Palo Alto Networks Deploys AI Model to Slash Code Debugging Times

Palo Alto Networks部署AI模型大幅缩短代码调试时间 首席执行官Nikesh Arora预测，AI将在五年内消除传统软件用户界面，并迫使企业对其系统进行重构，将数据基础设施转变为高度自动化的公共效用层。 Palo Alto Networks首席执行官Nikesh Arora探讨了诸如Claude Mythos等先进AI模型如何在短短几周内发现积累数年的代码漏洞，从而重塑网络安全格局并改变攻防双方的力量对比。他概述了SaaS领域的结构性洗牌，预测了分析型软件的消亡，以及自主智能体对企业用户界面的根本性重构。最后，Arora分享了关于驾驭企业大规模扩张、推动公司市值向万亿美元迈进，以及利用AI驱动的运营效率大幅提升未来业务利润率的见解。 在Nikesh Arora八年的任期内，Palo Alto Networks实现了大规模扩张，市值从170亿美元增长至2380亿美元。 (0:00) AI被用作普及智能和协调人类故障模式的工具，使庞大的员工团队能够以前所未有的一致性和协同度开展工作。 (0:47) Palo Alto Networks在“极限模式”下使用名为Mythos的先进AI模型，仅用六周就发现了相当于五年工作量的代码库漏洞。 (0:47) 先进自主AI功能的成本正在迅速下降，预计将在三个月内扩散至开源或外国模型中，从而推高网络威胁矩阵。 (5:15) 防御自动化、AI驱动的网络攻击，需要现代企业记录并留存十倍以上的网络安全数据，以妥善区分正常行为与恶意行为。 (5:15)

Palo Alto Networks CEO Nikesh Arora argues that artificial intelligence is fundamentally dismantling traditional SaaS architectures and automating vulnerability discovery, while warning that defense lags behind offense due to crippling false-positive rates. Rather than shrinking technical organizations, Arora contends that harnessing AI requires deep operational re-engineering, surging data infrastructure, and expanded technical hiring. Asymmetric Cyber Threats and the False-Positive Bottleneck (18:26) During an early-access trial with the advanced AI model Mythos, Palo Alto Networks discovered software vulnerabilities across its own codebase in six weeks that human engineers would have taken five to seven years to uncover, incurring token costs in the low millions. In persistent-thinking modes, such models can autonomously daisy-chain flaws into end-to-end network intrusion pathways. Arora warns that comparable offensive capabilities will reach open-source or foreign models within three months, tilting the advantage toward attackers who only need to locate single unpatched flaws in legacy or operational systems. However, deploying frontier models for enterprise defense remains hindered by high error rates. Arora highlights an unpublicized industry crisis: models like MSO have demonstrated false-positive rates as high as 30%. While tolerable in offensive red-teaming, a 20% to 30% error rate paralyzes security operations teams and commercial workflows by demanding manual triage of nonexistent threats. Arora asserts that massive post-model engineering is required to reduce false positives near zero before autonomous defense is viable. The SaaS Shakeout and the Disappearing Interface Arora divides enterprise software into three distinct categories facing unequal futures. First, analytical SaaS applications that merely extract and evaluate data from primary databases are rendered obsolete, as enterprises can query internal data repositories directly using foundation models. Conversely, data infrastructure layers such as Snowflake, Databricks, MongoDB, and Oracle remain undervalued, driven by the mandate for enterprises to store ten times more operational and security telemetry to feed AI systems. Third, foundational systems of record like Salesforce and Oracle face a five-year re-engineering cycle centered on eliminating human user interfaces. Characterizing visual interfaces as a historical compromise for manual data entry, Arora projects that autonomous backend agents will listen to workflows, extract structured milestones, and populate records directly. This shift reduces manual human error, streamlines audit trails, and compresses multi-seat software spending into chat-driven agent interfaces. AI Utility Layers, Software Commoditization, and Hardware Limits (25:25) Rather than a single monolithic superintelligence dominating the enterprise, Arora envisions AI operating as a variable utility layer where enterprises dynamically route tasks according to cost and complexity—paying premium rates for high-reasoning models on diagnostic problems while routing routine tasks to cheaper models. Concurrently, standard application vendors are losing pricing power as enterprise clients realize small internal engineering teams can replicate $500-a-month software seats using generative tools, accelerating software replacement cycles. Despite cloud expansion, low-latency requirements and computing economics preserve a critical role for on-premise hardware, particularly in latency-sensitive sectors like financial trading. Arora identifies physical manufacturing backlogs for GPUs as the critical supply-chain bottleneck, noting that while accelerated depreciation tax incentives drive domestic facility investments, establishing a fully independent domestic hardware supply chain will require roughly a decade of sustained commitment. M&A Realignment, Operational Margins, and Technical Hiring Growth (28:51) Palo Alto Networks has evolved its corporate development strategy from acquiring point products to amortize across its sales force into targeting structural architecture shifts, exemplified by a recent $25 billion acquisition in identity security. Arora projects that enterprises capable of mastering AI-driven internal efficiencies can push gross margins into the 90% range and operating margins into the 40% range, unlocking significant valuation advantages over subscale competitors that cannot afford similar optimization. Rejecting the prevailing corporate thesis that generative AI will drive net workforce reductions, Arora expects technical staffing to increase. Because the technology demands an end-to-end restructuring of corporate workflows, customer interactions, and data governance, Palo Alto Networks currently employs more technical personnel than it would have without AI, with management monitoring a 6-to-12-month horizon for enterprise deployment patterns to settle. While frontier model weights can now be distilled in 24 to 48 hours and fit onto portable storage, making regulatory freezes impractical, enterprise adoption remains bounded by high error rates, sprawling unmanaged open-source code, and the vulnerability of distributed edge systems in commercial infrastructure.

Palo Alto Networks 首席执行官 Nikesh Arora 指出，人工智能正深刻重构网络安全与企业软件格局：AI 虽然将数年期的漏洞排查压缩至数周，但也带来了攻击链自主化与严重的误报瓶颈；与此同时，传统分析型 SaaS 与图形界面正加速消亡，推动行业价值向底层基础设施、垂直业务应用以及极致的组织运营效率迁移。 网络安全攻防失衡与高误报瓶颈 (18:26) 在测试尖端 AI 模型 Mythos 期间，Palo Alto Networks 仅以数百万美元 Token 成本，在六周内便查出了内部代码库中原本需人工耗时五至七年才能发现的安全漏洞。Arora 证实，该模型在具备持续尝试能力的“极速模式”（Ultra Mode）下，能自主将分散漏洞串联为全新的端到端攻击路径。Arora 预警，这类前沿能力预计在三个月内就会扩散至开源或海外模型。攻击者仅需利用老旧工业运营技术（OT）遗留的微小漏洞即可造成严重破坏，使防御方陷入疲于应对的补丁悖论中。 然而，AI 模型的致命缺陷在于畸高的误报率。Arora 透露，部分模型在安全测试中的误报率高达 30%，在通用企业场景下也常达 10% 至 20%。如此高的误报会导致防御团队将宝贵资源浪费在虚假警报上，使当前技术呈现“极利于攻击、极不利于防守”的失衡状态。企业必须在后模型工程中将误报率压降至接近零，并通过保留比现状高出 10 倍的安全数据来构建组织记忆，方能实现有效抵御。 软件产业格局重塑与用户界面消亡 Arora 将软件行业划分为三类并预测了传统 SaaS 的淘汰路线。第一类仅用于数据采集与分析的“分析型 SaaS”已被判死刑，企业可直接通过大语言模型对原始数据池进行自然语言查询；例如有企业取消了 20 个软件席位中的 17 个，改用 Slack 配合 Claude 查询，直接削减 90% 的开销。第二类是以 Snowflake、Databricks 和 Oracle 为代表的基础软件，因承载企业未来激增 10 倍的数据需求而严重被低估。 第三类工作与记录系统（如 Salesforce 和 Oracle）则面临长达五年的底层重构，其标志是用户界面（UI）的彻底消亡。Arora 认为传统 UI 迫使人类手动录入数据是技术史上的重大资源浪费。未来自主智能体将通过监听会议自动提取关键节点并同步至后台系统，不仅消除人为录入误差、提升审计合规性，还能使跨系统的五人协同缩减为单一中央对话窗口。 模型实用公用化与垂直应用层博弈 (14:52) Arora 认为未来 AI 不会被单一庞大的超级模型垄断，而是演化为按需采购的“分级公用事业层”（Utility Layer）。企业将根据任务复杂度匹配成本，例如花费 10 美元调用高智商模型处理复杂诊断，而将日常客服分配给廉价的基础模型。由于最先进的模型权重已被压缩至闪存盘大小，且能在 24 至 48 小时内完成蒸馏，单纯依赖模型参数已难以维持长期壁垒，软件厂商原有的定价权正迅速被削弱。 因此，核心利润池正在向能够解决特定业务难题的应用层转移。由于多数企业缺乏独立开发基础设施的能力，行业亟需专为垂直领域定制外围保护机制（Harness）与企业记忆的应用商。与此同时，大量初创企业正在利用生成式 AI 复制昂贵传统软件的核心功能，并转为纯用量计费模式，加速对现有应用软件利润池的分流。 物理硬件底座与经济软肋的现实防御 (25:25) 在国家安全层面，Arora 强调当下的实际脆弱点并非受严密保护的核心基础设施或情报机构，而是广泛分布的中小型机构（如基层医疗诊所）。目前 89% 的网络入侵依然由于凭据失窃或弱密码等低级疏漏引起；此前 Change Healthcare 遭遇勒索软件攻击导致全国独立诊所瘫痪的事件，暴露出分布式商业系统才是经济中最脆弱的软肋。 在基础设施端，尽管公有云渗透率极高，但对超低延迟和高吞吐量有极致要求的金融机构（如高盛与摩根大通）仍高度依赖本地硬件。受供应链产能严重积压影响，全球 GPU 与内存制造均面临交付瓶颈。尽管首年 100% 资本支出折旧等税收政策极大地刺激了本土算力基建投资，但 Arora 预计建立完全自主独立的本土硬件供应链至少需要长达十年的持续推进。 并购重组范式转移与技术用工反常识 (28:51) 伴随市值突破 2000 亿美元，Palo Alto Networks 将并购策略从原先依托现有销售网络摊销渠道成本的模式，升级为针对关键技术拐点的战略重构，近期便完成了一笔 250 亿美元的身份安全收购。Arora 提出，企业若能借助 AI 优化内部流程、将毛利率提升至 90% 且净利率提升至 40%，便能以无与伦比的运营效率溢价收购并重组低效同行，开辟全新的扩张空间。 针对 AI 将引发科技业大裁员的普遍预期，Arora 提出了相反见解。他指出，由于 AI 倒逼企业从底层彻底重构所有业务流程，带来的工程转型任务极为庞大，因此 Palo Alto Networks 当前雇佣的技术人员数量反而创下历史新高。整个行业仍需经历未来 6 至 12 个月的沉淀期，以摸索出 AI 在大型企业组织中切实可行且具备经济效益的落地模式。 Arora 的观点表明，AI 在展现指数级效率潜力的同时，也使网络防御与软件商业模式陷入容错率极低的严峻考验。如何在高达 30% 的模型误报率下维持防御有效性、在模型快速扩散的背景下抵御针对分布式软肋的攻击，以及在长达十年的硬件供应链瓶颈中完成业务重构，仍是整个行业面临的长期现实挑战。

## Background 本期视频由著名科技商业播客《All-In Podcast》于 2026 年 6 月 27 日在 YouTube 平台发布。节目特邀网络安全巨头 Palo Alto Networks 首席执行官 Nikesh Arora 做客，围绕 Claude Mythos 惊人的漏洞挖掘实测、分析型 SaaS 面临的生存危机、基础模型公用事业化趋势以及企业迈向万亿市值的并购战略展开深度探讨。 ## TL;DR Nikesh Arora 披露前沿模型 Claude Mythos 仅用 6 周时间便挖出了人类团队需耗费 5 至 7 年才能定位的代码漏洞，警示自动化漏洞利用能力在公开网络扩散仅剩约 3 个月窗口。他断言依赖报表加工的分析型 SaaS 已经走向消亡，未来交互界面（UI）将被智能体全面取代，价值将不可逆地流向底层数据基础设施与具备记忆控制平面的新型应用层。针对模型竞争，他指出前沿模型高居 30% 的误报率使其成为黑客利器却是防御方噩梦，大模型的终局是按需调用的公用事业化基础设施。此外，他还力挺专有硬件与底层供应链的不可替代性，并分享了如何借助 AI 极致优化运营杠杆、吞吐重构软件资产以冲刺万亿市值的商业飞轮。 ## Atmosphere & style 全场对话节奏紧凑明快、信息密度极高，兼具顶级科技领袖的前瞻视野与硅谷老友聚会般的机智幽默。Nikesh Arora 表现出极具压迫感与决断力的实操商业洞察，在抛出“分析型 SaaS 已死”和误报率危机等硬核行业猛料时冷静而犀利，同时又在面对“职业经理人 CEO”争议与 Uber 董事会话题时游刃有余地调侃接梗。主持团与嘉宾之间充满高频次的笑点碰撞、捧哏起哄与思想交锋，让整场对谈在严峻的网安攻防警示与宏大的万亿商业图景之间自如切换，听感刺激且张力十足。 ## Bulletin - [01:21](https://www.youtube.com/watch?v=hObRMv6qCi0&t=81s) Palo Alto Networks 实测 Claude Mythos 极限模式，仅用 6 周时间便挖出了人类团队需 5 至 7 年才能定位的代码漏洞量，且能自主将离散漏洞串联为全新攻击路径。 - [06:49](https://www.youtube.com/watch?v=hObRMv6qCi0&t=409s) CEO Nikesh Arora 断言分析型 SaaS（analytical SaaS）已经走向终局，企业更倾向于用大语言模型直连底层基础设施数据，将原有高额席位制采购削减高达 90%。 - [18:26](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1106s) 大模型高达 10% 至 30% 的误报率使其成为黑客攻击的利器，却沦为防守方的虚警灾难；产业真正的壁垒在于通过工程适配套件（Harnesses）将误报率压低至近乎 0%。 - [11:32](https://www.youtube.com/watch?v=hObRMv6qCi0&t=692s) 全网 89% 的网络违规与入侵源于最简单的凭证失窃而非高深漏洞，缺乏防护纵深的中小型医疗及供应链机构是引发系统瘫痪的最大脆弱点。 - [28:22](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1702s) 针对 AI 导致失业的普遍担忧，Nikesh 提出反共识判断：底层系统与安全架构的彻底重写，反而促使 Palo Alto 雇佣了公司历史上规模最大的技术研发团队。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=hObRMv6qCi0&t=0s) Google 实现了信息分发的平权，而本轮 AI 浪潮的核心逻辑在于实现企业智力与执行力的平权，大幅拉齐员工个体技能短板。 - [01:21](https://www.youtube.com/watch?v=hObRMv6qCi0&t=81s) 前沿模型在“极限推理模式”下展现出惊人的链式攻击生成能力，全球代码库暴露在这种扫描下的测试成本仅在数百万美元级别。 - [04:15](https://www.youtube.com/watch?v=hObRMv6qCi0&t=255s) 具备同等攻击能力的 AI 工具向开源和黑客阵营扩散的时间窗口仅剩约 3 个月，防守方正面临软硬件供应商打补丁与自研代码审计的双重挤压。 - [04:15](https://www.youtube.com/watch?v=hObRMv6qCi0&t=255s) 网络防御正在从单点打补丁转向构建全企业的“记忆与上下文”（Memory & Context），企业需收集多达 10 倍的安全遥测数据以实现智能免疫。 - [06:49](https://www.youtube.com/watch?v=hObRMv6qCi0&t=409s) 依赖图表和数据汇总的分析型 SaaS 正被大模型直接调取底层数据所瓦解，企业软件价值正大幅向底层数据基础设施软件转移。 - [09:00](https://www.youtube.com/watch?v=hObRMv6qCi0&t=540s) 企业软件繁琐的图形用户界面（UI）将被淘汰，未来的“工作系统”将由无感监听与智能体自主闭环，大幅重塑人效与审计跟踪机制。 - [11:32](https://www.youtube.com/watch?v=hObRMv6qCi0&t=692s) 现实中 89% 的网络安全事件源自账号密码凭证泄露，中小医疗机构等供应链环节的脆弱性远比顶尖国家级机构更容易引发经济停摆。 - [14:06](https://www.youtube.com/watch?v=hObRMv6qCi0&t=846s) 大模型终将走向“公用事业化”，企业将根据任务难度动态采购不同智商算力，产业链的超额利润正不可逆地向具体业务应用层转移。 - [17:39](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1059s) 基础模型权重由于物理体积小且易被快速数据蒸馏，难以单凭权重本身构建超过 3 至 6 个月的长期技术护城河。 - [18:26](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1106s) 高误报率是通用大模型进入关键业务生产环境的最大拦路虎，企业落地胜负手在于将 10%~30% 的误报率工程化压制到 0.01% 甚至 0%。 - [25:55](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1555s) 硬件仍是承载低延迟高吞吐数据流最具成本效益的物理底座，华尔街金融机构等对延迟极度敏感的核心业务无法完全迁移至公有云。 - [25:55](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1555s) 芯片产业的短板在于制造工厂与物理组装产能而非前端设计，美国本土想要重构完整硬件供应链至少需要 10 年持续投入。 - [28:22](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1702s) 未来的跨界并购逻辑将由“产品叠加渠道”演进为“AI 极致运营杠杆”，率先跑通 AI 内部提效的企业将具备降维整合低效标的的能力。 - [28:22](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1702s) AI 并没有消灭工程师岗位，全行业遗留系统的推倒重来与架构重写反而催生了历史上最庞大的技术人员需求。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=hObRMv6qCi0&t=0s) 迈向万亿市值与 AI 对企业智力的平权 作为网络安全巨头 Palo Alto Networks 的掌舵人，Nikesh Arora 即将迎来执掌公司的第八个年头。节目开场时，主持人回顾了 Nikesh 的傲人战绩：他刚接手时公司市值约为 170 亿美元，而录制当天早晨已飙升至 2380 亿美元。面对主持人的打趣——跨过千亿美元门槛后实现下一个 10 倍增长迈向万亿美元俱乐部反而更容易，Nikesh 幽默回应道："借你吉言"（From your mouth to God's ears）。回顾在 Google 任职的十年经历，Nikesh 指出 Google 搜索的本质是实现“信息平权”（democratizing information），而如今 AI 所掀起的技术海啸，核心逻辑则是实现“智力平权”（democratizing intelligence）。 智力平权对大型企业运转方式的重塑是颠覆性的。Nikesh 以具体业务部门为例：公司市场部有 250 名员工，过去各人产出的方案水平参差不齐，如今借助 AI，可以让这 250 人中 90% 的工作输出保持高度统一与高水准；而在拥有 5000 名直面客户的销售与技术支持团队中，过去最致命的管理痛点在于不同员工能力悬殊，客户往往抱怨“我一定要找 Joe 来对接，因为他懂怎么解决问题，而 Jim 根本搞不定”。有了 AI 的加持，这 5000 名员工在与客户交互时能够输出几乎完全一致的专业表现。这种能力拉齐不仅将重构整个商业运营形态，也将彻底改变企业软件生态。 - **市值跨越**：Nikesh Arora 执掌 Palo Alto Networks 近 8 年，公司市值从约 170 亿美元跃升至 2380 亿美元。 - **核心哲学**：Google 搜索实现了信息分发的平权，AI 正在实现企业专业智力与执行力的平权。 - **组织效率拉齐**：AI 消除个体技能短板，使 250 人的市场团队产出 90% 保持高水准一致，让 5000 人的客服与销售网络提供同等质量的专业交互。 ### [01:21](https://www.youtube.com/watch?v=hObRMv6qCi0&t=81s) Claude Mythos 震撼实测：6 周挖出 5 至 7 年的代码漏洞 在谈及近期轰动业界的 Anthropic 前沿模型 Claude Mythos 时，Nikesh 透露 Palo Alto Networks 是极少数率先获得内测访问权限的企业之一。他直言，人类程序员在过去 50 年中不断写出充斥着缺陷的糟糕代码。作为顶级网络安全企业，Palo Alto 向来以全行业最高标准严苛测试自身代码，但 Mythos 的表现彻底刷新了他们的认知：在短短 6 周的测试期内，Mythos 挖掘出的安全漏洞数量，如果换算成传统人类工程师与工具的排查速度，通常需要耗费 5 到 7 年的时间。面对主持人的惊愕追问，Nikesh 郑重确认这一结论绝无水分（"Mythos was not oversold. It was legit"），而且测试对象正是 Palo Alto 自己的代码库。 更具颠覆性的是前沿 AI 模型在安全渗透上的逻辑链条。Nikesh 解释道，如果将模型置于“极限模式”（ultra mode，即具备持久思考与反复试错的深度推理状态，直到得出结果前绝不放弃），AI 展现出了惊人的“漏洞链式串联”（daisy-chain vulnerabilities）能力。这意味着 AI 不仅能单点识别漏洞，还能将多个分散的次级漏洞像多米诺骨牌一样拼装串联，自主推导并开辟出一条潜入系统核心的全新攻击路径（attack path）。全球上千万开发者日以继夜编写的代码如果全部暴露在这种扫描下，AI 将能在几周内扫出人类原本需要 10 年才能察觉的隐蔽缺陷，而 Palo Alto 进行这次深度测试所消耗的 token 成本仅在低个位数百万美元（low millions）区间。 - **实测效能突破**：仅耗时 6 周即挖掘出人类团队需 5 至 7 年才能定位的代码漏洞量。 - **攻击链自动生成**：开启“极限模式”（持续推理试错）后，AI 能够自主将独立漏洞拼接串联，生成端到端的全新渗透路径。 - **测试成本与门槛**：整体 token 消耗成本仅数百万美元，且随着各大模型厂商推陈出新，算力与推理成本曲线正在迅速压低。 ### [04:15](https://www.youtube.com/watch?v=hObRMv6qCi0&t=255s) 攻防白热化竞速与仅剩 3 个月的时间窗口 针对主持人关于“若 Anthropic 没有克制而是直接将 Mythos 级别能力公开发布，是否会引发全球企业网络灾难”的提问，Nikesh 给出了极其紧迫的时间表：具备同等杀伤力的 AI 工具在公开网络或黑客阵营中泛滥成灾“最多只要三个月，甚至可能现在已经悄然存在”。当嘉宾指出行业原先预测中国模型或开源生态追平这一水平大概还需要 6 个月时，Nikesh 反驳称开源或竞品新版本（如 4.8、5.5 系列）已经在逼近类似能力。更可怕的是，网络攻击从来不需要攻破最难的核心防火墙；只要利用 AI 顺藤摸瓜，扫描运行过时 OT 代码（运营技术代码）的工业控制系统或边缘设备，黑客就能轻而易举找到致命突破口。 当下的网络安全正演变为一场防守方与恶意攻击者之间的生死竞速：安全人员必须赶在黑客利用 AI 武器化之前把漏洞全部修补好。Nikesh 坦言，防守方目前的进展“远没有达到应有的理想水平”。如今跨国企业的 CIO 每天都焦头烂额，各大软硬件供应商纷纷登门催促打补丁，因为供应商自己也用 AI 发现了成堆的历史漏洞；与此同时，CISO 们既要排查公司自研代码的漏洞，还要面对几乎无人能完全掌控的庞大开源软件依赖（IBM 为此甚至设立了 50 亿美元专项计划尝试修复开源安全）。Nikesh 强调，AI 时代的网络防御关键在于构建全企业的“记忆与上下文”（memory and context）。企业必须将过去采集的安全数据量级提升 10 倍，才能训练防御系统真正理解何为正常行为、何为恶意入侵。 ```mermaid flowchart TD accTitle: AI时代网络安全攻防竞速与数据防御架构 accDescr: 展示以Claude Mythos为代表的AI能力向外界扩散所引发的企业漏洞暴露，以及防守方依赖10倍上下文数据进行防御的闭环。 subgraph AttackRisk["AI 攻击侧威胁扩散 (3个月窗口)"] A["高推理能力 AI 模型<br/>(如 Mythos / 开源新模型)"] B["自动化漏洞挖掘<br/>与链式攻击串联"] C["薄弱目标突破<br/>(边缘设备 / 老旧工控 OT 代码)"] A --> B B --> C end subgraph DefensePressure["企业防守侧的三重补丁危机"] D["企业内部数字化资产"] E1["第三方软硬件供应商补丁潮"] E2["企业内部自研代码漏洞"] E3["无主维护的开源代码盲区"] D --> E1 D --> E2 D --> E3 end subgraph EnterpriseDefense["新一代防御体系"] F["全企业级记忆与上下文<br/>(Memory & Context)"] G["采集 10 倍安全遥测数据"] H["动态识别正常与恶意行为<br/>(构建免全量修补的免疫层)"] G --> F F --> H end C -.->|"威胁渗透"| D H -.->|"化解防御滞后"| DefensePressure ``` - **能力扩散倒计时**：Mythos 级别的自动漏洞利用能力预计在 3 个月内流入公开/地下网络，黑客攻击门槛断崖式下跌。 - **防守方的三重困境**：供应链厂商软硬件频繁预警、内部海量代码自查压力、庞大开源代码库（Open Source）的系统性脆弱。 - **防御核心要素**：单纯打补丁已难以为继，企业需沉淀 10 倍以上的运行与交互数据，建立全局上下文记忆以实现智能阻断。 ### [06:49](https://www.youtube.com/watch?v=hObRMv6qCi0&t=409s) 分析型 SaaS 终结论与基础设施软件的重估 当探讨起传统软件公司在 AI 时代的生存空间时，Nikesh 抛出了决绝的判断：“如果你是一家分析型 SaaS 公司（analytical SaaS），那你的生意已经彻底结束了（it's over）。”长期以来，以 Salesforce 等生态市场（Marketplace）为代表的 SaaS 行业构建了一种商业惯性：各类插件声称“帮你汇总并分析特定数据”，企业为此购买了海量按人头收费的增量模块。Nikesh 指出，企业根本不再需要第三方分析工具，因为企业可以直接让大语言模型直接对接底层数据进行即时分析，整个分析型增量软件的销售神话将土崩瓦解。 现场主持人分享了一个生动的真实降本案例：某公司采购了一款 20 个席位的传统 SaaS，平时根本没人主动登录系统；于是他们直接退掉了 17 个账号，仅保留 3 个账号作为数据通道接入 Slack 并打通 Claude。如今全员只需在 Slack 中使用自然语言提问交互，软件采购开销当场暴跌 90%。Nikesh 进一步指出，未来的杀手级场景是彻底打破数据孤岛——将销售人员的生产力数据、SAP 里的供应链库存数据全部汇聚到一处，直接向大模型发问：“哪个区域的销售业绩极佳但库存告急？我们需要把库存迅速调配到高产出团队所在区域。”在过去，回答这种业务问题需要横跨采购三款不同 SaaS 并在报表间反复比对；而现在，只需将数据沉淀在底层，由大模型统一推理。这一演进预示着分析型应用层 SaaS 将走向边缘，而提供数据承载与算力管道的“基础设施软件”（infrastructure software，如核心数据库与底层平台）正处于严重低估状态。 - **分析型 SaaS 走向消亡**：依赖数据报表、图表呈现与初级数据加工的寄生型 SaaS 将被直接调用数据的 LLM 全面取代。 - **席位制收费（Per-seat）坍塌**：企业转向“API/极少坐席 + Slack/LLM 交互界面”，企业软件开销可被削减 80%~90%。 - **跨孤岛统一决策**：企业不再依赖割裂的单点 SaaS，而是将 CRM、SAP 库存等多源数据集中，利用 LLM 进行端到端统筹查询与业务决策。 - **价值向底层转移**：应用分析层价值被侵蚀，负责存储海量数据与支持系统底层运转的基础设施软件成为真正被低估的资产。 ### [09:00](https://www.youtube.com/watch?v=hObRMv6qCi0&t=540s) 告别 UI：智能体如何重构企业“工作系统”与 SaaS 软件 在探讨企业软件的演进时，嘉宾们首先将目光聚焦于底层数据基础设施。像 Databricks、Snowflake、MongoDB 以及 Oracle 这类核心存储与数据管理软件依然不可或缺。随着企业对数据的渴求爆发，未来三年内企业所需存储的数据量预计将达到现在的十倍，任何能够协助收集、管理底层基础数据的系统都将持续繁荣。然而，夹在底层存储与最终用户之间的这一层——即所谓的“工作系统”（system of work）或“记录系统”（system of record），正在迎来剧烈震荡。以 Palo Alto Networks 内部拥有 6,000 名销售人员为例，传统模式下所有业务都严重依赖既有的软件界面与流程，但这种模式即将被彻底打破。 智能体革命的第一步，就是消灭图形用户界面（UI）。Nikesh Arora 直言：“企业软件和消费级软件的 UI，是我们这代技术人员做过最糟糕的事情。”在过去，企业耗费无数产品经理的精力去精心打磨界面，仅仅是为了让人类能够与界面背后的数据进行交互。然而，伴随行业投入数万亿美元构建智能体后端，未来的交互形态将发生根本改变：销售人员只需对着智能体说一句“根据我刚才的销售电话提炼要点，并录入追踪系统”，智能体便能自动在后台将一切同步至 Salesforce 或 Oracle。当冗余的 UI 被剔除，企业的“工作系统”将被彻底重新连线，销售代表打完电话即可直接了当说一句“搞定了”，后台所有繁琐的文书流程全部交由 AI 闭环。 这种工作流程的重塑正以一种被动、无感的方式发生。AI 已经在自动读取邮件、直接调取 Zoom 通话记录生成摘要、甚至自动生成精美的销售演示文稿，用户最终只需面对一个简单的对话窗口提出诉求。这不仅带来了真正的人效跃迁——在部分企业中甚至能实现“五人工作量归拢为一人完成”，更大幅优化了审计跟踪记录，因为所有数据不再经由人类手工录入和篡改，而是全程由智能体规范托管。在未来五年内，所有现存的“工作系统”类 SaaS 软件都必须面临从头到尾的重构。 ```mermaid flowchart TD accTitle: 传统企业工作流与智能体工作流对比 accDescr: 展示从繁琐人工 UI 录入向智能体自动化处理后台数据的转变过程 subgraph Legacy["传统模式：依赖 UI 与人工录入"] A["销售沟通与会议"] --> B["销售手动登录 CRM/UI"] B --> C["人工录入纪要与更新记录"] C --> D["后台数据库 / 记录系统"] end subgraph Agentic["智能体模式：无 UI 与自动闭环"] E["销售沟通与会议"] --> F["智能体被动监听与分析<br/>(邮件/Zoom/纪要)"] F --> G["智能体自主调用工具并校验"] G --> H["自动回写后台系统<br/>(无人工干预审计链)"] end ``` - **数据需求激增**：预计在未来 3 年内，企业所需存储与处理的核心数据量将增长至目前的 10 倍。 - **人效重构指标**：随着后台全流程代理化，传统企业中原本需要 5 个人协同处理的工作流有望被精简至 1 人管理。 - **无感数据捕获**：系统直接基于 Zoom 转录、电子邮件往来与自动生成的演示文稿生成记录，彻底省去了手工数据录入环节。 --- ### [11:32](https://www.youtube.com/watch?v=hObRMv6qCi0&t=692s) 真正的网络安全危机：凭证失窃与中小机构的系统瘫痪 当讨论延伸至国家安全层面时，主持人抛出了一个沉重的问题：像 Claude Mythos 这样的前沿模型如果落在敌对国家或恶意黑客手中，是否会成为制造一国经济瘫痪的武器？针对这种对超级 AI 武器的恐慌情绪，Nikesh Arora 给出了非常务实的反驳。他指出，现实中发生的数千起严重网络攻击或入侵，绝大多数根本不是因为对手攻破了多么复杂的加密防线，而是源于极其简陋的原因——多达 89% 的攻击仅仅是因为凭证失窃，也就是用户名和密码泄露。现场随即爆发了一阵调侃，主持人打趣自己的密码可能就是简单的“password”，Nikesh 则笑着接梗问他是不是在前面加了个美元符号，并戏谑道：“如果你加了 `$`，那你可比绝大多数人都领先了。” 在 Nikesh 看来，公众往往把目光聚焦在错误的地方。真正关键的国家安全机构与核心基础设施单位其实防备森严，他们不仅拥有全球顶尖的专家团队，还会将整个 IT 预算的 10% 毫不犹豫地砸在安全防护上，因此攻破像 PG&E 这类大型发电设施的难度极高。真正令人担忧的“阿喀琉斯之踵”，是遍布全美各地的诊所、牙医办公室等依赖通用打包软件的中小型机构。这些单位几乎没有任何安全纵深，一旦供应链软件被勒索软件攻陷，就会引发剧烈的连锁反应。 Change Healthcare 遭勒索软件攻击的惨痛教训便是典型例证。作为医疗支付清算系统的神经中枢，其遭遇瘫痪后直接导致全美各地的全科医生诊所被迫关停，逼得母公司联合健康集团（UnitedHealth）不得不垫付数十亿美元的紧急贷款以维持诊所的基本运转。这类能够制造大规模经济混乱的薄弱环节，才是真正需要警惕的威胁。面对这些深植于传统架构中的漏洞，业内并没有一蹴而就的“灵丹妙药”，唯一的解法是耗费漫长的时间逐步升级、修复并更换全社会的遗留系统，而这也无形中推高了网络安全行业的长期终局价值（Terminal Value）。 - **失窃凭证主导风险**：全网高达 89% 的网络违规与入侵并非源于复杂的漏洞破解，而是简单的账号与密码泄露。 - **安全预算投入对比**：国家级核心机构通常将 IT 预算的 10% 专项投入网络安全，防御坚固；而中小机构普遍缺乏防护。 - **Change Healthcare 事件冲击**：清算平台遭勒索病毒瘫痪后，全美大量私人诊所陷入业务停摆，联合健康集团为此垫付了数十亿美元的信用借款。 --- ### [14:06](https://www.youtube.com/watch?v=hObRMv6qCi0&t=846s) 模型走向公用事业化，真正的利润池在应用层 随着前沿大模型能力不断跃升，业内有人主张应当像对待核武器一样对高阶模型进行国家主义管制，甚至只允许 NSA 等情报机构或特定巨头访问。对此，Nikesh Arora 持有截然不同的看法。他认为大模型终将走向“公用事业化”（Utility Layer），未来的企业将像按需购买水和电一样“动态购买智能”。企业在处理日常任务时完全可以按需匹配算力与智商：解答简单问题只需调用 120 智商的模型，每次花费 1 美分；而攻坚复杂难题时才调用 250 智商的高阶模型，每次支付 10 美元。根本不存在一个“赢家通吃、无所不能”的通用模型能完美满足所有场景。 正因如此，当前 AI 产业链的利润池正在发生不可逆的转移：真正的超额利润并不在基础模型本身，而在能够解决具体业务痛点的应用层。OpenAI 与 Anthropic 自身的产品动向也印证了这一点——OpenAI 重点推介的是 Codex 的飞速发展，而 Anthropic 也在力推 Claude Code 以及法律、会计等垂直模型，双方都在疯狂向下切入应用层，因为“那才是利润的源头”。对于绝大多数普通企业来说，它们既没有能力也没有动力直接调用底层模型 API 去从头重写一套销售或人事系统，因此基础模型层向普通企业的收费空间非常有限。 关于 OpenAI 或 Anthropic 是否会像当年的微软 Office 那样统领所有组织级办公应用的争论，Nikesh 认为未来大概率会诞生一批全新的 AI 原生应用公司。这些公司将在不同的基础模型之间进行套利，并在模型之上构建专业的控制平面（Control Plane）、适配架构（Harness）与记忆机制（Memory）。对于拥有通用需求的五万家企业而言，由第三方专业应用提供开箱即用的智能化销售或 HR 工具，效率远高于每家公司自行研发。与此同时，传统 SaaS 厂商的定价霸权正在瓦解——一旦老牌软件商试图以高昂的许可费“绑架”客户，企业完全可以借助 AI 技术以更低的成本迅速更换供应商或自建替代模块，整个 SaaS 行业的利益格局正在被彻底改写。 - **智能的按需定价模型**：未来计算智力可按阶梯购买，普通任务调用 120 IQ 模型仅需 1 美分，极难任务调用 250 IQ 模型支付 10 美元。 - **应用层的关键技术栈**：新型 AI 应用公司依靠构建 Harness（适配框架）、Memory（记忆机制）以及 Control Plane（控制平面）来沉淀核心业务壁垒。 - **传统 SaaS 定价权丧失**：企业借助 AI 甚至仅需 10 名开发者即可替代原本收费上千万美元的专用软件，传统软件厂商滥用授权费迫使客户换道替代的案例屡见不鲜。 ### [17:39](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1059s) 模型权重的物理脆弱性与开源竞赛 关于前沿基础模型的竞争壁垒，业界始终存在激烈争论。Nikesh Arora 直言，所谓的“技术领先窗口期”极为脆弱，试图将前沿模型雪藏三到六个月以保持竞争优势根本行不通，因为在激烈的全球开源竞赛中，落后者很快就会追平甚至开源类似能力。他分享了一段令人心惊的对话：一家顶级前沿大模型公司的 CEO 曾亲口告诉他，他们最新、最强大模型的全部权重文件，实际上可以完整装进一个普通的 USB 闪存盘中。“这就是他们全部的核心知识产权（IP），”Nikesh 感慨道，而且外界通过数据蒸馏技术，仅需 24 到 48 小时就能把同等能力提炼出来并复刻发布。面对这种现实，指望单凭模型权重自身建立长达半年的护城河无异于痴人说梦。 - **模型便携性冲击**：顶级模型最核心的资产（模型权重）物理体积极小，一枚普通的 USB 闪存盘即可全部装下。 - **技术蒸馏速度**：通过知识蒸馏，全量数据与核心能力可以在 24 至 48 小时内被提炼并产出同等水平的模型。 - **开源全球化竞争**：将私有模型推迟 3 至 6 个月发布的策略毫无意义，开源生态会以极快速度填补技术空白。 --- ### [18:26](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1106s) 误报率困境：为何攻击者的利器却是防御者的噩梦 当被问及 Palo Alto Networks 拥有深厚专有安全知识、为何不亲自下场训练专属的前沿大语言模型（LLM）或专用小模型（SLM）时，Nikesh 抛出了一个行业内鲜有人触及的核心症结——误报率（False Positive Rate）。他反问全场：“谁知道 4.8 或 5.5 这些前沿模型的误报率是多少？”在场主持人和嘉宾无人知晓。Nikesh 随即披露，在安全漏洞扫描测试中，顶尖模型的误报率甚至高达 30%。这直接导致了非对称的结果：大模型对于黑客攻击是神器，但对于安全防御却是灾难。黑客只要有一次报准就能得手，但防御方若是按 30% 的误报率工作，就会每天疲于奔命去封堵根本不存在的“漏洞”，就像在没有来袭导弹时频繁拉响防空警报。 这种误报风险一旦进入实际的企业核心业务，后果将是致命的。如果一家保险公司直接使用 10% 到 20% 误报率的模型去自动处理理赔，企业将直接承受巨额资金亏损；如果梅赛德斯-奔驰宣布采用某款前沿大模型来做全自动驾驶，面对 10% 的误报率，没有任何一个理智的父母敢把孩子送进车里。Nikesh 强调，大模型行业真正的难点和价值壁垒，根本不是抢着发新模型，而是如何构建配套工程框架（Harnesses）和后处理训练，将 10% 到 20% 的误报率硬生生压低到 0.01% 甚至 0%。“在我的网络安全业务里，我要求的容忍度是 0%。” ```mermaid flowchart TD accTitle: 大模型从前沿研发到企业生产环境的转化漏斗 accDescr: 展示前沿基础模型的高误报率特性，如何经过工程套件与领域验证，收敛为企业级零容忍标准的落地过程。 subgraph Frontier["前沿基础模型层 (Frontier Raw Models)"] A["通用基础模型<br/>(高推理力 / 强泛化)"] --> B["高误报率困扰 (~10% - 30%)<br/>适合进攻渗透，防御端极易虚警"] end subgraph Engineering["企业工程转化层 (Harness & Post-Training)"] B --> C["领域知识增强与确定性校验"] C --> D["工程过滤套件 (Harnesses)"] end subgraph Enterprise["生产环境准入 (Production Reality)"] D --> E["极端低误报率 (&le;0.01% &rarr; 0%)"] E --> F1["网络安全实时拦截"] E --> F2["金融与保险自动化赔付"] E --> F3["自动驾驶核心控制"] end ``` --- ### [20:24](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1224s) 摇椅 CEO：点评 Waymo 与被低估的 Google 节目随后进入“摇椅 CEO”（Armchair CEO）的快问快答环节，主持人 Jason Calacanis 邀请曾任 Google 首席商务官（CBO）、软银集团总裁、现任 Palo Alto Networks CEO 的 Nikesh Arora 对各大巨头挨个进行犀利点评。Nikesh 率先打趣：“David Friedberg 刚才试图在创始人 CEO 和非创始人 CEO 之间划清界限，我到现在还在生闷气呢。”引得全场大笑。随后 Jason 抛出第一个公司 Uber，Nikesh 立即机警地挡了回去：“兄弟，我是 Uber 的董事会成员，我可不能对自家公司评头论足，Dara Khosrowshahi 是个极棒的 CEO，别想让我被开除。” 谈及 Waymo 时，Nikesh 给出了高度评价：“这车确实好使，体验太棒了。唯一的建议是他们应该更快、更激进地推向全球更多城市。”而谈到老东家 Google 时，Nikesh 坚定地认为市场严重低估了 Google。他认为 Google 拥有在 AI 浪潮中彻底胜出所需的全部底层资产，尤其是强大的企业级销售网络：“即使你做出了顶尖模型，你依然需要一支庞大的直销团队去说服全球企业客户采购和部署，而三大超大规模云厂商（Hyperscalers）拥有全行业最庞大的企业销售军团。” 期间，Friedberg 赶忙为自己先前的“职业经理人 CEO 与创始人 CEO”论调圆场，盛赞 Nikesh 是极罕见的反例：“就像《黑客帝国》里的救世主 Neo 一样，属于万中无一的异类”，不仅具备创始人级别的冒险精神，更能将别人的公司当成自己的生命去掌舵，比许多创始人还要稀缺。主持人与全场嘉宾立刻起哄调侃 Friedberg 这番滑跪“比 ChatGPT 还要会奉承逢迎”，现场气氛欢快热烈。 --- ### [23:20](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1400s) 利润池争夺战与变现最快的两大途径 话题转向当前大模型头部玩家之间的商业变现竞争。主持人指出，Anthropic 在企业级市场的年度经常性收入（ARR）增速似乎明显超过了 OpenAI，关键原因在于 Anthropic 全力押注企业级需求并推出了专属的安全能力。Nikesh 指出，当前行业正在疯狂争夺高价值“利润池”（Profit Pools），因为底层算力军备竞赛的资金消耗极其惊人，每建设 1 吉瓦（GW）的数据中心可能就要背负上百亿美元的收入压力。 在讨论哪些垂直领域能产生最丰厚利润时，嘉宾们提到了代码生成、新型底层数据库以及网络安全。Nikesh 补充道，各家大模型厂商都在极力证明自己的网络安全防护与测试能力，Anthropic 已经公开提供了具备网络安全能力的专用模型，OpenAI 和 Google 也在跟进，因为企业首席信息安全官（CISO）们迫切需要用 AI 来做代码审计。与此同时，下一波巨大的财富转移将发生在应用软件层——当前 SaaS 领域数百上千亿美元的存量市场正面临洗牌，新一代创业者直接利用生成式 AI 从底层重构业务骨干网，以极具杀伤力的颠覆性架构去抢夺老牌软件的市场份额。 当 Jason 询问商业世界里实现收入增长最快的两条路径是什么时，Nikesh 总结出了两条黄金法则： - **企业级市场的“平替软件”（Replacement Apps）**：在企业端，创造新预算极其艰难，但如果直接替换现有老旧系统，客户原本就有现成预算。拿更好的生成式 AI 工具去替代昂贵、笨重的老一代 SaaS（甚至从每月每座 500 美元降维打击为按调用量计费），客户为了降本增效会以最快速度完成迁移。 - **消费级市场的微额订阅（Consumer Revenue）**：在 C 端市场，从单个用户钱包里掏出每月 5 美元（类似 Netflix 订阅逻辑）的决策门槛极低，一旦形成网络效应与用户习惯，资金聚拢速度冠绝整个商业世界。 ### [25:55](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1555s) 硬件的不可替代性与全球供应链产能瓶颈 面对未来硬件需求是增是减的疑问，Nikesh Arora 给出了非常明确的判断：哪怕在今天，硬件依然是处理“低延迟、高吞吐量数据流”（low latency, high throughput bits）最具成本效益的手段。数据中心的核心本质就是管理高吞吐与极低延迟的数据流。正因如此，金融服务业一直是向公有云迁移最迟缓的行业——在华尔街，延迟增加就意味着真金白银的利润缩水。无论是高盛（Goldman Sachs）、摩根大通（J.P. Morgan）还是摩根士丹利（Morgan Stanley），所有顶尖金融机构都严重依赖专有硬件，根本无法脱离物理硬件完全在云端跑核心业务。Nikesh 回忆起当年给银湖资本（Silver Lake）做顾问时的情景，当时市场上一片悲观，人人都觉得“戴尔（Dell）已经完蛋了，没人再需要硬件”，而如今戴尔等硬件巨头的市值已经重新冲回数千亿美元规模。硬件非但不会消亡，反而是支撑算力运转最快、最不可替代的基础底座。 谈及硬件研发流程是否因生成式 AI 对芯片设计的加速而变短时，Nikesh 指出，当前硬件领域真正的“长板与瓶颈”（long pole in the tent）根本不是前端设计，而是实打实的后端制造与产能。如今市场上甚至连组装好的硬件机箱都拿不到，所有硬件元器件全线缺货排队。全球的代工厂产能全部被排得爆满，因为整个科技界都在疯狂抢购 GPU 模组板卡，以塞满全世界正在新建的 AI 数据中心。针对美国本土是否有能力承接并补齐这套供应链的疑问，Nikesh 直言至少需要 10 年时间，且必须配合坚定不移的自上而下的国家意志与资金承诺。不过令人欣慰的是，硬件行业眼下迎来了“一生难得一遇的大繁荣”（bonanza of a lifetime）。企业有充足的资本和底气敢于砸下 100 亿、200 亿甚至 500 亿到 1000 亿美元进行基建与建厂投资（例如存储芯片巨头在电视上高调承诺千亿级建厂扩产计划）。此外，美国税收政策中对资本开支（CapEx）的加速折旧机制（第一年可全额 100% 抵扣税款）也起到了巨大的推波助澜作用，驱动资本切实落地变现为物理产能。 - **金融机构的物理限制**：高盛、摩根大通等顶级机构无法承受云端网络抖动带来的延迟惩罚，“增加延迟即减少利润”，因此核心系统死守专有硬件。 - **供应链实际堵点**：芯片设计流程已被生成式工具极大提速，但物理层面的生产制造、元器件组装和晶圆制造能力全面超载。 - **美国本土制造周期**：在政策大力支持与自上而下推进的前提下，重建完善的供应链体系仍需至少 10 年。 --- ### [28:22](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1702s) 并购飞轮演进与 AI 驱动的万亿企业运营哲学 回顾 Palo Alto Networks 过去八年的跨越式增长，公司不仅有机增长极其强劲，更以快准狠的并购整合闻名，赢得了资本市场极高的信任度。在被问及是否会趁当前许多 SaaS 公司估值承压之际向网络安全之外更广阔的横向赛道出击时，Nikesh 拆解了公司并购策略的两阶段演变：在过去，团队执行的是“产品收购+渠道放大”经典战术——收购具备单点创新能力的产品型初创公司，重构其底层代码使之融入自身销售引擎；当原本向客户卖 1000 万美元产品的销售团队能顺带打包卖出 2000 万美元时，获客渠道支出（GTM spend）就得到了最高效的分摊。正是凭借这套打法，公司市值一路冲破 1500 亿美元。直到近期，团队预判到了身份安全（Identity）在智能体时代（Agentic AI）的关键转折点，因此迅速斥巨资完成并购，以此构筑未来防御壁垒。 然而，眼下正在出现的却是一个颠覆性的新机会：如果一家企业能够成为全行业利用 AI 运营业务最极致、最高效的范例，其营业利润率（Operating Margin）就能拉开同行数个身位（实现 90% 以上的毛利率和 40% 以上的净利率）。Nikesh 强调，一旦这套利用 AI 改造企业降本增效的代码被彻底破解，“你收购什么标的都不再重要”。目前市场上大部分中小规模（subscale）软件企业根本无力承担自身全面的 AI 化业务改造，而大厂一旦证明自己能接管一家利润率仅有 20% 的公司并迅速将其提效优化，华尔街和公开市场就会彻底放开限制。当主持人打趣提到他刚上任时资本市场极为悲观，怀疑“一个完全不懂网络安全、不懂企业软件的前谷歌高管怎么可能干成”时，Nikesh 笑着自嘲离开谷歌后能持续创业成功的高管确实“表现参差不齐”，但他用成绩予以了回击。展望未来 6 到 12 个月，整个行业将逐步看清 AI 在企业内部落地沉淀的真实形态。与大众普遍担忧“AI 普及将导致大裁员”的悲观叙事相反，Nikesh 提出了极具反共识的坚定观点：Palo Alto Networks 在技术研发人员上的雇佣规模不仅没有减少，反而比历史上任何时期都多，因为 AI 浪潮正在逼迫底层所有的业务架构推倒重写，“AI 正在倒逼一切经历彻底转型，若不是 AI 的到来，我们现在的技术团队人数绝不会有今天这么多。” ```mermaid flowchart TD accTitle: Palo Alto Networks并购与运营飞轮演进 accDescr: 展示Palo Alto Networks从传统产品并购渠道放大，转向AI赋能极致运营效率的新并购模式架构。 subgraph Phase1["传统并购飞轮 (0 到 1500 亿美元)"] direction TB A["收购前沿单点产品公司"] --> B["重写底层后端并融入统一平台"] B --> C["注入全球 GTM 销售引擎"] C --> D["客单价翻倍：客均 1000 万增至 2000 万"] D --> E["极度摊薄销售获客成本"] end subgraph Phase2["AI 驱动的万亿运营范式 (未来空间)"] direction TB F["利用 AI 重构底层业务流程"] --> G["拉开代差级利润率 (毛利90%+ / 净利40%+)"] G --> H["吞吐整合低效中小规模企业"] H --> I["突破单一网安赛道限制实现横向扩张"] end Phase1 -.->|"范式升级"| Phase2 ``` - **经典并购飞轮**：针对优秀产品团队进行后端架构再造，借助庞大的企业级销售网络实现交叉搭售，摊薄 GTM 成本，驱动市值从数百亿飙升至超 1500 亿美元。 - **AI 运营杠杆理论**：规模较小的 SaaS 公司无力独自完成 AI 架构重组；掌握 AI 原生化运营的大型平台能够通过改造低效标的释放巨大溢价空间。 - **关于技术人员的反共识洞见**：AI 没有消灭工程师岗位，反而因触发全系统的重构与底层安全重构需求，驱动 Palo Alto Networks 雇佣了创纪录数量的技术人员。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=hObRMv6qCi0&t=0s) From your mouth to God's ears — Nikesh Arora > [01:21](https://www.youtube.com/watch?v=hObRMv6qCi0&t=81s) Mythos was not oversold. It was legit — Nikesh Arora > [06:49](https://www.youtube.com/watch?v=hObRMv6qCi0&t=409s) 如果你是一家分析型 SaaS 公司（analytical SaaS），那你的生意已经彻底结束了（it's over）。 — Nikesh Arora > [09:00](https://www.youtube.com/watch?v=hObRMv6qCi0&t=540s) 企业软件和消费级软件的 UI，是我们这代技术人员做过最糟糕的事情。 — Nikesh Arora > [11:32](https://www.youtube.com/watch?v=hObRMv6qCi0&t=692s) 如果你加了 $，那你可比绝大多数人都领先了。 — Nikesh Arora > [18:26](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1106s) 在我的网络安全业务里，我要求的容忍度是 0%。 — Nikesh Arora > [20:24](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1224s) 兄弟，我是 Uber 的董事会成员，我可不能对自家公司评头论足，Dara Khosrowshahi 是个极棒的 CEO，别想让我被开除。 — Nikesh Arora > [20:24](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1224s) 这车确实好使，体验太棒了。唯一的建议是他们应该更快、更激进地推向全球更多城市。 — Nikesh Arora > [28:22](https://www.youtube.com/watch?v=hObRMv6qCi0&t=1702s) AI 正在倒逼一切经历彻底转型，若不是 AI 的到来，我们现在的技术团队人数绝不会有今天这么多。 — Nikesh Arora ## Mentioned Palo Alto Networks, Nikesh Arora, All-In Podcast, Chamath Palihapitiya, Jason Calacanis, David Sacks, David Friedberg, Anthropic, Claude, Claude Mythos, Claude Code, OpenAI, Codex, Google, Salesforce, Slack, SAP, Databricks, Snowflake, MongoDB, Oracle, Zoom, PG&E, Change Healthcare, UnitedHealth Group, NSA, Mercedes-Benz, Uber, Dara Khosrowshahi, Waymo, Goldman Sachs, J.P. Morgan, Morgan Stanley, Silver Lake, Dell, Netflix, IBM ## Video description (0:00) Palo Alto Networks CEO Nikesh Arora joins the Besties! (0:47) Claude Mythos found years of vulnerabilities in Palo Alto's code in weeks (5:15) Are cyber defenders losing the race against AI attackers? (6:50) Analytical SaaS is dead, so what survives the AI wave? (14:06) If models become a utility, where will the money be made? (20:35) Armchair CEO: Nikesh rates Waymo, Google, and OpenAI (28:22) Palo Alto's M&A playbook and the path to $1 trillion Follow Nikesh: https://x.com/nikesharora Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg Follow on X: https://x.com/theallinpod Follow on Instagram: https://www.instagram.com/theallinpod Follow on TikTok: https://www.tiktok.com/@theallinpod Follow on LinkedIn: https://www.linkedin.com/company/allinpod Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg #allin #tech #news [Source](https://www.youtube.com/watch?v=hObRMv6qCi0)

[0:00] Nikesh Arora: From your mouth to God's ears [1:21] Nikesh Arora: Mythos was not oversold. It was legit [6:49] Nikesh Arora: 如果你是一家分析型 SaaS 公司（analytical SaaS），那你的生意已经彻底结束了（it's over）。 [9:00] Nikesh Arora: 企业软件和消费级软件的 UI，是我们这代技术人员做过最糟糕的事情。 [11:32] Nikesh Arora: 如果你加了 $，那你可比绝大多数人都领先了。 [18:26] Nikesh Arora: 在我的网络安全业务里，我要求的容忍度是 0%。 [20:24] Nikesh Arora: 兄弟，我是 Uber 的董事会成员，我可不能对自家公司评头论足，Dara Khosrowshahi 是个极棒的 CEO，别想让我被开除。 [20:24] Nikesh Arora: 这车确实好使，体验太棒了。唯一的建议是他们应该更快、更激进地推向全球更多城市。 [28:22] Nikesh Arora: AI 正在倒逼一切经历彻底转型，若不是 AI 的到来，我们现在的技术团队人数绝不会有今天这么多。

Nikesh Arora: From your mouth to God's ears. Nikesh Arora: talk to Joe because he knows how to solve the problem and Jim doesn't. Nikesh Arora: not oversold Nikesh Arora: Not as well as we should be doing, which is great for our business, but that's a different story. Nikesh Arora: what good looks like and what bad looks like. Nikesh Arora: enterprise software and consumer software UI is the worst thing we did as technologists. Nikesh Arora: five people become one in a company dollar sign password [14:52] the profit pools are in applications that companies can use [20:24] Nikesh Arora: the cars work [20:24] Nikesh Arora: the first trillion-dollar company in our lifetime. [22:07] Neo in the Matrix [28:51] Nikesh Arora: the most efficient way for me to amortize my go-to-market spend. [28:51] Nikesh Arora: crack that code, then it doesn't matter what you buy. [30:33] the menu's open

[0:00] One of the biggest winners right now. [0:01] [music] The big daddy of the [0:02] cybersecurity space. Palo Alto Networks [0:04] is a now performer in the space. [0:07] CEO Nikesh Arora [0:09] >> This might [music] come as news to you, [0:11] but humans have been writing bad code [0:13] for a very long time. [0:14] >> I spent 10 years [music] at Google and [0:16] you know, Google search was [0:17] democratizing information. If you take [0:19] that analogy and think about what AI is [0:21] doing, AI is democratizing [music] [0:24] intelligence. [0:25] >> Money is a way to keep track. [0:27] >> Yeah. [0:27] >> It's not the goal. [0:29] >> You've been a CEO of Palo Alto Networks [0:31] for 8 years? [0:33] >> Coming up on 8 years this week. [0:34] >> And I think when you started it was $17 [0:36] billion market cap if I remember [0:38] correctly. [0:39] >> There about. [0:39] >> And this morning I checked it's $238 [0:41] billion. Which if you listen to what we [0:43] said yesterday, now that you passed 100, [0:45] you're more likely to actually 10X. So [0:47] the first 10X was actually much much [0:49] harder. So you're on your way to a [0:50] trillion dollars. [0:51] >> From your mouth to God's ears. [0:53] >> Well, I I think you are. Okay, so let's [0:55] just double click into what you see [0:57] because you are [0:58] sort of in a really interesting position [1:00] to see all of it. [1:02] You see the birth of AI. Maybe you see [1:05] you've seen the rise and fall of SAS. [1:08] All the models talk to you. You were one [1:10] of [1:10] >> The rise again, right? [1:11] >> The rise again. [1:12] Uh you were one of the first and the few [1:14] that got access to Mythos. So just Let [1:17] me just push the button. Go Nikesh, [1:19] start. [1:20] >> [laughter] [1:21] >> Well, uh first of all, thank you for [1:22] having me here. [1:24] I think AI is exciting. [1:26] I think it's exciting to see all the [1:28] stuff that's gone down in the last [1:30] possibly 24 months. [1:32] Um [1:33] I think Sarah just said it. They were [1:36] right in anticipating the huge amount of [1:38] compute that was going to be needed. So [1:40] all that stuff's going on. But [1:43] you can see [1:44] that [1:46] you know, this this notion which we [1:47] talked about briefly last time that AI [1:50] is really democratizing intelligence. [1:53] What that means is I have 250 people in [1:55] marketing. They produce varied forms of [1:57] output. Now you can get 90% of the [2:00] output to be consistent across those 250 [2:02] people. I have 5,000 people who talk to [2:05] customers. [2:06] There's My my failure mode is when 5,000 [2:09] people do different things, where people [2:11] say, "I want to talk to Joe because he [2:13] knows how to solve the problem and Jim [2:15] doesn't." [2:16] So now you can get 5,000 people to act [2:18] almost consistently in their [2:20] interactions with people on the other [2:22] side. So I think it's going to have a [2:23] phenomenal impact to how we run [2:26] businesses, how we operate. It's going [2:29] to change the entire landscape. [2:30] Now, [2:31] in that context, you touched upon [2:33] Mythos, and you know, Dave has been very [2:35] involved with this. [2:36] Mythos has shown us that all the bad [2:38] code that humans have written over the [2:40] last 50 years [2:42] can be assessed by AI and shown uh the [2:45] vulnerabilities can be shown. We tested [2:47] for 6 weeks, and in 6 weeks we found [2:50] what would have taken us 5 to 7 years. [2:52] >> Wow. Say that one more time. [2:54] >> In 6 weeks we found vulnerabilities [2:57] which would have normally taken us 5 to [2:59] 7 years to find. [2:59] >> So Mythos, but these are vulnerabilities [3:01] where? [3:02] >> Sorry? [3:02] >> These are vulnerabilities in your own [3:04] code base or in your customer or in your [3:06] own code? [3:07] >> Oh, wow. [3:07] >> So Mythos was not oversold. It was [3:10] legit. [3:11] >> The capabilities of AI in being able to [3:15] assess vulnerabilities in code are real. [3:18] Not just that, [3:20] if you put it on ultra mode, which is [3:22] persistent thinking, so it keeps trying [3:23] until it gets an answer, you can [3:26] actually daisy chain vulnerabilities, [3:28] i.e., finding a new attack path into [3:30] your company into your vulnerabilities. [3:32] Now, we pride ourselves as a top [3:34] percentile of companies that test our [3:36] code because we're in cyber security [3:37] business. If you take that and compound [3:40] that across all the companies that exist [3:42] in the world that write their own code [3:43] or the 10 million developers write code, [3:45] this thing is going to find stuff which [3:48] would have taken us 10 years to find. [3:50] >> How much did it cost? Like did you track [3:52] the token cost? Was it a hundred million [3:54] dollars, ten million dollars? [3:56] >> No, it was in the low millions. But [3:58] again, you know, the cost as Sarah said, [3:59] the cost curve is going to come down. [4:01] Already OpenAI has got a model which is [4:03] cheaper and more consistent. You know, [4:05] Anthropic's come out with another model [4:07] >> you buy the hype. [4:08] >> It's not hype, it's true. [4:10] >> It's That's the point. [4:11] >> the capabilities are [4:12] >> The You know that. [4:12] >> The capabilities are true. [4:14] >> Yes. [4:15] >> I mean, you saw IBM announce a project [4:16] for five billion dollars to fix open [4:18] source. That's the biggest problem. [4:19] >> What would have happened if Claude [4:21] didn't have the restraint and they put [4:23] it out in the public? Do you think it [4:25] would have been like a real attack [4:27] vector and caused chaos in corporations? [4:30] >> we're three months away, if not already [4:32] there, from this being available in [4:37] the wild. [4:38] >> Okay, open source. [4:39] >> Yeah, just three months. [4:41] >> Yeah. [4:41] >> Yeah, cuz I mean, we've been saying that [4:43] it's roughly six months away before [4:45] mythos-level capabilities are available [4:48] in Chinese models, you know, open [4:51] models, whatever. But you're saying it [4:52] could be three months. [4:53] >> Well, look, there's what, is 4.8 is [4:55] already out, 5.5 is already out. They [4:57] have similar capabilities. And look, you [4:59] don't need to [5:01] crack the hardest code to crack. Just [5:03] need to find a few vulnerabilities in [5:05] code that are out there. Just take an [5:08] Take an old industrial system which is [5:10] running, you know, OT code on the edge. [5:12] You can find that vulnerability [5:13] reasonably easily. [5:14] >> So, so we're in a race right now between [5:17] the cyber defenders finding these [5:19] vulnerabilities and patching them before [5:22] the cyber attackers do the same thing. [5:25] >> Yes. [5:25] >> And how do you feel like we're doing in [5:28] that race? [5:29] >> So, [5:30] not as well as we should be doing, which [5:32] is great for our business, but that's a [5:33] different story. [5:34] >> [laughter] [5:35] >> So, like, every company has to go look [5:37] at their code base and figure out where [5:39] the vulnerabilities are and fix them. [5:41] So, if you talk to CIOs today, their [5:43] biggest problem is all the vendors are [5:45] showing up saying, "Please patch my [5:47] piece of boxes that hardware that you [5:49] have please patch my code that you have [5:51] because I found vulnerabilities fix it. [5:53] While the CISOs are busy finding their [5:55] own vulnerabilities to fix their own [5:56] vulnerabilities and then this huge thing [5:58] called open source which nobody knows [6:00] quite how to solve. [6:01] >> So is it is it fair to say that it as [6:03] model capabilities go up [6:05] systemic business risk of large [6:08] enterprises also goes up? [6:10] >> On the cyber side, yes. [6:12] There are antidotes being built by [6:14] people like us and others where we're [6:15] going to provide some capability where [6:16] you don't have to patch everything. But [6:18] look, cyber has done something very [6:19] interesting around harnesses, memory and [6:22] context. Right? [6:24] The part we don't talk about here is [6:26] organizations don't have memory and [6:28] context of everything they do every day. [6:31] That's why you need to store a lot more [6:32] data [6:33] enterprise-wide to learn what good looks [6:36] like and what bad looks like. [6:37] >> Right. The same problem is in [6:38] cybersecurity. [6:39] >> We need to collect We need to collect 10 [6:41] times the data in the enterprise from a [6:43] cyber perspective to be able to [6:45] understand how to [6:47] defend ourselves against the AI [6:48] attackers. [6:49] >> Do you think that the traditional [6:51] companies like the SaaS businesses that [6:53] have existed [6:55] in this world, what is their place? [6:58] As all this knowledge becomes more [6:59] persistent and stored, what happens to [7:01] SaaS? [7:03] >> Well, [7:04] you see SaaS is as Bill said, SaaS is [7:06] different pieces, right? [7:07] >> Okay. [7:08] >> If you're an analytical SaaS company, [7:09] it's over. [7:11] >> It's over. What is an analytical SaaS [7:13] company? [7:13] >> Somebody that says I'm going to collect [7:15] a lot of data for you and analyze it for [7:16] you. I don't need you to analyze it for [7:18] me. I can run [7:20] models against data and analyze them [7:22] myself. So if you think about there's a [7:24] lot of every SaaS company has a [7:25] marketplace. You can buy Salesforce [7:27] marketplace. What do they say? You have [7:29] Salesforce data, I'm a marketplace app, [7:31] take me and I'll help you analyze the [7:33] data. I don't need you. [7:34] >> You don't need that. [7:35] >> I can just go run an LM against the [7:36] data. So the entire incrementality that [7:39] has been sold as incremental software [7:41] modules to all of us [7:43] doesn't need to be sold to us because [7:44] I'd much rather have LLMs running [7:46] against that. [7:46] >> Interesting you bring this up. We had an [7:48] instance with a SaaS product with 20 [7:50] seats. [7:51] Nobody was logging in and using it, but [7:53] the data was there. [7:54] >> Yes. So, we created like three accounts, [7:58] got rid of 17, connected it to Slack, [8:01] connected it to Claude, and now [8:02] everybody can interface it through [8:04] natural language and we've reduced our [8:06] bill by 90%. [8:08] >> Well, not just that, what are you going [8:09] to do next? They [8:10] Jason said, you're going to take data [8:12] from different products, put them in one [8:14] place, run the analytics against that. I [8:17] want my data for my sales reps, my [8:19] productivity data, my you know, [8:21] inventory data from SAP. I want it all [8:23] in one place so I can run analytics [8:25] against it and say, "Who's selling a [8:27] lot? Where do I have less inventory? [8:29] Let's build inventory in a region where [8:30] my sales people are extremely [8:32] productive." To run that query, you'd [8:33] have to have talked to three different [8:35] SaaS products. Tomorrow, you can put all [8:37] the data in one place. So, so that's [8:39] sort of category one. [8:40] >> Analytics SaaS today. The analytics [8:42] >> Okay, [8:43] category one analytics dead. [8:45] >> Yes. In the medium term, you get all [8:47] these bounces today and tomorrow that's [8:48] these are marginally irrelevant. [8:50] Infrastructure software undervalued. [8:53] >> Okay, what is infrastructure software? [8:55] >> Stuff that gives you databases. You [8:56] collect data into it. Stuff that allows [8:58] infrastructure to work, whether it's a, [9:00] you know, database software. [9:01] >> Databricks, Snowflake, like that. [9:03] >> Databricks, Snowflake, MongoDB, Oracle, [9:06] Oracle, all these things you need [9:08] core storage infrastructure, core data. [9:10] You're going to need 10 times the data [9:11] stored in an enterprise than we have [9:13] today. Right, 3 years. 10 times. [9:15] So, anything that helps you collect [9:17] infrastructure data, manage it, you [9:19] need. [9:20] I think the category in the middle is [9:21] called, let's call it system of work [9:24] or system, you know, of record, people [9:26] call them. [9:28] Those are deeply embedded in the way [9:29] businesses work. I have 6,000 sales [9:31] people, they know how this works. What's [9:33] going to happen is step one, we will [9:35] take away UI and let agents do the work. [9:38] UI [9:39] enterprise software and consumer [9:41] software UI is the worst thing we did as [9:44] technologists. [9:45] >> You had a couple of examples of this. [9:46] You told me this story, I don't know if [9:48] you want to repeat it, of this one [9:49] company they tried to hold you hostage [9:51] on a license. [9:52] >> Yes, that was analytical SAS, so that's [9:53] over. [9:54] >> just pointed AI at it and you just [9:55] >> Yes, we just got rid of them. That's a [9:56] different issue. But, I mean, think [9:58] about it. Today, we spend our lives [10:01] having product managers design UI so all [10:03] humans can interact with data behind the [10:05] UI. [10:05] >> Yeah. [10:06] >> If all like if you believe agents are [10:08] going to work, and I say, I just tell an [10:09] agent, look, figure out from my sales [10:11] call, figure out the key points, and go [10:13] post it into [10:15] you know, whatever sales tracking system [10:17] I have, whether it's Oracle or [10:18] Salesforce, right? An agent conceptually [10:20] should be able to do it. we're [10:21] spending trillion dollars building these [10:22] agentic backends, we need these agents [10:25] to be able to do it. If that happens, UI [10:27] goes away. [10:28] If UI goes away, I can rewire my system [10:32] of work. [10:34] Right? I have sales guy should have to [10:35] say, I had the sales call, do all the [10:38] paperwork and all the that needs to [10:39] happen in the back of the company, and [10:41] just [10:41] I'm done. [10:42] >> Right. [10:43] >> If I can change the way [10:45] work happens, which is where you will [10:47] get true efficiency, where five people [10:48] become one in a company, [10:50] all these SAS software that does system [10:53] of work needs to be re-engineered for [10:54] the next five years. [10:55] >> And it's also happening passively, which [10:57] is really interesting. It's looking at [10:59] email, it's automatically taking the [11:01] Zoom transcript and summary. So, the [11:03] sales system of record is now like, you [11:05] don't even need to input it. It's like, [11:07] I already have the Zoom call notes, I [11:09] have the deck the deck was made, the [11:12] sales deck was made by AI. It's just [11:16] we're we're all going to be looking at a [11:17] chat window and just saying, here's what [11:19] I want. [11:20] >> Your audit trail becomes a lot better [11:21] because humans are not touching your [11:22] data. It's always being managed by [11:24] agents, so I think the whole system of [11:26] work, system of record, gets reinvented [11:29] in the next five years. [11:30] >> Yeah, there's no data entry. That's an [11:31] interesting point. Yeah. Let's talk [11:32] about national security for a second. I [11:34] just want to maybe zoom out. So, [11:38] the one side of Mythos, as you said, is [11:39] like the value that it has to you and to [11:41] your and to enterprises. [11:44] The red team version of Mythos is where [11:47] foreign state actors or you know, can [11:50] essentially create economic havoc inside [11:52] of a country. [11:53] >> Yes. [11:54] >> As these models escalate in their [11:55] capability, what do you think should [11:57] happen when these models are ready? [12:00] >> You know, the sad truth is, you know, [12:01] here there's a few thousand [12:03] breaches or attacks that happen. [12:06] They happen for pretty rudimentary [12:07] reasons. It's not because somebody [12:09] cracked a hard to crack thing. It [12:11] happens because 89% of attacks happen [12:13] because credentials get stolen. [12:15] >> Or your username and password. [12:16] >> That's it. [12:17] >> I bet my password is password. [12:19] >> Yeah, I'm sure it is. Did you have [12:21] dollar sign? [12:21] >> Dollar sign password. [12:22] >> Fantastic. Well done, see? You're [12:23] already ahead of everybody else. So, [12:25] 89% of breaches happen because of simple [12:27] things. So, I don't think we need more [12:28] models to go crack this stuff. Now, we [12:30] will need these models can attack [12:32] critical infrastructure and things we [12:34] try and protect from a national security [12:35] perspective. Yes, we need defenses [12:37] there. I'm not worried about [12:40] the national security part being [12:41] protected because they're very on it. [12:43] They're the right people. They spend 10% [12:45] of their budgets on IT on security. I'm [12:47] worried about the small offices across [12:50] the country where they're using some [12:52] piece of packaged software and you're [12:53] running a dentist's office or doctor's [12:55] office. Remember when then Change [12:57] Healthcare got breached? [12:59] >> Every physician's office shut down. [13:01] >> Shut down and it's ransomware. [13:02] >> Because of ransomware to Change [13:03] Healthcare. That's was the clearing [13:04] system. That's when [13:06] United Health had to actually have give [13:08] billions of dollars of credits to the [13:10] physicians to be able to run their [13:11] businesses at that point in time. [13:13] >> That's what one should worry about. It's [13:14] less about [13:15] >> the big nuts will get cracked. [13:17] >> about cracking some PG&E power [13:20] generation facility. It's more economic [13:22] chaos. Yes. [13:24] And so, what what what do we do? [13:28] >> I don't think there's a sort of a silver [13:30] bullet. I think this will take time. I [13:33] think this will basically take a while [13:35] until every system gets upgraded, [13:37] renewed, fixed over time. I just think [13:39] it increase the terminal value of the [13:41] industry, right? [13:43] >> Do you think that there's a world in [13:45] which these models become so good that [13:48] you could see yourself advocating for [13:50] more nationalism around how they're [13:53] controlled and [13:54] how they're managed and how they're [13:57] where we point them? Or do you think [13:58] there should be a maybe a set of these [14:00] models that never see the light of day [14:02] that only the NSA and other folks get [14:04] have access to or guys like you? [14:05] >> I have a slightly differentiated view [14:07] about models and how they will evolve [14:09] versus what we heard earlier from an [14:11] open AI perspective. [14:13] I think [14:14] I still believe models are going to [14:16] become a utility layer. [14:18] You'll be able to buy intelligence on [14:20] the [snorts] fly. [14:22] Or you can say, "I don't need a 180 IQ [14:24] person to go do this task. Give me a 120 [14:27] IQ and I need a 250 IQ to do this task. [14:29] I'll pay $10 for this or for this I'll [14:31] pay 1 cent." So, I don't know there's a [14:33] one-size-fits-all [14:35] give you the most up-to-date model to [14:37] answer my customer call saying, "Sorry, [14:38] sir. I have no idea how to solve your [14:40] problem." [14:41] So, [14:42] I think models will get differentiated [14:43] from utilitarian perspective. [14:46] Um so, if you look at already what's [14:48] happening in the market, right? The [14:50] profit pools are in applications, not in [14:52] models. More Sarah talked about Codex [14:56] running away. She didn't say [14:57] Open AI is running away. She just Codex [14:59] is running away. Just say that's the way [15:00] I'm sure Dario says Claude code is [15:02] running away. So, you're seeing that [15:03] >> They're attacking profit pools. [15:05] >> They're attacking profit pools because [15:06] that's where the money's going to come [15:07] from. The profit pools are in [15:08] applications that companies can use. The [15:10] profit pools are not in model usage by [15:12] companies because most companies have no [15:14] idea how to use the models. [15:15] >> at these companies in a way Open AI and [15:18] Anthropic as the new Microsoft Office [15:21] coming in and doing all applications, [15:24] all productivity software for [15:26] organizations. [15:27] >> No, I see there's going to be [15:28] application companies that are going to [15:30] arbitrage between models and solve your [15:32] business problem. [15:33] >> So, you still think they won't go to the [15:35] application layer? Because this is a big [15:37] debate. Should you engage with OpenAI [15:40] and train their systems to then take [15:43] your business from you? [15:45] And Anthropic keeps releasing their [15:47] legal model, their accounting model. And [15:49] it does feel like in order for them to [15:51] hit their revenue numbers, they might [15:52] need to do what Microsoft did, which is [15:55] release the Office product on top of the [15:57] operating system. [15:57] >> See, if I'm a company, I don't want to [16:00] write every piece of software myself. I [16:03] want my HR system software, which is [16:05] agentic enabled and AI enabled to be [16:08] delivered by some application company. [16:10] It'll be a new AI application company. I [16:12] want my sales management system built by [16:14] the new agentic AI sales force of the [16:17] world, whether it's sales force or [16:18] somebody else. So, I want applications. [16:20] Now, what [16:22] Sarah said is the profit pools are in [16:23] the application layer. That's why they [16:25] want to be the application layer. So, I [16:26] think we're still waiting for that layer [16:29] of companies to be invented or created [16:31] where applications will sit. Because [16:33] 50,000 uh companies need the same [16:35] application. Why would I build it [16:37] myself? It's highly inefficient. It's [16:39] silly for me to use OpenAI directly and [16:41] rewrite my entire sales system because [16:43] uh I'm smart. Right? I'm not. I want [16:45] somebody to do it for me. So, I think [16:47] that layer of companies is still not [16:49] fully formed. [16:49] >> Or so, we're going to be waiting for it. [16:51] >> control plane, a harness, and then [16:52] >> That's right. They will build the [16:54] harnesses and the memory into those [16:56] application layers. Now, the question is [16:57] how big is the application layer? Is it [16:59] one application? Is it Is one, you know, [17:01] enterprise application that does [17:03] everything? Or is it specialized [17:05] application? [17:05] >> and you kicked out this software vendor, [17:08] you did it because they were being [17:09] abusive in pricing. So, that [17:10] >> use a different vendor. [17:12] >> What's that? [17:12] >> We swapped out for a different vendor. [17:13] We just took more control. [17:14] >> Love it. So, it really is a pricing [17:16] issue. And and that's why the SAS [17:18] apocalypse in some ways makes sense. [17:21] They're not having pricing power because [17:23] you could say, "Well, I'll just put 10 [17:24] developers on this and I'll save $10 [17:27] million." Yes. [17:28] >> I think the part back to what Chamath [17:29] said about the regulation or whether you [17:31] want to regulate these higher-powered [17:33] models, the question is at some point in [17:34] time [17:36] when these newer models, which are even [17:37] more powerful, get built, they will come [17:39] at a different price point and they [17:41] might have to go through a certain [17:42] vetting process to understand what their [17:43] capabilities are. But I think we're in a [17:45] global race. [17:47] I don't think holding back our models [17:48] for 3 to 6 months is going to help us [17:50] any. Somebody else is going to put them [17:51] out in open source. [17:53] I I was [17:54] I was shocked to hear when I was talking [17:55] to the CEO of one of these model [17:57] companies. He says, "The entire weights [17:59] of their most recent model can fit on a [18:00] USB stick." [18:02] >> Say that again. The entire weights [18:05] >> model weights of their newest model fits [18:08] on a USB stick. That's the IP. [18:12] >> Yeah. [18:12] >> That's incredible because all the data [18:13] can be distilled in under 24 to 48 hours [18:16] and model comes out. [18:18] >> I'm curious [18:18] >> That's the IP. So, are you telling me [18:21] that [18:22] you know we can hold on to that for 6 [18:24] months? [18:25] >> Right. [18:26] We we have a debate about um how [18:29] difficult it is to make a frontier [18:31] model. Some companies are starting to [18:33] think about making frontier models using [18:35] their data advantage to to build their [18:38] own. [18:39] Have you thought about that at [18:40] Palo Alto because it [clears throat] [18:42] does seem like you have proprietary [18:43] knowledge on how security works. Could [18:46] you build your own large language model [18:47] or a VSML, a small language model that [18:50] would give you some advantage in the [18:52] future? [18:54] >> nobody talks about [18:56] is the false positive rates on the [18:57] models. [18:59] What is the false positive rate on 4.8 [19:01] and 5.5? [19:03] >> No idea. [19:04] >> You guys don't talk about it. You [19:04] should. The false positive rate on MSO [19:07] was 30%. [19:08] >> Oh, wow. [19:10] >> Right? [19:11] >> So, it thought it found something, but [19:13] it hadn't. [19:14] >> Yes. So, the problem is [19:16] it's great for attack, it's horrible for [19:18] defense. [19:19] Cuz you find 30 times 30% of the time [19:22] you find something that says, "I found a [19:23] problem." And you say, "Let's plug the [19:24] hole." Wait, there wasn't a hole there [19:25] in the first place. [19:26] >> No missile inbound. [19:27] >> Right. [19:28] >> Yeah. [19:28] >> So, now the same problem applies in [19:30] enterprise. If you use a If you use a [19:31] model without the right harnesses, the [19:33] right training, you could be running [19:35] into 10 20% false positive rates. Let's [19:38] use the model to pay I don't know, [19:40] insurance claims. [19:41] >> Yeah. [19:42] >> Oh, great. 10% 20% false positive. I [19:44] just lost money. [19:45] >> The sycophantic nature of these is [19:47] ridiculous, yeah. [19:48] >> So, so the problem is not who wants the [19:51] newest model. The problem is how do you [19:53] take that model with 20% or 10% false [19:55] positive and make it 0.01% false [19:58] positive. In my business, I want 0%. [20:00] >> Without losing the false negative. [20:02] >> Sorry? [20:02] >> Without losing the negative, the false [20:04] negative. [20:04] >> Yes, but it's like saying, "Hey, let's [20:06] take the new self-driving car. Mercedes [20:07] is going to use Opus 4.8 and you can [20:10] just sit in the car and it's going to [20:11] drive you." I'm not putting my kids in [20:12] that car with a 10% false positive rate. [20:14] Are you? [20:16] So, there's a lot of work that happens [20:17] post a model, which needs to happen to [20:20] make this thing [20:21] useful and effective in the business [20:23] context. [20:24] >> Let me slightly pivot for a second. You [20:26] were for a very long time the chief [20:28] business officer at Google. [20:30] You were the president of SoftBank. [20:32] Now you're the CEO of Palo Alto [20:33] Networks. So, let's play armchair CEO. [20:36] >> Armchair CEO. [20:37] >> Armchair CEO. [20:37] >> I'm still I'm still bristling from David [20:39] Feiberg trying to create a distinction [20:41] between founder CEOs and non-founder [20:43] CEOs. Just saying. [20:44] >> [laughter] [20:44] >> Just saying, David. [20:45] >> By the way, false positives. [20:46] >> Sorry? [20:47] >> And false negatives, too. [20:49] Give us what you would keep, what you [20:50] would change, and what you like about [20:52] the following companies. [20:54] >> This is going to get recorded and put [20:55] out there to say that really something I [20:56] don't know. [20:57] >> thoughts. You're one of the smartest [20:58] business people [20:58] >> don't like get to live with the glory of [21:00] these all-in podcast sessions. [21:02] >> Don't [21:03] >> Okay, ready? [21:03] >> people and roasting people. [21:05] >> ready? [21:05] >> Yeah, sure. [21:06] >> Okay. [21:07] >> [laughter] [21:08] >> What you keep, what you change, what you [21:09] like, what you don't like. Uber. [21:12] In a world of a [21:13] >> it, dude. I can't talk about my [21:15] >> on the board of Uber? [21:15] >> I'm on the board of Uber. I'm not going [21:16] to talk about Uber. [21:17] >> I didn't know that. Sorry. Okay. [21:19] >> Dr. Dara, he's the CEO. He's a great [21:20] guy. [21:21] >> Okay. [laughter] [21:22] Uh Waymo. [21:23] >> You're trying to get me fired. [21:25] >> Waymo. [21:26] >> What do I like about Waymo? The cars [21:28] work. It's amazing. [21:29] They should have more. [21:31] In many more cities around the world, [21:32] faster. I I I would say that at the rate [21:34] that I'm going to be fired. [21:36] >> Google red large. [21:38] >> I think Google's underrated. [21:40] I think it's going to be the first [21:41] trillion-dollar company in our lifetime. [21:43] I think they have all the assets that [21:44] are [21:45] that are needed to make this successful. [21:46] I think people underestimate. You can be [21:48] a model company, you still need to have [21:49] a sales force that convinces customers [21:51] to go out there and embrace these models [21:53] and buy them. And if you think about it, [21:55] three hyper scalers have the biggest [21:57] number of sales people out there. So, [21:58] they should [21:59] >> of why they're a little bit undervalued [22:00] is just the conglomerate nature is hard [22:02] to understand? [22:03] >> I don't know. You guys are smarter than [22:04] I am. I'm just a hired hand CEO. [22:07] >> I didn't say that. Reeboks said that. [22:09] Let's just be clear. [22:10] >> I know. I know. [22:11] >> I was I was providing a thesis on [22:13] recovery out of the SAS pack. Let's just [22:15] say. [22:15] >> Okay. Okay. Got it. [22:16] >> Just to be clear, [22:18] there's there's a way to segment that [22:19] basket. Okay? And you're not in that [22:21] basket. [22:22] >> I thought you were making a distinction [22:23] about how people who are founders CEOs [22:26] have [22:27] uh have the right to take more risk [22:29] and are allowed to take more risk. [22:31] >> saying that. [22:32] And I think and I and I think you you [22:33] provide a unique counterpoint to that. [22:36] And and there's not a lot of people like [22:37] you. I think the same would be true of [22:38] Jeff Weiner. And I think that there's a [22:39] few other [22:42] uh really great CEOs, but they are like [22:44] Neo in the Matrix type anomalies. And I [22:47] think you're one of those people. And [22:48] there's a very rare kind of personality [22:50] profile of someone that's willing to [22:52] take risk and take ownership of [22:53] something that wasn't theirs in the [22:55] first place and they make it theirs. And [22:57] uh it's a it's a extraordinarily unique [22:59] trait. Far more unique actually than [23:01] being a scalable founder. [23:03] >> That's an incredible save. [23:04] >> You're forgiven. [23:05] >> Yeah, good save. Incredible save. [23:06] >> back to Armchair CEO. [23:07] >> Wow, that was incredible. [23:09] >> He's more sycophantic than ChatGPT. [23:11] [laughter] He's like, "Actually, I'm [23:13] actually I'm actually the best." [23:15] >> Let's go Let's go back to Armchair CEO. [23:16] >> I'm liking this. He has Open AI more [23:18] often. Yes. [laughter] [23:20] >> They do sell faster. [23:22] >> Open AI. [23:22] >> They should sell faster, right? [23:24] >> They should sell faster. [23:26] >> I mean, I I you said it. Didn't you just [23:28] say it when you were Sarah was here that [23:29] >> And Anthropic seems to have improved [23:31] their ARR much faster than Open AI. [23:33] >> I mean, that's just the statistics. [23:35] >> They kind of went all in on enterprise, [23:37] and including specifically. [23:38] >> I think I think that's like [23:40] the the conversation right now is it's a [23:43] race to take over the profit pools. [23:46] If you are going to need tens and tens [23:47] of billions of dollars every year to get [23:49] What is that? 1 gigawatt is 10 billion [23:51] of revenue. [23:51] >> to build you? What are the most [23:52] What are the most exciting profit pools [23:54] then? [23:54] >> So, what are the most exciting [23:55] >> It costs 50. So, this is not a great [23:57] deal. [23:57] >> So, what are the most exciting profit [23:59] pools then? So, we [24:00] >> You've got coding. That's been the [24:01] breakout application over the past year. [24:03] It's massive. [24:04] You've got infrastructure, like you [24:06] said, the new databases. I think [24:09] cybersecurity is clearly one of them [24:11] because of the threats and patching [24:12] cycles so much more dynamic. [24:14] >> There's There's a slight difference in [24:15] in Yes, as you can see, these models are [24:17] trying to be the the enablers of better [24:20] cybersecurity, which is good because all [24:22] of us need to use them to test. And [24:23] you're probably going to see I mean, you [24:25] you saw Anthropic is already [24:27] uh made their cyber capable model [24:31] available generally, so that everyone [24:32] can use it. And Open AI has got one. I'm [24:35] sure Google has one, too. But they [24:36] understand this is a place where CISOs [24:39] or chief security officers want to use [24:40] it to test the code. So, this is another [24:42] profit pool. Uh I think we haven't seen [24:45] the onslaught against the application [24:47] software companies yet. I mean, there's [24:49] tens and tens of billions of dollars in [24:50] application software, which is waiting [24:52] to get reinvented, as we talked about. I [24:54] think eventually you'll see these people [24:55] saying, "What if I took this 40, 50, 100 [24:58] million dollar time down, I can build a [25:00] whole brand new backbone with a [25:01] generative AI, and it'd be so [25:03] differentiated that it will cause [25:05] customers to move." [25:06] >> We are seeing it as a playbook in the [25:08] accelerators [25:10] now. The year zero and year one [25:12] companies, people are coming to us with [25:13] the pitch, "This is a thousand dollar a [25:15] seat per year, [25:17] five hundred dollars a month seat SaaS [25:19] software. We can do it for less. We're [25:20] going to charge them based on [25:21] consumption. We're going to take 80, 90% [25:24] of the cost out as [25:25] >> What are the two fastest places to make [25:27] revenue? [25:27] >> The two fastest places to make revenue? [25:29] >> Yeah. [25:30] >> In enterprise a replacement apps. If you [25:32] replace something, I already have a [25:33] budget, it's easy. I take something bad, [25:36] I replace it something better, [25:37] I get money. So, replacement apps are [25:39] beautiful. If you can replace an [25:41] industry, replace the profit pool, it's [25:43] great. The second place is consumer [25:46] revenue. It's a lot easier to get five [25:48] bucks from a per user on a consumer [25:50] side. [25:50] >> Netflix. [25:51] >> So, that's where I mean, look at it. I [25:53] think we collectively probably pay more [25:55] on subscriptions per month than we ever [25:57] did historically, and you thought your [25:58] cable bill was high. [26:00] >> Yeah. [26:00] Do you think that you're going to end up [26:02] building more or less hardware in the [26:03] future if you had to guess? [26:05] >> Hardware, even today, is the cheapest [26:09] way to uh [26:13] manage low latency, high throughput [26:14] bits. [26:16] You still need a data center. [26:17] >> Yeah. [26:17] >> What's a data center doing is just [26:19] managing high throughput, low latency [26:21] bits. [26:21] >> Yeah. [26:22] >> That's why if you look financial [26:23] services is the most reluctant industry [26:26] to go to the cloud [26:28] because you increase latency. [26:29] If you increase latency, you reduce [26:31] profit. So, if you look at every of your [26:33] largest financial services company, [26:34] whether it's Goldman or JP Morgan, [26:35] Morgan Stanley, or [26:37] these guys, they're using hardware. [26:39] >> Try to get them to run their business in [26:40] the cloud, they can't because they will [26:42] have higher latency, they will lose [26:43] money. [26:44] >> Right. [26:44] >> So, hardware is still being made. I [26:46] mean, I remember when I used to advise [26:49] Silver Lake, and I had have Dell was [26:52] done. Nobody wanted hardware. I think [26:54] Dell's might be back to like a 3 400 [26:56] billion dollar market cap. So, hardware [26:58] is still going to be around. We're going [26:59] to need it. It's the fastest uh way to [27:01] move it. [27:01] >> Are our hardware development cycles [27:03] changing because of AI? Like are you [27:05] seeing a lot of like generative design [27:08] stuff moving in silica that historically [27:10] was manual and long cycle? [27:12] >> Yeah, but the long pole in the tent is [27:13] their design, right? The long pole in [27:15] the tent is production. Today, you can't [27:17] get a box produced because every [27:20] every [27:21] piece of hardware componentry is back [27:24] ordered. Everything's expensive and [27:26] every factory in the world is back [27:28] ordered because we're trying to build [27:30] all these GPUs based, you know, chip [27:32] cards for every data center in the [27:33] world. So, [27:34] >> Do you think the US is equipped to fill [27:35] that supply chain need? [27:37] Can we do that here? Or do you think [27:39] we're just done? [27:40] >> 10 years. [27:41] >> With a with a with a firm top-down [27:43] commitment. [27:44] >> Well, I mean, the good news is that [27:47] I think the hardware industry [27:50] is seeing a bonanza of a lifetime. [27:52] And generally, when you see a bonanza of [27:54] a lifetime, you can go commit 10, 20, [27:55] 50, 100 billion dollars. I mean, I've [27:57] seen a CEO on television committing a [27:59] 100 billion dollar plan to go build more [28:01] memory. [28:02] So, that's good. That means they have [28:04] the money to go put the money in the [28:05] ground, literally, to go build these [28:07] things for the future. So, I think that [28:09] gets us more certain that the fact [28:11] >> I think the tax incentive has a big has [28:13] a lot to do with that. The accelerated [28:14] depreciation on the the CapEx. You get a [28:18] 100% write-off in the first year, right? [28:20] Under the under the year. [28:21] >> Just a Just a final question as we wrap [28:23] up. You, over the last 8 years, [28:26] you've grown organically very [28:27] aggressively, but you've also been [28:28] pretty acquisitive. You'll, you know, [28:30] you'll take shots and they've generally [28:31] worked. So, you have a ton of permission [28:34] in the market. [28:35] When you hear what Bill Ackman said [28:37] about how there's this kind of [28:40] over beaten companies, there's a few [28:41] that get celebrated, that's a ripe pool [28:43] for you to pick from. [28:45] But some of that would require you to go [28:47] maybe a little horizontally far afield, [28:49] some would say. [28:51] How do you maintain the discipline or do [28:52] you see yourself at some point [28:53] considering things that are [28:55] not nearly so much right down the middle [28:57] of of cyber? [28:58] >> So, I'll tell you what. Um [28:59] until about a year and a half ago, we [29:02] used to buy product companies and throw [29:03] them into our go-to-market engine. [29:05] We could rewire their back end so they [29:07] can work better with our go-to-market [29:09] engine. So, for me, if I'm selling $10 [29:11] million to a customer, next time I go to [29:13] her later if I can sell them 20, it's [29:14] the most efficient way for me to [29:16] amortize my go-to-market spend, right? [29:18] So, that was the model. We played that [29:20] we ran that playbook to north of 150 [29:22] billion. Then we got to a point where it [29:24] says, "Oh, we see an inflection arriving [29:26] in identity. It's going to be important [29:27] from an agentic perspective, security [29:29] perspective." So, we bought a $25 [29:30] billion company, which we closed 3 [29:32] months ago. [29:34] Um [29:35] now it's actually a very different [29:37] opportunity has presented itself. [29:39] And the different opportunity sort of [29:40] goes like this. If you can be the best [29:43] at leveraging AI to run the most [29:46] efficient enterprise business in the [29:47] world, [29:48] your operating margin can be far in [29:50] excess of the industry. [29:52] And if you can if you can crack that [29:53] code [29:54] >> Gross and net, you're saying? Gross in [29:55] the 90s, net in the 40s. [29:56] >> Yeah, if you can crack that code, then [29:58] it doesn't matter what you buy. [30:00] >> Yeah. [30:00] >> I think the problem right now is the [30:02] execution problem. Most subscale [30:03] companies cannot afford to go optimize [30:07] their company and run it better. [30:09] So, if we can run our company much [30:11] better than everybody else and have a [30:12] higher operating margin, then the street [30:14] will say, "Fine." If you take something [30:15] at a 20% margin, make it a [30:17] >> Your first M&A was really tough, no? [30:19] Like they were pretty skeptical and then [30:21] you kind of shoved it in their face. [30:22] >> pretty skeptical when they found a guy [30:23] who didn't know cybersecurity, didn't [30:25] know enterprise show up, who worked at [30:26] Google, and their you know, the track [30:27] record of people leaving Google and [30:29] being successful out of Google is still [30:31] >> Yeah. [30:32] >> varied. [30:33] >> So, basically you're saying the menu's [30:34] open. And [30:35] >> I I think we need the next 6 to 12 [30:37] months to figure out how this AI settles [30:39] down and how can we use that effectively [30:41] in enterprises? I think if you think [30:43] about it, uh you know, the [30:46] the the people keep hoping that less [30:48] people we need to run companies. I [30:50] actually have a counter view. I think [30:51] we're going to have more people at Palo [30:53] Alto on the technology side than we've [30:54] ever had before because I think AI is [30:56] causing [30:57] everything to ask for a transformation. [30:59] So, I have more technical people today [31:01] than I would have had if AI didn't [31:02] exist. [31:04] >> Ladies and gentlemen, CEO Palo Alto [31:06] Networks, Nikesh Arora. [31:07] >> Thank you, guys. [31:08] >> [applause] [31:09] [music] [31:11] >> Thank you, sir.