Meta Weighs Compute Sales to Offset AI Data Center CapEx

Meta考虑出售算力以抵消AI数据中心资本支出 Meta正考虑出售闲置的数据中心算力，以便在内部AI产品开发跟进的同时，将其高昂的基础设施支出变现。 主持人分析了Meta推出Meta Compute计划以出售云基础设施背后的战略，将短期的算力变现与个人超级智能及消费级AI功能等长期抱负进行了对比。讨论重点探讨了对Meta当前AI落地执行的不满——例如泛化的Instagram数据分析以及错失Agent智能购物机会——同时权衡了市场对其转向云业务的看多与看空观点。此外，本期节目还探讨了Meta可能进军预测市场、Google AI Overviews对网络流量的影响、DeepMind的早期历史，以及AI视频生成模型的最新进展。 Meta 计划推出代号为“Meta Compute”的云基础设施业务，对外出售闲置算力及 AI 模型，直接与 AWS 和 GCP 展开竞争，以缓解其巨额基础设施 CapEx 压力。 (0:02) 尽管宣布转型云服务，Meta 旗下应用仍缺乏突破性的消费端 AI 功能；测试显示 Instagram 内置的 Meta AI 仅给出第三方的通用建议，未能调用私有账户分析数据。 (6:02) Meta 目前仍未在广告流中落实无缝的 Agent 化一键结账购物功能，错失了利用自身应用生态缩短电商购买转化漏斗的机会。 (9:47) 投资者对 Meta Compute 产生分歧：空头认为转售闲置算力反映出行业算力瓶颈正在缓解，多头则认为这可能推动 Meta 大举扩张 CapEx，打造超越广告收入的新增长引擎。 (73:23)

Meta Platforms is exploring a cloud infrastructure business to sell compute capacity and AI models, attempting to monetize heavy capital expenditures. However, TBPN hosts argue that leasing excess computing power underscores Meta's struggle to translate its massive infrastructure investments into compelling consumer AI features. Compute Monetization and Neocloud Market Impact (0:02) Meta's proposed "Meta Compute" initiative seeks to sell access to its AI models and idle data center capacity, placing it in competition with cloud hyperscalers like AWS and Google Cloud Platform. The plan emerged via a Bloomberg leak rather than an official enterprise rollout, yet the prospect of generating returns on hundreds of billions in capital expenditures (CapEx) initially bolstered Meta's stock. Conversely, neocloud shares dropped because Meta is both an anchor client and a formidable potential rival. Tech investor Amit outlined two divergent industry implications. In the bearish scenario, Meta having surplus compute implies that industry-wide compute constraints may be overstated, which could lead Meta to curb future infrastructure spending and negatively impact semiconductor makers. In the bullish scenario, Meta aggressively scales a public cloud platform using rapid deployment methods—such as modular GPU tents—to build a major revenue engine beyond advertising, driving long-term CapEx and chip demand even higher. Execution Shortfalls in Consumer AI Features (0:57) Despite aggressive spending, Meta has not delivered a breakout consumer AI experience across Instagram, Facebook, or WhatsApp. The hosts evaluate Muse Spark as a competent benchmark performer that remains underwhelming for end users, while Meta Vibes is characterized as an uninspired Midjourney wrapper. The Ray-Ban Meta smart glasses have found product-market fit, but primarily due to the integrated camera and form factor rather than native AI utility. Furthermore, image generation on Instagram remains neglected, and the "Man of Steel" project is reportedly being unwound. To illustrate this product gap, a host tested Meta AI inside Instagram for custom account growth analytics. Instead of evaluating proprietary engagement metrics, the tool returned generic advice citing a public blog post from third-party tool Buffer. Creators still rely on direct-to-camera videos from Instagram head Adam Mosseri rather than an automated AI advisor. Additionally, Meta has neglected basic agentic workflows, such as one-click ad checkouts that keep users within the app, missing obvious conversion gains for advertisers. Short-Term Compute Offloading versus Superintelligence (71:16) Pivoting toward cloud services contrasts with CEO Mark Zuckerberg's stated priority of building "personal superintelligence," prompting internal strategy staff to joke about Zuckerberg taking a loss. The hosts compare this transition to Meta's previous hardware pivot, questioning whether underlying model development will be mothballed much like early VR initiatives. Skepticism also remains over whether Meta's core advertiser base of direct-to-consumer and mobile gaming brands will naturally convert into buyers of cloud inference tokens. Alternatively, the hosts suggest the move could be a short-term holding pattern similar to SpaceX's compute deals with Google and Anthropic. Selling capacity lets Meta monetize idle hardware to appease investors while consumer AI use cases mature. Because Meta owns the physical infrastructure, it retains the flexibility to exit 90-day leases once internal model demand surges. Even so, the hosts emphasize a cultural disconnect: Meta possesses world-class research talent but lacks an agile product organization capable of rapidly shipping consumer-facing AI. Prediction Market Exploration and Search Traffic Erosion Reporting from NPR's Bobby Allen revealed that Meta considered acquiring prediction market startup Kalshi before opting to develop its own in-house forecasting application to compete with Polymarket and Manifold. Kalshi's leadership publicly dismissed Instagram as mindless "brain rot," while the hosts questioned whether adding gambling-adjacent prediction mechanics would introduce dangerous regulatory scrutiny to Meta's core social platform. The hosts also reviewed search ecosystem disruptions documented by Carnegie Mellon and the Indian School of Business. Their study found that Google AI Overviews appeared in approximately 41% of surveyed searches, causing a 40% reduction in outbound organic clicks and a 35% jump in zero-click searches—a development publisher Josh Marshall described as an emerging "Google AI oligarchy" eroding the open web. Meta's move toward selling compute reflects a pragmatic attempt to offset historic capital spending, but it highlights an unresolved challenge: converting world-class AI infrastructure into differentiated consumer applications before market patience wears thin.

据 TBPN 节目分析，Meta 计划通过“Meta Compute”向外部出售闲置计算能力和模型接口，以平衡其千亿美元级别的基础设施资本支出，但此举引发了外界对其战略重心转移以及内部消费级 AI 产品交付能力的广泛质疑。 基础设施商业化：Meta Compute 与算力转售 (0:02) 彭博社报道披露，Meta 正在筹划名为“Meta Compute”的云基础设施新业务，拟向外部客户出租其自研 AI 模型访问权限及底层裸算力，直接进军由 AWS 和 Google Cloud Platform（GCP）主导的云服务市场。该计划的核心目标是将数据中心和先进芯片建设所形成的冗余产能商业化变现，以此缓解数百亿美元规模的资本支出（CapEx）压力并提升投资回报率（ROI）。消息引发了新兴云（neocloud）概念股的抛售行情，因为 Meta 正从其重要买家转变为直接竞争对手；与此同时，资本市场对 Meta 基础设施支出有望实现商业回收表现出积极情绪。 节目分析指出，与 SpaceX 在筹备 IPO 期间高调公布与 Anthropic 达成数十亿美元交易相比，Meta 的算力转售计划目前仅停留在媒体爆料阶段，尚未公布具体的大客户合作协议。主持人认为，这一举措很可能是 Meta 的短期务实过渡策略：由于 Meta 拥有物理基础设施资产，其可以在出租闲置容量以安抚华尔街的同时保留战略灵活性，一旦内部产品需求激增，便可在 90 天内终止外部租约收回算力。 应用落地断层：个人超级智能承诺与终端体验脱节 (0:57) 尽管首席执行官 Mark Zuckerberg 与 Meta 管理层此前将核心愿景定义为实现“个人超级智能”，但节目批评 Meta 至今未能在 Facebook、Instagram 和 WhatsApp 等核心产品中落地具备突破性的 AI 杀手级功能。在现有模型和工具中，自研模型 Muse Spark 被评价为基准测试表现尚可且适合作为 API 供应商底座，但终端用户体验平淡；Meta Vibes 被指责仅为 Midjourney 的套壳应用；“钢铁侠（Man of Steel）”项目据传正被逐步缩减；而 Meta Ray-Ban 智能眼镜虽获得市场认可，其驱动因素主要在于集成摄像头硬件，而非背后的 AI 能力。 主持人的个人测试案例凸显了 Meta AI 与实际应用场景之间的割裂：在 Instagram 内向 Meta AI 询问如何根据现有账号数据优化 Reels 表现时，AI 未能调取专有数据提供个性化诊断，反而输出了引用第三方社交工具 Buffer 博客的通用建议。创作者真正需要的是融合高级模型与第一手平台数据的分析引擎；此外，Meta 迟迟未在其应用矩阵内打通基于智能代理的一键广告结算和无摩擦电商购物流，未能有效利用技术手段缩短转化漏斗。 多空博弈：云巨头雄心与商业基因冲突 (73:23) 针对 Meta 涉足云业务，科技投资人 Amit 给出了截然对立的多空分析。空头观点认为，Meta 对外转售闲置算力表明全行业可能并不像此前预期的那样受制于算力短缺，这不仅打击了 CoreWeave 和 Iron 等新兴云服务商，而且若 Meta 依赖闲置资源展业，未来可能缩减整体 CapEx，进而对半导体制造产业链构成利空。相反，多头观点指出，若 Meta 依托在模块化帐篷中快速部署 GPU 等数据中心敏捷交付能力，全力打造成熟云平台与头部厂商全面竞争，将驱动资本支出持续扩大，形成超越广告主业的全新增长曲线。 然而，节目强调 Meta 的组织文化与商业基因面临深刻挑战。内部战略团队私下将近期的模型变动戏称为“扎克伯格认输”，并将其类比为元宇宙硬件向实用眼镜收缩的研发周期。尽管 Meta 拥有世界顶级的人工智能科研机构，但缺乏能够快速迭代消费级功能的敏捷产品组织；更重要的是，Meta 现有的庞大销售与客户服务体系主要面向移动游戏和 D2C 电商广告主，能否将广告主成功转化为采购算力 Token 的企业客户仍存在高度不确定性。 战略试探与行业变局：预测市场及搜索生态重构 在核心社交与云计算业务之外，Meta 的外部扩张尝试引发了合规争议。NPR 记者 Bobby Allen 报道称，Meta 曾考虑收购预测市场初创平台 Kalshi，后转向评估自研预测应用。然而，Kalshi 管理层曾公开指责 Instagram 为“大脑腐蚀（brain rot）”，节目主持人亦警告，在 Meta 面临全球立法与反垄断严密审视的背景下，贸然将带有博彩属性的预测交易引入生态系统犹如触碰“有毒的金蛋”，极易招致高额监管风险并威胁基本盘。 与此同时，大模型对开放网络生态的冲击正在被学术研究所证实。卡内基梅隆大学与印度商学院的最新论文显示，Google AI Overviews 在约 41% 的搜索请求中被触发，导致向外跳转的自然点击量下降约 40%，零点击搜索激增 35%。这种被媒体形容为“AI 寡头化”的流量截流现象，进一步印证了大型科技平台在垄断基础设施和分发入口后，正在深层次重塑数字经济的利益分配格局。 综上所述，Meta Compute 的立项折射出 Meta 在承担天量算力支出压力下的短期财务防御姿态，暴露出其在尖端科研储备与消费端杀手级应用转化之间的严重脱节。Meta 未来能否真正建立云业务护城河，不仅取决于华尔街对其资本支出的耐心，更取决于其能否证明自身的广告客户生态能够平滑过渡为算力消费者，并在核心社交应用中交付真正具备生产力价值的原生智能体验。

## Background 该内容由科技播客栏目 TBPN 于 2026 年 7 月 23 日发布在 YouTube 平台。根据页面简介，这是由 John Coogan 与 Jordi Hays 主持的“Diet TBPN”精华栏目，致力于在 30 分钟内浓缩呈现当期直播节目的核心亮点，聚焦硅谷前沿科技议题与重磅商业动态。 ## TL;DR 本期节目深入辩论了 Meta 计划推出“Meta Compute”对外出租 AI 算力的重磅传闻，剖析了其从“个人超级智能”宏大愿景倒退为云服务商背后的千亿美元 CapEx 回报压力与产品落地困境。主持人们不仅辛辣吐槽了 Instagram 内嵌 Meta AI 的鸡肋表现以及智能体电商（Agentic Shopping）的缺位，还将此轮大模型狂潮与当年的元宇宙周期进行了透彻对比。此外，节目还探讨了 Meta 涉足 Kalshi 预测市场所面临的合规“毒金蛋”风险、学术研究揭示的谷歌 AI Overviews 对开放网络流量的毁灭性吞噬，并趣味回顾了 DeepMind 早期以白菜价贱卖的科技史料。 ## Atmosphere & style 节目的整体基调兼具敏锐犀利的产业洞察与极具感染力的美式科技脱口秀风格，两位主持人在硬核商业逻辑拆解与幽默捧哏吐槽之间切换自如。讨论充满生动形象的比喻与一线创作者的真实体感，既有对巨头战略摇摆和产品平庸化的无情讽刺，也有对算力错配和宏观资本走向的严谨推演。结尾处从严肃沉重的 AI 寡头反思骤然滑向对自家周边的荒诞吹捧，整场对话节奏明快、张弛有度，带给听众一种仿佛置身于硅谷咖啡馆旁听顶尖内幕人士畅聊的酣畅体验。 ## Bulletin - [00:02](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=2s) 彭博社等媒体披露 Meta 正计划推出“Meta Compute”，通过对外销售原始 AI 算力及托管模型直接竞争 AWS 和 GCP，消息引发 CoreWeave 等新云概念股应声重挫。 - [05:43](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=343s) Meta 产品端被指未能有效消化庞大算力储备，Instagram 内置 Meta AI 仅能给出通用模板建议，而备受期待的 Agentic Shopping 一键结算闭环也迟迟未见落地。 - [13:20](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=800s) 华尔街对 Meta 卖算力产生两极分化解读：空方担忧行业面临“算力错配”并重创 Neo Cloud 估值，多方则看好 Meta 打造第四大云巨头并维持半导体 CapEx 扩张。 - [19:47](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1187s) 最新学术研究量化揭示，谷歌搜索中 41% 的查询会触发 AI Overviews，直接导致外链自然点击量骤降 40%、零点击搜索激增 35%，开放网络正加速迈向流量绝境。 ## Key takeaways - [00:02](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=2s) Meta 计划启动 Meta Compute 向外部企业出租 GPU 算力及大模型访问权限，试图为数千亿美元的基础设施 CapEx 寻找更快的短线投资回报路径。 - [01:04](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=64s) 从扎克伯格曾宣扬的“个人超级智能”转向底层云算力租赁，折射出 Meta 内部当前缺乏能在短期内充分消化自有算力储备的消费级杀手应用。 - [01:04](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=64s) 雷朋智能眼镜（Ray-Ban Meta）目前的 PMF 本质上仍停留在便携拍摄与随身相机硬件，其所搭载的 AI 交互系统并未展现出决定性的智能化吸引力。 - [05:43](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=343s) 创作者实测表明 Meta AI 在 Instagram 内部表现平庸，空有海量第一方后台数据，却只能输出引用第三方博客的公关式烂大街运营建议，缺乏数据工程闭环。 - [09:09](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=549s) Meta 在“智能体购物”（Agentic Shopping）领域动作迟缓，尚未借助 MCP 与 Shopify 生态打通免跳转 Safari 的应用内端到端 AI 结算转化路径。 - [11:16](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=676s) 算力分租或为一种务实的短期对冲财务举措，Meta 采取类似苹果的观望策略，保留在 90 天内解约收回算力全力押注爆款 AI 场景的灵活性。 - [11:16](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=676s) Meta 拥有世界顶级的 AI 研究团队，但相对缺乏能快速将前沿模型工程化并打造成爆款消费级产品的 AI 产品组织架构。 - [13:20](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=800s) 空方分析师认为 Meta 兜售闲置算力证实了行业属于“算力错配”而非真实瓶颈，CoreWeave 等 Neo Cloud 大客户订单受阻将对半导体供应链形成压制。 - [13:20](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=800s) 多方观点则认为若 Meta 严肃构建全栈公有云业务，其仅次于 SpaceX 的高密度数据中心快速部署能力将驱动其持续加码硬件资本支出。 - [14:53](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=893s) 当前大模型热潮被类比为当年的元宇宙与 VR 周期，高昂自研模型若因推理需求疲软而被战略冷藏，Meta 最终可能仅留存窄门硬件与算力底座资产。 - [18:04](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1084s) 传闻 Meta 曾考虑收购真金白银预测市场平台 Kalshi，但引入具博彩属性的金融工具无异于捡起“带毒的金蛋”，恐将核心高估值广告母体拖入全球监管泥潭。 - [19:47](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1187s) 卡内基梅隆大学等机构的研究证实，谷歌 AI Overviews 显著加剧“零点击搜索”，科技巨头用全网内容训练 AI 却切断原创外链流量，加速开放网络瓦解。 - [20:29](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1229s) Demis Hassabis 在 2010 年曾以区区 500 万美元投后估值出让 DeepMind 多数股权，成为科技创投史上极具戏剧性与唏嘘感的早期并购注脚。 ## Detailed walkthrough ### [00:02](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=2s) Meta 启动 Meta Compute：转向出售云端算力 节目一开场，主持人便抛出了科技界的一则重磅新闻：Meta 计划进军云基础设施领域，推出“Meta Compute”计划。根据彭博社等媒体的报道，Meta 正在制定计划，将通过向外部销售 AI 算力以及其基础设施上托管的各类 AI 模型访问权限，直接与亚马逊 AWS、谷歌云 GCP 等传统云计算巨头展开竞争。面对在数据中心和昂贵 GPU 芯片上砸下的数千亿美元资本支出（CapEx），Meta 希望通过对外变现过剩算力来收回巨额投资成本。消息一出，市场反应剧烈且耐人寻味——所谓“新云”（Neo Cloud）概念股纷纷下跌，因为 Meta 既是这些算力托管公司的大客户，现在又摇身一变成为了潜在的强力竞争对手。 ```mermaid flowchart TD accTitle: Meta算力商业化转变路径 accDescr: 展示Meta从千亿美元资本开支到向外销售算力及可能引发的市场连锁反应 subgraph CapEx["巨额AI投入"] A["数百亿美元基础设施<br/>数据中心与先进芯片"] end subgraph Transition["Meta Compute 计划"] B["自用AI产品与模型<br/>（个人超级智能）"] C["过剩算力对外出租<br/>（Meta Compute）"] end subgraph Impact["市场影响"] D["与 AWS / GCP 展开正面竞争"] E["新云（Neo Cloud）市场震荡下跌"] F["华尔街对短线投资回报（ROI）表示认可"] end A --> B A --> C C --> D C --> E C --> F ``` - Meta 考虑销售的内容包括底层原始算力容量（Raw Compute Capacity）以及自家托管的 AI 模型 API。 - 资本市场对这一举动初期表现积极，Meta 股价大幅上扬，显示出投资者对其为巨额资本开支寻找变现出路的认可。 --- ### [01:04](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=64s) 战略困惑：从“个人超级智能”到“云服务商” 针对 Meta 的这波操作，两位主持人展开了深度辩论。一方直言整件事非常令人困惑：作为 Meta 商业模式的长期坚定支持者，大家普遍认为 Meta 拥有近乎完美的广告机器与社交产品矩阵；然而，豪掷千亿购买 Neo Cloud 算力与自建硬件，如今却转手做起了云租赁生意，这让人对 Meta 的顶层产品战略产生怀疑。如果向外卖算力是目前最好的投资回报率（ROI）手段，是否意味着 Meta 内部根本拿不出能在短期内消耗掉这些海量算力的杀手级消费端产品？ 扎克伯格（Mark Zuckerberg）此前描绘的宏伟愿景一直是“个人超级智能”（Personal Super Intelligence）——一个深度融入手机端社交图谱、能为用户生活和创作提供全方位赋能的私人专属 AI。主持人表示，自己本是这一愿景的忠实信徒，但现实中 Meta 拿出的东西却反差极大：除了基准测试表现尚可、但 API 尚未掀起波澜的模型（如 Llama 衍生分支与 Muse Spark），就是类似“Meta Vibes”这类被戏称为 Midjourney “套壳”的小修小补。更具讽刺意味的是，前一天行业内还在讨论谷歌向 Meta 表示“我们已经没有多余算力能租给你”，转天 Meta 却传出算力过剩、要对外转售的传闻。如果 Meta 能像某些先锋公司那样直接官宣拿下顶级前沿实验室（如类似 Anthropic）单月数亿美元的采购大单，故事会更加好看，但目前的泄密传闻更像是在仓促填补回报缺口。 - 雷朋智能眼镜（Ray-Ban Meta）虽然实现了一定程度的 Product-Market Fit（PMF），但主持人们指出，该产品当前的核心卖点依然是随身相机与拍摄功能，而非眼镜本身搭载的 AI 智能系统。 - 尽管 Instagram 在图像质量上保持了较高的水准，几乎没有被粗制滥造的“AI 垃圾内容”（AI Slop）泛滥侵蚀，但 Meta 却迟迟未能推出像样的原生图像生成与扩散编辑功能。 --- ### [05:43](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=343s) 创作者困局：极其平庸的 Meta AI 与 Mosseri 的真身答疑 为了检验扎克伯格所谓的“个人超级智能”在实际场景中的能力，主持人分享了自己在 Instagram 上的真实测试经历。作为拥有丰富内容和细粒度后台数据的创作者，他直接在 Instagram 内嵌的 Meta AI 中提问：“我应该多发布什么内容才能让账号涨粉？”他原本期待 AI 能深度结合自身账号的具体播放量、互动率等后台数据给出量身定制的高阶分析，没想到得到的回应却极其机械和通俗。 Meta AI 竟给出了类似于搜索引擎搬运来的模板化建议，甚至直接引用了第三方社交媒体管理工具 Buffer 的博客文章：告诫用户要注重可持续有机增长、不要搞“互粉列车”（Follow Trains），并给出诸如“Reels 的曝光量比轮播图（Carousels）高 36%”、“使用数据分析工具看看哪类内容能转化粉丝并加大投入”、“与同垂直领域的小创作者合作”、“多回复评论并在 Bio 中优化关键词”等烂大街的公关常识。主持人对此深感无奈：“我问你该怎么涨粉，你却让我去看自己的数据分析，但你明明就拥有我全部的后台数据！”这说明 Meta 空有强大的基础模型和最精准的第一方数据，在产品工程化上却没有将模型与用户自身的数据闭环接通。 两位主持人调侃道，与其向 Instagram 里这个“传统古旧版”的内置聊天机器人求教增长秘诀，眼下获取 Instagram 运营指南最靠谱的途径，反而是去看 Instagram 负责人亚当·莫塞里（Adam Mosseri）亲自出镜拍摄的 Reels 短视频。比如莫塞里最近就亲自解答了一个经典迷思——“如果一条帖子发出去反响不好，应不应该删掉重发？”莫塞里直白地回答：不要重发，因为算法大概率只会给出相同的结果，而且那些第一遍就没点赞的粉丝如果再刷到一次，只会对你的内容产生更强烈的反感。这种来自高管直接坦率的面对面解答，反衬出平台内部 AI 助理在实际落地中的尴尬与缺位。 ### [09:09](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=549s) 创作者工具与智能电商：Meta 缺失的 AI 产品化落地 主持人们探讨了 Meta 旗下应用（Family of Apps）目前在 AI 产品化上的缺位。以 Instagram 为例，负责人 Adam Mosseri 虽然经常在平台上分享见解，但那些建议往往是千篇一律的通用内容。如果能拥有一个由 Meta AI 增强、针对个人账号定制的“Adam Mosseri 大脑”，充当社交媒体的 AI 领航员（Co-pilot），对广大创作者乃至普通用户都会极具吸引力，因为几乎每个人都希望自己的帖子能多获得一些播放量。 另一个巨大的落地盲区是“智能体购物”（Agentic Shopping）。在去年的 Meta Connect 大会上，主持人们曾向 Mark Zuckerberg 当面问及这个方向——戴着具备 HUD 和 AI 功能的 Meta Ray-Ban 智能眼镜，看到一双鞋说一句“帮我下单这双”，系统便自动完成购买。当时 Zuckerberg 承认这是可能的未来方向之一，但令人费解的是，Meta 至今甚至没有尝试过消除繁琐的点击链路：在看到广告时，用户本应点击一下就由 Agent 代为结账、在 Meta 应用内部确认信息，而无需跳转至 Safari 外部浏览器窗口。这种缩短漏斗、提升转化率的方案对广告主、Meta 和用户都是三赢，而且当前计算机操作（Computer Use）、MCP（Model Context Protocol）以及 Shopify 等生态工具已逐渐成熟，Meta 却没有在这方面进行敏捷的 A/B 测试，这被视为主播口中其缩减“超级智能”（Super Intelligence）宏大叙事的一个缩影。 ```mermaid flowchart LR accTitle: 理想的 Meta Agentic Shopping 链路 accDescr: 对比传统跳转 Safari 的外链购物与由 Agent 代劳的应用内一键闭环购物 subgraph Current ["现存繁琐路径"] A["看到 Feed/Stories 广告"] --> B["点击跳转外部 Safari 浏览器"] B --> C["手动填写地址与支付信息"] end subgraph Proposed ["Agentic 理想闭环"] D["看到广告 / 眼镜识别物品"] --> E["AI 智能体直接调用 MCP/Shopify 接口"] E --> F["Meta 应用内一键确认直接结账"] end ``` --- ### [11:16](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=676s) 算力分租策略与产品创新的观望态度 对于 Meta 传出将其闲置算力对外出租的消息，主持人们分析这可能是一种务实的短期财务举措。类似于 Google 与 Anthropic 之间的协议，这些合作通常并非长达五年的死约，而是灵活性极高的短期合作。Meta 可以借此向公开市场释放出“管理层并非盲目非理性扩张”的信号：既然自身产品线尚未完全就绪，不如先将多余的算力出租变现，待自身产品开发成熟后再逐步收回算力自用。 在产品开发逻辑上，Meta 似乎正像当年的 Apple 一样采取“后发制人”的观望策略。Instagram 目前并没有面临用户极度渴求 AI 功能的迫切压力；在搜索栏塞进 AI 并没有从根本上提升体验。与其盲目创新，Meta 更倾向于等待市场上真正验证出某个极具爆发力的 AI 场景，再在 90 天内解除算力租约，将庞大的基础设施调转枪头全力支持该功能。Meta 的核心基因从来不是在产品概念上完全原创，比如近期 Instagram 推出的类 Stories 聊天侧栏短图文功能“Glimpses”，以及此前类似 Midjourney 贴纸工具的“Meta vibes”，都显示出 Meta 拥有顶尖的研究团队（Research Org），却相对缺乏一个能迅速将前沿研究变成爆款的 AI 产品组织（Product Org）。 --- ### [13:20](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=800s) 华尔街的多空辩论：Neo Cloud 暴击与 CapEx 走向 围绕 Meta 或将转型对外出售算力并进军云业务的传闻，科技投资人 Amit 提出了极具代表性的多空双向视角： - **空方观点（Bearish）**：如果 Meta 拥有多余算力以至于可以对外充当新型云业务出售，意味着行业可能根本不存在算力瓶颈。这对 CoreWeave、Nebius 等新兴 GPU 云厂商（Neo Clouds）而言将是毁灭性打击——Meta 怎么可能再给 CoreWeave 订单，反而自己成了它们的直接竞争对手。此外，以闲置算力作为新业务基石说明其实际算力需求不及采购预期，未来 Meta 可能会大幅削减资本支出（CapEx），这将对整个半导体板块（Semis）构成巨大的利空。 - **多方观点（Bullish）**：如果 Meta 是下定决心要严肃构建与 Google Cloud Platform（GCP）、AWS、Azure 竞争的第四大完整云巨头，那么仅仅依靠广告业务已不足以支撑其野心。为了提供全套云服务并持续快速搭建数据中心（Meta 在帐篷里快速部署 GPU 机柜的能力仅次于 SpaceX 和 AWS），Meta 反而可能被迫持续加大 CapEx 投入，从而对底层半导体硬件供应链带来超预期的利好。 二级市场对此消息的第一反应十分剧烈：Meta 股价受到正面提振，而 Neo Clouds 概念股则因迎来巨头竞争而遭受重创。科技策略团队 Zephyr 甚至讽刺道：“Zuck finally takes the L”（扎克终于认输了）。 --- ### [14:53](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=893s) 元宇宙翻版？大模型算力错配与硬件留存 这轮 AI 超级智能（Super Intelligence）与自研基础模型的狂热，被主持人们比作 Meta 当年疯狂下注的元宇宙（Metaverse）和 VR 战略。如果大模型最终也像 Meta Quest 一样遭遇战略收缩与冷藏，那么在这一轮 AI 浪潮中，Meta 能否留下像 Meta Ray-Ban 那样真正成功的硬件产品？ ```mermaid flowchart TD accTitle: Meta 战略浪潮与收缩周期对比 accDescr: 对比元宇宙周期与 AI 周期中 Meta 的庞大投入与最终落地的实用级产品 subgraph MetaverseCycle ["元宇宙 / VR 周期"] M1["全面豪赌 Metaverse / VR"] --> M2["业务收缩、战略冷藏 (Mothballed)"] M2 --> M3["最终杀手级留存：Meta Ray-Ban 智能眼镜"] end subgraph AICycle ["大模型 / 超级智能周期"] A1["全面自研超级智能模型 / 万卡集群"] --> A2["推理需求不符预期 / 闲置算力出租"] A2 --> A3["可能留存：窄门硬件、专业音频/图像模型或云业务"] end ``` 关于 Meta 能否顺势做大 B 端推理业务，主持人们产生了分歧： - 一种观点认为，Meta 服务着全球数百万家中小企业，掌握广告投放与获客全渠道，其庞大的销售与大客户管理（Account Management）团队完全有能力向企业客户售卖推理与交付服务。 - 另一种观点则对此表示怀疑：一家在其平台上买量投放的手游公司或 D2C 电商，绝不会自然而然地转而向 Meta 购买大模型 Token；云业务极其强调底层基础设施和协议标准，与 Meta 熟悉的高毛利社交广告网络存在本质区别。 分析师 J Yoon 犀利地指出，行业目前并非真正的“算力过剩”，而是严重的“算力错配问题”（Compute Allocation Problem）：Meta 和 xAI 之所以考虑出租算力，是因为它们的自研模型缺乏足够的外部推理需求，导致算力集中在缺乏内部消化场景的玩家手中。此外，NPR 的 Bobby Allen 爆料称，在开发自己的预测市场应用之前，Meta 曾认真考虑过收购知名预测市场平台 Kalshi，这再次印证了 Meta 看到新趋势先谋求收购或抄袭跟进的经典打法。 ### [18:04](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1084s) 预测市场的金钱诱惑与“毒金蛋”监管风险 关于科技巨头布局预测市场的传闻，主持人展开了深入探讨。此前有人推测相关平台可能只是搭建一个类似 Manifold 的纯预测与积分竞技平台，让人凭借预见未来的能力一较高下；然而最新消息表明，他们并未尝试收购 Manifold，反而试图收购 Kalshi。这清晰地释放出一个信号：他们大概率选择走受金融激励驱动的真金白银路线，正如 Polymarket 和 Kalshi 的打法。这种策略符合 C 端消费级产品追求高利润与爆发式增长的逻辑。有趣的是，Kalshi 联合创始人兼首席执行官 Tarek Mansour 此前还在公开开炮，指责 Instagram 完全是“大脑腐烂”（brain rot），宣称用户在 Kalshi 上度过的每一分钟，都等于从 Instagram 的精神内耗与垃圾内容中夺回了一分钟。对此，主持人们调侃道，在大众眼里这两者可能根本是半斤八两，甚至赌博式产品的杀伤力还要更胜一筹。 由此引发的核心争论在于：为了预测市场那点潜在利润，冒着被全球监管机构围剿的风险，真的划算吗？当前相关平台已经在全球无数领域面临口诛笔伐与反垄断调查，如果再把博彩与下注功能强行塞进主流产品中，无异于引火烧身。主持人打了个极其生动甚至略带荒诞的比喻：这就好比你手里已经拥有了一只估值极高、能源源不断下金蛋的“摇钱母鸡”，现在你在远处看到了一颗新的金蛋，但那其实是一颗“带毒的金蛋”。如果你为了捡起这颗毒金蛋而把它带回农场，极可能会直接毒死你原本那只下蛋的母鸡。 - **两种截然不同的策略路径**：一种是以 Manifold 为代表的趣味性、无资金赌注的积分预测社区；另一种则是以 Kalshi 和 Polymarket 为代表、具备真实资金激励的合规或离岸博彩预测市场。 - **监管反噬的隐忧**：在主流消费产品本身已面临诸多立法者审视的情况下，引入具备真钱赌博性质的预测功能，可能会让整个核心业务面临巨大合规灾难。 ### [19:47](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1187s) 谷歌 AI Overviews 冲击：零点击搜索与开放网络的终结 随后，话题转向了生成式搜索对整个互联网生态的毁灭性打击。知名移动广告分析师 Eric Seufert 近期分享了卡内基梅隆大学（Carnegie Mellon University）与印度商学院（Indian School of Business）联合发布的一篇重磅学术论文。该研究量化分析了谷歌搜索中引入的 AI Overviews（AI 摘要功能）对网站流量的实际冲击。研究发现，在所观察的谷歌搜索查询中，有大约 41% 触发了 AI Overviews；而一旦该功能被触发，流向外部第三方网站的自然搜索点击量（organic clicks）直接骤降约 40%，几乎呈现出近乎一比一的吞噬效应。 不仅如此，AI Overviews 的存在使得用户只看搜索结果页而不点击任何链接的“零点击搜索”（zero-click search）概率激增了约 35%。主持人们感叹，照此趋势发展，所有内容创作者和媒体正在被迫迈向“谷歌归零”（Google zero）的绝境。政治新闻网站 Talking Points Memo 的创始人 Josh Marshall 甚至为此撰写了一篇深度长文，直指《谷歌 AI 寡头与开放网络的终结》（Google AI oligarchy and the end of the open web）。科技巨头利用全网内容训练 AI，再用 AI 直接给出答案切断原作者的流量血脉，开放网络的生存空间正在被无情挤压。 - **AI Overviews 触发率**：在受测的谷歌搜索中占比约 41%。 - **自然点击暴跌**：触发 AI 摘要时，外链自然点击量下滑约 40%。 - **零点击搜索增加**：AI 摘要使零点击搜索的发生概率提高约 35%。 ### [20:29](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1229s) DeepMind 早期贱卖往事与 Veo 视频模型 聊到谷歌，主持人们顺势翻出了一段令人唏嘘的科技史料：2010 年 12 月，Demis Hassabis 正在为一家名为 DeepMind 的初创团队东拼西凑几百万美元的早期资金。对于这几年才入行的年轻创业者来说，当时的数字简直匪夷所思——甚至会怀疑当年是不是少打了一个零，以为融资规模起步就是 2.5 亿美元。然而事实残酷而滑稽：作为当今全球最顶尖的技术领袖之一，Demis Hassabis 当年是以区区 500 万美元的投后估值，硬生生卖掉了 DeepMind 半数以上的股权。 主持人们忍不住打趣调侃，这算不算是科技史上最经典的“手软抛早了”（paper hands）。虽然有人推测如果 DeepMind 当初坚持完全独立走下去是否会有另一番景象，但也有人反驳指出，当年几乎所有早期风投机构其实都极力反对将其出售给谷歌，甚至一度在创投圈流传着“千万别卖给谷歌，谷歌是个大坏蛋”的论调。不过事后来看，谷歌在很长一段时间里扮演了极为负责任、资金雄厚的后盾角色。随后主持人顺带聊起了上手体验谷歌 Veo 视频生成模型的感受（无论是 Veo 3 还是 Veo 4）。他们评价道，如今在手机端就能直接跑出这种画质确实令人惊艳，但画面细节依然无法做到与现实完全真假难辨，“看似近在咫尺，实则仍有很长的路要走”。 ### [21:56](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1316s) Codex 赞助、革命性周边上新与节目收尾 节目进入尾声，主持人借着展示身上衣服的契机插入赞助商口播，介绍了 Codex 这款为 AI 智能体工作流打造的强大协作工作区，无论是用于编写代码、分析数据、生成内容还是自动化业务流程都能提供助力。接着，演播室的气氛彻底转向了插科打诨的欢乐模式。Jordi Hays 被现场听众起哄在镜头前转个身展示身上的最新周边服饰，两人随即透露，团队刚刚收到了首批大约 201 件全新周边现货。 关于这批周边，主持人们开始了一本正经的胡说八道。他们卖关子表示，这不是普通周边，而是一款“人类几个世纪以来一直试图创造的革命性产物”，经过长期的潜心攻关，团队终于迎来了“重大历史突破”。在把观众的胃口吊到极致后，他们宣布这批周边产品很快就会正式上架发售，届时会优先把购买链接直接甩进直播间聊天区回馈铁杆粉丝。最后，主持人们感谢了观众的长期耐心陪伴，呼吁大家在 Apple Podcasts 和 Spotify 上打出五星好评，并订阅官网 TBPN.com 的每日邮件通讯，在一片轻松欢快的“明天见，爱你们”中结束了当天的直播。 ## Notable quotes > [01:04](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=64s) 我们已经没有多余算力能租给你 > [05:43](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=343s) 我问你该怎么涨粉，你却让我去看自己的数据分析，但你明明就拥有我全部的后台数据！ > [05:43](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=343s) 如果一条帖子发出去反响不好，应不应该删掉重发？ — Adam Mosseri > [09:09](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=549s) 帮我下单这双 > [13:20](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=800s) Zuck finally takes the L — Zephyr > [14:53](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=893s) 算力错配问题 — J Yoon > [18:04](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1084s) 大脑腐烂 — Tarek Mansour > [18:04](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1084s) 用户在 Kalshi 上度过的每一分钟，都等于从 Instagram 的精神内耗与垃圾内容中夺回了一分钟 — Tarek Mansour > [18:04](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1084s) 带毒的金蛋 > [19:47](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1187s) 谷歌归零 > [19:47](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1187s) Google AI oligarchy and the end of the open web — Josh Marshall > [20:29](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1229s) 千万别卖给谷歌，谷歌是个大坏蛋 > [20:29](https://www.youtube.com/watch?v=dUBFpPlmuy0&t=1229s) 看似近在咫尺，实则仍有很长的路要走 ## Mentioned Meta, Meta Compute, John Coogan, Jordi Hays, Mark Zuckerberg, Adam Mosseri, Eric Seufert, Josh Marshall, Demis Hassabis, Tarek Mansour, Amit, J Yoon, Bobby Allen, AWS, Google Cloud Platform (GCP), Microsoft Azure, CoreWeave, Nebius, Google, DeepMind, Anthropic, xAI, SpaceX, Shopify, Kalshi, Polymarket, Manifold, Buffer, Instagram, Reels, Stories, Glimpses, Ray-Ban Meta, Meta Quest, Llama, Muse Spark, Meta Vibes, Midjourney, AI Overviews, Veo, Codex, MCP (Model Context Protocol), Talking Points Memo, Carnegie Mellon University, Indian School of Business, Bloomberg, NPR ## Video description Sign up for TBPN’s daily newsletter at TBPN.com Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after. Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella. TBPN is made possible by: Ramp - https://ramp.com Public - https://public.com Cisco - https://www.cisco.com Console - https://www.console.com CrowdStrike - https://www.crowdstrike.com Figma - https://www.figma.com MongoDB - https://www.mongodb.com NYSE - https://www.nyse.com Railway - https://railway.com Shopify - https://www.shopify.com/ Follow TBPN: https://TBPN.com https://x.com/tbpn https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231 https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235 https://www.youtube.com/@TBPNLive [Source](https://www.youtube.com/watch?v=dUBFpPlmuy0)

[1:04] 我们已经没有多余算力能租给你 [5:43] 我问你该怎么涨粉，你却让我去看自己的数据分析，但你明明就拥有我全部的后台数据！ [5:43] Adam Mosseri: 如果一条帖子发出去反响不好，应不应该删掉重发？ [9:09] 帮我下单这双 [13:20] Zephyr: Zuck finally takes the L [14:53] J Yoon: 算力错配问题 [18:04] Tarek Mansour: 大脑腐烂 [18:04] Tarek Mansour: 用户在 Kalshi 上度过的每一分钟，都等于从 Instagram 的精神内耗与垃圾内容中夺回了一分钟 [18:04] 带毒的金蛋 [19:47] 谷歌归零 [19:47] Josh Marshall: Google AI oligarchy and the end of the open web [20:29] 千万别卖给谷歌，谷歌是个大坏蛋 [20:29] 看似近在咫尺，实则仍有很长的路要走

[6:02] Meta AI: Buffer recommends focusing on sustainable organic growth instead of quick hacks like follow trains. Post reels consistently since they get 36% more reach than carousels... LMAO, Zuck finally takes the L.

[0:02] Meta is selling computer They're getting [0:04] out of the computing business. They [0:06] said, "We don't need computers anymore [0:08] to do what we need to do. We don't need [0:10] them. We're going to be selling them." [0:11] Meta Platforms is developing plans for a [0:13] cloud infrastructure business to sell [0:16] access to AI computing power and models [0:19] competing with industry leaders like AWS [0:21] and GCP. [0:23] The company is considering selling [0:24] access to various AI models hosted on [0:27] its existing AI infrastructure as well [0:28] as raw computing capacity as part of its [0:31] Meta Compute initiative. Meta plans to [0:33] generate revenue from excessive [0:35] computing power could help return its [0:38] investment in AI infrastructure which [0:39] includes hundreds of billions of dollars [0:41] spent on data centers and expensive [0:43] chips. And so, lots of reactions to [0:45] this. The The Neo Cloud market is [0:47] selling off oddly enough. Meta has a [0:49] bunch of Neo Cloud contracts. Some of [0:51] those companies are selling off because [0:53] now they're a buyer and also a [0:54] competitor. Uh lots of different takes [0:57] about, you know, Meta finding its [0:59] footing, finding something that [1:00] justifies the massive uh CapEx. Of [1:03] course, Meta's [1:04] >> confusing, John. [1:05] >> Yeah? Is it? [1:06] >> I mean, the whole thing. The whole [1:07] thing. [1:08] >> What's confusing about [1:09] >> Uh I think that [1:11] I think it is practical what they're [1:13] doing. [1:13] >> Yeah. [1:14] >> But as somebody that [1:16] that uh you know, I would say overall [1:18] has been a big, you know, cheerleader [1:20] for Meta. I think it's truly the best in [1:22] in my view. It is a perfect business. [1:25] Um [1:26] it doesn't give you a lot of confidence [1:28] in like the strategy overall. If they're [1:30] signing these Neo Cloud deals worth tens [1:33] of billions of dollars, they're built, [1:35] you know, spending hundreds of billions [1:37] of dollars. And yeah, they can make the [1:40] argument that these type of um [1:43] like doing any type of Neo Cloud deals [1:44] themselves is just good business. It's [1:47] It's just like how It's just the best [1:48] way to get ROI today. It doesn't give [1:51] you a lot of confidence that there's [1:53] near-term products on the horizon for [1:55] Meta that are going to be able to [1:56] utilize that capacity themselves, which [1:59] is [2:00] clearly been their strategy. I don't [2:02] Mark [2:03] and and the team have never [2:06] have never said we want to be in the [2:08] cloud business. They've talked about [2:10] >> Yeah. [2:11] >> They've talked about the possibility of [2:12] it. [2:13] >> Yeah. [2:13] >> But the stated goal of MSL is personal [2:16] super intelligence. We don't know what [2:18] >> Which I was a fan of and you I think you [2:19] were a huge fan of. You were like a [2:20] manas on your phone going around your [2:23] social networks. That's my biggest bull [2:25] case for all of this. Like it's very [2:28] there are so many different applications [2:30] that I can imagine being a daily driver [2:32] of in the Meta family of apps. Uh [2:35] oddly none of that has really been even [2:38] tried in my opinion. It feels like a [2:40] little bit early to call it. [2:42] >> so far is Muse Spark. [2:46] >> Good on benchmarks, like decent, you [2:48] know. [2:48] >> Again, not anything to that that anyone [2:51] should really get that excited about. [2:53] >> As a, you know, [2:54] >> As a [2:55] >> as an API provider. [2:56] >> Exactly. Exactly. They did announce that [2:58] they were going to release a VN API. I [3:00] don't think they have. They might still. [3:02] >> a good model, sir, but I don't think it [3:03] will have a lot of demand. And then [3:06] we've seen Meta vibes, which was a [3:07] mid-journey [3:08] >> Yeah. [3:09] >> wrapper. [3:09] >> Even if Muse Spark is not on the super [3:13] giga frontier, can it be good enough to [3:16] get some work done inside Meta family of [3:19] apps? Like it should. I would imagine [3:21] yes, but they just haven't found that [3:23] that killer feature. There are plenty of [3:26] applications that are AI powered. There [3:28] are plenty of models out there that have [3:30] found their footing without being on [3:32] the, you know, super intelligence path [3:34] or on that particular curve. [3:36] >> it's interesting because yesterday we [3:37] were talking about the story where [3:38] Google had been telling Meta like, "Hey, [3:40] we don't have the capacity [3:41] >> Yeah. [3:42] >> for you." And here Meta is uh with with [3:46] plenty of capacity themselves. I don't [3:48] think we can read too much into this cuz [3:49] it's just one article from Bloomberg. [3:52] Uh I think it will matter a lot who the [3:55] potential buyers of compute are going to [3:58] be. If it's There's a number of [3:59] companies that I think the market would [4:01] get excited about. But if they're [4:03] actually just going and trying to [4:05] compete as [4:06] >> Yeah. It's weird that it came as a leak [4:08] about around like a plan to sell compute [4:11] as opposed to just what SpaceX did where [4:13] it was just like boom, huge contract [4:15] with Anthropic, lots of excitement going [4:17] into the IPO. Like that was such a [4:19] perfectly massaged story as SpaceX [4:23] entered the public markets. That would [4:25] have been great if they just said, "Hey, [4:27] we have a Frontier Lab that's paying us [4:29] a billion dollars a month now." And like [4:32] it's going to show up in earnings next [4:33] quarter. Like get ready. But the stock [4:35] market loved it. Like the stock is way [4:37] up and I don't know if it's way up [4:38] because they see it as a huge growth [4:40] area for Meta. Now, is it that crazy [4:43] >> Well, I think that's what people have [4:43] been wondering [4:44] where's the ROI going to come from for [4:47] this hundreds of billions of dollars to [4:48] spend? And up until now there's been no [4:52] obvious place that it's going to come [4:54] from, right? They had the Man of Steel, [4:55] that's being unwound. [4:56] >> Yeah. [4:57] >> Uh they have the the deal with, you [4:59] know, Midjourney and Vives, that's, you [5:01] know, unclear. [5:02] >> Yeah, yeah. [5:03] >> There's It seems very obvious that [5:04] they're going to be able to integrate AI [5:06] into their glasses over time. [5:08] But the glasses that have product market [5:10] fit today are more of just like the [5:13] Uh I I think the product market fit is [5:15] really with the camera. Not the [5:19] intelligence combined with a pair of [5:21] glasses. [5:21] >> Yeah, which is also surprising that we [5:23] haven't seen a diffusion model and image [5:24] model. Instagram in general is pretty [5:27] It's It's remarkably slop-free, at least [5:29] my feed is. I don't see a lot of viral [5:31] slop images. But you can imagine [5:34] AI-powered features, background [5:36] replacement, a lot of that stuff being [5:38] AI-enhanced and people receiving that [5:41] very positively and enjoying that. I was [5:43] using Adobe products recently and like [5:46] in Photoshop they have an integration [5:48] with Gemini Nano banana and you can plug [5:51] in all the different image models and [5:52] you can actually use the tools in very [5:54] interesting ways. I keep going back to [5:56] that idea of like the personal super [5:58] intelligence. What do I actually want? I [5:59] ran into an interesting [6:02] conundrum the other day because I've [6:05] been sort of disappointed by the lack of [6:07] AI in Meta apps which I know I'm like [6:08] the only person that feels that way [6:10] because the general vibe on Instagram is [6:12] like extremely anti-AI. Meta does have [6:15] granular data about every reel I post. [6:17] Some do better than others. [6:20] I went to Meta AI in the Instagram app [6:23] and I asked, "What should I do more of [6:25] in order to grow my account?" This would [6:27] be useful sort of a high-powered [6:30] analytics tools personalized analytics [6:32] tools for what I do. But I got this very [6:35] generic LLM response. We might be able [6:38] to pull it up. I think it's in I shared [6:41] the image in the timeline. Let me see if [6:43] I can pull it up here. It said, "Buffer [6:45] recommends focusing on sustainable [6:48] organic growth instead of quick hacks [6:50] like follow trains. Post reels [6:51] consistently since they get 36% more [6:54] reach than carousels and 125 [6:57] >> It's so funny to be referencing it's [6:58] referencing a blog post from a social [7:00] media management SaaS company. [7:02] >> Exactly. [7:02] >> Like you would think that if you had I [7:04] mean and clearly they're not focused on [7:06] this use case, but you would think that [7:08] there would be an opportunity to give [7:11] creators personal and super intelligence [7:13] to just keep them creating on the [7:15] platform. [7:15] >> Yeah, it said, "Use Instagram analytics [7:17] to see which content converts viewers [7:18] into followers and double down on it." [7:20] Like that's what I asked you to do. [7:22] >> [laughter] [7:22] >> I said, "What can I do better to grow my [7:24] following?" Like you have all the data [7:26] about what does well. I don't want to go [7:28] and look at this reel got 5,000 views, [7:31] that one got 50,000 views. What's the [7:32] difference here? I want you to do that. [7:35] Collaborate with micro creators in your [7:36] niche for authentic cross-promotion [7:39] rather than paid ads. Engage with [7:40] responding to comments and testing out [7:42] trials. It also said to optimize your [7:44] profile with clear keywords in the bio [7:47] and maintain a consistent visual brand. [7:49] It's like a lot of that I'm already [7:50] doing, a lot of that is is is is old. [7:53] And so, I'm just like sort of [7:54] disappointed that say what you want [7:56] about Muse Spark and its benchmarks or [7:58] whatever. Like clearly, if they wire [8:00] that model up to the users' data, they [8:04] should be able to integrate [8:05] appropriately and they just haven't [8:07] productized it properly. And maybe if I [8:09] get managed get an API integration, like [8:11] I could get there. But like it should [8:13] just live in the Instagram search box, I [8:15] imagine, since they already have an LLM [8:16] there, it's just a legacy one. On the [8:18] product side, it's just not it's just a [8:20] not enough to get to okay, there's crazy [8:23] demand for AI within the app. [8:25] >> Did you ask it if the shareholder value [8:27] is three eggs is the goose value? [8:29] >> [laughter] [8:31] >> That just sounds like a like a jumble of [8:33] words. Is that even a correct sentence? [8:35] I don't know. But [8:37] >> Sounds correct to me. [8:38] >> I mean honestly, like the best place [8:39] It's very ironic, but the best place for [8:41] Instagram growth hacks these days is [8:43] just Adam Mosseri's [8:45] front-facing videos. He takes to [8:47] Instagram, he uses Instagram very well, [8:49] he goes direct, posts reels about all [8:51] sorts of things. So, today I got [8:52] serviced one. If you post something and [8:54] it doesn't do well, should you delete it [8:56] and then post it again? And he says, [8:57] "No, because the algorithm will give the [8:59] same result and your followers who saw [9:01] the first one didn't like it, they're [9:03] going to see it and like it even less [9:04] the second time they see it." He just [9:06] answers a bunch of common questions, he [9:07] does Q&A's, and it's actually the best [9:09] way to communicate and get insights into [9:11] how the Instagram platform works. But [9:13] it's not personalized. Like he's just [9:15] giving generic one-size-fits-all advice. [9:18] I want the Adam Mosseri brain enhanced [9:21] with Meta's AI tailored to my account. [9:24] That's what I want. And like a social [9:27] media co-pilot. But maybe I'm in the [9:29] minority here. I I don't know. May- [9:30] maybe I'm the only one that would want [9:31] this. I feel like [9:33] tools for creators would be a great way [9:35] to to launch this. Even like a small [9:37] user, [9:38] they almost everyone almost always cares [9:41] about how I'd love to get a couple more [9:42] views on whatever, even if they're just [9:44] using it and in primarily a consumption [9:47] tool. And then there's also agentic [9:49] shopping, which we asked Mark Zuckerberg [9:51] about at Meta Connect last year. This [9:53] idea that, you know, the Meta Ray-Ban [9:54] displays, they have the head they have [9:56] the HUD, they have AI, you should be [9:59] able to look at a pair of shoes and say [10:01] order me those and he sort of like [10:04] gestured towards that being one [10:06] potential possible possible future, but [10:09] it's crazy to me that we haven't even [10:11] seen them really try to remove at least [10:14] one click from the shopping experience, [10:16] uh, store some more of your data, [10:18] shorten the funnel, increase conversion [10:20] rates. That's good for brands, that's [10:21] good for companies that advertise on [10:23] Meta, it's good for Meta, it's good for [10:25] users, uh, and it just doesn't feel like [10:27] that's where the energy has been, uh, in [10:30] terms of productizing AI within the [10:32] family of apps. And so, um, it feels [10:35] like e-commerce will see agentic [10:37] shopping happening. We're getting [10:39] closer, computer uses getting better, [10:41] APIs are there, and MCP servers and [10:43] Shopify has a bunch of tools for this [10:45] stuff, but, uh, it's weird that Meta [10:47] hasn't even been experimenting there and [10:49] we haven't seen like, oh yeah, like they [10:51] launched a thing where if you see an ad, [10:54] you can click a button and the agent [10:56] will try and go check out with you and [10:58] then just confirm the details within the [11:00] Meta app and you don't actually have to [11:01] open up the Safari window. [11:03] I don't know. Maybe that works, maybe [11:04] that doesn't. At least run the [11:05] experiment. Maybe they have. I don't [11:06] know. Maybe I just missed it. Yeah, this [11:07] feels like a wind down of like the super [11:09] intelligence ambitions that Zuck was [11:10] gesturing towards last year, but I don't [11:12] know. May maybe this is in the path, [11:14] maybe maybe this is just a temporary [11:15] thing. [11:16] >> Yeah, and and and, uh, you know, if you [11:18] look at the SpaceX deals with Google and [11:21] Anthropic, they were they weren't like [11:24] five-year deals, right? They were [11:26] shorter-term opportunities for both [11:28] sides to get out and [11:30] >> [clears throat] [11:31] >> I think that Meta could easily do [11:32] something like that where they could [11:34] make, again, this is like the practical [11:36] decision and say like, "Hey, we actually [11:38] do have way more capacity than we need. [11:41] >> Mhm. [11:41] >> We plan on being able to utilize it [11:44] fully over time, but we need to get our [11:46] products sort of ramped up and so in the [11:48] meantime, why not sell that capacity and [11:50] signal to the market that we're not [11:52] entirely irrational? [11:53] >> Yeah. Yeah. [11:54] >> I think Meta can actually wait [11:56] until like the the features that they [11:58] need to implement are like extremely [11:59] obvious. [11:59] >> Mhm. [12:00] >> Right? Like Apple waited a long time to [12:01] actually implement these things even [12:02] though everyone was like basically [12:03] begging. I I don't know if people are [12:05] like begging for AI features in [12:06] Instagram. And maybe that'll happen in [12:08] the future and they can just basically [12:09] just wait while like, "Oh, we're not [12:11] really sure how to implement this [12:12] stuff." Yeah, you can put in the search [12:13] bar or whatever, but it doesn't improve [12:14] the experience that much. But maybe you [12:16] know, in a year or two there's going to [12:17] be some feature that like, "Wow, [12:18] everyone is really wants this in [12:20] Instagram." [12:21] By then, you know, they'll have all the [12:23] compute necessary. They can you know, 90 [12:25] days before they implement it. They they [12:26] can, you know, get out of the the lease [12:28] or whatever. [12:29] >> Yeah. Yeah. [12:30] Yeah, it's interesting. I feel like they [12:32] just need to do more product [12:33] experimentation. Like the core the core [12:35] Instagram team, they launched the the [12:38] what's it called? Glimpses or something? [12:39] It's like a shorter even shorter version [12:41] of stories, like lower. [12:43] It like sits in the like that little [12:44] side bar of the chat. Like they are [12:46] launching different features across the [12:49] family of apps, but that culture like it [12:51] feels like they have a research [12:52] organization, but they don't have an AI [12:54] product organization that's actively [12:56] productizing things as quickly. I mean, [12:58] Meta vibes for you know, I I even though [13:00] it was like this white label of [13:01] Midjourney was [13:03] well, at least it was launched fast. At [13:04] least they got the feedback. [13:05] >> I don't think they've ever been like the [13:06] the company that like really innovates [13:09] on on product, right? There's going to [13:10] be some [13:11] some feature that everyone's like, "Wow, [13:12] this is really great for AI." And then [13:14] they can then just integrate it into [13:16] their apps. They have all the capacity [13:17] to serve it by that point. [13:19] >> Yeah, I don't know. I don't know. I [13:20] don't know where it goes. Amit is [13:23] investing [13:24] over there, shares his perspective, two [13:26] perspectives on the meta news. Uh he [13:29] says, "Bearish, the bearish take is if [13:31] meta has access compute that they are [13:33] willing to sell the new cloud business, [13:35] doesn't that mean we aren't compute [13:37] constrained? [13:38] Isn't this really bad for neo clouds? [13:40] Why would meta give a deal to CoreWeaver [13:42] Iron if they just sell the compute [13:44] themselves? Furthermore, wouldn't they [13:46] cut CapEx because idle compute as the [13:48] basis for a new business means they [13:50] don't need as much compute as they [13:51] bought, which means CapEx should come [13:53] down. That would be bearish for all [13:55] semi. The bullish take is if meta is [13:57] building a cloud business, even if they [13:59] are using idle compute, which means they [14:01] aren't compute constrained, they might [14:03] end up spending more on CapEx to compete [14:06] with GCP, AWS, and Azure. Uh like if [14:09] they realize that selling compute and [14:10] services on top of a meta cloud is [14:12] better than just ads, then wouldn't they [14:14] end up having to spend in the same way [14:16] that Google, Microsoft, Amazon do in [14:18] order to build out a full cloud [14:19] business? They do have a lot of [14:20] capabilities in terms of spinning up [14:22] data centers quickly. Uh maybe not quite [14:24] as quickly as SpaceX and AWS, but [14:26] they're certainly like near the frontier [14:29] of that capability in terms of putting [14:31] up GPUs in tents. Um so, he says more [14:34] CapEx would be good for semi. So, what [14:36] do people think? Where do we land on [14:38] this? The market is certainly reacting [14:40] positively to meta and negatively to the [14:42] neo clouds because there's new [14:43] competitor in town. Um [14:45] anyway, [14:45] >> Zephyr and the strategy team are going [14:48] pretty hard. They say, "LMAO, Zuck [14:51] finally takes the L." [14:53] >> Okay, so if this is a if this is a [14:55] metaverse style side quest and the and [14:59] the super intelligence meta trained [15:01] models are the VR of this cycle for [15:06] meta's attempt at a new business [15:08] creation, [15:09] uh and the meta trained models get sort [15:13] of mothballed in the way that the meta [15:14] quest and the VR strategy got [15:16] mothballed, what is the meta Ray-Bans of [15:20] MSL? [15:21] Like what would what would remain? [15:23] Because when they wind down one of these [15:25] projects, sometimes you get a Meta [15:28] Ray-Bans, which is pretty good business [15:29] and growing and and cool and like [15:31] there's development there, but it's a [15:33] much more narrow focused. And I would [15:35] say image model, but they they seem to [15:37] >> Wait, wouldn't it just be this like Neo [15:38] Cloud business? [15:39] >> Maybe, yeah. [15:39] >> That's what's going to stay on. And it's [15:40] like profitable. [15:41] >> as Meta [15:42] Ray-Bans though. I want a consumer [15:45] product, it's a consumer company. [15:46] >> Yeah, and it's interesting because you [15:47] can you can make the case that [15:51] Meta could build [15:53] an amazing [15:55] inference business, right? They already [15:57] serve millions of businesses globally. [16:00] They could make a pretty compelling case [16:03] for how like, "Hey, we help people [16:04] acquire customers and we help people [16:07] deliver products and services over [16:09] here." [16:10] >> Yeah, we've talked about that before. I [16:11] don't know that I buy that that the fact [16:13] that they have like [16:14] you know, every single mobile gaming [16:16] company and D2C e-commerce business on [16:19] is actually flows over to, "Well, now [16:22] get your tokens from us." Like [16:24] >> Yeah, just I just I just think that like [16:25] people have it in their head like, [16:26] "Okay, Meta is a consumer business." But [16:30] they have massive sales teams, account [16:33] management teams. They know how to get [16:35] in front of customers. They've got to [16:37] the point where, you know, again, they [16:40] they're not operating a commodity [16:41] business, right? Having their you know, [16:44] social networks. Cloud is is closer to [16:47] one, [16:48] but [16:50] anyways, I'm interested to see I'm I'm I [16:52] imagine they'll have to come out with uh [16:55] their own kind of news around this [16:57] pretty quickly so that they're not [16:59] sitting in limbo with just one kind of [17:01] rumored article floating out there and [17:04] and people are speculating. [17:05] >> The Meta Ray-Ban display is good example [17:08] of like or Meta Ray-Bans good example of [17:10] like there's there's probably a piece of [17:12] consumer hardware that's AI enabled like [17:14] the ring or something where or just [17:17] getting really good at voice models or [17:19] just getting really good at image [17:20] models. And if they're a little bit more [17:22] narrow and constrained, they could [17:23] probably completely dominate that. But [17:25] trying to do super intelligence and and [17:29] coding agents and it's a little it's a [17:31] little scattered potentially. [17:32] >> J Yoon says, we are still massively [17:34] short compute. Meta and XAI are selling [17:36] compute because there's no inference [17:37] demand for their models. It's a compute [17:40] allocation problem. Too much compute in [17:42] the hands of players with no internal [17:43] use for it. Not a compute surplus [17:46] problem. [17:47] >> More Meta news. [17:50] >> that? [17:51] >> Apparently, according to Bobby Allen [17:53] over at NPR, Meta considered buying [17:55] Kalshi before it's developing its own [17:58] prediction market app. That is uh sort [18:01] of a classic Meta playbook. This sort of [18:04] puts to bed your theory that uh they [18:06] might be just making a a you know uh [18:10] a cloud-based prediction markets where [18:12] you compete for your ability to see the [18:14] future. Um [18:16] I uh I mean I mean Manifold is that, [18:19] right? [18:19] >> Yeah. [18:20] >> So it's like a it's like a pretty big [18:22] platform. [18:23] >> Yes, but I was saying there's a chance [18:26] right now based on the reporting that it [18:27] could be the Manifold strategy or it [18:29] could be the Polymarket Kalshi strategy. [18:31] And the fact that they didn't try and [18:33] acquire Manifold, they tried to acquire [18:35] Kalshi sort of signals like, "Hey, [18:38] they're probably going the the [18:40] financially incentivized route." Which I [18:43] I think fits with your thesis that, you [18:44] know, it's in consumer, it's profitable [18:47] and growing very fast. And also, Terek [18:50] from Kalshi was taking shots at [18:52] Instagram saying it's brain rot. Saying [18:54] that like every minute you spend on [18:55] Kalshi is a minute that you're not [18:56] spending brain rotting on Instagram, [18:58] which is like, okay. A lot of people [19:00] would say that these are like equivalent [19:01] or maybe one is worse than the other. [19:03] >> Is the potential profit pool worth the [19:06] risk of all the attention you're going [19:08] to get from lawmakers globally by [19:10] integrating like betting into the [19:13] product that is already [19:15] under attack on like a million different [19:18] fronts, right? It feels like you're [19:20] >> With the movie coming out and stuff, [19:21] it's like you're jumping straight to the [19:23] first [19:23] >> you have a golden goose, right? [19:25] >> [laughter] [19:26] >> And the goose is getting valued. The [19:27] goose is getting valued. I'm going to [19:29] just keep going back I'm going to just [19:29] keep going back to the [19:30] >> You love these slides. [19:31] >> Um the goose is valued. It's producing [19:34] golden eggs and you see another golden [19:36] egg, but it's almost like a poison [19:37] golden egg. And if you bring it over [19:40] might to the farm it might kill the [19:42] goose. [19:42] >> yeah. [19:43] >> It might kill your your main goose. [19:44] >> Potentially. [19:45] >> And so it feels feels risky. [19:47] >> Yeah. Google AI overviews decreased [19:49] outbound organic clicks by 40%. Eric [19:52] Seufert is sharing a new paper by [19:54] researchers at Carnegie Mellon and the [19:56] Indian School of Business [19:58] finds that AI overviews were triggered [19:59] in roughly 41% of observed Google [20:02] searches and when triggered reduced [20:03] outbound organic clicks by about 40%. [20:05] That seems like one-to-one. The presence [20:08] of AI overviews increased the likelihood [20:09] of a zero-click search by roughly 35%. [20:13] Everyone is going [20:14] everyone's going Google zero at this [20:16] point. And the the founder Josh Marshall [20:20] of Talking Points Memo wrote about [20:23] Google AI oligarchy and the end of the [20:25] open web. It's an interesting read, but [20:28] >> [snorts] [20:28] >> go through it another time. [20:29] >> December 2010, Demis [20:32] was scraping together a couple million [20:34] bucks [20:35] >> This is incredible. [20:36] >> company called DeepMind. [20:38] This [20:39] is for a founder that just like, you [20:42] know, had their uh started their career [20:45] in the last few years. This is like in [20:47] inconceivable. [20:48] >> Mhm. [20:49] >> They're like wait, you mean 250, right? [20:53] >> [laughter] [20:53] >> It's like no. [20:54] >> 250 [20:55] >> Sold half the company at 5 million post [20:59] as one of the most elite [21:00] >> Can you imagine what it would be like if [21:02] it was [21:02] >> technologists in the entire world. [21:03] >> stayed independent this whole time? I [21:05] don't know. Maybe maybe maybe there'd be [21:06] another another path or something. [21:08] >> Are you saying they sold too early? [21:10] >> It seemed like they sold too early. I [21:11] don't know. I [21:13] Hard to say. Hard to say. [21:15] >> Paper hands? [21:16] >> Paper [21:16] >> Paper hands Pete. [21:17] >> Paper [21:18] >> [laughter] [21:18] >> Oh, [21:19] no, no, no, no. I think all the VCs [21:21] didn't want to sell. Especially because [21:23] the whole Google thing, that was like [21:24] the whole story was [21:26] was don't sell to Google. Google's like [21:27] the bad one. [21:29] Which is funny because like Google's [21:31] been very responsible and and you know, [21:34] great great company. [21:35] Um I was I was playing around with VO3 [21:37] or VO4. Where where are we in the VO [21:39] models? Um so good and yet still not [21:44] indistinguishable from reality. We're so [21:47] so close and yet so far if you want to [21:50] generate a video these days. [21:53] Um it is cool that you can do it on your [21:54] phone though. So I like it. And Jordi, [21:57] people were asking you to do a little [21:58] bit of a little bit of a spin for [22:00] everyone. Show off the new merch you got [22:02] there. I'll tell everyone about Codex on [22:05] the arms. Codex is a powerful workspace [22:07] for getting work done with AI agents. [22:10] Whether you're writing code, analyzing [22:11] data, creating content, or automating [22:13] business workflows, or just trying to [22:15] just trying to look good. Codex [22:18] >> We we just got our first shipment of I [22:20] think around 201 of our new merch [22:24] product. [22:24] >> Oh, okay. I know the one you're talking [22:26] about. [22:26] >> And that one is going to be available [22:28] online and [22:28] >> That's the one. [22:30] >> No, no, no, no, no. Not the one not the [22:31] one you're thinking of. [22:32] >> Okay, okay, okay. [22:33] >> No, that one will also be available [22:35] >> After all of that [22:36] >> one we have one that we have one that's [22:38] a little bit silly. [22:38] >> It's groundbreaking. [22:39] >> It's groundbreaking. It's it's it's it's [22:42] a product that humanity has been trying [22:45] to create [22:46] for centuries. [22:47] >> Yeah. [22:48] >> And we did it. We did it. Uh and [22:52] >> We had a breakthrough. [22:53] >> We had a big breakthrough. [22:54] >> We had a big breakthrough. [22:54] >> Uh it took a long time. [22:55] >> It did. [22:56] >> Uh but but yeah, no, the merch will [22:58] finally be for sale and we will make the [23:01] it will be dropping in the chat [23:03] >> Oh, yeah. [23:04] >> first [23:04] >> Yeah. [23:05] >> We're not exactly sure exactly when. [23:07] >> Yeah. [23:08] >> But we will give it to you guys first. [23:11] And thank you for the patience. [23:13] >> Yeah, we appreciate [23:13] >> But on that note [23:14] >> We'll see you tomorrow. Leave us five [23:15] stars on Apple Podcast and Spotify. Sign [23:17] up for our newsletter at tbvpn.com. [23:19] We'll see you tomorrow. Goodbye. [23:21] >> We love you.