STRATECHERY'S BEN THOMPSON WARNS OF 2029 GLOBAL AI CHIP CRUNCH

Stratechery 创始人警告：2029年全球或爆发AI芯片荒 台积电保守的扩产策略叠加自主 AI Agent 对算力密度的巨大需求，将在本年代末引发全球严重的硬件短缺。 在本期内容中，Stratechery 创始人 Ben Thompson 深入剖析了数字媒体、聚合理论与人工智能的演进格局，对比了 Meta 基于用户画像的广告模式与 OpenAI 的路径。他阐述了 Agent 电商未来发展的三个层级，同时警告受台积电保守资本支出影响，全球预计在 2029 年出现严重的芯片短缺。最后，Thompson 还讨论了按席位计费的 SaaS 估值增长停滞的现状，评估了 Google、Apple 和 Amazon 等科技巨头独特的企业性格与战略，并探讨了优质付费订阅通讯的商业逻辑。 得益于受日本影响的城市布局、混合功能分区以及迅捷的外卖生态，台湾（特别是台北）提供了无可比拟的生活便利性，不过 Uber Eats 上幽灵厨房的兴起对当地传统的餐饮文化造成了冲击与蚕食。 (1:00) 在互联网领域，受零边际成本影响，话语权从供应方转移至需求聚合方，正如 Booking.com 聚合了酒店资源，自身却又被 Google 进一步聚合。 (4:58) OpenAI 基于上下文的横幅广告模式存在严重缺陷且易引发利益冲突；平台更应采用类似 Meta 的整体画像模型，在完全独立于实时对话界面的位置展示相关产品。 (11:14) 广告转化效果极大地取决于用户的使用心智与视觉媒介形态；在 Twitter 等以获取信息为导向的平台上，用户对广告排斥度较高，而在 Instagram 等被动浏览视觉信息流的场景中，用户转化率极高。 (14:27) TikTok 彻底抛弃了传统的社交图谱，转而押注基于算法的兴趣图谱，从而颠覆了社交媒体行业格局，将平台演变为一种高度个性化的“电视”。 (22:05)

Ben Thompson argues that while artificial intelligence expands consumer aggregation and agentic commerce, attempting to monetize conversational interfaces with direct ads fundamentally misunderstands user posture and incentive alignment. He contends the tech sector faces compounding structural shifts: SaaS valuation compression, unbundling dynamics in digital media, and physical semiconductor manufacturing bottlenecks centered at TSMC. AI Advertising Flaws and Aggregation Mechanics (11:14) Thompson views OpenAI's ChatGPT as a premier demand aggregator, yet criticizes its contextual banner ads as a flawed approach that creates perceived conflicts of interest. In his view, direct ad placement inside chat prompts risks making answers appear sponsored while ignoring user posture. Whereas social feeds capture passive, receptive attention, chat clients involve active, task-oriented users who resist direct text ads. Rather than in-chat banners, Thompson advocates a profiling model similar to Meta's, where platforms build holistic user understanding over time to serve relevant ads outside conversational flows. This dynamic gives Google a structural advantage over OpenAI: Google can channel Gemini's conversational insights into external properties like YouTube and Search, whereas OpenAI lacks external ad inventory and risks degrading the user experience by monetizing directly inside ChatGPT. Agentic Commerce and the Hyper-Competition Threat Thompson outlines three developmental stages for agentic commerce: replacing checkout forms with just-in-time interfaces, enabling natural-parameter discovery, and deploying persistent user profiles that anticipate needs. He notes Meta already operates an early version of anticipatory commerce, functioning as an autonomous agent where merchants set target acquisition costs and algorithms find buyers. However, Thompson cautions that automated research brings severe economic tradeoffs. When technology grants every consumer instant, exhaustive comparison capabilities, markets risk collapsing into pure price competition that decimates product categories. Furthermore, algorithmic buying tends to over-index on quantifiable specifications at the expense of qualitative, unmeasurable craftsmanship—echoing how sports analytics can over-optimize measurable statistics while ignoring intangible dynamics like chemistry. SaaS Headwinds and Sovereign Media Tradeoffs (44:44) Assessing public software markets as of January 2026, Thompson disputes that software-as-a-service is dead, but highlights severe valuation compression. Because enterprise SaaS depends on seat-based pricing, slowing headcount growth stalls revenue expansion, forcing markets to value companies on earnings per share rather than revenue multiples. Concurrently, the internet and AI enable individual entrepreneurs to operate niche businesses, bypassing legacy enterprise sales stacks in favor of self-serve or custom tooling. In media, Thompson observes the limitations of the sovereign creator model he pioneered. While direct monetization allows low-overhead operations to thrive independently, it cannot easily fund intensive investigative reporting. Additionally, efforts to re-bundle independent writers through platforms like Substack face adverse selection: top-tier writers with strong direct billing have no economic incentive to join a shared bundle, leaving platforms unable to replicate traditional media packages. Fabrication Bottlenecks and Amazon's Silicon Gamble (66:05) Thompson identifies physical fabrication capacity at Taiwan Semiconductor Manufacturing Company (TSMC) as the primary bottleneck for AI expansion. Because fab construction involves billions in fixed depreciation, TSMC rationally moderates capital expenditures to prevent catastrophic price collapses from overcapacity. After cutting capex post-2022, TSMC's conservative expansion could trigger a severe global compute crunch around 2029 unless hyperscalers actively subsidize alternative foundries at Intel or Samsung. Meanwhile, cloud providers navigate conflicting silicon strategies. Amazon seeks to repeat its Graviton CPU playbook with its Trainium AI training chips, absorbing hardware internally to establish a structural cost advantage before selling externally. However, Thompson points out that this playbook worked for CPUs during a period of architectural stabilization from 2005 to 2025; pursuing proprietary silicon amidst today's rapid generational leaps in AI hardware risks leaving providers behind Nvidia's cutting-edge systems. Thompson concludes that physical fab limits, enterprise pricing friction, and consumer psychology constrain the AI transition far more than raw model improvements. As AI individualizes software and content, he argues that un-commodifiable shared experiences and verified authenticity will capture premium value, while pure commodity providers face razor-thin margins.

科技分析师Ben Thompson认为，AI时代的商业竞争核心在于聚合与广告机制的重构，但受制于台积电产能与算力跃升的物理现实；同时，传统按人头付费的SaaS模式与媒体分发结构正面临不可逆的范式转移。 AI时代的聚合理论与广告重构 (4:58) 在互联网的零边际成本环境下，商业权力从供应端转移到需求聚合者，例如Booking.com虽聚合酒店却被Google聚合。Thompson指出，OpenAI最有价值的资产是ChatGPT这一直接触达用户的终端，即便停更模型也具备巨大商业价值，但硅谷对广告变现的道德迟疑削弱了其商业化进程。他批评OpenAI基于单次对话上下文的横幅广告是最低限度的妥协，容易引发用户对核心回答中立性的质疑。 针对AI商业化路径，Thompson主张借鉴Meta的长期用户画像模式，而非依赖对话框内的即时搜索词匹配。平台应在多轮交互中构建整体认知，将广告分发至对话外部场景。Google的理想路径是保持Gemini界面的纯净，将AI提炼出的精准意图转化为YouTube和搜索广告的高变现率；反观缺乏外部广告生态的OpenAI，将库存直接塞入ChatGPT可能会损伤用户体验。 从表单自动化到主动预测的代理商业 针对AI代理驱动的商业模式，对话提出了由浅入深的演进三阶段：第一阶段是“即时生成界面（Just-in-Time UI）”，AI直接接管繁琐的结账表单；第二阶段是超越传统关键词匹配的参数化发现，允许用户按温度、风格等复杂现实条件检索商品；第三阶段则是基于历史行为与偏好的持久化画像。Thompson指出，第三阶段在很大程度上已由Meta的智能广告引擎实现，广告主只需设定客单预算，算法即自主完成供需匹配。 然而，AI驱动的商业决策也带来了“过度优化”的经济陷阱。如同全面透明的互联网导致地方报纸陷入纯粹竞争一样，AI带来的极度透明可能将日用品推入无利润的残酷竞价中。Thompson借用Steve Jobs对硬件规格与感官体验的论述，指出AI倾向于基于可量化参数做决策，从而剥夺消费品的“灵魂”与非量化溢价，不过像Amazon Basics这类无品牌基础日用品的普及也确实切实降低了大众的生活成本。 SaaS模式危机与小池塘经济学 (44:44) 针对2026年1月公开市场上“SaaS已死”的讨论，Thompson分析认为，SaaS的核心危机并非归零，而是按人头（seat-based）收费的扩张模型在企业人员增长放缓甚至缩减时受到挤压，估值逻辑从高增长倍数转向每股收益（EPS）。然而，作为系统记录核心（Systems of Record）的深层基础设施（如Workday）依然稳固，甚至AI前沿机构Anthropic也选择采购而非自行编写内部HR系统。 与此同时，AI正在降低微型企业的运营门槛，推动垂直领域的“小池塘”经济繁荣。创作者与创业者不再需要模仿成熟巨头，而是在细分领域维持高粘性。以媒体为例，Thompson开创的独立订阅模式正是通过技术杠杆压低运营开销，以自营轻资产形态换取高利润率。他指出Substack在早期设计中让作者保留独立Stripe账户并拥有订阅者数据，导致后续搭建“统一订阅包（Bundle）”时陷入严重的逆向选择难题。 物理算力瓶颈与2029年芯片危机 (66:05) AI扩张的核心制约正从算法转向台积电（TSMC）的晶圆代工物理产能。建厂成本极高且折旧占比接近99.9%，若盲目扩产将引发严重的产能过剩与价格崩盘，因此台积电缺乏激进扩张的动力。在缺乏同代逻辑制程竞争对手的格局下，台积电成功将供应不足的风险转嫁给客户，甚至在ChatGPT引发投资热潮后的两年内连续削减了资本支出。 由于晶圆厂扩建周期与自主代理所激增的计算密度之间存在脱节，业界预计全球将在2029年前后遭遇严峻的高性能芯片短缺。Thompson强调，超大规模云服务商不能仅依靠向台积电加价或寄望地缘政治庇护，而必须积极投资扶持Intel与三星的代工能力，以打破单一巨头的垄断瓶颈。 巨头执行力博弈与云端定制芯片分歧 (83:46) 在科技巨头的执行特质上，各方呈现显著的结构性矛盾：Apple精于硬件制造与工业设计，但在平台治理与软件体验上较为严苛；Microsoft擅长平台与企业级整合，即便单一产品平庸也能通过套件分发击败垂直工具；Google凭借庞大搜索现金流表现出极强的冗余试错与迟缓抗性；Meta在工程管理与开源生态上执行力极强，但高度依赖用户注意力停留，若ChatGPT挤占原本属于Instagram的时间，其商业根基将受到侵蚀。 在底层基础设施方面，Amazon在通用计算时代通过Graviton芯片确立了内部验证并向外输出低成本的成熟路径。然而在AI专用芯片Trainium的战略上，由于大模型算力正经历跨代爆发式跃升，而非2005至2025年通用CPU的渐进演进，过度押注自研定制芯片可能导致AWS在争夺当前Nvidia顶尖算力资源的竞争中错失先机。 无论是广告范式的重构、按人头计费SaaS的重估，还是代理商业对商品经济的穿透，技术演进在赋能个体的同时也瓦解了旧有的聚合中介。然而，上游代工产能的物理制约与定制芯片追赶算力跃升的滞后性，表明这场软件与商业模式的变革仍受制于严苛的硬件周期与基础设施现实。

[11:14] Ben Thompson: The bare minimum easiest solution. [17:53] Ben Thompson: Apple saved Facebook from itself. [17:53] Ben Thompson: An advertising-based model is fundamentally incompatible with being a platform. [23:51] Ben Thompson: Fairly insane to have a primarily information source controlled by your chief geopolitical adversary. [28:28] Ben Thompson: Burnout's not a function of how much work you're doing, it's doing work you don't enjoy. [29:58] Ben Thompson: In the short term, technology often makes old business models even more powerful before it kills them. [40:15] Mark Zuckerberg: The largest and most successful agent in the world today is Facebook advertising. [44:44] Ben Thompson: Silicon Valley Inc. [52:24] Ben Thompson: Did it all wrong. [56:27] This is not going to work on the internet. [73:13] Ben Thompson: I'm a world's biggest semicolon user. [82:47] Ben Thompson: The price of becoming a platform is making a great product. [82:47] Ben Thompson: It might take them a really long time to get there, but when they get there, you're doomed.

[0:01] Ben Thompson is the founder and author [0:03] of Stratechery, the newsletter that [0:04] everyone in tech reads to make sense of [0:06] what's happening. He's also early to the [0:08] premium newsletter model that's become [0:09] very popular in media nowadays. For many [0:11] [music] years, he ran Stratechery as a [0:13] solo founder in Taiwan. [0:15] Cheers. Good to see you. Cheers. [0:20] It feels like people in San Francisco [0:23] have not properly discovered Taiwan as a [0:24] tourist destination. Like, do you agree [0:26] with that characterization? Or what's [0:27] your recommendation? [0:28] >> ask me about Asia. By the way, [0:30] [clears throat] the way I always [0:31] characterize Taiwan is there's lots of [0:32] great places to visit in Asia. And I [0:35] would also put Japan on the list. But, I [0:38] I you know, I like to think I was in [0:39] Japan before it was cool. Yeah, nothing [0:40] against Japan. It's just [0:41] >> Well, the whole thing with Japan is [0:43] going to Japan pre-smartphone was a [0:44] completely different experience than [0:46] going there post-smartphone. [0:48] Like, you think, "Oh, the subway [0:49] system's amazing, the trains." Try [0:51] navigating that with no smartphone and [0:53] nothing's in English. Like Like Japan [0:55] used to be very low on English. Uh it's [0:57] still lower than places like Taiwan. [0:59] >> low. Yeah. And Japan it just has the [1:02] the way to visit Japan is you just walk. [1:05] Like, don't go to like set destinations. [1:09] Um where the way I would talk about this [1:10] is places to visit, but the best place [1:12] to live is undoubtedly Taiwan. The one [1:14] word everyone says for Taiwan sounds not [1:18] that impressive, but the word is [1:20] convenient. Mhm. Like, it is the most [1:22] convenient place to live. So, part of [1:24] that is actually [1:25] >> 7-Eleven has really good food. Well, [1:26] it's actually downstream from the [1:28] Japanese cuz Taiwan was a Japanese [1:30] colony for the first 50 years of the [1:33] 20th century. And it's laid out a lot [1:36] like [1:37] Why is it great to walk in Tokyo? Cuz [1:38] Tokyo's all mixed-use, right? That's how [1:40] Taipei is as well. You have these big [1:43] blocks where the exterior will be [1:45] commercial and the interior of these big [1:47] blocks is all residential and the first [1:49] floor is all like small shops or [1:51] restaurants, things like that. So, [1:52] wherever you live, you basically have [1:54] access to everything all around you. [1:56] But, the the I think the downside as a [1:59] tourist is it's kind of an ugly city. [2:01] Like Taiwan's kind of notorious for just [2:04] these [2:05] dumpy, dilapidated buildings. Then you [2:07] go inside and they're like palatial on [2:09] the inside. Like Taipei is very, very [2:11] rich. It's like in the top 10, I think, [2:13] as like number of billionaires in the [2:15] world or something like that. All [2:16] downstream from building out China, [2:17] things like that. It's a very beautiful [2:19] country. [2:20] From Taipei, 30 minutes to the ocean, 30 [2:22] minutes to the mountains. East coast is [2:24] amazing. But if people listening to this [2:26] are visiting, like I feel like one thing [2:27] they should do is it's a mistake to try [2:29] and use Yelp or anything like that too [2:30] much cuz like you should maybe just try [2:32] and go to a night market and follow your [2:34] belly and see what looks good. And [2:37] there's like a lot of excellent street [2:38] food. And so that'll be one thing is to [2:40] don't try to over plan. [2:42] >> Well, here's the problem though where [2:43] tech has made it worse, I would argue, [2:45] as a tourist. Which is uh Taiwan is [2:48] arguably the greatest Uber Eats market [2:49] ever. Because there there's amazing [2:52] options. It's all delivered by scooter, [2:53] so it's always like 10 minutes to get to [2:56] get dinner and things along those lines. [2:58] I think you were going to ask me about [2:59] difficulties living in the States. Not [3:01] having access to that is definitely one [3:03] of them. But the problem is that it's [3:05] such a huge market now that I think [3:08] there are fewer and fewer restaurants [3:11] in that a lot of these places actually [3:13] just straight up closed their [3:14] storefronts or just ghost kitchens, [3:15] basically. And all they do is just make [3:18] Uber Eats orders all day. I see. So, I [3:20] mean, famously, yeah, the the restaurant [3:22] economy in places like Taipei had been [3:24] really good, but it's gotten worse [3:25] because people are eating in more with [3:27] Uber Eats and stuff like that. I think [3:28] so. I think so. As far as the like [3:29] walking around and just like stuff on [3:31] No, there's still plenty of places. It's [3:33] still great. But there are like there's [3:35] a number of restaurants that I used to [3:36] always take people to, like holes in the [3:38] wall that I knew were super good beef [3:40] noodles or something. And I remember a [3:41] couple times like, "Oh, there you can't [3:43] actually go eat there anymore." But then [3:44] they're still on Uber Eats. That's a [3:45] bummer. It's like a separate problem [3:46] than the the San Francisco problem at [3:47] restaurants is that nobody drinks [3:49] anymore. And so the restaurants can't, [3:51] you know, they've lost a major revenue [3:52] source. [3:53] >> you you had to get on [3:54] >> [laughter] [3:54] >> Exactly, we're just trying [3:56] Exactly, we're trying just first hand to [3:58] to to fix this. [3:59] >> Be the change you want to see in the [4:00] world. Should people visit places beyond [4:02] Taipei or how should they visit? Oh, for [4:03] sure. Yeah, I know I Taipei is great. [4:05] It's great to walk around. Taipei 101 [4:08] which is obviously very much in the news [4:10] these days with with the the scaling. [4:12] But you can go up the elevator on the [4:13] inside. [4:13] >> You go up there because there's a [4:15] massive ball at the top. [4:17] >> damper, yeah. Yes, which is amazing. [4:18] It's like like I do If you If you're [4:20] into like engineering, that's actually a [4:22] very underrated thing. National Palace [4:23] Museum is amazing. But the whole the [4:25] East Coast in particular is is [4:27] incredible. You like you drive There is [4:29] a train, but the driving you can drive [4:31] like you're driving on the coast like [4:33] Yeah, it's like the Lost Coast of Hawaii [4:34] kind of. [4:34] >> Exactly. The problem there's an [4:37] incredible gorge called Taroko Gorge. [4:39] That's closed, right? [4:39] >> That was really yeah, messed up by an [4:41] earthquake a couple years ago. So I [4:43] don't know if it's even reopened yet. [4:44] But I used to take people to that all [4:46] the time cuz it's it's world-class. [4:48] Yeah, yeah. It is impressive that they [4:50] said, "We're going to build the tallest [4:51] skyscraper in the world in like a very [4:53] frequent earthquake region." So Yeah, [4:55] it's a skyscraper. It worked out well [4:57] for the Netflix. [4:58] >> Yes. So you've talked a lot over the [5:01] years about aggregation theory and [5:03] really popularized this idea where [5:05] pre-internet, often power would live [5:07] with the supply, whereas on the [5:10] internet, because of the different, you [5:12] know, marginal cost dynamics and things [5:13] like that, [5:15] uh [5:17] power will rest with the demand [5:18] aggregators. And so booking.com is a [5:22] much bigger company than any hotel [5:23] chain, you know, something like that. [5:25] And booking.com is the particularly [5:26] interesting one because they aggregate [5:28] all the hotels, but they are also [5:30] aggregated by Google. So they're like [5:33] Google's biggest customer even as [5:34] they're also on the other side. [5:36] >> I feel like booking.com is a very [5:37] underappreciated success story in tech. [5:40] You know, they're a European company, [5:41] kind of much quieter in a lot of ways, [5:43] but like if you invested a dollar in [5:45] booking.com and a dollar in Google 20 [5:47] years ago, you made much more money as a [5:49] booking.com shareholder. And I think [5:51] people don't appreciate that fact. It's [5:53] a very well-run business. [5:54] Um, where where I was going with this is [5:57] how [5:59] does aggregation theory apply to AI? How [6:02] does one need to update the framework? [6:04] TBD to a certain extent. I mean, this is [6:06] part of the huge, [6:08] probably the the the one of the most [6:10] angsty debates that I have internally [6:12] generally, which is OpenAI's sort of [6:14] welfare going forward. I put forward a [6:16] few years ago [6:18] that [6:20] actually OpenAI could stop making models [6:23] and be one of the most valuable [6:24] companies in the world just because of [6:26] ChatGPT. That's their most That's their [6:28] most valuable asset. [6:30] And part of the problem issue that they [6:32] have is [6:34] that was definitely the case in 2023, [6:36] 2024. [6:38] Um, but they never You have to actually [6:40] build the business model around that. [6:42] And I've, you know, I think fairly [6:44] famously, who is based on all the tweets [6:46] that I got when they announced that [6:47] they're going to launch ads, have been [6:49] losing my mind about this fact for a [6:50] long time. And I think this is [6:52] interesting. I'd actually be curious [6:53] your view of this. [6:56] Which is there's this mindset in [7:00] the Valley [7:01] of this skepticism of advertising. And [7:03] people have sort of like internalized [7:05] that it's it's bad and evil. Do you do [7:07] you do you do you sense that? Do you [7:08] feel that? [7:10] I agree there's kind of a knee-jerk [7:11] skepticism of ads. And like, look, you [7:14] know, when I I mean, I'm a YouTube [7:15] Premium subscriber. When I see someone [7:17] you know, watch a video without Premium, [7:19] I'm aghast. Yeah, yeah, it's like, what [7:21] are [laughter] you What are you doing [7:22] with your life? And so, I like I get the [7:24] knee-jerk reaction. And that's right, [7:26] but at some level it's kind of the the [7:27] anti-advertising company. And like, [7:29] we're we're the opposite form of [7:31] monetization. But, [7:33] I don't know. I have no particular, um, [7:35] issue. I think it's like a very [7:37] efficient form of monetization. It makes [7:39] a lot of sense for certain products. [7:41] Um, and so, I think it's just different [7:43] strokes for [7:44] >> Yeah, I think it's [7:45] you're you're I think you're with [7:46] Stripe. You're you're on the you're on [7:48] the skepticism side. I think ads are [7:49] amazing. [7:50] And uh, and you know, I'm talking my [7:52] book a little bit. Stratechery has [7:54] gotten tremendous traction just by not [7:56] hating ads. Even though I'm not an ad [7:58] model myself, right? [7:59] >> but you're a paid model. [8:00] >> Uh, well, it's actually it's funny. I [8:02] actually think I got a lot of traction [8:03] over the years by talking about ads when [8:05] no one else was despite the fact it's [8:07] the most important business model that I [8:08] have. [8:09] And I look back in all my early writing [8:11] about ads was terrible. I had no idea [8:12] what I But I would just by virtue of [8:14] talking about it, it was it was helpful. [8:16] Uh, the the reality of advertising is [8:20] number one, [8:22] people if you're making a product in the [8:24] world, like Stratechery is very [8:26] fortunate. Like it is definitely a new [8:29] model or a new internet-native model in [8:32] that I have subscribers in like 200 [8:35] countries, right? Like I'm literally the [8:36] whole world is my market. [8:38] Uh, and Stripe obviously helps me also [8:40] got buying Yeah, a few in the Vatican, [8:41] you know, they're following along with [8:42] the I could go check. You know, I I I [8:45] would put the odds are very high that I [8:47] do have at least one one subscriber in [8:49] the Vatican. Um, and but where I [8:52] benefited from is I was a massive [8:54] beneficiary of social media, [8:55] particularly Twitter. And that back in [8:57] the early days, good days of Twitter, if [9:00] you want to say whatever it was, uh, [9:01] there was currency in sharing smart [9:03] links. And so if I was a regular [9:07] provider of links that people felt made [9:09] them smart, so they would share them and [9:10] talk about them and be sort of back and [9:12] forth. And so that solved my customer [9:14] acquisition [9:16] issue. [9:17] The reality though is most because the [9:19] other thing about about content actually [9:21] is a point that I I should have come [9:22] back to [9:24] is it's a something to talk about. It's [9:26] it's something that's [clears throat] [9:26] commonality sort of sort of between us [9:28] that we can both read the same thing. We [9:30] can both have opinions on it, sort of [9:31] react to it. [9:33] The [9:35] sort of product that I buy on Instagram [9:37] is not what I'm I'm not going to like [9:39] post about it or talk about it, but it [9:41] can be tremendously beneficial. And they [9:43] in theory have this those these small [9:46] businesses or whatever they might be [9:49] or Chinese suppliers or whatever, they [9:50] have the same opportunity which is sell [9:52] to everyone in the world. They just need [9:55] a way to tell people about it. And as [9:56] someone who buys way too many products [9:58] off of Instagram, and by the way, one of [10:00] the great things about we back to the US [10:02] is the Instagram ads are unbelievable. [10:05] They're pretty good in Taiwan. They're [10:06] so much better than the US. So I like [10:07] oh, this is the best part of living in [10:09] the richest country in the world. [10:11] It's amazing. Like I've like what what's [10:13] this native content? Give me more ads. [10:15] Which by the way, Facebook is very happy [10:16] to do over the last 6 months. [10:19] Lots of ads these days. But [10:22] I get stuff that I never would have [10:23] thought of. I didn't even know about and [10:25] it's great. It's amazing and it's a real [10:27] benefit to me as a consumer who [10:30] subscribes to for sure subscribes to [10:31] YouTube Premium and looks down on people [10:33] who don't, but [10:34] finding things that I didn't know about [10:36] that make my life better. So as a user, [10:39] I'm benefiting. As a rich user, I'm [10:40] benefiting. [10:41] As [10:43] the world is 6 billion people, most of [10:45] whom do not have disposable income that [10:47] I have, much less you have, much less [10:49] anyone else in San Francisco has, [10:51] they get the same experience I do. And [10:53] something for AI like when you think [10:56] about particularly when it's so costly [10:58] to provide and the free product is so [11:01] much worse than the paid product, [11:05] of course it's a win for them to be able [11:07] to get access like so to have a mission [11:09] of believe that AI makes the world [11:11] better and to not embrace ads, No, no, [11:14] no. I I agree that ads being an [11:15] efficient form of monetization. What do [11:17] you think is the right way for consumer [11:19] AI apps to do ads? Like you know, [11:21] ChatGPT just announced that they're [11:22] doing ads and they're doing them well, [11:25] no, they're they're doing them as a very [11:26] separate experience to the answer. No, I [11:29] just [11:30] No, this is why it's so bad. This is why [11:32] I'm so frustrated with them. So, what [11:34] they're doing is the bare minimum [11:35] easiest solution. [11:37] >> banner ads, basically. It's banner ads, [11:38] but it's based on the context of the [11:39] conversation. [11:41] And the problem is that like they [11:42] released like their ad principles, [11:44] right? Which is [11:46] our ads do not influence your answer. [11:48] If you're using the easiest possible way [11:50] to target ads, which is based on the [11:52] context of the conversation, we're going [11:53] to show you a roughly relevant ad. [11:54] Number one, your market's way smaller [11:56] because you have to hope someone starts [11:57] a conversation that matches the [11:58] inventory you have. [12:00] Number two, you're getting into a [12:02] my t-shirt answers questions that my [12:04] t-shirt is raising sort of situation [12:06] where if the ad is clearly connected to [12:09] the answer, you're going to raise [12:11] suspicion in the user's minds about [12:13] about what sort of the connection is. [12:14] So, [12:15] I would prefer [12:17] if the ads had nothing to do with the [12:19] answer. [12:20] The way you get there is you build a [12:22] meta style understanding of the user and [12:25] show them stuff that's relevant to them [12:27] like in Instagram. The best Instagram [12:29] ads don't have anything to do with the [12:31] stuff I'm surfing. It's from meta's [12:33] understanding me broadly. [12:34] >> So, are you saying that like the AI ads [12:37] should be more like Facebook ads than [12:39] Google ads? And you know, right now the [12:42] focus is on doing targeted ads that are [12:43] related to the prompt, whereas instead [12:45] it should all be profiling the user and [12:47] who this person is and what their [12:48] interests are. [12:48] >> Yes, I think that would be better. It [12:50] would it would present less conflict of [12:51] interest, less uncertainty amongst the [12:53] user. And it's it's a model that I think [12:56] I'm I'm not the world's biggest fan of [12:58] search ads precisely for the reason why [13:01] they work so well. I think there's a lot [13:03] of search ads out [13:04] >> between organic [13:05] >> No, there's cannibalization. Like why is [13:06] it that I have to buy my own name in [13:08] search ads, right? Because someone else [13:10] will like go go in there and you're [13:12] getting a harvesting a quick on the ad [13:13] that would have been there sort of [13:15] organically, which is fine. It works. [13:17] The search is providing a lot of value, [13:18] but the the the challenge, obviously, is [13:22] they only have one space for inventory, [13:24] which is in ChatGPT. Well, so in the [13:26] defense of Google ads, that everyone [13:28] complains about the branded search and [13:30] yeah, you're paying for cannibalization, [13:31] but [13:32] Google pays so much attention to search [13:36] quality that the sponsored listings [13:40] themselves have a [13:43] a ranking of them like a relevance [13:45] ranking of them. And so it's really just [13:47] like the Yellow Pages where you know, [13:48] you need to pay to be listed in the [13:49] Yellow Pages. [13:50] >> I I I I I don't I'm not I don't again, [13:52] I'm the ad lover here. [13:54] I just think that meta ads are more [13:56] broadly valuable because they're showing [13:58] me stuff I didn't know that I wanted. [14:00] >> But if the AI apps are to generate a [14:02] profile of you, does that profile [14:05] include the content of all your [14:07] conversations? [14:07] >> is the thing. [14:08] >> And isn't that the same thing? [14:09] >> this is the thing. So Demis is out there [14:10] saying, "Wow, I can't believe they're [14:11] adding ads. We're not going to do that." [14:13] Which is hilarious because the entire [14:15] Gemini [14:16] DeepMind [14:18] apparatus [14:19] what is it funded by? [14:21] Sure. Yeah, it's the Google ad machine. [14:23] >> It's funded by ads. And that actually is [14:25] probably the ideal model. So, it's [14:27] actually very funny. I was actually in [14:28] New York City last year. [14:31] I was meeting actually was meeting with [14:33] someone in the office in in the shared [14:35] office or across the hallway was like a [14:36] was a hedge funder or someone. And they [14:38] came over and and he's like, "Oh, long [14:40] time reader. [14:42] You're responsible for our worst [14:43] decision ever." [14:44] >> [snorts] [14:45] >> And I'm like, "What?" He's like, [14:46] "Putting money in Twitter." I'm like, [14:48] "I've never said to put money in [14:49] Twitter." Like that's always been a [14:50] terrible company. Like I I [14:52] stopped covering them cuz it was such a [14:53] bad business. And [14:55] and I'm like, "Oh, no, I remember what [14:56] it was. It was when they bought MoPub." [14:59] And my theory my problem with Twitter [15:00] advertising has always been that it was [15:03] actually was very textual. I think text [15:05] doesn't work as This is a problem for [15:06] the All this applies to the chat [15:08] clients. Text isn't the best interface [15:10] for ads. Obviously, visuals are [15:11] generally better. There's also a [15:12] posture. [15:14] If I'm on Twitter, I'm like, I'm ready [15:16] to do battle. I've walked in. I'm or I'm [15:18] searching for information. If I'm on [15:20] Instagram, the whole point of like [15:21] seeing an ad is I don't really care what [15:24] I'm seeing right now. I'm wasting time. [15:25] You're actually in a much better [15:26] posture, I think, to absorb just like [15:28] TV. Like, you're you're sort of [15:30] absorbing Can absorb the ad. [15:32] And Twitter's like bad for that. So, but [15:34] but Twitter, because it's an [15:35] interest-based network, at least in [15:37] theory, it should be able to understand [15:39] a lot about you above and beyond, [15:42] theoretically, having pixels and and [15:44] SDKs sort of all over the web. And so, [15:46] my theory with the MoPub acquisition, I [15:47] thought it was a great acquisition. Cuz [15:49] like, oh, they can harness signal from [15:52] Twitter and manifest it in other apps [15:55] through this sort of MoPub MoPub [15:57] network. Now, Twitter was incompetent, [15:58] so they did nothing with MoPub, gave it [15:59] to Apple 11, who's now ridden MoPub to [16:01] like like the top of the world. But, [16:03] that was sort of my thesis. [16:06] And I think that could apply that to AI [16:07] as well. I think the ideal outcome for [16:10] Google is they never put ads in Gemini, [16:13] but they understand so much about you [16:16] because of what you do in Gemini that [16:19] they can then manifest that through ads [16:21] on YouTube, through ads on Google, [16:23] through ads on their other properties. [16:24] And the challenge for OpenAI is they [16:26] only have one place for inventory, which [16:27] is which is in ChatGPT. Okay, so you're [16:30] saying this Google could use uh Gemini [16:33] to just improve the targeting of the ads [16:35] across the Google properties, and then [16:36] maybe if you want to have ads in Gemini, [16:40] uh I don't think you need to ever put [16:41] ads in Gemini. [16:41] >> But but just if you did, you would also [16:43] have the profile that Google has of you [16:45] from across the Yeah. [16:46] And you don't need to have ads that are [16:48] like making the user feel weird because [16:51] why are you showing me ads about what [16:52] I'm asking about? [16:53] >> Okay, but in the scenario you're just [16:55] describing for Google, wouldn't that [16:57] have the same just like the, you know, [16:58] "Meta is listening to your microphone" [16:59] conspiracy theories, where when the [17:01] targeting is too good, people get [17:02] concerned? Like, wouldn't you get [17:04] similar issues? [17:05] >> a made-up concern. [17:06] >> No, it it it But, it it Sure, but like [17:08] people people have this. And [17:11] wouldn't you have a similar issue where [17:12] if you're using the Gemini data to make [17:16] better ads, [17:17] wouldn't ultimately the targeting will [17:18] be too good and people will find it [17:20] weird. I mean, I think that that's a [17:22] bridge that every tech company would be [17:23] happy to cross if they came to it. [17:25] >> I see. It's a thing that people say when [17:26] you have very good targeting. I think [17:27] there's a lot of like I honestly think a [17:29] lot of the there's a real stated versus [17:32] revealed preference about a lot of this [17:33] stuff. The reality is is people like you [17:37] even the US and they go you could pay [17:38] for Facebook or you could show people [17:40] would rather see the ads. [17:40] >> Yeah. I think most people don't care. In [17:42] this [17:43] a lot of tech and this sort of ties into [17:45] the skepticism of ads. Yeah. [17:47] It's sort of in a wee town there there's [17:50] a wee regulators. Everyone's thinking [17:51] about these very theoretical things. [17:53] Isn't it a bit of the challenge banner [17:55] blindness where [17:57] Instagram advertising works so well [18:00] because it's a picture feed and it's [18:02] showing you pictures. [18:03] >> No, the amazing thing is some of the [18:03] pictures are like commercials. So this [18:05] is where Facebook with an AI app you're [18:06] like looking for an answer and you don't [18:08] want to look at the banner. No, it's a [18:09] huge concern. And this is one of the [18:11] great ironies of meta / Facebook is the [18:15] extent to which the and of course [18:18] Mark and everyone that hates Apple for [18:20] lots of I think very justifiable [18:22] reasons. But Apple saved Facebook from [18:25] itself. [18:26] >> [snorts] [18:26] >> Like back in the day remember Facebook [18:28] platform and there's like Facebook [18:30] payments and all this sort of thing and [18:32] Mark has always wanted to build a [18:34] platform. And if you're just an app on a [18:37] phone [18:38] you can't build a platform. [18:40] And the problem is that I think being an [18:43] advertising based model is generally [18:45] incompatible with being a platform. The [18:47] whole point of a platform is you're just [18:49] letting something else shine, something [18:51] else bring to the surface. You're just [18:52] you're you're the support structure for [18:55] somebody to take over. So you get an [18:56] operating system is not about the [18:58] ideally it's the application on top of [19:00] it that that's you're using. When [19:02] Facebook was forced to not be a platform [19:04] but just be an app suddenly [19:07] they could be fully into being an [19:09] advertising thing. And think about a [19:11] Facebook ad. Even back in the day when [19:12] it was a feed ad or a story ad. Mhm. [19:14] Literally your entire device [19:17] is all an ad. [19:19] And somehow it's not a banner that's a [19:21] little thing on the edge. [19:22] >> Yes. They literally have achieved [19:24] permission from users [19:26] >> Yes. to take over your entire device [19:28] Yes. [19:28] >> to show you a full screen ad [19:30] every 5 seconds. It's amazing. And they [19:32] and they were forced into it by Apple. [19:34] Do you Okay, this reminds me of I don't [19:36] want to come back to the AI um dynamics, [19:38] but this reminds me of a view I've had [19:41] this I'm curious for your thoughts on [19:43] which is [19:45] often when tech companies become really [19:47] big they become really big just because [19:49] the core idea works [19:52] better than [19:54] even the founders could have realized. [19:56] And so Meta's a really uh big company [19:59] because they uh have a feed and the feed [20:03] got really big and they like were very [20:04] smart along the way where they bought [20:05] Instagram and they're like incredibly [20:07] targeted good at monetization. [20:09] >> But it just it turns out people spend a [20:10] lot of time and many people, you know, [20:12] the P times Q of that with the feed and [20:14] they monetize it very well and that's [20:15] what got really big. And same with [20:17] Nvidia. It just turns out that the GPU [20:19] market has got really big and they sell [20:21] a lot of GPUs. And so [20:23] maybe founders because they're often [20:26] like high-powered individuals who want [20:27] to have lots of new ideas, they're often [20:29] thinking about the next thing or like [20:31] what the second act or the third act is [20:33] and everyone wants to invent invent an [20:34] AWS, but I'm curious what you would say [20:36] to the idea that just [20:38] generally it's making the core thing [20:40] really big and there's more orders of [20:42] magnitude at the top than you thought. [20:43] Yes, I think that's always the case. And [20:45] I think that [20:46] sometimes people end up making something [20:49] that [20:50] they didn't want to make and they [20:52] continually push back. I think Meta is [20:53] the perfect example. [20:55] I I my impression is [20:58] Mark's not very interested in ads. He's [21:01] had very good people along the way that [21:03] have helped him build these ad products. [21:05] I think Meta has suffered from that [21:07] because he has not been front and center [21:09] fighting [21:11] for actually ads are good. They are [21:13] societal good. They get They are the [21:16] driver of all the consumer surplus that [21:18] tech throws off. Like the president uses [21:21] the same search engine as the guy on the [21:23] street or the same AI or sort of [21:25] whatever it might be. That's because of [21:26] ads. And the uh Probably not. The [21:29] president probably uses like a Palantir [21:30] search engine or something. Yeah, it's [21:32] probably worse. Uh Google [21:33] Google has slept a lot there. [21:35] There's so much junk online. But the um [21:39] I don't know I don't know if Donald [21:40] Trump ever searched. I don't know. [21:41] That's a good question. [21:42] >> [laughter] [21:42] >> But uh but he's not made that case. And [21:44] I think Meta has suffered because of the [21:47] failure to make that case. And then you [21:49] get things like we're going to do the [21:50] metaverse. We're going to do XYZ. It's [21:52] always coming back we'll be a platform, [21:54] be a platform. And Meta is an [21:55] entertainment company. I wrote this like [21:57] years ago about the like I think it was [22:00] this actually [22:02] It was simultaneously a good call and a [22:04] bad call. [22:05] Do you remember that Paul Krugman quote? [22:07] Uh the internet's not going to be very [22:09] big. Or the fact machine cuz people [22:11] don't have much interesting to say. [22:13] I actually defend that quote. [22:15] And because it's actually true. [22:18] Most people actually don't have that [22:20] much interesting to say. And I brought [22:22] up that quote around 2015 by saying this [22:24] is a fundamental limiter on Meta's [22:27] long-term potential. [22:29] As long as they think of themselves as a [22:31] social media company, [22:33] they're going to run into a problem with [22:35] their feeds becoming insufficiently [22:38] interesting over time. [22:40] Now, uh the move from kind of peer [22:42] content to [22:44] um Well, so that was that was a if I [22:46] might say so myself, a very brilliant [22:47] insight. [22:48] The bad insight was my prescription, [22:50] which was they needed to do more with [22:52] professional content makers, like more [22:54] funding of like the BuzzFeeds of the [22:55] world and share revenue, all that. Cuz [22:57] actually it was user-generated content [22:58] that was [22:59] >> answer is what TikTok did, which is [23:02] TikTok is not a social network at all. [23:04] It is a harvesting and YouTube you know [23:07] the same sort of idea. It's like [23:08] personalized TV. Yeah. What actually [23:10] matters, and this is a key thing, people [23:12] get hung up on relative numbers. What [23:14] matters is absolute numbers. [23:16] So, it is better to have 0.1% of your [23:19] content is good [23:21] if your content is like in the billions [23:23] or trillions as opposed to oh 10% of [23:26] your content is good, but you only have [23:27] 100 pieces of content. That's actually [23:29] worse even if you have a better hit [23:30] rate. And so, [23:32] spurring lots of creation [23:34] writing the algorithms to to capture the [23:36] good stuff, put up there. That actually [23:38] solves the Pac-Man and fax machine [23:40] problem. And [23:42] Facebook was blindsided by that. They [23:43] were so stuck on their identity of being [23:46] a social network that they let TikTok [23:49] take this huge chunk. [23:51] It was their blind spot. Speaking of [23:53] TikTok [23:54] I feel like you don't write about [23:55] ByteDance that much and I'm curious just [23:57] what your thoughts are on ByteDance from [23:59] here and the TikTok sale and everything. [24:01] I mean what a what a what a mess. Uh [24:04] I had to make a decision a long time [24:05] ago. I I wrote about Chinese companies [24:07] more previously. And I think there's [24:09] uh [24:11] it's too Number one, I have to decide [24:13] what what I'm going to be able to cover [24:14] and what I'm not. Uh I'm not in China. I [24:17] was in I was in Taiwan. It is a [24:18] different internet. And there was too [24:22] much sort of uncertainty and unknowns [24:24] just in general about a lot of Chinese [24:26] companies. I would write about them [24:28] occasionally in the context of US tech [24:29] companies. So, I think I wrote about [24:31] like WeChat and what it meant for the [24:33] iPhone's relative competitive position [24:34] in China, how it's different than sort [24:36] of other countries. I think that sort of [24:37] held up pretty well. [24:39] Wrote about TikTok in the context I mean [24:40] more about TikTok I think in the context [24:43] particularly in the context of Meta. Um [24:44] TikTok came up around the same time as [24:46] the was was Quibi. [24:49] Which Quibi was the example of that. [24:50] Quibi was like it was actually right [24:52] that there was room for a mobile [24:54] entertainment product. It was totally [24:56] wrong about the content acquisition [24:57] strategy. So, even if they had a hit [24:58] rate was higher, their total volume was [25:00] way too small. [25:01] >> I follow them, but not super closely. [25:04] It's just a hard market to understand, [25:06] and the uh But, like TikTok's very [25:09] relevant to the US market. So, TikTok uh [25:11] TikTok. I So, I wrote [25:14] the TikTok war. [25:16] Basically making the case that the [25:18] problem with TikTok, and back then [25:19] everyone was talking about user data. [25:21] >> Mhm. [25:22] Who cares? Like, the the whole user data [25:24] thing People have this view of like the [25:26] East German Stasi, and like folders like [25:30] going through people's data. Like, these [25:31] are like vector databases with like [25:33] numbers that no human can parse. Like, [25:35] they're they're It's really quite [25:36] anodyne. It's just really to target ads. [25:39] And um I was very skeptical about that [25:42] being a forcing function in terms of [25:45] forcing divestiture, or whatever it [25:46] might be. The issue I had was [25:50] the algorithm. [25:51] And I noticed [25:54] I think it was when the Hong Kong [25:55] protests happened. And Daryl Morey, the [25:58] then GM of the Houston Rockets, tweeted [26:01] like free Hong Kong or something like [26:03] that. And there's a huge meltdown, and [26:04] the NBA games got canceled. [26:06] And I noticed that on TikTok, and this [26:08] was from I tested it from Taiwan and via [26:10] VPN from the US. [26:12] Uh if you search for [26:14] every single NBA team, you got NBA [26:15] clips. [26:17] Except for Rockets, you got nothing. Oh, [26:18] that's funny. And I'm like [26:20] >> demonetized. There is there is there is [26:22] there is a thumb on the scale here. And [26:23] I started talking about it then. And my [26:26] I did support the ban of TikTok, or the [26:29] forced sort of divestiture from China. [26:32] Because it seems fairly insane to have a [26:34] primarily information source controlled [26:36] by your chief geopolitical adversary. [26:39] >> Yeah, it's the same way that there's [26:39] rules over TV station ownership, and [26:42] it's like not wildly different. And so, [26:44] I I [26:46] You know, [26:47] everything's a trade-off. Of course, I'm [26:49] pretty well known for being a a stark [26:51] defender of free speech and against [26:53] censorship. And my issue wasn't Tik Tok [26:55] per se, it was [26:57] tech the reality of China is [27:00] the founder of Tik Tok of ByteDance is [27:01] long gone because he got called to the [27:03] carpet for ByteDance showing a little [27:06] too much of what people liked, which is [27:07] mostly like hot girls dancing and being [27:10] insufficiently like showing the right [27:12] things that the party wanted. It's in [27:13] China's like the reality is China [27:16] has [27:17] the price of doing business is they get, [27:20] you know, they're on somewhere on the [27:22] the control structure they can tell you [27:24] what to do. And this just seemed like a [27:25] very foolish thing to tolerate. Uh [27:28] unfortunately, [27:29] um [27:30] the US political process, or [27:32] fortunately, maybe the reality is is the [27:34] US [27:36] process and system is such a mess [27:39] can anyone really truly impact it over [27:41] time? [27:42] Uh the way that it shows up messily is [27:44] we somehow do pass the law banning Tik [27:46] Tok and [27:48] it didn't get banned and now it is sold, [27:51] but China still controls the algorithm. [27:53] So, I think it's a big disaster. It's [27:54] also like what what can I say about it? [27:56] I said my piece. Yeah. Um we ended up in [27:58] the worst possible case, which is we [28:00] violated property rights and we did all [28:02] this stuff that's ridiculous and we [28:04] probably bartered XYZ for ABC and we [28:08] didn't get the most important thing, [28:09] which was control the algorithm. Has [28:10] that not happened as part of the sale? [28:12] No, ByteDance still controls the [28:13] algorithm. I didn't know that. Yeah. Hm. [28:16] Good job by us. That does seem like it [28:18] was the point of the spinout. Well, the [28:20] data was always the [28:21] was always the most Yeah, yeah. uh [28:23] salient political point. And so, when I [28:25] wrote about it, that was my point. It's [28:26] like I don't care about the data. The [28:27] issue is the algorithm. [28:28] >> Yes, yes. And um unfortunately, that did [28:30] not get made. I should have written [28:31] about it more, but it the anything like [28:33] all the politics stuff, there was a [28:35] period I mean, thank god for AI. There [28:37] was a period I wrote like when I wrote [28:39] the Garrison theory a couple weeks [28:40] later, I wrote like something about [28:41] regulation. I'm like, this is going to [28:43] drive a bunch of regulatory issues and [28:45] anti-trust things and and all these bits [28:48] and pieces. When that actually happened [28:49] the late end of the last decade, of [28:51] course I was writing about it and I was [28:53] watching like congressional hearings, [28:54] all sort of thing. And that is the [28:56] closest I came to quitting and burning [28:58] out. I think burnout's not a function of [29:00] how much work you're doing, it's doing [29:02] work you don't enjoy. And at one point [29:04] I'm like, all right, either I quit or I [29:06] stop covering congressional hearings. [29:08] So I decided stop covering congressional [29:10] hearings. I only wrote about anti-trust [29:13] stuff because it was super prominent. [29:15] And I've been much happier ever since. [29:17] And maybe that's part of the price of [29:19] just not writing about that is maybe I [29:21] should have pushed on the TikTok thing [29:22] more. [29:24] That's interesting. I said my piece. [29:25] >> Yeah. [29:26] Is Stratechery very widely read in DC? [29:29] It is. Yeah. [29:30] Uh yeah. [29:32] It is. Uh [29:33] Sometimes it's gratifying. It's great [29:36] when you get called and asked for your [29:37] opinion [29:38] or you know, you get certain responses [29:40] or you see impact. It's less gratifying [29:42] when you get yelled at and people are [29:44] are mad at you, but [29:46] fortunately [29:48] the key thing to succeeding on the [29:49] internet is something I have in spades, [29:51] which is a very high level of [29:53] disagreeableness. So you can yell at me [29:55] all you want. I'm [29:56] not going to change my mind. Okay, but [29:58] getting back to [29:59] aggregation theory as it pertains to AI. [30:01] Like a simplistic view you could have is [30:03] that the AI apps are the new [30:05] aggregators. And so a huge amount of [30:08] economic value will accrue to them and [30:10] that's it. [30:12] You could also say that that's too [30:13] simplistic in a bunch of ways because [30:15] one of like we're saying the, you know, [30:17] booking.com you expect it to return you [30:20] hotels that you should book, but you [30:23] expect a little less of a commercial [30:24] incentive from the AI apps. And this is [30:28] like a little more of an abstract [30:29] technology where it's actually not [30:31] trivial to insert all of the, you know, [30:33] commercial incentives in the right way. [30:35] And you could you come up with various [30:37] objections. And so Well, I think the ad [30:39] model is is is is probably the way to [30:41] start. [30:41] >> Mhm. Which is I just talked about before [30:43] sort of the lean in versus lean back. [30:46] Ads are very tied into human psychology [30:48] and like what you're sort of tapping [30:49] into and people's response to that and [30:52] how do you make something creative and [30:53] in the short term [30:54] you know [30:55] technology often makes old business [30:57] models even more powerful [31:00] before it kills them. Right? So you have [31:02] something like [31:04] you know [31:05] suddenly you're a newspaper I used to be [31:07] limited to my geographic area. [31:09] >> Right. Now I can reach the whole world. [31:10] >> Yes. Oh wait a few years later everyone [31:12] can reach the whole world. [31:14] >> Yes. I mean pure competition I'm [31:15] screwed. And [31:17] that is certainly a concern about this [31:20] model. If you get to a world of say [31:22] agentic commerce and the agents are just [31:25] buying the right thing. Like and this is [31:27] I think this is also something that has [31:28] driven a lot of tech skepticism of ads. [31:30] People in tech tend to be [31:33] fairly nerdy, fairly obsessed. [31:36] Uh they're doing a ton of research to [31:37] find sort of the exact right thing. [31:39] >> Yes. Why would I need to tell me what to [31:41] buy when I've you know researched it for [31:42] 2 hours? That's right. Yeah. But and so [31:45] like ads have no effect on me. Well what [31:48] if that sort of obsessive [31:52] deep dive approach is now trivially [31:54] available to everyone because AI is the [31:55] one actually doing it. [31:57] Now where where do ads sort of function? [32:00] And I think this is definitely a bit of [32:02] a be careful what you wish for [32:04] scenario. [32:05] >> Mhm. Because [32:08] what this entails is [32:11] of course more transparency [32:14] more details, more understanding sounds [32:16] good. What it actually entails is sort [32:19] of perfect competition. Which is a very [32:21] sort of brutal game that can just wipe [32:22] out entire categories. That's you know [32:25] that's basically what happened in [32:26] newspapers in many respects. So that's [32:28] that's number one. Number two in a world [32:31] of this sort of world you're sort of by [32:35] definition [32:37] anchoring on whatever specifications or [32:41] whatever can be measured can be put [32:43] down. And you had the old sort of Steve [32:45] Jobs adage about feeds and speeds versus [32:47] like the feel of something where the [32:49] intersection of leverage and what the [32:51] what the [ __ ] does that mean? And it's [32:52] just like, well, [32:54] uh what it actually means is there's [32:55] things that can't be measured and that [32:57] don't go on an Excel spreadsheet. And [32:59] everyone you talk to [33:02] acknowledges this. They say, yes, [33:03] there's things that can't be measured. [33:04] And the way it actually plays out in [33:06] practice is only the things that get [33:07] measured sort of matter. I think this [33:09] has happened to a this this is a huge [33:11] problem sports analytics is a great [33:12] example of this. Where there I think [33:15] basketball is my favorite sport. There's [33:16] a lot that goes into basketball and [33:18] winning that is somewhat hard to wrap [33:22] your hands around. It's not like [33:22] baseball where it's very measurable. [33:24] They baseball is very measurable. [33:26] I do think there's aspects about [33:28] clutchiness and stuff that I'm I'm a [33:30] that I don't know that are properly [33:32] measured. But around basketball for sure [33:34] there's the interaction, way teams play [33:36] together, and how your effort on [33:38] or your involvement on offense can [33:40] affect defense or sort of back and [33:41] forth. And and you see again and again [33:43] like to take my, you know, I like Daryl [33:45] Morey. Um I think there's a reason his [33:47] teams haven't won. They've [33:50] over optimized at the expense of some of [33:52] these these other issues. [33:54] And [33:57] if you can't measure them, [33:58] they tend to get devalued. And in a [34:00] world of AI-mediated everything, how [34:03] many things that can't get measured [34:06] fall by the wayside because [34:10] you know, we end up with it's very [34:11] utilitarian sort of goods that have no [34:14] soul to them. [34:15] Um [34:16] sort of a silly sort of thing to worry [34:18] about in some respects. It's like or it [34:19] sounds silly. But [34:22] I'm a human [34:23] and I anticipate liking and preferring [34:27] the humanity of [34:29] things of all sorts in the long run. [34:32] But you could say this, um [34:37] like [34:39] e-commerce aggregators, like Amazon and [34:41] lots of others, have led to [34:44] you know, fairly anonymous manufacturers [34:47] of lots of everyday goods, uh the kind [34:50] of Amazon Basics type stuff, at a much [34:53] lower price point than [34:55] they, you know, were at previously, at [34:58] still perfectly good quality. And isn't [35:00] that isn't that fine? Like [35:02] >> it brings up the base level [35:06] for everyone. Like your basic consumer, [35:09] the access of items they have to the [35:11] like everyone There's no soul in an [35:12] Amazon Basics power adapter, and that's [35:14] fine. Everyone thinks back to like, "Oh, [35:15] my washing machine was so much better in [35:17] the 1960s." And it's like, yes, that's [35:19] true, and also far fewer people had [35:20] washing machines. And so, I'm now like [35:24] making the opposite argument sort of [35:25] against myself. [35:26] >> in your everyone person [35:28] >> [laughter] [35:28] >> says. Let's switch back and forth. You [35:29] can actually you can change sides of the [35:30] booth. Um you mentioned agent commerce. [35:33] Where you know, we obviously are our [35:35] biggest initiative and had our [35:36] announcement with OpenAI, um [35:38] uh back in October. Where do you think [35:40] that goes? Like how do you see agent [35:42] commerce playing out? I mean, the the [35:44] contrast between uh your and OpenAI's [35:46] announcement and sort of Google's [35:47] announcement, I think is is pretty [35:49] interesting and speaks to [35:52] what the companies are driving for. [35:53] OpenAI wants to be the place you do [35:55] everything. They want to be like the [35:56] aggregator. I think a critic would say [35:58] people compare them to Netscape. I think [36:00] the better analogy, if you're an OpenAI [36:02] skeptic, would be AOL. [36:04] Uh where they want to be the sort of [36:06] like the interface for everything that [36:08] you might do, and it would goes through [36:10] their channels. And Google, [36:12] just as they were relative to AOL, is [36:14] like, "Actually, we want to we want to [36:18] make equip everyone, knowing that if [36:21] everyone is capable, we are the greatest [36:23] beneficiaries because we still marshal [36:26] sort of the the sort of front end demand [36:28] sort of in that regard. Now, how does [36:30] that actually manifest in terms of [36:31] commerce? The funny thing about tech is [36:34] I don't think it will manifest in terms [36:36] of airplane tickets, which is everyone's [36:38] example. [36:38] >> [laughter] [36:39] >> Like everyone can never think of a [36:40] better example than that. [36:42] But what what is the AI going to buy? [36:46] What is it going to get? [36:47] I don't know. I think I would like to [36:49] think people will want to have agency in [36:52] their buying decisions. [36:54] But then again, I you know, we have like [36:56] assistants whether it be like, you know, [36:58] for work or whatever it might be and [37:00] they make buying decisions that we're [37:02] necessarily not involved in and that I [37:04] think is a good predictor precursor of [37:07] what people will ideally like do I [37:08] really need to know? Actually, I have [37:10] very strong paper towel ideas. That was [37:11] going to be my That was going to be my [37:13] But once I once I had set [37:15] can that be sort of monitored and done? [37:17] Um so, I don't know. I think this is a [37:18] very unsatisfying answer other than to [37:20] say [37:21] it has big implications on things like [37:24] advertising and on things like like is [37:26] that a going to be a viable business [37:27] model going forward? [37:29] Uh what margins are going to be [37:30] available? Is there going to be perfect [37:32] competition? Things along those lines. [37:34] Okay. Let me try [37:37] this on you for agent to commerce and [37:38] I'm curious to have you critique us, [37:40] which is sort of how I see things [37:41] playing out. [37:42] I think [37:44] some skepticism is triggered by people [37:47] pitching a very far end state with a lot [37:51] of agentic autonomy. And so it's like, [37:53] please book me a honeymoon in Japan and [37:56] all the activities or it's like no one [37:58] actually does things that way. Whereas [38:00] actually, you should go from the bottoms [38:02] up in some very basic building blocks [38:04] where um [38:06] step one is just replacing filling out [38:09] web forms. That's an activity that [38:11] sucks. No one likes us. And so imagine [38:13] you find the winter jacket you like and [38:15] you copy the URL into chat GPT and just [38:17] say, please buy this for me. And that's [38:19] a much better experience than going and [38:21] around a site you've never been to [38:23] before. And so there's just the agents [38:25] doing kind of tooling you something on [38:26] your behalf, and everyone can agree with [38:28] that. Maybe it clarifies, there's [38:29] multiple colors, which one do you like? [38:30] But it's just replacing filling out form [38:33] fields. This is by the way one thing [38:34] that I am very [38:37] A lot of people are skeptical of this, [38:38] but I am very optimistic about, which is [38:41] I call it just-in-time UI. Mhm. Exactly. [38:43] The the like [38:44] >> It's a better UI. That's right. Okay. So [38:46] that's like level one is a better UI for [38:49] for kind of doing it and actually [38:50] wanting to know. Okay. Then level two is [38:52] better discovery and search. It is crazy [38:55] that we've gotten this far in e-commerce [38:57] with keyword-based search. Like [38:59] keyword-based search works really well [39:01] when you're buying a book that you know [39:02] the name of. It's like I want to go buy [39:05] this particular title. Um for winter [39:08] jacket, it's like I don't know, I want [39:10] it's like a puffer, like what's it [39:11] called? And so instead, you want to be [39:13] able to say, "I'm looking for a jacket. [39:15] This kind of I'm going to this place. [39:16] It's going to be this cold." You know, I [39:18] like these kinds of things whatever. And [39:20] so step two is just better search and [39:24] the ability to search with parameters [39:26] that like no existing search UI lets you [39:30] specify the temperature of the place [39:31] you're going to actually get, you know, [39:33] a jacket of appropriate warmth. That's [39:35] obviously with a jacket one of the core [39:36] things. And so better search UI is kind [39:38] of level two from our point of view. [39:40] Right, which I think is already sort of [39:41] manifesting. Exactly. You're already [39:43] seeing it. And like in the early usage [39:44] of the kind of ChatGPT buying [39:45] experience, I think that's one of the [39:47] like super cool features. And then level [39:49] three, which we haven't really seen play [39:51] out yet, is this idea again of a [39:53] persistent [39:55] profile of the user that anticipates [39:57] their needs. Exactly. It's like I want [39:59] to be able to just pin things I like as [40:02] I go along. Or maybe, you know, if I can [40:03] share my browser history, or maybe if I [40:05] can just like, you know, share a [40:07] Pinterest board. I'm just like, "These [40:08] are some styles I like. Give me a good [40:10] winter jacket for the cold based on [40:11] that. Here are some photos of me, you [40:13] know, based on this." And so it's Or I [40:15] have an even better I have a even better [40:17] idea. Imagine if you were using chat GPT [40:20] and it's circa October 1st. [40:23] And there's an ad for a great winter [40:24] jacket [40:26] that is perfectly suited to me because [40:28] they've been understanding my interests, [40:30] they understand the context of where I [40:32] am. I'm not searching for winter jackets [40:34] because I don't plan well. I really it's [40:35] going to get cold and then I'm searching [40:37] for winter jacket and I'm [40:38] But what if it could anticipate that and [40:40] show me an ad at the right time when I [40:41] need to see it? [40:42] >> Okay, maybe that's level four is like [40:44] the That's what I mean. Why aren't they [40:45] building [laughter] This is my whole bit [40:46] before. Like this is like this is why [40:48] they're so late. They should be shipping [40:51] that this year. You're only shipping [40:52] that this year if you start your ad [40:54] product two or three years ago. This is [40:56] doable today. This is what meta ads are. [40:58] You need to be on more You need to watch [41:00] more reels. I I I Like just the I bought [41:02] more ski equipment this year that I [41:04] don't need. Just because it just shows [41:06] up I'm like and I'm moving back to [41:07] Wisconsin so I [41:09] I'm buying stuff for the house. [41:11] >> Yes, yes. I get all those ski hangers. I [41:12] bet those would be great. That sounds [41:14] very useful. They're still in the box. I [41:15] haven't actually put them up but Yeah, [41:17] so there's there's a limit to kind of [41:18] with just um [41:21] kind of banner ad type experiences to [41:23] what you can do. Whereas I think the [41:24] search thing is is very powerful. But [41:26] yeah, I'm curious what you think of kind [41:27] of step one, just the very act of [41:29] checking out [41:31] uh and then or level one, the very act [41:33] of checking out, level two, better [41:34] search, and then level three [41:37] defining your own embedding space of [41:39] preferences. No, I I completely agree [41:42] with that approach. I just think you [41:43] underrate the extent to which level [41:45] three has already been built. Actually, [41:47] I one thing that that um Mark Zuckerberg [41:49] said on a couple earnings calls ago that [41:51] I thought was very astute is uh we get [41:54] hung up on technological definitions for [41:56] like what is an agent. [41:58] And he's like, "Actually, the largest [42:00] and most successful agent in the world [42:01] today is Facebook advertising." Which is [42:04] exactly right. Facebook advertising, [42:05] people have it in their head that you go [42:07] and you put in like demographics and [42:09] your targeting and stuff. You know, [42:10] >> a autopilot-ish. Yeah, what you do is [42:12] you go in and you say uh [42:14] this acquiring a customer for this is [42:16] worth $10 to me. I'll spend up to $10 [42:19] and they will deliver you a customer for [42:22] $10. [42:22] >> Yes. They'll their margin will actually [42:24] increase because they'll make sure they [42:26] deliver at exactly $10. They can do it [42:27] for more. And they actually make more [42:29] money, you get exactly what you asked [42:31] for. Uh it's I think the extent to how [42:34] powerful this already works. They're [42:36] just stuck on the 50% of my ads work, [42:38] some don't. I don't know which ones. [42:39] >> Yes, yes. Uh Facebook [42:41] they all work. [42:41] >> Yes. [42:42] I feel like a bunch of new very big [42:44] successful companies will be created in [42:47] AI-powered e-commerce. It just feels [42:50] like a different enough product space. [42:52] But what what you're talking about [42:54] retailers, merchants, or or agents like [42:56] that? [42:57] >> about discovery and um kind of the [42:59] demand side. They're also probably [43:01] retailers. Yeah. Well, I mean I [43:02] certainly think I think that the the [43:04] part that I think would be new, which is [43:06] a bit which you were maybe talking [43:08] about, is this real anticipatory aspect, [43:11] which is right now what is amazing [43:14] about, to go back to meta ads, is it [43:17] helps merchants who have a very [43:18] specialized product [43:20] find customers that they never would [43:22] have found otherwise. But there's the [43:24] inverse of I need a very specialized [43:26] product, how do I find what it is, which [43:28] I think you were referring to before. [43:30] But to what extent can that not just be [43:31] an in-the-moment I need this specific I [43:34] remember I needed a [43:36] uh a server, a piece for my a rack to [43:39] mount my this router cuz I had like an [43:41] extra I didn't want to buy a whole new [43:43] thing or whatever. I had this extra [43:44] router. And of course there's some guy [43:45] in Australia that does 3D prints that [43:47] perfectly matches this on Etsy or [43:49] something. And it was great. I found [43:51] this random guy. I'm sure he made a [43:52] bunch of money selling me a $40 piece [43:54] that cost him $2 to make. Good for him. [43:55] >> Yeah. Yeah. Uh but [43:58] what if the an AI should be capable of [44:01] anticipating that need. So it's not I [44:03] Oh, I have a need, let me go find it. [44:05] It's like [44:06] I know you're going to need this. [44:08] >> Yes. And let me acquire it and that [44:10] would be very powerful. [44:11] >> Mhm. [44:15] In his excellent newsletter, [44:16] Stratechery, Ben Thompson argues that [44:17] whoever controls the customer [44:18] relationship shapes the entire [44:20] ecosystem. And in mobile apps, there's [44:22] always been this interesting tension [44:23] here with in-app purchases where [44:25] historically app developers had an [44:26] intermediary relationship with their [44:28] customer through app store policies. [44:31] Over the past year, that layer has [44:33] started to open up. Mobile developers [44:35] can choose what they use for in-app [44:36] payments. Developers now have more [44:38] freedom, but with that freedom comes new [44:41] challenges. [44:42] App store payments were previously [44:43] handling a whole bunch of different [44:44] tasks like payments, tax, fraud, [44:46] disputes, all bundled together. [44:49] Stripe managed payments is built for [44:50] this new world, handling all that [44:52] operational complexity for you with [44:54] Stripe at the merchant of record. [44:56] And with our new app-to-web flow, [44:58] customers can check out in seconds in [45:00] experience that feels native. It [45:02] converts like in-app payments, but it [45:04] runs on Stripe. [45:08] The public markets [45:10] indicate [45:12] as of January 2026 that [45:14] SaaS is cancelled. [45:16] Are they right? [45:19] I think it's probably a mix. [45:21] Uh the [45:23] I think the [45:25] One of the brilliance of American [45:27] business is [45:30] Actually, this is one of my theories [45:32] about why the Europeans are so gung-ho [45:33] about data privacy and regulation [45:36] is because they so often interact with [45:39] European companies. [45:40] >> [laughter] [45:40] >> So, like I was in Paris Paris a a couple [45:43] years ago or a couple years ago. And of [45:45] course, going on a tourist trip, going [45:47] to going to the Louvre, going to the [45:48] Museum of Modern Art or whatever it is [45:49] like or not whatever it was, [45:51] seeing a bunch of museums. They all have [45:53] their own home-grown registration [45:54] systems. [45:56] And they're collecting so much data. And [45:57] they're widely insecure. [45:58] >> it's like, yeah, what's your age? What's [46:00] your pet? What color is like it's like [46:02] why do you need to know all this [46:04] information? They're all non-standard [46:06] forms. They need This is where you need [46:08] AI to fill all this sort of thing in. [46:10] And it's like what Like, there's like [46:13] this theoretical idea in their head if [46:15] we capture this data could be useful. [46:17] So, they built these homegrown things in [46:18] like the 2000s that they are horribly [46:21] insecure. And I use them. I'm like, [46:24] "Where's the regulator? Can someone put [46:26] This is ridiculous." So, I get I get the [46:27] mindset. [46:29] US companies don't do that. Like, US [46:31] companies are so good I think one of the [46:33] big strengths of of US business culture [46:36] is understanding and I think about this [46:38] personally. This is where I give life [46:39] advice. What's the number one mistake [46:41] people make when they're young in [46:43] particular? [46:44] They focus on their weaknesses. They're [46:46] like, "I have to ameliorate my [46:47] weakness." Like, no, what you do is you [46:49] double down on your strength, you get [46:52] richly rewarded for that, and then you [46:53] hire someone to take care of your [46:55] weaknesses. Right? Like, I'm I'm a big [46:56] believer in the Getting Things Done [46:58] system. Like, great book Getting Things [47:00] Done. Even if you don't use the system, [47:01] the book is really good. Lots of great [47:03] insights. There's this whole thing like [47:04] tickler files and like all this sort of [47:06] thing. It's amazing system. I'm [47:08] completely incapable of managing the [47:10] system on my own. [47:11] So, there's a Mac app called OmniFocus [47:13] that is completely built around the [47:14] system that [47:16] I don't have a license for. My assistant [47:18] has a license for. And I text him stuff [47:20] and his job is to maintain my Getting [47:23] Things Done file because I can't do it. [47:24] What I What do I do? Actually, my wife [47:27] is very very optimized around I write [47:30] three pieces a week, I do an interview, [47:31] and I do three podcasts. And all my [47:33] focus and energy needs to be on that. [47:36] And if I do that, that will make a lot [47:38] of money and I can pay to fix all my [47:39] problems sort of elsewhere. And I think [47:41] American business does this very well. [47:43] They don't waste time and energy on [47:45] stuff they're not good at. They double [47:47] down on what they're good at and they're [47:49] focused on the upside, not on their cost [47:52] centers. Probably result of the very [47:53] large market in the US. I I so. And just [47:55] the the competition It would be in a [47:57] very large uh common market. So, you you [48:00] like you you don't have your you go back [48:02] to like newspapers. They have lots of [48:03] home-grown stuff. Like with your [48:04] publication online, if you're if you're [48:06] like me on the internet, [48:08] I get paid to comment on the big tech [48:10] companies. It's probably the most [48:11] competitive market on Earth, right? Like [48:13] lots of people have takes on the big [48:14] tech companies. Uh and so, [48:16] you have to be super focused. Given [48:19] that, [48:20] that speaks to the enduring value of [48:22] just paying someone to manage these [48:25] these business functions from a software [48:26] perspective. Now, there's a lot of like [48:29] there's a [48:30] there's a lot of SaaS applications. Not [48:32] sure they're all sort of strictly [48:33] necessary and and worth the price. Uh I [48:36] like to think [48:37] people talk [48:38] tech having a big five. I would say [48:39] there's a big six. [48:41] The six was Silicon Valley Inc., which [48:43] is basically this cookie-cutter VC goes [48:46] to this founder addressing this specific [48:48] business case with a SaaS business [48:50] model. Everyone likes to they get to [48:51] talk about changing the world. It's [48:53] actually the most predictable thing yet. [48:54] That's why VC returns compressed, but [48:56] because they're also very predictable in [48:58] terms of like the sort of engine going. [49:00] A big problem there is they're all [49:01] seat-based. Like anyone seat-based that [49:03] has some of a vestigial that just [49:05] there's going to be probably fewer [49:06] seats. [49:07] Uh and then if the replacement is more [49:10] small-scale, [49:12] uh [49:13] ideally, there's lots of you know, the [49:16] internet in general has [49:18] writing or content is a good example. [49:21] There used to be sort of you want to be [49:23] in the big pond. And everyone in the big [49:25] pond sort of ate. If you if you had a [49:26] job at Conde Nast at one of their [49:27] magazines, like you lived life well. If [49:29] you wrote like for for a magazine uh uh [49:31] uh [49:31] Today, if you want to be a writer, I [49:33] give advice to people that that want to [49:35] be content producers all the time. [49:36] >> Mhm. And I'm like, "Look, you don't want [49:37] to be you don't want to be in a pond [49:39] with me." [49:40] Right? Uh Bill Simmons is like the like [49:43] first internet sports writer. And you [49:45] don't want to be doing a Bill Simmons [49:46] impression on the internet because he [49:48] got there first. And what did you want [49:51] to do is you want to make your own pot. [49:52] The internet enables the creation of a [49:54] million different ponds. So you would do [49:56] get to define your own pond, be the only [49:58] biggest fish in that pond. That's how [50:00] you succeed. [50:02] To the extent AI makes that I think this [50:04] is the upside case is AI makes that [50:06] possible for more than just content. For [50:08] all sorts of businesses to be lots of [50:10] smaller scale individual entrepreneurs [50:13] or small teams. All of whom don't really [50:16] fit in the salesforce driven seat based [50:19] model for a lot of these so there might [50:22] be a big return to self-serve. [50:25] You know, the or maybe they'll just roll [50:27] their own because their their needs [50:29] aren't that large. [50:31] So that's more of a larger structural [50:32] change. But [50:36] The problem is it's fine to say [50:38] businesses will be okay as they are. [50:41] If you're eliminating the growth, that's [50:43] the big problem. I think that's the [50:44] biggest issue for all the compression. [50:46] >> By head count growth? [50:47] Just growth in general. Yeah, like [50:50] >> What's what's if if these are just [50:51] stable businesses with astronomical [50:54] stock based compensation that is [50:57] predicated on we're going to be very [50:58] large. [50:59] >> Yes. I can see two critiques you might [51:00] have of [51:02] kind of the software space and why [51:04] everything's traded down. One is where [51:07] you know, everyone's just going to use [51:08] cloud code to rebuild their own version [51:10] in house and so as software mode is less [51:13] and the second is actually just that [51:15] many of these products price on a per [51:17] seat basis and so if you're growing head [51:19] count less [51:20] Or shrinking. [51:21] >> Exactly. Yeah. On the first like [51:24] Anthropic just installed Workday. [51:26] >> Yeah, I know. [51:27] >> [laughter] [51:27] >> Famously. [51:27] >> Exactly. So I don't think I don't think [51:30] you know, we're cloud coding [51:31] >> Systems of record are are are is that [51:34] that's the category that is definitely [51:35] not safe. [51:35] >> seem to be you know, we see this with [51:37] Tribe building as well. Like I don't [51:38] think anyone's cloud coding one of those [51:40] systems of record anytime soon. They'll [51:42] use Workday. I don't know what to make [51:43] of the second criticism, but again it [51:45] just feels like for a very broad and [51:48] deep system of records, it's kind of [51:50] hard to make the [51:51] argument that the business is somehow [51:53] impaired versus a year or two ago. [51:54] >> Right, but but I think but I think that [51:56] that's my point though is [51:58] people saying they're going to zero are [52:00] wrong. [52:01] But if the assumption is you're fine, [52:04] Mhm. but you're not going to be growing [52:07] indefinitely. Like that shift from [52:10] thought of as being a growth company to [52:12] being a stable I see. That's a haircut [52:16] and again it's combined with these whole [52:17] compensation structures Yeah, you're now [52:19] valued on EPS rather than revenue or [52:21] something. So, yeah. Um can we talk [52:24] about your business and Substack? Sure. [52:27] So, you were very early to the I mean [52:31] the sovereign writer [52:32] um [52:33] concept. Uh I think you were one of the [52:35] first premium newsletters. Uh I think [52:38] so. Yeah. [52:38] >> Well, well, so [52:40] there's two predecessors to talk about. [52:42] One is just on Wall Street in general. [52:44] There's a long history of Yeah. facted [52:47] out newsletters and things like like [52:49] Grant's Interest Rate Observer might be. [52:51] Uh the difference there is those were [52:52] very expensive and a very small [52:54] addressable market. [52:55] >> Yep. So, uh the difference for Substack [52:58] is it's much cheaper and the market's [53:00] much larger. [53:01] >> Yep. Uh the other person that deserves a [53:04] call out, which I think was the first [53:05] person to do it before me, otherwise I [53:08] think I was the first, was Andrew [53:10] Sullivan who did uh I didn't realize he [53:13] had a paid newsletter. He did for like a [53:14] year. Huh. [53:16] The problem is he did it all wrong. [53:18] You're doing it wrong. [53:19] Uh he was like he would turn out like 50 [53:22] posts a day, right? Just about a [53:24] gazillion different things. He totally [53:25] burned out and like all all all all sort [53:27] of stuff. But um that was happened to be [53:29] a great fit for the advertising model [53:30] back in the day cuz you would always go [53:32] back there and just always be new stuff [53:33] and I'm sure you had a gazillion I'm [53:35] sure you drove a gazillion impressions [53:36] for The Atlantic especially when he was [53:38] with them. He went independent. He was [53:40] pretty successful. I think he had me did [53:41] like around a million dollars or [53:43] something like that. But it was a this [53:44] very leaky paywall. It was like after [53:46] like 35 posts then you'll hit like a [53:48] paywall. And there's a bit where you're [53:49] like you're punishing your worst users. [53:51] It's very easy to get around. [53:53] Um and so but he was actually very [53:55] inspirational in how I thought about the [53:57] model. [53:58] In that he was hailed as a failure [54:00] because he like burnt out and quit. But [54:02] like he made a million dollars, right? [54:03] Like that's just pretty good. [54:05] Uh I wanted from the and just think [54:07] about the psychology of this. [54:10] When I started Substack I had a [54:11] gazillion ideas of things to write [54:14] about. And I limited myself to writing a [54:16] max of two times a week. [54:19] And the reason is I had the subscription [54:21] model in mind. [54:22] And when I added the model I didn't want [54:25] it to be I'm taking stuff away and now [54:27] you have to pay. [54:29] I'm like you like this so much if you [54:31] pay you can get more. And so I always [54:33] wanted to be you're paying to get more [54:35] sort of sort of aspect. And I think [54:38] that probably mattered more at the [54:39] beginning especially because the model [54:40] was new. [54:42] You know, I had I had my metric I looked [54:45] at was people who visited Substack on [54:47] days I didn't post. [54:49] Cuz they were people that were going [54:50] there hoping I had posted that day and [54:52] they were leaving disappointed. And so [54:54] in this case usually previously a [54:56] paywall would disappoint people that [54:57] they hit it. In this case the paywall [54:59] would [55:00] relieve their disappointment cuz they [55:02] they they could now get what they [55:03] wanted. And so I'm like if I can capture [55:05] X percentage of these visitors it'll be [55:07] very good. Um [55:09] One week goal, one week goal, one month [55:11] goal. Failed to reach all of them. [55:13] What happened was [55:15] I I I actually thought it was good not [55:16] going to work. I was like if I go back [55:17] to like teaching English or something [55:18] like that. [55:19] The but it sort of grew and grew and [55:22] grew and at six months actually hit my [55:24] one year goal which was a thousand [55:26] subscribers, thousand true fans. You [55:27] know, it's a hundred thousand dollar run [55:29] rate. [55:29] And I posted a little [55:31] >> how long to get to a thousand subs? [55:33] >> Six months. Okay. [55:34] And I posted a little note saying, "Hey, [55:36] you know, model works. We you know, my [55:38] goal is a thousand for a year. I've [55:40] already reached it." And [55:42] this is the only step change in [55:44] subscribers we've ever had. In the next [55:45] 24 hours, I got 250 new subscribers. 25% [55:49] increase. [55:50] What they were was I had identified [55:53] those people who wanted to be [55:54] subscribers. [55:56] They just didn't trust that it was going [55:57] to work. I was going to go out of [55:58] business and take their money. And so, [55:59] once they realized I wasn't going [56:02] anywhere, then they all they all signed [56:04] up. And then those people signed So, I [56:05] actually had my my metrics were right, [56:08] but I didn't properly calculate [56:10] the uncertainty of people's fear of [56:12] losing their money. [56:13] Uh I So, I'm very grateful that now [56:16] people decide up for stuff all the time, [56:17] right? Uh you know, I Of course, I'll [56:20] probably, you know, go to my grave being [56:21] most well known for Substackery, but I [56:23] am equally proud of the model and that [56:26] lets people make a living doing this. [56:27] How far do you think this model [56:30] can go? Like again, it [56:33] the the defining characteristic to me [56:34] seems like the unbundling. Like, you [56:37] know, maybe 30 years ago you would have [56:39] been writing for a publication, whereas [56:41] now it's unbundling. It's the direct [56:43] relationship with your subscribers. It's [56:46] direct monetization and generally paid. [56:48] I mean, there might be some ad-supported [56:49] component as well, paid. Obviously, kind [56:51] of Substack has proven that this has [56:53] very broad applicability, but how far do [56:56] you think this goes versus traditional [56:58] media bundling? There I think there's a [57:00] couple interesting angles to this. Uh [57:02] number one, I think people, including [57:05] people in tech, serially underrate how [57:07] large the internet is. [57:09] You know, like the one of the biggest [57:09] pushback I got when I announced like the [57:11] Substackery paid product was from [57:13] VCs. I won't say who. So, I love you, [57:15] Ben. [57:16] This doesn't not going to work on the [57:18] internet. Um [57:19] the uh So, and my bit about ponds [57:22] before, [57:23] I don't know that we've scratched the [57:24] limit of how many ponds [57:26] can be sort of built in the world and [57:28] you can sort of occupy. And and the [57:30] other part of this, the critical piece [57:32] of this, and AI is actually an important [57:34] factor here, is [57:37] the key to the model is your costs. So, [57:40] you need just as technology enables you [57:42] to reach everyone, you need to leverage [57:45] technology to keep your costs very, very [57:47] low. And so, for the first several [57:49] years, for sure, it was just me. So, as [57:51] long as I could feed my family, [57:53] I was I I was fine. And this is This is [57:56] the problem for the traditional [57:58] like media companies. Their cost [58:00] structures were were not internet cost [58:01] structures. They were They were [58:03] predicated on on much higher revenue. [58:06] And there is, you know, And this is it [58:08] it's interesting to think about and talk [58:09] about this, cuz a lot of this it is like [58:11] not really applicable. It's like before. [58:13] I write about ads a lot, and I'm not in [58:14] that business. In this case, I write [58:16] about VC high-scalable companies, but my [58:20] actual business is very sort of boutique [58:22] and small and artisan in that Yeah, [58:25] that's right. In that regard. And this [58:27] is a super important point is managing [58:29] your costs. [58:30] If you manage your costs appropriately, [58:32] then the possibilities But that also [58:34] means there are some things that don't [58:36] work with this model. Like your [58:37] traditional classic investigative [58:39] journalism 6-month sort of piece, it it [58:43] did it's not well supported by this. [58:46] What did support that was the bundle. [58:48] Having lots of different writers in one [58:50] publication all together. [58:53] And the thing I worry about, I wonder [58:56] about, is [58:58] bundles are good for everyone involved, [59:01] and no one wants to be a part of it. [59:04] So, TV's the classic example. Why did we [59:07] have a TV bundle? Because you had like [59:10] So, in I think it started in [59:12] Pennsylvania. So, you have like a [59:13] television station in Philadelphia, then [59:14] you have the Allegheny Mountains. And [59:16] you have a bunch of towns there that [59:18] want to get the signal from [59:19] Philadelphia, but they can't get a good [59:21] signal. So, they band together, they put [59:23] up a big tower to get the signal, they [59:24] run actual cable from that tower to all [59:27] their houses. And all of cable [59:29] television started in small-town rural [59:31] America to get TV from the big cities. [59:34] And Ted Turner comes along and like I [59:35] could just broadcast directly to these [59:37] towers. This would be amazing. And you [59:39] get the model and suddenly every but you [59:41] had a fork geographic forcing function. [59:44] And you ended up with all these [59:45] companies with the best business model [59:47] in the world. They everyone paid them [59:49] whether they watched you or not and they [59:50] made a ton of money. And what happened [59:52] the moment they could do something [59:53] different? [59:55] Well, I mean I could also go directly. I [59:56] could I could stream directly. And [59:58] there's just something about business [59:59] and like cuz it's it's a it's almost [60:01] like you have to be forced into the game [60:04] the optimal game from like a game theory [60:07] perspective. And the moment you can [60:08] desert everyone always deserts even if [60:11] it's the best thing. Yes. And I so I [60:13] wonder so I do wonder can bundles ever [60:14] be optimal? [60:15] >> yeah to your world? Because [60:17] in theory there should be bundles. [60:20] Substack should be a bundle. [60:21] Like you should be able to pay one fee [60:23] and get everything. But they you know, [60:25] they started I think a mistake Substack [60:27] made and I'm a huge Substack fan just to [60:29] be clear. I've I've made this claim [60:30] before and I'm getting mad about it, but [60:31] is they characterize themself as being [60:33] totally writer-friendly. [60:36] And I think that was a mistake cuz it's [60:37] impossible for them to be ultimately [60:39] writer-friendly cuz the most [60:40] writer-friendly setup is running on open [60:42] source software like on your own server. [60:43] Like then no one can do anything to you. [60:45] And by definition [60:47] >> going to say like the most [60:48] writer-friendly thing is to have like a [60:49] humming consumer business. [60:51] >> It would be, but the problem is so they [60:52] they all their initial terms what [60:55] Mhm. [60:56] did made the bundles impossible. And all [60:59] the individual publishers owning their [61:00] own subscribers, having their own Stripe [61:01] account, and all these sorts of bits and [61:02] pieces. [61:03] >> on, isn't there like a I feel like [61:04] there's a well-trodden path in tech here [61:06] where OpenTable started as like entirely [61:08] it started as on-prem, purely software [61:10] for restaurants, and then they added the [61:12] customer discovery layer on top of it. [61:14] Shopify started as a solution just for [61:16] merchants, and then they added the Shop [61:18] Pay kind of network effect layer for [61:20] consumers on top of it. Can they what [61:22] what actually [61:24] >> even with those businesses though? Like [61:25] the Shop Pay bit is nice. It's not the [61:27] driver of the business. It's a pretty [61:29] cool part of the business these days [61:31] >> I like it, but it's but at the end of [61:32] the day, the vast majority of shop [61:34] interactions are I see an ad on [61:36] Instagram and I go there and I get shop [61:38] the shop button, which is incredible one [61:40] of the greatest things. [61:41] >> no, but yeah, but then it makes the [61:42] Shopify offering to merchants so much [61:44] more compelling cuz you get the Shop Pay [61:45] network. And then you're just from one [61:47] of these [61:47] >> I'm skeptical that's the driver of the [61:48] business. I think it's a nice to have. I [61:50] think the but But can't Substack just [61:52] add a Substack Prime bundle on top of it [61:55] and merchants can or writers can choose [61:57] to opt in? [61:58] >> yeah, the problem is the merchants who [62:00] will drive make that bundle valuable [62:02] have no incentive to join the bundle cuz [62:04] they could make more just monetizing [62:06] users directly. So imagine I'm on [62:09] Substack. How much more revenue does [62:12] Substack have to give me for me to trade [62:15] $15 a month from my subscribers for a [62:18] smaller amount from [62:21] or whoever's part of this this So the [62:23] problem is they have to really pay me [62:24] off to be a part of it. [62:26] Meanwhile, everyone who doesn't have any [62:28] subscribers, of course they'd love to be [62:29] in the bundle. So you So this is why the [62:31] geographic forcing is tough. [62:33] >> to me. I I think it's solvable at the [62:36] beginning. No, sorry. It feels solvable [62:38] now. [62:39] >> think a counterexample is something like [62:41] Spotify. Spotify is arguably the best [62:43] bundle on the internet. But the reason [62:46] why they were able to assemble the [62:48] bundle is because they only needed to [62:49] negotiate with like four four entities. [62:52] And so it's interesting because on one [62:54] hand that limits Spotify's upside [62:57] because those entities are able to [63:00] negotiate such a large share of [63:02] Spotify's revenue. On the other hand, [63:04] that's also why Spotify was possible cuz [63:07] they only needed to negotiate with four. [63:08] Yeah, yeah, yeah. [63:09] >> And if you're trying to to like get [63:10] every artist on Earth, [63:13] well you Well, of course I got to get [63:14] Taylor Swift. Okay, good luck with that. [63:16] Oh, all the small fry will sign up. Like [63:19] and the music's unique in particular [63:20] because you music the moment a song [63:22] comes out is now part of the back [63:24] catalog. And actually people only ever [63:26] listen to back catalogs. [63:27] So that it's a pretty unique industry in [63:29] that regard. But that is a bundle that [63:30] formed [63:31] but I think it's cuz there's only four [63:33] players. Mhm. Okay. [63:35] Um [63:36] how do you use AI in writing Strictly [63:38] these days? [63:41] I think it probably replaces what I used [63:43] to do a lot of on it's a much more [63:45] efficient cooling. Like so I really [63:49] the most gratifying articles I write is [63:52] when I write about a topic that I [63:54] usually don't and then someone from that [63:56] industry is like wow I did that's that [63:58] was good [64:00] cuz you're you're always worried [64:01] >> Yeah, this imposter syndrome. [64:03] I mean fortunately I don't really have [64:06] That's why it all works. No what's [64:07] what's the what's the mechanism where [64:09] you're reading something about your area [64:12] >> Yeah, and then you realize Yeah, that's [64:13] right. It's like this is totally wrong [64:15] and then you trust everything else. I am [64:17] yeah, I don't want to trigger [64:18] government's [64:19] uh ambition amongst anyone. Yeah. So I [64:22] would if people ask me like I hate the [64:24] book question like what books do you [64:25] read? [64:26] I read a lot of books but they're very [64:27] targeted. Like I will I'm a very very [64:29] fast reader. So sometimes I'll write an [64:31] article I know there's a pertinent book [64:33] and I will just read the whole book in [64:34] the morning and like that [64:36] really contacts [64:37] >> in there. Um [64:38] and but but in general I really want to [64:40] make sure I fully understand a space [64:43] particularly if it's new that I'm [64:43] writing about. This is partly though why [64:45] I have a big competitive advantage. I've [64:47] been thinking about tech since I was in [64:49] junior high school and I've been writing [64:51] about tech for 13 years. [64:53] >> Yes. So I've already done so much [64:55] preparatory work that anyone starting [64:57] from scratch it's like [64:58] it's hard. But so so I want to dive into [65:00] it. [65:01] I'm one of the world's greatest Googlers [65:03] I'd like to think I know every sort of [65:05] parameter and how to find and it [65:07] and so I think I can say pretty [65:09] authoritatively Google has gotten worse. [65:11] And I don't think it's Google's fault. I [65:12] just think there's just it's harder What [65:14] is Google's fault is they got so biased [65:16] towards recency. [65:18] And so you have to be super like [65:19] diligent but [65:21] AI is incredible for this. Like just [65:23] sort of getting background, making sure [65:26] you understand an issue, the ins and [65:28] outs of it, [65:29] how things work. You can query stuff, [65:31] dive deeper. So that is by far my number [65:33] one use case. I do not always but I will [65:36] sometimes ask it to This is where I like [65:38] ChatGPT. [clears throat] This is like I [65:40] type in BB edit as an integration. This [65:42] is also like I very annoyed by and I'm [65:45] very sensitive to the coy nature. Oh, [65:47] this is really great. But no, that's not [65:49] what I'm asking for. I can watch it [65:50] actually go in and find stuff. [65:53] So I do not use it to actually generate [65:57] any any exact content but I can say [66:00] targeted research and then critique. [66:02] Yes. Those would be the the two biggest [66:04] use cases. [66:05] Um you've written a lot about the TSMC [66:09] break. This idea that the limiting [66:13] factor on all AI expansion is basically [66:19] the rate of TSMC capacity expansion [66:21] because almost AI all AI chips are [66:24] fabbed at at TSMC. [66:28] It seems like [66:30] as you look at the AI space and [66:31] everything interesting going on [66:33] so you know if we're [66:35] mostly chip constrained right now which [66:37] would not have to be the case you could [66:38] be power constrained and stuff but if [66:40] you're [66:41] chip constrained [66:43] there's a population of people who want [66:46] to expand very quickly the AI labs, [66:48] Nvidia, people like that and then yeah [66:50] famously which you know TSMC would just [66:52] more conservative in how it expands. [66:55] Why is that? Like why why does the [66:58] market signal not cause them to build [67:00] out fab capacity faster? Because the [67:02] risks for fabs are basically larger than [67:05] for anyone else. [67:07] You're you're spending billions and [67:09] billions of dollars on a fab that if [67:12] it's not fully utilized, if you end up [67:15] with too much capacity, number one, [67:17] those all those all your costs are [67:19] locked in. Like basically 99.9% of the [67:22] cost for a fab is depreciation. Which [67:24] you're paying the depreciation no matter [67:25] like or you already sure you're paying [67:27] it in cash flow obviously, but it's on [67:28] your accounting statement no matter [67:28] what. Uh so the fabs can be extremely [67:33] profitable. TSMC's margins are higher [67:35] than ever, but they can very quickly tip [67:37] over into [67:39] having a huge problem. And then once [67:41] it's already built, these fabs can run [67:43] for a long time. So you that excess [67:46] capacity is a [67:48] depresses prices for years to go. I [67:51] mean, we see this in memory all the [67:52] time. Memory famously goes through these [67:54] cycles. Like what's going to happen? [67:56] We're going to Believe it or not, we are [67:57] going to have too much memory capacity [67:59] in a few years because we have such a [68:01] shortage right now. Like Micron just [68:03] announced like their a huge new fab in [68:05] Singapore, like right? And everyone's [68:07] going to do that. And which is And so [68:09] but why is this happening memory? [68:11] There's three competitors in memory. If [68:13] Micron doesn't do it, SK Hynix will. If [68:15] SK Hynix doesn't do it, Samsung will. [68:17] And so [68:19] you have a dynamic where [68:21] a healthy dynamic, which is the fabs [68:25] know better, but they can't help [68:27] themselves. And so they take on the risk [68:30] and they build these fabs. The problem [68:32] we have with logic is that TSMC doesn't [68:35] have that pressure. And so they're [68:37] actually behaving rationally. TSMC is [68:40] giving up potential long-term revenue. [68:43] But the downside for a fab in particular [68:46] is so large [68:48] that they don't want to realize that [68:50] downside. Can they not pass the risk [68:52] onto the customer where it's like, "You [68:53] are going to pay for the entire fab?" [68:56] That's probably what they need to get [68:57] to. And so Apple famously [69:00] did a lot of this uh of prepaid and [69:03] particularly when TSMC was sort of [69:05] expanding hugely in the 2010s and they [69:08] maybe need to get even more explicit [69:10] about that. [69:12] But I think the better solution and the [69:15] cheaper solution for the hyperscalers in [69:18] the long run [69:19] would be to do what is necessary for [69:22] TSMC to get better. Then you get it for [69:23] free. Yeah. You don't need to prepay it. [69:26] So there's this risk that's out there. [69:27] There's risk of overbuilding. [69:29] >> Mhm. Right now TSMC is shifting all that [69:33] risk [69:34] to the hyperscalers, to Nvidia, to [69:36] Apple. [69:37] And the way it map and the reason why [69:39] they get away with it is because the [69:41] risk is foregone revenue. Yeah. It's [69:44] money you don't make. And worse than [69:45] that, it's money you don't make four or [69:47] five years down the road. [69:48] And everyone like what does every [69:51] company say on their earnings call right [69:52] now? We could have made more, but we [69:54] don't have enough supply. Yeah. And if [69:56] you think it's bad Why is it bad right [69:58] now? ChatGPT comes out. Every [70:00] hyperscaler starts investing like crazy. [70:02] What does TSMC do? They actually [70:04] decreased their capex year over year. [70:06] Two years in a row. They did They There [70:08] was no market response from TSMC to the [70:11] ChatGPT moment. Now they increased to 41 [70:14] last year. They're going up to like 60 [70:16] this year. But even that increase to 60 [70:17] is a less percentage increase than than [70:19] last year. I think we're looking at a [70:22] massive short shortness in chips in 2029 [70:26] or so. And particularly as the other [70:28] thing the compute density of AI is so [70:31] much larger, right? If you have an agent [70:33] out doing stuff, it's doing so many more [70:36] computations in a limited amount of time [70:38] than [70:39] me and my Googling is even humanly [70:42] possible to do. And all these lookups [70:44] and so we have a CPU shortage, too. And [70:46] Intel Intel [70:48] shut down some of their CPU plants, [70:50] right? [70:51] So it's the whole semiconductor [70:54] the [70:55] I just think it's a it's it's a big [70:56] problem and we're shifting to for a long [70:59] time like how can we get an alternative [71:01] to TSMC? For geopolitical reasons. [71:04] And the truth is it's kind of like the [71:05] bundling thing. It's really hard to get [71:07] companies to buy insurance particularly [71:10] when the insurance is [71:12] number one you have a like everyone else [71:14] wants somebody else to do it, right? [71:15] Who's going to be the one to go go and [71:16] make the sacrifice? But also it might [71:19] not happen. China might not attack [71:20] Taiwan. [71:21] And also [71:23] as long as it doesn't happen, it's super [71:26] sub-optimal to go somewhere else cuz [71:27] TSMC is better, right? And their [71:28] customers it's not just their their fabs [71:30] better, their customer service is better [71:31] and they have all the IP blocks you need [71:32] and they've done this before and you [71:34] have existing relationship and they'll [71:35] punish you cuz they have limited they [71:37] have control cuz they like they're [71:40] they're not going to fulfill all their [71:41] orders right now because there's so much [71:44] demand. And so they can pick and choose [71:46] sort of who who and so people are [71:47] scared. They don't want to go anywhere [71:48] else. [71:49] And [71:50] and the uh [71:52] So how are we going to solve this [71:53] problem? And I think [71:55] I actually wrote on the front page of [71:57] Stratechery about this this week which [71:59] is basically the same thing I wrote in [72:00] an update but this was a like [72:02] this somebody needs to wait yeah [72:04] somebody needs to the hyperscalers in [72:06] particular need to appreciate I think a [72:08] massive crunch is coming and it's now on [72:12] them to get Intel up to speed, to get [72:14] Samsung up to speed, to get a credible [72:16] alternative. Yes, in theory you could [72:19] pay TSMC [72:20] >> reasons or for shortages? [72:21] >> No, we'll get the geopolitical reasons [72:22] for free. Okay. [72:23] >> I think there's massive economic reasons [72:25] to do so which is all the revenue you're [72:28] going to be foregoing in 2029 if you [72:30] don't do it now. And then we'll happily [72:32] get geopolitical [72:34] insurance for free. But if TSMC are the [72:36] best, like isn't it rather than like [72:38] stand up Intel which seems hard, isn't [72:40] the answer to just again pre-pay for an [72:42] extra fab build-out for [72:44] >> this is like how do we feel in tech [72:45] about ongoing operational costs as [72:47] opposed to putting in some money up [72:48] front and fixing the problem [72:49] permanently. The the market structure is [72:52] a problem. You're dealing with a [72:53] monopolist. [72:54] And like not a like a mean monopolist [72:57] per se. [72:57] >> the That's the [72:58] >> [laughter] [72:58] >> problem. [72:59] >> right? And they actually have not [73:01] arguably not raised prices nearly as [73:03] much as they should have. But the The [73:05] reality is there's a market structure [73:07] problem that is going to impact the [73:09] hyperscalers, and it behooves them, I [73:12] think, to fix the structure. [73:14] Otherwise, the cost of [73:16] insuring the or overcoming that are just [73:19] going to be be larger and larger. This [73:21] seems like the topic you have felt [73:22] strongest about in the past year or two. [73:25] I felt pretty strong with Apple Vision [73:26] Pro. Okay, fair. What was your take with [73:28] Apple Vision Pro? They finally showed an [73:29] NBA game, and they kept changing [73:31] cameras. They're applying 2D television [73:35] production techniques to an immersive [73:36] technology. Let me sit courtside. I I [73:38] the TSMC break seems like a bigger deal. [73:40] Um Oh, probably. I have some rapid-fire [73:43] questions, or I'm not going to say [73:45] rapid-fire necessarily, but um [73:47] uh more uh a collection of disconnected [73:51] uh questions for you. I'll connect them. [73:52] That's That's what I do. [73:53] >> Great. Um [73:55] how should schools do homework now that [73:56] AI exists? [73:59] I think they should [74:01] incorporate it, and they should probably [74:03] do in-person exams, like uh or in-person [74:06] like the [74:09] Yeah, I mean, you [74:10] it's silly to try to crush it out. I'm [74:12] very opposed to these like AI detectors, [74:14] which they don't work. It's I mean, [74:16] probably I'm particularly sensitive to [74:17] it because obviously a lot of my prose [74:19] is in these models. [74:21] >> Um No one like My thing is that was uh I [74:23] wasn't an M dash user, but um I'm a [74:26] world's biggest semicolon user. [74:27] >> Yeah. Uh which I was a world's [74:29] >> M dash user all along. [74:30] >> Oh, fortunately, the models don't seem [74:31] to have really incorporated the the [74:32] semicolons. So, I haven't been that [74:33] influential. But um But yeah, no, you [74:36] you want kids to use it. Yeah, yeah. Cuz [74:38] whoever can use AI most effectively in [74:40] their jobs going forward is going to [74:41] have a big advantage. [74:43] Um So, there's probably some return to [74:47] you know, more [74:48] in class being more important. I think [74:50] the This might This is my view on [74:52] content generally. This is I just [74:54] Uh [74:55] I think there's a world in which not all [74:58] content, but some content is more [75:00] valuable than ever. [75:02] Because AI is a perfectly individualized [75:03] experience. What you read is not [75:05] necessarily what I read. So, stuff that [75:07] we both read [75:08] is actually compelling. And I'm very [75:11] interested in [75:13] figure out how to leverage that to be [75:15] sort of beneficial to people in the long [75:17] run. And what can you get from school [75:19] that you can't get elsewhere, right? [75:21] Like I can read the notes. I can read [75:23] XYZ. But there's being in class, having [75:26] a discussion about it, like actually [75:28] interacting, being pushed on these sorts [75:29] of things. [75:31] All this is a sort of a beautiful [75:32] theoretical depiction of what school [75:34] might be that is probably very far [75:36] removed from the reality. But [75:39] identifying [75:41] things that are common experiences are [75:44] going to be more and more valuable. [75:45] Yeah. Common content, common classroom [75:48] time, live events, like shared [75:51] experiences, [75:52] because anything that's individualized [75:54] is just going to be completely [75:56] swallowed. Yes. Do sports teams become [76:00] more valuable in an AI abundance future? [76:02] >> becomes more valuable. That's what I'm [76:03] thinking a lot about in this far as my [76:04] business, right? Like the [76:07] There's some aspects of [76:08] tens of thousands of people reading the [76:09] same thing every day that is actually [76:11] really powerful. There's something [76:12] interesting there. The possibility of [76:14] doing live events where people can come [76:16] together. [76:17] Uh I think a lot about community. [76:20] I think no one's ever really solved [76:21] community around content. Like a message [76:23] board or comments are not You actually [76:25] get very bad dynamics. There's a few [76:26] people that dominate it. [76:27] >> Totally, yeah. Uh but what is great is [76:30] if we're in a group chat and you share [76:31] an interesting article and you have a [76:32] discussion about that. [76:34] Uh so, there there's a lot of stuff [76:35] around that I think is really [76:36] interesting and that I'm thinking a lot [76:37] about. [76:38] What do you think of what's going on in [76:39] crypto these days? What's crypto? No. [76:42] [snorts] Uh I've always been a crypto [76:43] defender. Like just because digital [76:46] scarcity is fundamentally interesting. [76:48] >> Mhm. It's probably even more interesting [76:50] to this point in a world of infinite [76:52] content. Which you thought we had [76:53] infinite content before, now we have [76:54] infinite content on steroids. [76:56] Not just the 6 billion humans typing [76:58] away, but agents generating stuff sort [77:01] of constantly. And in that world, I [77:04] think crypto as an identifier of [77:09] authenticity is going to be more and [77:11] more important. Like [77:12] at the end of the day, I want the [77:14] original. [77:15] I don't want a reproduction. [77:17] >> Mhm. And [77:20] I'm optimistic about humans' ability to [77:24] create value where it seemed impossible [77:26] it ever exists. I'm literally a [77:27] professional like podcaster and content [77:29] creator and get paid a lot of money to [77:31] do it. [77:32] Imagine explaining that to someone on [77:33] the farm worried about the the [77:35] automation. [77:36] >> Speaking of that, you mentioned this a [77:39] majority of Stratechery [77:43] consumption is now in the audio form [77:45] rather than the written form. I was [77:46] going to say [77:46] >> as I can tell. I don't do like but the [77:48] more than well more than half of my [77:49] subscribers are subscribed to the audio [77:51] form. [77:52] >> It's quite interesting. I added the pod [77:54] This is actually where I started [77:55] building my own software. I was begging [77:57] everyone to support paid podcasts. There [78:01] was dedicated paid podcasts and there [78:02] was like written ones and no one would [78:04] do it. So of course I had to just hire [78:07] engineers and build it myself. [78:09] At which point it obviously was the [78:11] right thing to do and now everyone does [78:12] it. Whatever, that's my fate in life, I [78:14] guess. But the [78:16] yeah, people love it. Like the [78:17] interesting thing is I'm not sure it's [78:19] been good for my business. [78:20] Why? [78:21] Oh, cuz people don't share. [78:22] >> news is is I think it drives retention. [78:25] Cuz people would build up emails cuz [78:27] like it feels like a lot of work and [78:29] then you say I haven't read this in [78:30] ages, I guess I'll unsubscribe. Whereas [78:31] you know, they just consume it 7 minutes [78:33] or 8 minutes. The problem is they They [78:35] share. Yeah. Audio content is not [78:36] shared. [78:37] >> Totally. I listen to in the car on the [78:38] way home from work and uh that's great [78:41] and then I [78:43] Never think about it ever again. [78:44] Exactly. [laughter] [78:45] But it's great for me cuz I can write [78:46] the same I can say the same thing the [78:47] next day. You're like, "Oh, that was a [78:49] very insightful comment." You didn't [78:50] even know I said it yesterday. Yeah. We [78:52] if we reason about what sectors are [78:55] going to be important [78:56] down the road, [78:58] you know, for the AI buildout, [79:01] energy is um [79:03] is going to be a big a big deal and the [79:05] ability to actually power the data [79:06] centers that are coming online. Uh that [79:08] may be a bigger constraint going forward [79:09] than even chips. Uh robotics are clearly [79:12] going to be a big thing. [79:14] It seems like China is doing better on [79:16] energy and better on robotics and is [79:20] catching up on chips, [79:22] doing okay on, you know, the AI models, [79:24] but [79:25] does that mean China's potentially very [79:26] well positioned for the coming wave of [79:29] tech trends? I mean, I think any country [79:31] that is capable of actually building [79:33] things is well positioned. But then [79:35] again, the counter argument of I could [79:38] support sort of put a silver lining on [79:40] it is [79:42] the challenge, the trick going forward [79:44] and to sort of defy the doomers, as it [79:47] were, is actually creating new sorts of [79:50] value, new sources of value in a way [79:53] that humans are uniquely capable of. And [79:55] that is by definition a sort of an [79:57] innovation story. It's a freeing up [80:00] resources from things that can be done [80:01] by machines [80:03] >> Mhm. [80:04] to more productive It's It's having a [80:05] consumer market that pulls out that sort [80:08] of innovation that makes it possible [80:10] that to write a newsletter or podcast [80:13] and actually pay for it. And so there is [80:16] a scenario where China [80:18] is well positioned to win [80:21] the total commodification of everything, [80:25] which [80:26] doesn't have much margin. Mhm. And the [80:29] actual value creation and what makes [80:32] humans humans and [80:35] generates the value that I think people [80:37] in AI are skeptical can be created. [80:40] Despite the fact 90% of us use working [80:42] agriculture and like 1% do. [80:44] Uh for some reason that's not going to [80:45] repeat. Uh [80:48] if you want to be optimistic, that's the [80:50] sort of thing that America has always [80:51] done well. Mhm. [80:53] What's your Stripe feedback for us? [80:56] Oh, where to start? Uh [80:58] I mean I am obviously it's hard for me [81:00] to drive on Stripe cuz I'm not biased [81:01] cuz I was very early. I think you you [81:03] introduced the billing API in 2011, [81:07] which was a direct spur for wanting to [81:09] Stripe and thinking this is a business [81:10] model that's possible. So, very very big [81:12] thumbs up on that. [81:14] The [81:16] You didn't warn me about this. I should [81:17] have thought about this. [81:18] >> [snorts] [81:18] >> Uh actually you do have you have one [81:20] huge issue that I was just dealing with. [81:22] Oh yeah, ACH. Mhm. Your ACH [81:24] implementation is a is [81:26] if someone can go in and if I try to add [81:28] on to an ACH plan, so let's say I have a [81:30] team, cuz that's where you use ACH with [81:32] like large companies, and they want to [81:34] add someone on, [81:35] if that add-on fails, the entire plan [81:38] gets canceled. [81:39] >> Mhm. So, we have to build a bunch of [81:40] logic to handle that independently. [81:42] Okay. So, [81:43] >> a very detailed specific problem that [81:45] we're facing. [81:46] >> Buggy ACH subscription interactions. [81:48] Okay, that's a good one. Uh there's I [81:50] there's definitely more. I'd have to go [81:52] back and think about it. But I mean I do [81:54] think the [81:55] you know, we didn't talk about [81:55] stablecoins in this sort of area of of [81:58] I've always been a big skeptic of like [81:59] microtransactions. [82:00] >> Mhm. [82:01] Because the problem it goes to the [82:02] investigative reporting thing. [82:04] You can't build something sustainable if [82:07] you're only monetizing on the back end. [82:09] And the only way to do that is to have a [82:11] very large market, which this is what [82:12] YouTube is. YouTube is a bunch of [82:14] speculative video makers hoping that [82:16] they'll get enough views that the ads [82:18] will pay for it. And they're such a [82:20] large scale and they monetize their ads [82:21] so effectively that it works. There's no [82:23] market like this for written content or [82:25] like podcast content. And you can't [82:28] People are like, "Oh, will you pay for [82:29] one article?" I'm like, "No, what you're [82:30] paying for when you [82:32] when you pay for me is you're paying for [82:34] my ongoing production. I'm making a [82:35] promise to you I'm going to write [82:36] something every day, and you're paying [82:38] for that promise. You're not paying for [82:40] the actual content. The content is is a [82:42] byproduct of that. [82:44] The question for AI and microtransaction [82:47] is [82:49] you have all these labs paying people [82:51] all over the world to generate data. And [82:54] we're like, you know, we see here you're [82:55] like a radiologist, you can pay $300 an [82:57] hour. That's what I saw some article [82:57] about it. And there are all there are [82:58] all these sliding scales. [83:00] And they're all duplicating work. This [83:01] is because they're all you know, [83:03] everyone feels this incentive I can get [83:04] differentiation. What we clearly need is [83:07] some sort of market mechanism [83:10] for [83:11] data generation that in the long run [83:14] will replace what we're getting from [83:16] journalistic enterprises, which are even [83:18] more doomed than ever before. So, how do [83:20] you generate uh like where you're paying [83:22] directly for content, and then AIs can't [83:25] get it and can [83:27] you can build a large market like [83:28] YouTube. [83:29] >> Yeah. That people will speculatively do [83:31] it trusting that they'll get paid [83:32] because the market is large enough. [83:33] That's what needs to happen. Like a lot [83:36] of things, there's this massive valley [83:38] of how do we get from here to there? [83:40] Um [83:41] but we'll see. I know Cloudflare is [83:43] trying to push on that, so we'll see [83:44] what happens. Yeah. Last question, how [83:47] would you rate the execution of the [83:48] major tech companies? [83:50] Like the big five? Yeah, sure. [83:53] Apple [83:55] traditionally very strong. Their [83:57] manufacturing hours remains amazing. [83:58] Like the iPhone Air that just The alarm [84:01] went off and make sure I turn this news [84:02] off. Uh the greatest smartphone ever [84:03] made. Really? Oh, yeah. [84:05] >> even seen one. Oh. Oh. [84:08] It's It's thin. Is the battery life [84:09] good? Good enough. Yeah, that's what I [84:12] thought. It's bad. What's that? It [84:13] sounds like it's bad then. No, it's [84:14] fine. Uh you know, I actually forgot my [84:16] external battery and it's sitting here [84:17] okay now. And I'm back in my Scots after [84:19] wearing jeans cuz it's cold and it [84:20] slides right in. [84:22] It's fan actually I love it. I'm very [84:24] devastated to hear they might not make [84:26] it regularly. [84:29] Obviously Apple's Apple's software has [84:31] gotten pretty rough. Their relationship [84:32] with the I mean the Apple is so [84:34] interesting because [84:35] the reality is when it comes to [84:36] platforms [84:37] you have to build the price of becoming [84:40] a platform is making a great product. So [84:42] Apple gets platforms so they make great [84:44] products and they're terrible stewards [84:45] of the platforms. Microsoft is a great [84:47] platform steward but they can't make [84:50] good products so they never get the [84:51] permission to sort of have big platforms [84:54] which is sort of a tragedy there. [84:56] Um [84:57] but Apple it's um you know it's an old [85:01] company driven by managers not founders [85:03] and maybe they you know [85:06] the AI is Siri got as bad as it was is [85:08] obviously really bad but at the end of [85:10] the day [85:11] we still need devices. They're still [85:12] better than anybody else so they'll [85:14] probably be okay. [85:16] Google [85:18] I've had the hardest time understanding [85:19] Google [snorts] [85:20] in part because I think Google does a [85:22] lot of stuff [85:24] suboptimally. [85:26] Almost everything I feel like they do [85:27] suboptimally. [85:29] But I think that lack of [85:32] like there's Apple can be super [85:33] optimized. [85:34] But I think it's their lack of [85:36] optimization that actually makes them [85:38] maybe the most resilient of all the tech [85:40] companies because they never they never [85:42] get so exactly doing what they should do [85:45] and they have all this extra fluff and [85:47] doing things and gazillion science [85:49] projects but because their their their [85:51] core business model is so good and [85:53] throws off so much cash they can just [85:55] sort of like be sort of [85:58] very flexible and I've come to [86:01] appreciate that about them. Everything [86:04] that frustrates me in analyzing them [86:06] actually has this hidden benefit of [86:09] resiliency and strength and [86:11] adaptability. And you know they're like [86:13] the the [86:15] the amorph what's this what's the the [86:17] slime that just slow like and if they're [86:19] coming in your direction, like you're [86:21] actually in big It might take them [86:24] a really long time to get there, but [86:25] when they get there, you're doomed. Um [86:27] so, Microsoft Microsoft is always I've [86:29] got a lot of mileage around about [86:31] Microsoft. [86:33] Everyone is just sitting in the SaaS [86:34] era. All these companies are like, "Oh, [86:36] the Microsoft's sucks. We're going to [86:38] make the best-of-breed product." And [86:40] guess what? Uh startups in Silicon [86:41] Valley, they want to buy all the [86:42] best-of-breed products and they have the [86:44] ability to send together. [86:46] Joe in managing the tire shop doesn't [86:49] care about that. He just wants this crap [86:51] to work and to work together. And if [86:54] it's all mediocre, but it kind of works [86:55] together, that's better [86:58] than best-of-breed. And Microsoft is [87:00] just squashing these companies to grow [87:01] and boom, just hit the hit that [87:03] Microsoft wall again and again. Is that [87:04] going to persist in the AI world? It's [87:06] probably tied to the SaaS question sort [87:07] of before in some respects. They their [87:09] distribution and and power there, you [87:11] know, remains sort of substantial. [87:13] Meta's probably in my experience [87:16] been the best execute. There's I mean, [87:18] there's usually see stuff like [87:19] interacting with PR or executives. Like [87:22] they just run such a tight ship. [87:24] Yeah, that that's always been very [87:26] impressive to me. I think the Again, I [87:29] think their ad model is underrated. The [87:31] trick with them is [87:33] keeping engagement. Like that that's [87:35] what makes the whole thing go. They've [87:37] done a decent job of that. Hours spent [87:40] in ChatGPT hours not spent on Instagram [87:43] or not spent and I think that that's an [87:45] underrated area. And I think they're [87:47] kind of betting that [87:49] what? That's all fine and well today, [87:52] but in the long run, this is an [87:53] infrastructure game. We have cash flow [87:55] to fund it and [87:58] OpenAI doesn't. I think OpenAI might be [88:00] a bigger threat to Facebook than Google. [88:01] Something worth considering. But [88:04] uh [88:04] Facebook is obviously clearly spending [88:06] to to meet it. So. [88:08] Now, Amazon. [88:10] Amazon There's a lot of There's a lot of [88:12] fab capacity and power being spent on [88:14] Tranium that one wonders could be better [88:16] spent on on other chips, but we'll see [88:19] what happens. Aren't people happy with [88:20] the Tranium chips? [88:23] The degree to which Amazon optimized [88:26] uh [88:27] cloud computing I think is under [88:28] appreciated. [88:30] When you're operating a commodity [88:32] market, so there's two ways to succeed, [88:34] right? You can have a differentiated [88:35] product where you can charge a high [88:36] margin or you can have a lower cost [88:38] structure in a commodity market where [88:41] the the [88:42] price floor is the market price, but [88:45] your cost structure is lower than your [88:46] competitors. So, that's where you make [88:47] your margin. That was how Microsoft That [88:49] was how Amazon dominated the cloud. They [88:51] their their their cloud was way more [88:54] optimized than anyone else's. The whole [88:56] Nitro architecture and like just the way [88:58] they architected everything doing a lot [89:00] of their own chips, shifting to [89:01] Graviton. I think the thing with [89:02] Graviton their sort of their ARM CPU is [89:05] they could Who's the number one customer [89:08] for Graviton? Amazon itself. And so, [89:11] they can move all their loads to that, [89:13] optimize it, build all the software [89:14] libraries, and then start offering it on [89:16] a cost good basis to others. That's the [89:18] playbook that they're trying to roll [89:19] with Tranium where the number one [89:20] customer Tranium in the long run is [89:22] Amazon, but then they develop all the [89:24] capabilities around it for [89:26] other people for it to be attractive to [89:27] other people at lower prices, and they [89:29] they have that structurally smaller cost [89:32] structure. The problem is [89:35] that works when you sort of leveled off [89:38] in performance, right? Amazon executed [89:41] this model between 2005 and 2025. Of [89:44] course, processors got faster in that [89:45] time, but not it wasn't like the '80s or [89:48] '90s when every leap was massive. Does [89:51] that [89:52] Does that work in a relatively new [89:54] market when there's massive leaps to be [89:56] made generation on generation? And they [89:59] have Nvidia servers. Do they have as [90:01] many as they could? [90:02] >> Yeah. Uh because they're on the strategy [90:06] um [90:07] probably not. Ben, thank you. Thank you.