Model pricing comparison

Meta’s pricing is amazing – it targets eliminate Claude’s gross profit if Muse Spark capability can catch up.

Model Input ($/1M tokens) Output ($/1M tokens)
Meta Muse Spark 1.1 $1.25 $4.25
Claude Sonnet ~$3 ~$15
Claude Opus ~$5 ~$25

With cached input.

Model Input ($/1M) Cached Input ($/1M) Output ($/1M) Cached discount
Meta Muse Spark 1.1 $1.25 $0.15 $4.25 88%
Claude Sonnet $3.00 $0.30 $15.00 90%
Claude Opus $5.00 $0.50 $25.00 90%

However, compare that with DeepSeek.

Model Input / 1M Cached input / 1M Output / 1M
DeepSeek V4 Flash RMB 1 / $0.15 RMB 0.02 / $0.003  RMB 2 / $0.3
DeepSeek V4 Pro  RMB 3 / $0.44 RMB 0.025 / $0.0037 RMB 6 / $0.89

Zhipu could see some trouble…

 

Model Input / 1M Cached input / 1M Output / 1M
Zhipu GLM-5.2 RMB 8$1.18 RMB 2 / $0.3 RMB 28 / $4.13
Meta Muse Spark 1.1 $1.25 $0.15 $4.25
DeepSeek V4 Pro  RMB 3 RMB 0.025 RMB 6

Also recall that I was mentioning the bifurcation in AI models and AI infra.

Which Buffett stock performed well during doc-com bubble?

1/ Washington Post

In 2000, the company spent about $125 million developing Kaplan, online education and Washingtonpost.com.

Kaplan became a success.

Kaplan revenue grew roughly 76% in two years, while it swung from a meaningful loss to profitability. Test preparation benefited as the weak job market sent more people back to graduate school, while Kaplan Higher Education produced record enrollment, revenue and operating income.

 

2/ H&R Block

Core tax business is a recurring, mostly cash-paid service that was largely insensitive to technology spending or stock-market valuations.

H&R Block owned Option One, a large non-prime mortgage originator. As the Federal Reserve cut rates following the dot-com collapse, refinancing activity surged. << an early subprime-housing trade

 

3/ Wells Fargo

Wells Fargo surged in 2000 as money rotated away from expensive technology stocks toward reasonably valued, profitable financial companies. Even after its 20% decline in 2001, an investor entering at the end of 1999 still made roughly 24% through 2002, versus a 38% loss for the S&P 500.

Its underlying banking businesses also remained healthy: Wells Fargo reported growing net interest income and improving earnings in parts of its business during 2002.

The AI death, zoombies, and unsustainables

Some of the notable AI services have been sunset, both in the US and in China.

In March 2026, Sora shutdown was officially announced. The consumer web and app experiences were shut down on April 26, 2026, while the developer API will end on September 24, 2026.

In Sep 2025, 秒鸭相机 (owned by Alibaba) has dissolved the team and news broke out in May 2026.

Baidu’s 文心一言 / Yiyan / Ernie Bot commercialization (membership-based) has not worked; Baidu announced the sunset of paid membership in Feb 2025 and began refund in Apr 2025. The chat function of frontier LLMs in China is basically free for all, but the paid subscription had previously cost about RMB 59.9 per month.

 

Yiyan service itself has been moved/remade/rebranded.

 

There are many more.

e.g. in China

Product Company Type Date/status What happened
Glow MiniMax AI companion March 2023 Removed following complaints involving sensitive content and privacy. It never returned under the Glow name; MiniMax later launched 星野. (21 Economic Net)
Trik AI Xiaohongshu AI image generation Late 2023 Shut after controversy and litigation over artists’ work; Xiaohongshu subsequently launched 奇域AI. (36Kr)
未伴 Tencent Music/QQ Music AI companion and social app June 2024 Removed amid regulatory scrutiny of AI-companion content. Mainland services were discontinued rather than merely renamed. (21 Economic Net)
腾讯翻译君 Tencent AI translation March 13, 2025 Web, app and mini-program all ceased operating. Translation functionality increasingly became part of 元宝 and Tencent’s larger products. (36Kr)
万物追踪 MiniMax AI information monitoring Around April 2025 Entered public testing in January 2025 and was comprehensively delisted less than three months later. (36Kr)
鹿班智能设计 Alibaba Automated advertising/design June 30, 2025 All services ended after Alibaba had already stopped purchases and renewals. (36Kr)
奇域AI Xiaohongshu Chinese-style AI images June 30, 2025 Standalone service closed; some functionality was transferred to Xiaohongshu’s 镜构 mini-program. (36Kr)
腾讯智影 Tencent AI video and digital humans June 30, 2025 Officially described as a platform “renovation,” but login ended, user data was cleared and its official account was later cancelled. This is best treated as an effective indefinite closure. (36Kr)
异世界回响 Soul AI character companion Around September 2025 Reported withdrawn/stopped as Soul reduced its standalone AI-companion experiments. The exact final server date was not clearly published. (36Kr)
狸谱 Independent startup AIGC anime and fan-art community October 31, 2025 Officially ceased operation despite reportedly reaching around one million monthly active users during its peak. (36Kr)
冒泡鸭 StepFun/阶跃星辰 AI role-play companion November 5, 2025 Formally stopped operating after exceeding six million cumulative downloads, according to third-party app data cited by Chinese media. (36Kr)
讯飞写作 iFlytek AI writing November 30, 2025 Officially terminated; users were instructed to export their content before closure. (36Kr)
XEVA / X Eva Xiaoice/小冰 AI clones and companionship November 30, 2025 Ended all operating services. (36Kr)
Wow AI Meituan AI friend and character chat December 18, 2025 Servers closed and users could no longer access their accounts. (36Kr)
纸飞机AI ModelBest/面壁智能 AI companion Reported exited in 2025 Included in the 2025 shutdown census, although no precise public server-closure date was located. (36Kr)
momoAI 清杉信息 AI companion Reported exited in 2025 Listed among discontinued or withdrawn AI-companion products; detailed closure documentation is limited. (36Kr)
我在AI 有零科技 AI companion Reported exited in 2025 Reported as withdrawn or no longer maintained; no clear public final-service notice was found. (36Kr)
AiU 爱哟 Zhipu AI/智谱 AI companion Reported exited in H2 2025 Standalone companion experiment was withdrawn as Zhipu concentrated resources elsewhere. Exact date is unclear. (36Kr)
Lumi 噜米 禹幻科技 AI emotional companion Reported exited in H2 2025 Included in the companion-app withdrawal wave, but its final backend-shutdown date was not publicly documented. (36Kr)
梯子AI Baidu AI search March 4, 2026 Shut after operating for roughly six months; its site and app-store listings disappeared and functions moved into the 文心 app. (Sina Finance)

Some have become “zombies”.

Product Company Condition as of June 30, 2026
胃之书 Independent team Had gone at least five months without an update by late 2025; user-feedback groups were reportedly nonfunctional or muted. No formal shutdown announcement, but it appeared close to natural abandonment. (36Kr)
AskXBOT 澜码科技 The company experienced unpaid wages, staff reductions and severe financing problems. AskXBOT appeared stalled, although the company website remained accessible, so it cannot be called a clean server shutdown. (Sina Finance)
Vega AI Independent team Included in the Chinese-media census of delisted or long-unmaintained AI applications. I found no reliable official closure notice or exact date. (36Kr)
开搜AI 快搜-related team Listed as delisted or long-term unmaintained, but public evidence is insufficient to establish a precise final shutdown date. (36Kr)
晓象 Startup project Reportedly encountered major operating and financing problems. Its product activity appears substantially stalled, but evidence does not support a precise “servers shut on X date” claim. (36Kr)
小星星AI陪练 星星趣弹 Included in the 2025 inactive-app census, but third-party download pages continued to appear. I would label it unconfirmed/inactive, rather than fully closed. (36Kr)

Many can be found in the US.

Product What it did Shutdown date What happened
Huxe Turned emails, calendars, webpages and prompts into personalized AI audio briefings/podcasts May 28, 2026 Removed from the app stores and stopped functioning. The founders said the team was moving on and would discontinue the product. (TechCrunch)
Rewind Mac app that continuously recorded and made users’ screen and audio history searchable December 19, 2025 Following Meta’s acquisition of Limitless, the Rewind app’s screen and audio capture was disabled. Limitless’s Pendant service continued for existing users in supported regions, so only Rewind and other non-Pendant functionality were sunset. (Limitless)
Yara AI AI mental-wellness and therapy-style chatbot By November 2025 The founder voluntarily wound down the service, citing concerns that AI chatbots were not safe enough for people experiencing serious mental-health situations. Its website now carries a farewell notice and directs users to other resources. (Yara AI)
Dot by New Computer Personalized AI friend, confidante and companion October 5, 2025 New Computer sunset Dot after its co-founders decided to pursue different directions. Users were given one month to export their conversations and other data. (TechCrunch)
Woebot consumer app CBT-based mental-health chatbot June 30, 2025 The direct-to-consumer Woebot app was retired; existing accounts became inaccessible. Woebot Health itself did not entirely disappear and continued working on institutional and provider-oriented applications. (Woebot Health)
Builder.ai Claimed to use AI to make developing apps and websites easier Operations collapsed in May 2025 The Microsoft-backed company entered insolvency after financial problems, debt defaults and frozen or seized cash left it unable to continue operating normally. Its main U.S. entity subsequently filed for bankruptcy in June. (Financial Times)
Humane Ai Pin services Screenless wearable assistant for AI queries, calls, messages and cloud-stored content February 28, 2025 Humane sold key assets to HP and switched off its servers. Calling, messaging, AI responses and cloud access stopped working, leaving most Pins substantially unusable. (The Verge)
Coqui AI commercial operations AI voice generation, speech synthesis and voice cloning January 2024 The startup and its hosted commercial services shut down amid funding and monetization difficulties. The underlying open-source TTS code survived through community forks, so the technology itself did not vanish. (GitHub)
Artifact AI-personalized news reader created by Instagram’s co-founders Wound down during 2024 Artifact announced that its market opportunity was insufficient and began winding down its app. Its core news reader remained alive longer than originally planned, before Yahoo acquired its technology and incorporated elements into Yahoo News. (Medium)
OpenAI AI Text Classifier Website that attempted to identify whether text was written by AI July 20, 2023 OpenAI withdrew it because of its low accuracy rate. It was a particularly clean example of a product being shut down because the underlying AI simply did not work reliably enough. (OpenAI)
Neeva consumer search Subscription-based, ad-free search engine that added generative-AI answers and citations June 2, 2023 Neeva could not attract enough users to make the consumer search business sustainable. The search website closed, user data was deleted and the company’s technology and team subsequently moved to Snowflake. (Search Engine Land)

 

Bifurcation in AI models and AI infra

Bifurcation is coming.

SOTA (State-of-the-Art) models will collaborate / use the best hardware, i.e. latest GPU from Nvidia, HBM from SK Hynix, etc.

SOTA model war is between OpenAI and Anthropics, plus Google.

Meanwhile, Microsoft and Meta will be another group.

This group will prioritize TCO (Total Cost of Ownership) and use another group of hardware suppliers, like those offered by Qualcomm, announced in Qualcomm Investor Day.

置身钉内?

一个方向看似正确的 AI 办公产品,如何在组织目标、商业压力、产品基因和真实用户需求的拉扯中走偏?

ONE 想做“让事找人”的 AI 工作入口,但最终被做成了“让组织更容易看见、追踪、管控人”的工具;它失败的关键不在 AI 技术本身,而在产品定位、组织机制和用户立场错了。

核心观点:
// 产品发心过载
ONE 同时想服务用户、证明钉钉 AI 化、承载无招回归后的战略战役、消化商业化和集团指标。目标太多,导致产品无法保持一个清晰主线。

// To B 的根本矛盾没有解开
钉钉的付费方是老板/管理者,真实高频使用者却是员工。老板要“可控、可视、确定性”,员工要“边界、缓冲、自由度”。ONE 试图同时满足两边,最后实际更偏向管理者。

// AI 放大了钉钉原有的强管控基因
钉钉早期成功依赖“已读未读、强触达、组织确定性”。ONE 把这些能力 AI 化后,用户感受到的不是效率提升,而是更细颗粒度的监控。

// 用户真实反馈被误判
内测和大客户灰度阶段,管理层反馈很好,但基层员工强烈抵触。团队一度把这种抵触理解成“不适应新工具”,而不是产品方向本身有问题,这是文章认为最致命的误判。

// “信息收敛”变成了“信息轰炸”
ONE 原本想用卡片减少消息混乱,但各业务线都往 ONE 接入内容,最后变成“原有钉钉消息 + ONE 卡片”的双重轰炸。

// 发布会倒排工期摧毁了敏捷迭代
项目不是从 MVP、小范围验证、真实数据纠偏开始,而是围绕固定发布会日期高举高打。负面数据出现后,也没有真正修正顶层定位。

// 组织结构让问题被持续放大
产品、设计、研发、商业化、各业务线之间权责不清;业务线为了 AI 指标纷纷接入;一线用户反馈向上传递时被淡化,最终形成“上层看到正反馈,一线承受真实反噬”的断层。

// 它也并非毫无价值
文章承认 ONE 的方向踩中了办公 AI 的趋势:主动式信息流、AI 进入工作流、基于组织上下文做服务,这些都可能是未来方向。失败的是落地方式和用户立场。


无独有偶,硅谷也在反复证明同一个结论:当企业试图用数据和 AI 把人的工作过程彻底量化,最后得到的往往不是效率,而是反感、抵触和讽刺性的“表演生产力”。

Microsoft 的 Productivity Score 是一个很典型的例子。它原本想帮企业理解员工如何使用 Microsoft 365、Teams、Outlook 等工具,提高组织协作效率。但因为它能暴露个人层面的使用数据,很快被外界批评为“员工监控工具”。最后微软不得不调整产品,弱化个人级别可识别数据,改成更聚合的组织洞察。这个案例说明了一件事:员工并不是天然反对效率工具,但他们不接受被这样管理。

Amazon 内部统计 AI 使用量的做法也很讽刺。公司想推动员工使用 AI,于是开始看谁用得多、谁 token 消耗高、谁更积极拥抱新工具。但指标一旦被看见,就会被表演。有人开始为了显得自己更 AI-native 而刷 token、制造调用量。最后据报道,Amazon 撤掉了内部 AI 使用排行榜。企业以为自己在衡量生产力,实际衡量到的是员工表演生产力的能力。

Meta 的鼠标和键盘追踪则更直接。据报道,Meta 的内部项目会在员工工作电脑上记录鼠标移动、点击、键盘操作,甚至偶尔截图,用来训练 AI agent 理解真实办公行为。员工反应很强烈:内部帖子的高赞评论是“这让我非常不舒服,怎么退出?”;公告下最常见反应是愤怒表情。Meta 最初明确表示工作电脑无法 opt out,后来在压力下只开放了非常有限的临时暂停机制。它没有真正取消,但已经暴露出一个事实:当公司把员工的工作过程本身变成训练数据,员工首先感受到的不是效率,而是被采集、被观察、被占有。

这些案例和《置身钉内》里的钉钉 ONE 放在一起看,重点就很清楚了:这套逻辑正在失败。

钉钉 ONE 看见了未读、待办、遗漏和事项流转,却没看见员工需要缓冲和边界。

Microsoft 看见了工具使用行为,却撞上了隐私反弹。

Amazon 看见了 AI 使用量,却催生了刷指标。

Meta 想看见鼠标和键盘背后的工作过程,却激起了员工对数据采集的愤怒。

它们共同说明:工作不是越透明越高效,人也不是越可测量越好管理。一旦企业把“过程数据”误当成“真实贡献”,员工就会开始保护自己:要么关闭入口,要么消极抵抗,要么表演指标,要么彻底不信任系统。

所以,钉钉 ONE 的教训并不孤立。它只是更完整地呈现了一个趋势的终局:当 AI 办公产品从“帮人工作”滑向“帮组织看人”,失败几乎是迟早的事。

隐私不是效率的敌人,自由也不是管理的漏洞。很多时候,它们恰恰是一个工具能否被长期信任、长期使用的前提。

How Nvidia stock is like Apple and what may come next

1/ Both companies have a history before becoming the center of the world.

Apple: 1980

Nvidia: 1999

2/ Right before the new era, world is experiencing shocks.

Apple/Great Financial Crisis: iPhone was introduced in 2007; iPhone 3G was introduced in 2008

Nvidia/Russia-Ukraine war: H100 was revealed in Apr 2022

3/ Both become largest company in first few years of key product launch

Apple: 2011, 3 years after iPhone 3G

Nvidia: 2024,  2 years after H100


What comes next for Apple?

  • Samsung competition
  • iPhone 5 disappoint
  • China 2013 315 boycott
  • iPhone shipment peaks
  • innovation capability questioned
  • new business lines (watch) considered disappointment at first

Storms might be coming for Nvidia.

..

How high can ads ARPU go?

Bilibili, a younger Chinese user focused video website generates advertising revenue at about 7.5 rmb per DAU per month, slightly above $1 ARPU per month.

That is less than half of Kuaishou, which is above 16.5 rmb per DAU per month, or about $2.4 ARPU per month.

Kuaishou’s DAU size is near 4x Bilibili, so ads revenue is nearly 8x larger.

But these are nothing compare with Meta, which commands about $30-$40 ARPU per month in North America.

So Kuaishou is only like 1/10.

Douyin is the highest in China.

Using $50bn pure ads in China and 800mn DAU gives about $5 ARPU per month.

On top of that there are e-commerce, local services, payments, live-streaming etc. so overall Douyin app ARPU is a lot higher than pure ads.

Just comparing ads revenue: Douyin DAU 3-3.5x Meta NA but ads ARPU is 15% of Meta, which will give you about 50% of Meta NA ads rev of ~$100bn.

Why CFOs may not be good CEOs

I believe excellent CEOs need a set of unique skillsets – they need to be able to unite people, rally the morale, be bold and innovative in strategy.

Good CFOs are very good at numbers, very analytical when presented a well-defined question. They are also responsible for financing and executing M&As which can be strategical.

I think excellent CFOs can bring CEO-like value when they proactively pursue M&A deals or actively manage acquired companies. However, this is not a base case.

I could be wrong, but more often than not, CFOs don’t really need to be innovative or think outside the box to be okay – they are fine to be pragmatic and consistent, and sometimes expected to be so.

When CFOs become CEOs, there might be an attendance to deliver near-term numbers rather than focus on real value creation.

It could also go wrong if they rely too much M&As – bad deals will cost a fortune and even mediocre deals carry opportunity costs.

Alibaba used to have Daniel Zhang as CFO and he became CEO in 2015. When he stepped down in Sep 2023, the annualized return during his tenure is at best bond-like.

That’s why I don’t feel good about Mixue Group – it announced that its CFO Zhang Yuan will become CEO, effective immediately. Mixue founders will stay as co-chairman though.

Let’s see.

Two values of e-commerce and Meituan

There are two core values provided by “e-commerce” platforms, transaction and discovery.

If you know what you want, search in Taobao, and put an order, Alibaba provides the “transaction” value.

If you just want to order McDonald’s food delivery, Meituan also provides the “transaction” value.

What shopping mall provides and what Taobao used to be famous for is that consumers are just wandering around / 逛 – this is how the “discovery” value can be provided.

“Discovery” is deserved to earn higher, and almost as a cut to final sales – businesses believe they win this incremental order because of you. In “transaction” mode, businesses believe they have already had the order; thus transaction providers can only earn the fulfillment fee.

In the transaction mode, the lowest transaction cost provider will naturally win more busines.

In the discovery mode, the provider that can generate more “new sales” can win more business.

So Tencent mini-program / mini-shop is good for luxury brands as the they just need a place to list. Consumers almost already know these brands. The lower than transaction cost the better.

Douyin, Xiaohongshu are better in discovery mode as users spend a lot of time in their app “wondering”. The apps are more likely than others to create incremental demand – users don’t know they want it until they see it.

Previously, transaction mode and discovery mode are served in the same place like Taobao.

This is also true for Meituan in food delivery or in-store dining. Chinese consumers scroll Meituan app for look for food delivery or dining ideas. Even if they are already in the restaurant or know what to order, they also go to Meituan to make related transaction.

Meituan also provides pre-defined sets for many in-store dining – this is tricky. Is this discovery mode? Consumers don’t know what to eat even they are already in the restaurant.

But I think it’s still more transaction mode. Think about the fixed tasting menus provided at a Michelin restaurant. Meituan doesn’t help create incremental demand here if I browse and purchase the $199 set via the app – the value is transaction or like credit cards.

All I want to say is that if Meituan will be like Taobao, competition will be from two fronts – one is by providing lower transaction cost, another is by stealing the discovery cake.

拭目以待

Where does Alibaba’s bullishness come from?

Alibaba recently outlined a 5-year target to $100bn AI+Cloud revenue, however many investors are either not convinced or skeptical of future monetization.

Why Alibaba has $100bn target in the first place?

Besides the delayed AI boom in China with potential breakthrough in domestic AI chip production, what are the drivers?

1) Amazon CEO Andy Jassy outlined $600bn goal in 10 years recently

$100bn would be about 700bn rmb and matches the AWS “number”.

2) Within Alibaba, there could be potential internal politics consideration

E-commerce is still the most profitable core business; Jiang Fan leads the e-commerce efforts, spending billions in quick commerce etc.

Eddie Wu leads the AI+Cloud and is CEO of Alibaba. This is growing fast but still not as profitable.

These two people might be fighting for resources and territory within Alibaba.

AI+Cloud needs to be bold.

700bn rmb rev with 15% EBITA margin will be ~100bn rmb EBITA.

This could be matching Alibaba’s China e-commerce adj. EBITA in 2026, which dropped from ~200bn though, due to massive subsidy.

For Jiang Fan, quick commerce is a way to paint growth picture for overall china e-commerce business.

Otherwise, Eddie Wu might get more of Alibaba’s incremental spending, which then will almost ensure Wu’s rise and Jiang’s fall.

Jiang was also smart in not picking the fight with PDD, which is probably harder to win. The fight with Meituan is not easy, but if you have to pick one to fight with between PDD and Meituan, Meituan is the choice.

Tencent also has already made the choice – it sold Meituan stake and kept PDD stake.

Fighting with PDD is more defending vs. fighting with Meituan is winning new businesses. KPIs will look more exciting in the new business.