增量，直接与百度进行利润分成（Profit sharing）。 ```mermaid flowchart LR accTitle: 传统 IT 采购模式与 AI Agent 商业模式对比 accDescr: 对比传统IT采购以技术层成本预算为主，与当下AI Agent面向CEO以业务增量进行利润分成的演进。 subgraph TraditionalIT["传统 IT / 工具软件模式"] TechBuyer["对接对象：CTO / CIO<br/>定位：工具与成本中心"] ITBudget["从内部艰难申请 IT 预算<br/>按许可证/人头收费"] TechBuyer --> ITBudget end subgraph AIAgentModel["AI Agent 结果导向模式"] CSuiteBuyer["对接对象：企业 CEO<br/>定位：顶层驱动的增长引擎"] ValueShare["降低闲置耗损 + 提升运营收入<br/>按任务效果或利润增量分成"] CSuiteBuyer --> ValueShare end ``` 这种商业变革给销售端带来了质的飞跃。以往缺乏 AI 时，哪怕 CFO 本人亲自陪同拜访客户，坐下来谈判的对象往往只是对方的 CTO 或 CIO，软件被视作纯粹的“成本中心”，需要在企业内部艰难争取 IT 预算；如今，百度团队直接与企业的最高决策者——CEO 面对面。对于 CEO 而言，利用 AI 优化业务流程是一把手直接抓的顶层战略项目，他手里握有实权与预算，并且由于 Agent 能在多个港口或业务场景间高复用落地、降低单点定制边际成本，客户既有极强的付费意愿（Willingness to pay），更有实打实的付费能力（Ability to pay），构建了与传统 IT 采购截然不同的正向商业循环。 ### `47:00` 数字员工与主播永动机：“我们会被替代吗？” 针对主持人关于“未来是否所有行业都会转向按完成任务结算或按节省开支分账”的追问，何俊杰介绍了百度的另一条重要业务线——数字员工（Digital Employees）。他幽默地类比道：如果主持人两周内连续录制 10 到 20 期播客，肯定会精疲力竭，因为人类的精力和知识储备终究有其天然的生理边界。主持人听到这里立即逗趣打断：“等等，你这是在暗示我们要被取代了吗？因为我们会累，而 AI Agent 不知疲倦？”现场顿时洋溢着半开玩笑的紧绷感与欢笑。 何俊杰随即澄清并指出，若能以科学严谨的方式沉淀专有行业知识（Know-how），数字员工在特定标准化、长周期的场景中确实能展现出无可比拟的执行质量与极高效率。他以中国庞大的电商直播行业为例： - 顶流主播或 KOL 往往极具表演感染力，但人类肉身无法支撑 24 小时连轴转，各区域消费者的购买作息也截然不同。 - 百度的数字员工解决方案能协助电商商家突破生理与地域局限，不仅能提供 24 小时全天候无间断的直播带货服务，还能无缝切换普通话、英语等多种多语言。 - 更精妙的是，数字员工还能根据不同国家受众的文化背景，在直播中精准穿插当地的幽默段子与本土俚语，实实在在地赋能海内外商家，创造全时段的跨境电商增量营收。 ### `50:00` 数字人直播带货的落地与“防剁手”AI Agent 谈到大模型在电商领域的实际落地，数字人直播带货正在迅速铺开并展现出惊人的效果。数字人主播不仅能不知疲倦地持续推销商品，更关键的是在问答环节表现极为流畅，能够自然应对来自形形色色用户的海量、随机提问。这背后的核心正是基础大模型的深厚通识与推理能力在起作用。百度在商业化变现上也采取了多元策略，例如根据商家带货效果的提升以及不同维度的指标来进行收费赋能。 听到这里，主持人 Tracy 幽默地分享了一段自己当年的网购“惨痛经历”。她回忆自己最初使用淘宝时简直患上了“淘宝精神病”（Taobao psychosis），曾经连续 24 小时疯狂下单，结果直接导致她和丈夫在仅有 500 平方英尺（约 46 平方米）的小公寓里塞下了整整三张大沙发。百度 CFO 何俊杰（Henry He）打趣回应道，针对这种场景，用户显然需要另一个专门的买手 AI Agent 来帮忙把关和筛选真正需要的商品，这本身就是另一个绝佳的变现商业切入点。 ### `51:15` 百度昆仑芯分拆上市与中立芯片生态战略 随后，主持人将话题引向整场访谈的重磅环节：“既然这是在彭博投资大会（Bloomberg Invest）现场，也是 Odd Lots 在亚洲的首次现场录制，不妨趁此机会曝个大料——百度是否打算分拆其芯片业务？” 何俊杰幽默地表示终于到了“聊钱”的环节，并坦言自己的职业背景其实非常特殊：他最初并非科班金融出身，而是“机缘巧合下自学成为 CFO”，他的本科和硕士专业完全是芯片设计工程。加入百度后，他深刻意识到百度的昆仑芯（Kunlunxin）产品性能极为优秀，无论是在推理算力方面，还是在支持公司日活跃用户规模增长与整体业务运转上，都发挥了中流砥柱的作用。 何俊杰明确证实，根据目前的公开信息，百度已经就旗下芯片资产在香港联交所分拆上市提交了机密申请，目前整体上市流程正在按计划顺利推进中。他强调，分拆芯片业务旨在释放其被低估的独立商业价值。芯片不仅仅是纯粹的硬件堆砌，更是极其复杂的全套生态系统： ```mermaid flowchart TD accTitle: 昆仑芯独立分拆与生态赋能架构 accDescr: 展示百度芯片业务独立后在生态中性、开发者迁移和商业化方面的全景图 subgraph Independent["独立上市定位 (香港)"] SPIN["昆仑芯独立分拆"] NEUTRAL["确立第三方中立平台身份"] SPIN --> NEUTRAL end subgraph Ecosystem["生态系统协同"] DEV["软件开发者"] SUP["供应链伙伴"] CUST["外部客户与测试场景"] CUDA["CUDA 栈平滑迁移支持"] end subgraph Synergies["全栈 AI 协同"] MODEL["文心大模型 (Ernie)"] CLOUD["百度智能云"] end NEUTRAL --> DEV NEUTRAL --> SUP NEUTRAL --> CUST DEV --> CUDA SPIN -. 保持深层协同.-> MODEL SPIN -. 强化基础设施.-> CLOUD ``` - **强化独立中立定位**：成为独立的上市公司后，外部客户不再会将昆仑芯视为竞争对手的技术附属品，而是看作一个中立、开放的芯片供应商，从而更愿意在其自身业务场景中进行深度测试与大规模采购。 - **构建软硬件开发者生态**：芯片的成功依赖客户、供应商和广大软件开发者的同频共振。独立运作更有利于打造紧密协同的软件生态，让开发者能轻松将原有的 CUDA 代码栈无缝迁移至昆仑芯平台，大幅降低适配成本。 - **与主营业务互利共生**：分拆不仅不会削弱芯片与百度云、文心大模型的深层协同，反而能通过更活跃的外部生态进一步强化百度的全栈 AI 护城河。 访谈在热烈的掌声中结束，主持人对百度 CFO 何俊杰的深入分享表示由衷感谢，并为 Odd Lots 首次亚洲线下节目的成功录制画上了圆满句号。 ## Notable quotes > `03:40` 在底层架构上，正从硬件基础设施向实际应用层转移；在模型范式上，正从基础大模型（Foundation Models）向智能体（Agents）进化。 — 何俊杰 > `03:40` 全行业新增的 Token 消耗需求中，已有高达 80% 均来自推理（Inference）端。 — 何俊杰 > `05:40` 与用户闲聊（Chatting）跨越到动手干活并彻底解决问题（Completing）。 — 何俊杰 > `20:00` 工程问题，而非理论上的量子跃迁。 — 何俊杰 > `26:15` 在昨天的基础模型根本不可能知道 Joe 和 Tracy 今天在节目里聊了什么，必须通过实时搜索来捕获瞬态信息。 — 何俊杰 > `30:00` 恭喜财报表现，但你们 80% 到 90% 的收入依然来自传统搜索吧。 — 何俊杰 > `37:15` 在回答这个问题之前，其实汽车是我的个人兴趣之一。Joe 可能知道，我其实是一名持有专业执照的业余赛车手。 — 何俊杰 > `37:15` 颠覆传统的汽车所有权模式（Car Ownership）。 — 何俊杰 > `40:00` 在物理世界中穿梭并自主挣钱的实体 AI Agent（Physical AI Agent）。 — 何俊杰 > `47:00` 等等，你这是在暗示我们要被取代了吗？因为我们会累，而 AI Agent 不知疲倦？ — Tracy Alloway ## Mentioned 百度 (Baidu), 何俊杰 (Henry He), Joe Weisenthal, Tracy Alloway, 彭博 (Bloomberg), Odd Lots, IBM, 昆仑芯 (Kunlunxin), 文心大模型 (Ernie), 百度智能云, 萝卜快跑 (Apollo Go), 谷歌 (Google), Waymo, Gemini, TPU, 爱奇艺, Anthropic, Dario Amodei, OpenClaw, LMSYS Chatbot Arena, Uber, Lyft, Grab, 李彦宏, 伐谋 (FAMO), 香港联交所, CUDA ## Video description <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true" data-pm-slice="1 1 []">In the China tech space, Baidu is now a full-stack player in the AI industry. The company makes its own chips, has its own AI models (Ernie), its own cloud system, and it's integrating AI into its self-driving car business, Apollo Go. But before all this, Baidu was known for being China's leader in search. Things, obviously, have changed a lot since the company was founded in the late 1990s. In today's episode, we speak with Baidu CFO Henry He about the company's AI ambitions. He talks to us about maximizing token spend, how Chinese tech firms are thinking about safety and alignment, the global robotaxi competition, and how the core search business fits into the company now.</p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true">Read more:<br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><a href="https://www.bloomberg.com/news/articles/2026-06-22/chinese-ai-stocks-rally-on-demand-optimism-and-policy-support?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">Chinese AI Stocks Rally on Demand Optimism and Policy Support</a><br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><a href="https://www.bloomberg.com/news/articles/2026-06-25/us-seeks-ai-partnership-with-eu-on-regulation-supply-chains?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">US Seeks AI Partnership With EU on Regulation, Supply Chains</a></p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true">Only <a href="http://Bloomberg.com" data-inline-card="" data-card-data="">http://Bloomberg.com</a> subscribers can get the Odd Lots newsletter in their inbox each week, plus unlimited access to the site and app. Subscribe at&nbsp; <a href="https://www.bloomberg.com/subscriptions/oddlots?in_source=oddlotspodcast" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">bloomberg.com/subscriptions/oddlots</a></p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true"><a href="http://bloomberg.com/subscriptions/oddlots" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">Subscribe to the Odd Lots Newsletter</a><br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><strong data-prosemirror-content-type="mark" data-prosemirror-mark-name="strong">Join the conversation:</strong> <a href="https://discord.gg/oddlots" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">discord.gg/oddlots</a></p><p>See <a href="https://omnystudio.com/listener">omnystudio.com/listener</a> for privacy information.</p> [Source](https://podtrac.com/pts/redirect.mp3/tracking.swap.fm/track/UVBrz8bN8aM2Xe47PEPu/traffic.omny.fm/d/clips/e73c998e-6e60-432f-8610-ae210140c5b1/8a94442e-5a74-4fa2-8b8d-ae27003a8d6b/f5c05315-7559-4fa4-a6da-b47400f7cb90/audio.mp3?in_playlist=982f5071-765c-403d-969d-ae27003a8d83)

[3:40] 何俊杰: 在底层架构上，正从硬件基础设施向实际应用层转移；在模型范式上，正从基础大模型（Foundation Models）向智能体（Agents）进化。 [3:40] 何俊杰: 全行业新增的 Token 消耗需求中，已有高达 80% 均来自推理（Inference）端。 [5:40] 何俊杰: 与用户闲聊（Chatting）跨越到动手干活并彻底解决问题（Completing）。 [20:00] 何俊杰: 工程问题，而非理论上的量子跃迁。 [26:15] 何俊杰: 在昨天的基础模型根本不可能知道 Joe 和 Tracy 今天在节目里聊了什么，必须通过实时搜索来捕获瞬态信息。 [30:00] 何俊杰: 恭喜财报表现，但你们 80% 到 90% 的收入依然来自传统搜索吧。 [37:15] 何俊杰: 在回答这个问题之前，其实汽车是我的个人兴趣之一。Joe 可能知道，我其实是一名持有专业执照的业余赛车手。 [37:15] 何俊杰: 颠覆传统的汽车所有权模式（Car Ownership）。 [40:00] 何俊杰: 在物理世界中穿梭并自主挣钱的实体 AI Agent（Physical AI Agent）。 [47:00] Tracy Alloway: 等等，你这是在暗示我们要被取代了吗？因为我们会累，而 AI Agent 不知疲倦？

[4:12] Henry He: 80% of incremental token demand today is driven by inference rather than pre-training. [9:50] Henry He: one-person company [15:30] Henry He: impossible triangle [40:00] Henry He: a physical AI agent on the road [43:45] Henry He: Daily Active Agents [47:15] Joe Weisenthal: Wait, are you saying we're going to be replaced? [47:15] Tracy Alloway: Taobao psychosis

[0:00] Whatever your goal, trade show giveaways, client gifts, or team gear, 4imprint has the promo products to match. With thousands of options, from apparel and drinkware to tech and totes, it's easy to find the right fit for your brand and budget, with standout choices at every price point. And with their 360-degree guarantee, you can be 4imprint certain your order will show up just right, right on time. Explore more at 4imprint.com. 4imprint. 4certain. So there's a lot of noise about AI. But time's too tight for more promises. So let's talk about results. At IBM, we work with our employees to integrate technology right into the systems they need. Now, a global workforce of 300,000 can use AI to fill their HR questions, resolving 94% of common questions. Not noise. Proof of how we can help companies get smarter by putting AI where it actually pays off. Deep in the work that moves the business. Let's create smarter business. IBM. When you're running a business, the best days are the ones where priorities stay on track. For midsize and large companies, risk can affect multiple parts of the organization at once, from property and liability to cyber and regulatory challenges. At that level managing risk becomes an ongoing discipline At the Hartford the focus is on helping businesses manage risk before it turns into something more disruptive And when losses do happen, that work is paired with insurance coverage shaped by years of underwriting, risk engineering, and claims experience. Learn more at thehartford.com slash risk mitigation. Policies provided by Hartford Fire Insurance Company and its property and casualty affiliates, Hartford, Connecticut. Bloomberg Audio Studios. Podcasts. Radio. News. Hello and welcome to another episode of the Odd Loss Podcast. I'm Joe Weisenthal. And I'm Tracy Alloway. So we were in Hong Kong recently. That was a lot of fun. Nice to be back. Yeah, I haven't been back for four years. Not much has changed, actually. I was kind of surprised. Less than you expected? Less than I expected. But I am really glad we went back because obviously one of the big talking points in markets right now is competition between U.S. versus Chinese AI. Yeah And we finally got a chance to talk to a couple of high level executives of Chinese tech companies who are actually making all the big capital allocation decisions when it comes to the AI race Right It felt like we in this moment where there been the I mean, the way I think about it, there's been Chinese Internet giants, but they concentrated on China. Yeah. There's been American Internet giants that basically had the rest of the world. And whether we're talking about AI or self-driving cars, we're going to see the first sort of like real head to head battle on Internet companies specifically. And whether like competing for playing the same game on some of the same markets. And of course, we know American companies can use AI models built by China, et cetera. And so there are all kinds of options for people. So it's like really interesting to see like, OK, this clash is actually like it's happening. It's a good time to talk to a Chinese tech executive for sure. That's right. So the reason we were back in Hong Kong is because we were at the Bloomberg Invest Conference, though we also threw an odd lots trivia night while we were in Hong Our first non-U.S. overseas All Thoughts Quiz 9. Yeah, that was a lot of fun. And I'm sure we'll come back and do that again. But we were at the Bloomberg Invest Conference. And so we had the chance to speak with the CFO of Baidu, Henry He. So check it out. We truly have the perfect guest. We're going to be speaking with Baidu CFO, Henry He. So Henry thank you so much for coming on All Thoughts Thanks for having me And it a great season I think in Hong Kong And definitely it great to see both Joe and Tracy Thank you Very nice of you to say So why don we start with this You know, obviously, I feel like half the conversations are probably about AI these days. But within AI, Baidu is a full stack player, right? You have cloud, you have the application layer, you have your own chips, and of course, your own model. As the CFO, you might have to think about prioritization, et cetera. Is there one layer of the stack that you feel is a must win for Baidu when you think about resource allocation? Is there a layer where it's like, OK, this is an area where we have to win? Thank you so much. And I think you'll probably put the tough question in end. But I think it's probably the most difficult question to start with. So I think the very unique thing today is I think the entire AI has been shifting from infrastructure to applications and from model to agents. I think that's actually the backdrop. I think within that, frankly speaking, right now, it's very difficult to say at this moment which part is the must-have, because in my view, the chip is infrastructures. You need to have a grid model to bridge the capability. The cloud is a deployment of that capability. And obviously, the monetization and all [5:00] IOI questions, especially for the people like me at CFO, we focus on that, is on application layers. So without any of that, this IOI cannot work. So to answer a question, I think the key thing, if I have to pick one, is cloud. Because cloud at this moment is a platform. You can not only host Ernie, which is our own model, but also I can work here very open to hosting other models. And the MyChip, which connecting to my cloud platform, can also help on inference. because 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, but given the tough question, if I want a big one as a student, A, B, C, D, I want a big number C, which is my cloud. I'm going to ask you a question which I think is going to become standard for financial journalists in the same way we ask about headcount and expansion plans. What's your token budget? Is it bigger than Joe's? You mean the token? The token budget for Baidu. Yes. Or how do you measure, I'll ask it in a slightly different way, how do you measure what you're gaining from your token spend? How do you measure productivity? Yeah, sure. I want to categorize probably in two buckets One is we consume computing power to reach a higher level of technology standard You know, the AGI, the how good model perform, and also how harness can be designed to deliver better results. So these are R&D efforts. However, as also a tech house, we also deliver those know-how to our external clients with different verticals. I think if I'm measuring our internal consumption of the token, I really want to look at how better and how efficient our technology can be developed. So that's on one side. However, on the other end, what I think the ROI is more relevant is how many real tasks that OpenClaw and, for example, our own application called DualMate also is a real agent, human digit and other things can do the task. So I think these two different measurements are important in a way that right now, there are two things better than last year. One is a foundation model getting much better. And number two is the framework, i.e. OpenCloud and other things can link up the foundation model capability to the real world task. From chatting on something to doing something and completing on something I think completion part of the tokens is more important today And I think consumption internally we actually encourage the people to do that But think about that Even last year or year before everyone is actually beefing up the R budgets I think that budget is always there. But the completion new task is more important. But let me just press you on this question a little bit further. So let's say Tracy and I are, let's say we worked for Baidu in the same department. I don't know, some department of yours. would we have identical token budgets or would you have one way of saying you know what joe tracy is actually finding ways to get more value out of ai than you are so i'm going to increase her budget like do you make decisions like that and do you have measurement techniques to see like this person really should get 10 times the token budget of another person because they have figured out how to get a lot of juice from the squeeze so to speak you know joe given the question asking I think probably next time I'll give you more tokens. I think right now the technology evolves very fast. I think that's the beauty part of AI. So we don't want to constrain by ourselves by before thinking through something, we just install certain policy by saying, these are the employees, we define the token number by the titles or the seniorities I don think that the way it works So I think we want to be more open and more nimble in a way that giving enough tokens to the individuals to empower their internal R efforts. But on the other end, we do have a lot of efforts to make sure the token costs become dramatically coming down. I think the cost is coming down very fast. Before you even think about getting a policy, maybe the unit cost is coming down half in like a few weeks. So we need to Think about the speed and the cost and output efficiency, these three parameters as a package, not only on the number of tokens, that's one thing. On the other end, you'll feel very interesting facts. So right now, we are recruiting a lot of younger talents today, even for Baidu, which is 20 plus years, let's say the public company. So my feeling is the kids are actually getting smarter than people expected. So they will not waste the tokens you give to them. So they have a sensible judgment about what are the tasks they need to prioritize because they are working on agents and the models. The model actually helps them also prioritize all the tasks they have. So I think the power of the technology today is it is not only the tools. So that's actually the key concept I want to mention. It is not only a tool, it's a mindset. And the mindset becomes automatic and more intelligent. That if people can work on that well and have [10:00] new relationship with agents and with a model. Some of the old questions we kind of struggle ourselves will be kind of diminished and less important. Okay, so no token maxing at Baidu. But since you brought up talent, one thing I'm very curious about is we know the competition for the top engineers is so intense right now. And in the US, we see these headlines where engineers are treated like sports stars, you know, they're being traded for millions of dollars or whatever. What is Baidu's pitch to top talent? Like if you're trying to attract someone to the company, what is it you say to them that makes them want to work for Baidu versus another top tech firm? Yes, great question. So let's bring a different perspective. I think, you know, we are a technology company. Previously, I think the priority is we empower our clients to be more intelligent. We give them more technology tools to help them to remove a move from the traditional IT to the cloud environment, such as that. But right now, I think AI, especially for the big corporation like us, also changes us as well. We need to think about the cultural change, organization change, not only as an organization and a company but also how AI empowers ourselves So it actually equally important to do something for clients versus think about new tools affecting ourselves So Trace is right. I think there are a few things we actually make a lot of different thinkings and some of the new initiatives. First of all, we're probably among a few companies in China still very open, even increasing the campus recruiting and the focus on the younger talents. And number two, recently, we also tasked the senior people not only look at, you know, the current reporting structure, but also in the real mentor relationship with the younger growing piece of the human capital in a company. But more importantly, I think, is really about giving people more autonomy to work in a company. So more trust and more autonomy and give them more real work. And, you know, there's one concept called a one-person company, right? So we are very happy to working with one person company because they actually use our AI tool very nicely and they're willing to pay a lot of revenue to our products, given the quality. However, within the company, we also encourage people to be the one person team so they can actually use the agents to work a lot of internal tasks. So internally, we have a little tool called a Doodoo, right? In Chinese a very kind of nickname which is actually our internal kind of open cloud similar tools and which actually enhancing people efficiency And to a point I think give me more people more autonomy more trust, and more room to grow and attracting new talents. I think equally or kind of very important to change ourselves, but also, you know, reporting lines and organizing structure need to come with it to make sure that people can deliver the results. And the last note I want to mention, the key thing is that people see the application is important. They can work on the full stack in Baidu, which is very unique value in the China tech space. You know, it just occurred to me, American companies are kind of becoming more Chinese in the sense that they're doing more vertical integration. Like that's sort of a long history here of sort of the whole thing. And now we see one of the phenomenon is that every American company, like they want to even start designing and selling their own chips, their own silicon, which is something that you're doing. You have your own business and you've had it for a while. And I'm trying to wrap my head, like, what is the rationale? How much is it about just wanting to be able to control your own fate more? And so wanting to like control more of the supply chain versus having a chip that optimally aligns with the model that you working on Because those are distinct priorities So what is the real rationale for having custom silicon Yeah I think thanks for the tech trend in the past kind of year and two So if you look at entire computing power consumed, for example, last year, most of the consumptions actually relating to the pre-training of large foundation model. But right now, you know, many of them go into the inference and the completed tasks. and if you look at the different stacks, I think right now we are actually fitting to an incremental growing piece of the market, which is well-defined with a clear boundary, which is not focused on pre-training for a very super-scale foundation model. However, the inference application is important. So to your point, I think our chip product, supporting our cloud, focusing on the inference and application is a unique way of we see the positive network effects. I think that's where the area we want to invest. And also, given the issue you mentioned, I think within that defined areas, we feel pretty confident regarding all the issues you mentioned. And because on the supply and demand side, we can find the good match within the emerging market category, within the inference and application markets. [15:00] Whether you're planning a big tech event, launching a new campaign, or just stocking up on team gear, finding the right promotional products makes all the difference. 4imprint offers thousands of options, from on-trend apparel and premium drinkware to tech, totes, and giveaways, so you can find the right fit for any audience, purpose, or budget. You can customize it all, your logo, your message, your look, and many items come with no setup charge to help you save. And if you're really watching the bottom line, you'll find standout choices at every price point so you can make a real impact while staying on budget. Plus, you'll get expert help, fast turnaround times, and their 360-degree guarantee. So you can be 4imprint certain your order will arrive on time and look exactly right. Whatever your goal, 4imprint makes it easy to find your perfect promo match. Explore the possibilities today at 4imprint.com. 4imprint. 4certain. Support for the show comes from public.com. If you're actively involved in your portfolio, you probably catch yourself repeating the same actions. Buying the dip. Manually sweeping idle cash Putting on a hedge On public you can now create AI agents that handle all these tasks on your behalf Just describe what you want to do in plain English Like, if the VIX hits 25, buy a put option on the S&P 500. Or, if my cash balance goes above $20,000, move the excess into my direct index. You approve of the workflow, and your agent handles the rest. Monitoring the market, watching for your conditions, and executing your strategies exactly as defined. An investing platform driven by your intent, not just your clicks. You can also get full read and write access to your account via the public API. Go to public.com slash market and fund your account in five minutes or less. That's public.com slash market. disclosures. So there's a lot of noise about AI, but time's too tight for more promises. So let's talk about results. At IBM, we work with our employees to integrate technology right into the systems they need Now a global workforce of 300 can use AI to fill their HR questions resolving 94 of common questions Not noise proof of how we can help companies get smarter by putting AI where it actually pays off deep in the work that moves the business. Let's create smarter business, IBM. So hypothetically, you could try to do everything, right? The full stack. And I guess the capital investment you would need to do that is also hypothetically unlimited at this moment in time. And we hear these crazy numbers in the U.S. about the hyperscalers spending hundreds of billions of dollars this year alone. But when you're looking at the different parts of the business, so you have a very mature internet search business, and then you have everything that you're doing with AI, including infrastructure, how are you actually allocating capital? And then how are you actually, I guess, balancing that with returning capital at the same time to shareholders? So I called it impossible triangle. So I kind of scratch my head every few months, every few weeks, depends on also see headline numbers on news with other peers as well. So sometimes I make a little bit kind of hesitating to make a statement, but I just want to tell the facts and tell the views. I want to separate them out. So in the recent quarter earnings we mentioned we actually solved partially on this impossible triangles One is our operating profit increase almost doubled on the Q base And number two, our cloud revenue grew about 79% on a YOY basis, which is almost double of the YOY growth rate for the cloud market in China. And number three is since Q3 last year, our operating cash flow has turned positive. So positive operating cash flow, incremental operating profits, and the higher growth in the market. However, my CapEx is not seeing kind of double, even third multiple times. So I think the way of resolving that is as a salesman or as a management team of a heavy CapEx investor, AI tech company right now, need to find a way on one hand, really drive the growth, but also keep the density of the investment into AI in a reasonable pacing. However, when you do that, you need to keep a conscious regarding ROI and ROI in mind to look at the entire cash cycle. For example, every dollar we spend today, we'll probably need to wait for another probably 20, 30, 40 months, depends on the category, to get full cash back. And during that frame, obviously there's price hikes, some memories, there's difficulty on IDC centers, and a huge spending on servers. So my point is, as I say, on every project, you need to look at the entire lifecycle, not only [20:00] one time, but also the pacing important because the foundation model R&D always taking a few months, right? So these are the things you need to keep in mind. But my statement today is, as Baidu, we want to invest probably in a more responsible manner to the shareholders, but do not diminish our ambitious to investment into AI. Keep the right density is important. But given the results for this quarter, I think we kind of resolve that at least for this quarter. So hopefully we can keep on working on that. And maybe, you know, half year later, when we check on this point, we can still keep on the same pattern, you know, high growth, less dollar spend, but better IOI. I think that's probably the angle we want to achieve. Can I ask you something I'm very curious about? You know, I'd say the heads of the American AI labs maybe have varying degrees of AI psychosis. They have a lot of worries that the what they would call alignment research, et cetera. Do you work on similar things or do you have the same concerns? And you also have AI psychosis. No but do you speaking of trying to make money do you invest in or how much do you invest in what they would call AI safety or alignment and essentially making sure that the models that you building don go rogue and always work on behalf of human flourishing Is that a thing that you allocate capital to? Yeah, that's a great question. So there's an emerging area, for example, in this data sanity and all the kind of post-training efforts need to work on that. Alignment obviously is one of that. But my point is, if you look at this new concept of harness, it's not only about pre-training and getting more on the leaderboard, but also more importantly to measure the robustness and all the things you mentioned. I think in the context in China tech sector, the engineering has to be and has been a good competitive advantage. So the harness from the data flywheel to the alignment, to the data quality and the labeling, I think the entire ecosystem has been robust for if you think about even in the mobile internet work, right? So as As simple as data labeling to the alignment checks and post-training and SFT, I think this kind of the full chain of the capability in terms of the talents and the pool of resources and the cost of data sanity and all the checks has been in my view a little bit kind of more efficient in the way that the ecosystem has been in place there So the cost efficiency has been there So my view is this is engineering not a theoretical quantum leap right So on that, the engineering capability from the China tech world and industry has been there with the key elements I mentioned, right? Talents, lower costs, more efficient. I think these are the few things I just want to point out actually can help solve the issue. But as I mentioned, the things evolve very quickly, right? So we don't worry too much about the issue you mentioned in local market. Yeah, so this is interesting. I'm curious. I want to press further on this because the American AI lab is very anxious about this. And they published these model reports. And it says things that in the chain of thought, we were able to see that 4% of the time, the model was able to identify that it was being tested. And therefore, it changed its behavior in response to recognizing that it was tested. and this is a sign of potential misalignment. Are you doing the same sort of research and spending to establish that, again, the models work for people and don't have a rogue goal, so to speak? Yeah, sure. I think right now, if you look at this, right, so we are also part of the open source community So many of the good models especially publishing recently also will publish their thoughts about that So we kind of follow the newest thoughts but also doing our own tests as well. So overall, I think people in the open source community today, in my view, is very collegial. So people still want to do a better model, frontier model for everyone globally, not really on one country or two different places. So actually related to this, I'm going to ask something. Maybe it's slightly sensitive, but I think it's very important. So So in the U.S., the AI companies, even as they talk about safety, they're basically self-regulating, right? Like they choose to put out these reports and judge their own models and things like that. In China, tell me if I'm wrong, but it feels very different. It feels like the government is more hands-on when it comes to AI. China has been very explicit about this as an area. National security, national strategy. So you're operating in an environment where you're firmly embedded in China's technological and industrial policy. How does that influence the development of your AI models and your broader tech? So obviously we're not in position to comment on public policy, but I definitely happy to share some of my thoughts I have. I think in the world, in the China AI today, we believe we have a great group of very superior talents, not only [25:00] the engineers, but also people actually design the framework. So that's actually very important because it's not only about algorithms itself, it's about a whole system regarding infrastructures, regarding the data regulation, regarding the model and the cloud. I think given the past 10, 20 years in China, given this entire infrastructure has been upgraded to a level that is kind of world leading, I think the policy supporting to getting the moment we have today is already proven we have a proven path to leading not only the technology renovation and innovation, but also the way we monitor that into the stage we already have today. So that's my first point. My second point is, I think today the technology is growing very fast and the checks on the performance and on the data transparency and the rules regarding the data regulation, even without the AI model, even on the cloud age in the past kind of 20 years or five years, has been getting more robust. Because if you think about that, in the cloud environment, you have almost similar issues, right? Who own the data? Who use the data? Who can access that? But today it a new tool to actually cover all the things we are doing So I think it is not a new concept for the policymakers to think about It is a new model It is a new thing It is really a new and better tools to utilize and access resources we already building And the existing resources we're building on existing platforms, which has been 100% compliant, but also we have a lot of support from the policymakers, industry practitioners, academics. They're actually all contributing to that. So overall, my feeling is it's a very transparent and open environment, not only China, but also globally. And academia and industry practitioners actually contributing quite a lot of the good conversations to this environment. And my feeling is the policymakers through different channels has been very open, also listening to the new frontier issues and questions. I know this is a business conference and we want to keep things very professional here and not engage in gossip, etc. But I have like one sort of I'm just curious about something, which is if the American AI CEOs, the most hawkish on the sort of like chip exports on China stuff is Dario, who used to be a Baidu employee. Do you ever hear things in the office? Do people ever say like oh I remember that guy He was you know is there any Dario gossip that people talk about in the office from his stint at Baidu So that why I want to put the ball back to your coach I want to share another gossip, which you probably want to hear. So probably in the past kind of 100 days, right, has Open Cloud become very popular, right? And educate the market about how AI is really getting to the real task and the real world. Obviously in China, there are different ways, you have a new way of calling not only the claw, but other Niki names. So one day I saw Peter, who is the founder of the community, which drives the open claw to be prevailed, put Instagram saying he actually want to work with Baidu, because open claw is a tool, but the tool is kind of eating up all the capacity, i.e. the skills. So everyone is contributing to the skills, and the cloud is actually grab all the skills and do the work. So I think one day we are pretty happy. You know CEO Peter dropped us a note very in a positive way because he kind of on one side noticed that the search is an important capability of the skills i the skills in the open cloud environment So actually he asked Baidu to work with him to beef up the search skills in order for OpenCloud to do a better work to accessing the real information Because today the foundation models, one kind of carve out is every few months you train a new model and the model itself in his mindset doesn't have the real-time information. For example, the foundation model yesterday doesn't capture Joe and Tracy what we're talking about today. You need to have a new skills accessing the odd slots, what is happening in real time. That's right. So I think, you know, Google globally and Baidu for China, they are the powerhouse for searching the real-time information. So it has to be linking to the open cloud. So I think that's actually one of the things we're pretty much happy to do about. So, you know, next day, we ask our engineers to link up with Peter and we're actually part of this SKU marketplace doing pretty okay. And right now, I just want to share our foundation model called Earning 5.1. Right now, it's ranked as the globally number one in a text format of the global LIM arena and globally number five in the search skill capabilities globally in the LIM arena as well. So I think that's, I want to give you another gossip. So, but for the previous one, I probably can talk with you after this open session. Yeah, I know you didn't give us any Dario gossip, but implicitly, because I know that the open claw guy, you know, it was originally called open clawed and then anthropocent [30:00] sued him. And also he's got kind of annoyed because they didn't let the API users get full access. So I think that fellow who created OpenClaw is not the biggest fan of Dario's approach. So by giving us that answer, you at least gave us a little drama there. Thank you. You mentioned search a number of times already, and data is obviously very important to AI. We spoke with Grace Hsiao earlier in the week. She writes about AI on her sub stack. And I asked her if China has an edge when it comes to data collection. And she said she thought not really, because a lot of the data that's been collected was unstructured. And so it was hard to harness for AI model training and inference purposes. Can you talk a little bit more about how you did that at Baidu? Because you've got a lot of data, you're using it for Ernie. How is that transition process actually carried out? Yeah, sure. I'll probably talk about something the market has not noticed enough. And I'll talk about what's a real challenge, right? So it's always two sides of the story. I think I very happy to talk about Google versus what we think about Baidu in some certain formats A few things I think the markets not only the capital markets but also the industry has kind of undervalue a little bit regarding the same components we actually matching with the same structure of Google is monetizing and have this integrated capacity. So, first of all, Google has its own TPU, right, which power the cloud. So the GCP growth, the Google Cloud is growing faster, which is part of the reason it's a TPU. And so Baidu, we have our own trip department. And based on the public information, we recently did the public filing, one of the spin of these assets. So the trip, we have the same with Google. For the foundation model, we have our earnings, as I just mentioned. However, in the physical AI called applications or called work model, so we have our robot taxi called Apollo Go. Just want to share one number. I think in the market, sometimes I tell even my friends, I was kind of surprised that each week, including San Francisco and including all the cities Austin, Texas and US Waymo from Google delivered about 500,000 trips per week and in the last quarter Baidu's Apollo goal in a globally 27 cities delivered about 350,000 trips which is only about 25 fewer than Google the part of that is not only about robot tax it about how we using the data empower our own foundations models trainings and also do inference, but also have a lot of know-how regarding the multimodal contents and all the different things. And more importantly, if you look at the traditional search, right now on this quarter, also in the market, didn't notice that, you know, even my friends telling me, oh, Henry, congratulations for your earnings, but your search is probably still 80% or 90% of the revenue but the truth is for this border is declined to about 48%. So it's already below 50%. So the new growing area, for example, the digital humans and also our AI application software is becoming a powerhouse and growing very fast. So my point on that is, if you look at key components or the blocks from the chip, cloud, robot taxi for AI, physical AI applications and the software, and also one more thing I want to mention is Google linking with YouTube for the multi-model contents. And we actually have our controlled subsidy called IQ in China, which is actually over 50% of market share in China for certain long-form contents in China as well. So we also have our closed loop of the data fly well as well, probably at a different scale. But I think it still in the same format and the same model So my view is yes I think on one side it is right that certain data and elements are in their own kind of pockets But however, for Baidu, we still have access to those pockets, probably better than other peers. But I want to also very honest admit, right, given Joe is my girlfriend, I still owe him a little kind of gossip after the session. I just want to share my own challenge. Obviously, in China, you have different camps, right? Different camps, they kind of don't open up enough to share the data, which is reality known to the market for everyone. But my point is, right now, the agents and the foundation model become super smart. And it has a great push to move everything to a public cloud. It's actually helped resolving that issue to be accessing more information. Last note I want to share is, before AI can, the public cloud penetration in China is about 20-30% versus the US is kind of 90%. So that's why your comments I can totally understand because without AI, the gap is like this. But right now it's actually getting closer, but still there's a gap. But my confidence coming from this gap will further narrow because everything will be on cloud environments, everyone access real-time information, but also for Baidu, we have a full stack and each components given the value [35:00] Google Pass has been proven to be right and more efficient. We just want to follow the same pattern and access and the benefits from the different layer of the data itself. Whether you're planning a big tech event, launching a new campaign, or just stocking Thank you. expert help, fast turnaround times, and their 360-degree guarantee. So you can be 4imprint certain your order will arrive on time and look exactly right. Whatever your goal 4imprint makes it easy to find your perfect promo match Explore the possibilities today at 4imprint 4imprint 4certain Support for the show comes from public If you actively involved in your portfolio you probably catch yourself repeating the same actions Buying the dip, manually sweeping idle cash, putting on a hedge. On public, you can now create AI agents that handle all these tasks on your behalf. Just describe what you want to do in plain English. Like, if the VIX hits 25, buy a put option on the S&P 500. Or, if my cash balance goes above $20,000, move the excess into my direct index. You approve of the workflow and your agent handles the rest. Monitoring the market, watching for your conditions, and executing your strategies exactly as defined. An investing platform driven by your intent, not just your clicks. You can also get full read and write access to your account via the public API. Go to public.com slash market and fund your account in five minutes or less. That's public.com slash market. Paid for by Public Investing. Brokered services by Open to the Public Investing, Inc., member FINRA and SIPC. Advisory services by Public Advisors, LLC, SEC Registered Advisor. Complete disclosures available at public slash disclosures So there a lot of noise about AI but time too tight for more promises So let talk about results At IBM we work with our employees to integrate technology right into the systems they need. Now, a global workforce of 300,000 can use AI to fill their HR questions, resolving 94% of common questions. Not noise, proof of how we can help companies get smarter by putting AI where it actually pays off, deep in the work that moves the business. Let's create smarter business, IBM. People call it the Google of China, but separate markets, right? Google isn't in China. People in the US don't, by and large, use Baidu, as far as I know. But there's going to be cities now where there's going to be direct competition between Waymo and Apollo including London I think is going to be the first city where there going to be head bow And so this is exciting because I feel like OK the tech internet the consumer internet giants of the US and the consumer internet giants of China are for the first time going to really be competing in certain identical consumer markets. And so what I'm curious about is not who you think is going to win. I presume you think you're going to win. But like, what is the dimension upon which the winner will emerge? Will it be the quality of the application? Will it be who can produce and secure automobiles in volume? What is the most important dimension that will determine the winner either globally or in a specific city? So before getting there, just, you know, also the car is one of my hobbies, too. Just, you know, Joe probably knows. I'm actually a race car driver. So I got my license as well. So every time I actually drive a little bit. So it's quite nice, enjoyable, because I think that probably the remaining car you can still use gasoline and drive yourself. Probably 20 years from now. So before getting there, I think the key word is change the car ownership. Okay. My view is in my own calculation in US, as an example, right now, each mile increases [40:00] In insurance, gas price, parkings, on average it's about 60 to 80 cents per mile is the tipping points between rent a car versus own a car. So if it's cheaper than that, people will buy a car, but more expensive people will rent a car. So right now for global players, I don't want to name names, but the average robot taxi cost today, because the scale is still very small, is about $1 to $2.5 per mile. That's a range, about $1 to $2. So my point is this curve, just like agents getting more prevailing, is coming down very fast. So assuming at some point, five years, six years, whatever, 10 years, if globally the robotics could deliver average price per mile coming down to, let's say, 60 to 80 cents US per mile, then many people were thinking buying a car because today is very simple. You know, Joe and Tracy, you're probably in Hong Kong. You know, the parking is so expensive, even more expensive than the gas. And the gas in Hong Kong is also very expensive. So first of all, the car right now is all EV drive car. Number two, you don't have to buy parking because you and me can drive and the car can go out. And number three, why are we having this 40 minutes? My car actually can go out pick up passengers and it can make some money for me So it a new agent So that why I saying it a physical AI agent on the road to making money for myself So my view is RoboTaxi will change the human behavior of getting out in terms of behavior, pattern of transportation. So that's one thing, right? So we and Waymo and all other players globally are going to that direction. So that's my vision for the market going forward. However, as you mentioned about the market as a player in the near term competition, My view is right now that the market is due very early and the time is very high. So in the last quarter, we shipped our car in London. And as you know, both Waymo and us are starting to open the market in London. Hopefully next year you will see a car. And we have a good partner with both Uber and Lyft and also with Grab in Southeast countries. So next time you probably call a car from Uber or Lyft apps, you will get a Baidu's car. So I think it's actually helping increasing the services because one interesting note is the human driver probably don't work in the midnight in certain cities. But right now the car actually can work 24 hours. So it's expanding a new market. And right now it's still a very low percentage of penetration. So it's still going to have a lot of room to go. On the other hand to your question about the success factor I think the two things One is the technology needs to be cutting and improving and number two is operation efficiency right So it actually have a lot of work need to be operational driven For example how many locations you pick up passengers to be more efficient, right? The charging stations and all the different network designs is actually very important. But given we are working on this business for kind of 13 years so far, and I can tell you one interesting fact that globally, there are only two cities right now have over, you know, 1,000 cars in that scale, which is San Francisco and one city in China, which is Google operating in San Francisco and Apollo from Baidu operating one city in China. But I think our kind of partnership with both, you know, Lyft and Uber globally with different cities, I think has been very collegial because the demand is much higher than the supply. Joe, I'm going to admit something slightly embarrassing. Actually, you already know this, but I never learned to drive, partly because I grew up in Tokyo, and then I moved to a bunch of other big cities, and so I never needed to. And now I always joke that I'm basically, I'm never going to learn. I'm just going to hold out for the self-driving cars. So you know fingers crossed It coming I hope so I wanted to ask something about you know you mentioned agents a number of times and this seems to be becoming the hot new thing in AI And I know your CEO has talked about how one of the key metrics for Baidu is daily active agents. And my question is, how does that actually turn into revenue or return from a CFO's perspective? Because I understand with search, you know, you type something in, you see the ads, advertisers are paying you for that. But I'm very unclear how it works if the agent is actually going out and doing something. Yeah. So in the mobile internet where, you know, everyone look at, for example, the DAU, the daily active users, because that either fulfill information query demand and individual are the primary users for many of the mobile applications in app stores. So the DAU was the primary matrix to measure that. However, in a recent conference, our chairman and founder of Baidu, Robin, mentioned, I think, based on his leadership, that DAA, which is the daily active agents, are the new kind of matrix to defining the success of agents. So I kind of very much agree on that. The reason is, if you look at tasks, it's actually spread out into different verticals, right? So right now, it's [45:00] very difficult to find a new way to identify how much people are using, especially how much value coming out from using AI. So the agents today is basically can deliver a final task, not only using as a tool for human beings. The agent is smart enough to think about 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. So Tristan, to your question, overall, my thinking is the DAA will measure not only how many people using that, but also how difficult it is. To your question, the result-driven payment is coming up in a near trend. So I just want to share a few things. For example, right now we have three or four different key products. One of them in China called FAMO. It's really solving complicated issues for enterprises. It's very similar to AlphaGo, which actually in the previous years has to do the planning, but right now it's actually coming to the real world. So we install this agent to one of the biggest pot in China and help them deploy and planning for the shipments and the logistics It is saving the cost of the idle time and improving the revenue of their parts So the parts actually waiting to share a certain profit generation with us. So the key things I'm observing is in the previous meetings, even I'm in the CFO, but actually I'm attending a lot of the meetings to meet with clients. In the previous meeting without AI, most of the meetings we are talking with is the CTO and CIO of that company because it was a tool. It was a cost center. So they need to find a budget internally. Joe, you know, it's not easy, right? So they have their CFO and their CEO. But right now, most of the meeting we are having today is with the CEO himself. Because AI right now is not only about Baidu, it's really about helping our clients. So the client has to be a top-down level of the initiatives to really drive AI internally. So I think our sales process becomes relatively more efficient in the way that we get into the number one decision makers. He has a budget and he knows that driving the pot efficiency is important for his task. So he's willing to share certain economics with us. I think the customization is also diminished because right now the agents can be used different paths. You can repeating that success lower the unit basis of the cost and the agents become more real and the clients see the value and the profits So they have a higher willingness to pay and high ability to achieve that payment So I think these four cycles actually in the AI world is very different with the traditional IT This is actually an interesting question because I've seen debate on this within AI about what does revenue look like or what is a sales price look like? Because another thing people talk about is, for example, using an AI agent to say resolve an insurance claim or something like that. And then the AI provider gets paid on, say, like, you know, the number of successful claims resolved, etc. Are you bullish on that basic model where the payment is, as you said, OK, maybe they'll share revenue with you because they can measure that savings. is that the model that you see across a range of ai applications where it's like a sort of per task or sort of very clearly linked to the efficiency gain yeah we have another kind of line of business we call the digital employees which you know joe probably you have the similar experience that if you have one season of the podcast you're probably very energetic right if you do that like 10 20 times in two weeks it very exhaustive right because you need to think about Wait you saying we going to be replaced You share my views right Wait wait is that what you insinuating We going to get replaced because we get exhausted but the AI agent won't get exhausted? So my point is the humans, their motivation and the knowledge base have their own kind of territory, to be frank. But if you look at the conversation, look at the quality of the know-how, if you really tap in a good manner, of course, the digital human actually can deliver efficiency and a better execution quality. So one example I just want to share is e-commerce is a big industry in China. Yeah. And a lot of live performance is really selling the products. Yeah, it looks fun. But you know, the come out is the KOL cannot work like 24 hours, right? And people cannot buy stuff like 24 hours. But if you think about it, you have a great quality of the human employee can help the merchant owners to sell into different time zones and also can speak Chinese, speak English, and for the different parts of audience to have the little jokes from their own countries. It actually can help the e-commerce revenue. So that's why we have their own kind of product called Digital Employees. We actually help our merchant and the e-commerce store owners [50:00] really pushing on that and selling all the goods. And it actually can perform pretty well because on the Q&A sessions, on the questions, it's actually reacting to the random users asking a wide range of questions. That knowledge actually is very fluent in the way that for the foundation model, it is the way it works. So I think that actually has different user case. And we actually monetize by charging, for example, the result improvement and all the different things. I've tried to buy things for 24 hours straight before. I think when I first used Taobao, I think I had Taobao psychosis or something. And that's how my husband and I ended up with three couches in our apartment. That was about 500 square feet large. So my little suggestion, you need to have another agent help you to select the right products. That's another way you probably can monetize on that. Are you going to spin out the chips business? We're at Bloomberg Invest. It's our first live recording of Odd Lots. Let's break some news. All right. So it's coming to the money part. um so i was not to be frank i was not even trained by by finance i become self-help by accident so actually my both bachelor and master training was a chip designer myself so i think after joining baidu i definitely realized that baidu chip product has been really really high quality and really good for the inference for all the things we talk about helping our DAA to grow as well So based on the public information, we already filed the confidential filing for spinoff of our chip assets in Hong Kong. And we are doing that and processing that process on track. And that is one part of the assets we try to unlock at this moment. But however, as I mentioned, the cloud foundation model, they are all very important. So after spin off, we hopefully can enhance that ecosystem. And as you know, chip is not only the hardware. I deeply understand it is about ecosystems. We need to work pretty well with our customers and the suppliers and the software developers all at one goal. And I think to be a separatistic public company will help to achieve that goal, not only as hardware, but also the entire ecosystem as well. and our customers will view our chip products more neutral and independent products that can actually do more testing and more usage on their own cases. Yeah, I seem to recall reading that you've figured out a way that developers can easily port over their CUDA stack over to your stack without much trouble Henry thank you so much for coming on Avlas our first live recording anywhere in Asia Really appreciate it Again thanks for having me Thank you, Joe. Thank you, Tracy. Thank you. That was our conversation with Henry He, the CFO of Baidu, recorded live at Bloomberg Asia Invest. I'm Tracy Allaway. You can follow me at Tracy Allaway. And I'm Joe Weisenthal. You can follow me at The Stalwart. Follow our producers, Carmen Rodriguez at Carmen Erman, Dashiell Bennett at Dashbot, Kale Brooks at Kale Brooks, and Kevin Lozano at Kevin Lloyd Lozano. And for more OddLots content, go to Bloomberg.com slash OddLots. We have a daily newsletter and all of our episodes. And you can chat about all of these topics 24 7 in our discord discord.gg slash oddbots and if you enjoy oddbots if you like it when we talk to chinese tech executives then please leave us a positive review on your favorite podcast platform and remember if you are a bloomberg subscriber you can listen to all of our episodes absolutely ad free all you need to do is find the bloomberg channel on apple podcasts and follow the instructions there Thanks for listening Whatever your goal, trade show giveaways, client gifts, or team gear, 4imprint has the promo products to match. With thousands of options, from apparel and drinkware to tech and totes, it's easy to find the right fit for your brand and budget with standout choices at every price point. And with their 360-degree guarantee, you can be 4imprint certain your order will show up just right, right on time. Explore more at 4imprint.com. For imprint. For certain. When you're running a business, the best days are the ones where priorities stay on track. For midsize and large companies, risk can affect multiple parts of the organization at once, from property and liability to cyber and regulatory challenges. At that level, managing risk becomes an ongoing discipline. At the Hartford, the first thing is, [55:00] is on helping businesses manage risk before it turns into something more disruptive. And when losses do happen, that work is paired with insurance coverage shaped by years of underwriting, risk engineering, and claims experience. Learn more at thehartford.com slash risk mitigation. Policies provided by Hartford Fire Insurance Company and its property and casualty affiliates, Hartford, Connecticut. As industries evolve faster than ever, companies need an environment that accelerates strategic growth, and Michigan delivers on that promise. From emerging startups to global enterprises, Michigan offers what executives value most, a resilient, innovative ecosystem, diverse communities that attract top talent, and a quality of life that supports work-life balance. With our unified Team Michigan approach, businesses scale faster and compete at the highest level. Michigan, pure opportunity. Seize your opportunity at michiganbusiness.org.