Model companies mine; product companies refine
The best models capture the most traffic and revenue, but the mining window is shrinking: from two years to nine months to three or four months. Foundation model companies are going end-to-end into Agents, and app companies' cost disadvantage is structural.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Model companies are differentiating; only two still do general
Guangmi's most striking impression is that models are differentiating: going forward, maybe only Google Gemini and OpenAI will truly do general models, wanting every capability from multimodal to reasoning models to Coding to Agentic to 2C, and even Google wanting to do world models. Other model companies are differentiating into various domains — Anthropic into Coding and Agentic, Mira's Thinking Machines Lab wanting to do multimodal and next-generation interaction. This differentiation is not forced; it's an active bet.
— Guang MiAnthropic was pushed all-in by positive feedback
Anthropic has almost given up multimodal, given up 2C, and hasn't pushed an independent reasoning model. Guangmi judges the turning point was in the summer of 24 when Claude 3.5 was released; feedback from external developers gave the team a lot of positive feedback, and the whole team went more all-in on Coding and Agentic. He used the term "Reward Hacking" — the management team may also have been Reward Hacked by Coding. The return on this bet: at the end of 23 AR was probably less than $100 million, in 24 probably $950 million, some speculate 25 annualized revenue will exceed $12 billion, and even 26 could be $30-40 billion.
— Guang MiCoding alone can't support a grand narrative, so there's non-consensus here
Guangmi believes Anthropic's core is betting on Coding and Agentic. If Anthropic doesn't focus, and like Google and OpenAI does both multimodal and 2C, a sprinkle-pepper strategy is very dangerous. But conversely, if OpenAI and Google put their whole company all-in on Coding, Anthropic is also in danger. The problem is that when Google and OpenAI's CEOs or management are asked whether to all-in only on Coding as a single capability direction, the answer is No — everyone feels that doing only Coding is not enough to support a grand narrative. So there's still strong non-consensus on Coding.
— Guang MiCoding and world models are actually the same thing
Google DeepMind CEO Dennis described Google's world model Genie 3 as an infinitely extensible virtual womb that incubates AGI. Guangmi says Coding is also a virtual womb; world models and Coding are actually the same thing, just different paths. His argument: the difference between humans and animals is language and the ability to create or use tools; language and Code plus Agentic tool capabilities can let an Agent in the digital world perform all the behaviors a human does on a computer or phone — that's already AGI. Making a front-end website or PPT used to require many rounds of communication with a studio; now it's hundreds to thousands of steps of Agentic reasoning, and finally Coding generates HTML.
— Guang MiThinking Machines is the most expensive angel round in global history
Mira was previously OpenAI's CTO; her Thinking Machines Lab is the most expensive angel round in global history, raising $2 billion at a $10 billion valuation. Guangmi thinks three points are most critical: First, this is the strongest exit team in Silicon Valley in recent years, possibly stronger than today's xAI, more like a spin-off of OpenAI ChatGPT or Post Train plus core Infra team — Meta's Zuckerberg wanted to spend $1.5 billion to poach one person, that Infra Lead was directly rejected, not a single person was poached. Second, Mira is OpenAI's most core management, knows what's done well and what's done poorly. Third, Mira may be globally suited to be Apple's CEO, her product-first philosophy matches Apple.
— Guang MixAI's massive compute hasn't gotten payback
Musk's biggest bet before was a single massive cluster, that compute could bring fundamental model improvements. But Guangmi says currently there's no payback — massive compute can only be said to be the foundation for not falling behind, but it may not let Grok overtake. Grok today is still groping for its ecological niche: doing chat can't beat ChatGPT, doing search there's still Google and perplexity ahead, doing coding there's Anthropic, doing multimodal and world models, it feels like Grok today hasn't thought it through. Guangmi guesses xAI may very likely merge into Tesla as one within the next half year to a year, because an independent AI Lab without a big leg is actually very difficult.
— Guang MiHorizontal everything-bundle and vertical integration happen simultaneously
Guangmi says there are two most obvious trends today. One is the horizontal everything-bundle: ChatGPT, you spend 20 bucks or 200 bucks to buy an account, already includes chat, search, coding, agent, workspace PPT, equal to buying a dozen suites, which is very unfavorable for single vertical capability suites. The other is vertical integration: Google Gemini from TPU chips to Gemini models to agent applications above, plus Google Docs, Chrome browser, Android OS, YouTube video, can do super inheritance. So Google Gemini being a few months behind on model is no problem, especially since it's not even behind now.
— Guang MiThe mining window is shrinking: from two years to nine months to three months
Guangmi had a colleague specifically pull the window periods, and found the window is shrinking. Perplexity's window was two years — model companies themselves started adding web search on average two years late, but today ChatGPT added web, Perplexity is still maintaining high-speed growth, maybe still tens of millions of new users per month. Cursor got hot because of Claude's Sonnet, from Cursor starting to have heat to Claude Code release, actually just nine months. Minus's window was released this year in March-April, ChatGPT followed up with a similar product in about three to four months. From two years, one year, nine months, three months, the window is shrinking.
— Guang MiCursor's long-term competition is very pessimistic
Guangmi says if he were the Cursor team, today having to compete long-term with Claude Code, it's actually very pessimistic. Suppose Cursor can immediately replicate a product effect like Claude Code, then it has to face such a high loss rate — product pricing is 200 a month, but many users sometimes use thousands of dollars of tokens in a few days, even 1 to 3 or even 1 to 10 token loss rate, the cost is unbearable, easily dragged to death. And this is without talking about how to optimize model capabilities later; if Cursor wants to change some capability or data distribution below, the model training capability can't keep up.
— Guang MiBullish on Google again, because of scale effects
Guangmi's big change this time is being bullish on Google again. His observation is that ChatGPT and Google's final product forms will definitely converge, like Xiaohongshu (小红书) which integrates search, short video, and friends circle. Another guess is that ChatGPT will definitely do advertising later, because it recently hired a new commercialization CEO, previously Instagram's CEO, and before that Lead of Meta's ad system. Maybe in the future top users will be charged very expensively, 200 a month, 2000 a month, but paying users may be only 3% to 5%, over 95% of users are free, then it depends on advertising. Google is still the world's best advertising platform.
— Guang MiIn their own words · checked verbatim
Actually I think Coding is also a virtual womb, that world models and Coding are actually the same thing.
其实我觉得Coding也是虚拟子宫 就是世界模型和Coding其实是一件事
Guang Mi8:09
Model companies, they spent tens of billions of dollars on model training, like mining, mining this intelligence ore. Then you find, after I mined the ore, the front-end application Cursor is there making money hand over fist. Damn, model companies aren't stupid either, I'll definitely go forward to harvest.
模型公司 人家花了上百亿美金来做模型训练 就像挖矿一样 挖这个智能的矿 那你发现 我挖了矿之后 前端的应用 Cursor在那里卡卡卡的赚钱 我靠模型公司我也不傻 我肯定要往前去做收割
Guang Mi27:20
From two years, one year, nine months, three months, actually the window is shrinking.
从两年一年 九个月三个月 其实 窗口是在 这个缩短的
Guang Mi40:25
Cursor or coding companies that don't do models will have no advantage in the future. In the future, I think it's all about competing on cost advantage.
Cursor或者coding公司不做模型 未来是没有优势的 未来我觉得就是比拼成本优势
Guang Mi48:40
I think there's a very good summary, called Jewish finance, Chinese AGI.
我觉得有一个总结是非常好的 叫犹太人的金融 华人的AGI
Guang Mi58:49
Figures
| Thinking Machines angel round | $10 billion valuation, raised $2 billion | 10:10 |
| Meta poaching offer | $1.5 billion to poach one person | 10:10 |
| OpenAI public AR | $12 billion | 24:20 |
| Anthropic AR | $5-6 billion | 24:20 |
| Claude Code AR | May soon reach $1 billion, possibly $1.5-2 billion by year-end | 41:27 |
| Cursor public AR | $500 million | 27:20 |
| ChatGPT paying user ratio | 3% to 5% | 52:44 |
Glossary
- Reward Hacking
- Models or teams guided by short-term positive feedback, over-optimizing a certain direction.
- L4 experience
- End-to-end product experience that feels wow, with an aha moment.
- Everything-bundle
- One account includes multiple suites like chat, search, coding, agent, etc.
- Vertical integration
- Owning the entire chain from chips to models to applications.
- DPI
- Distributions to Paid-In Capital: the ratio of a fund's realized returns to invested capital.
How to listen
People watching AI investment and startups, especially founders and investors who want to know how app companies survive once foundation model companies do products.
After 55:53, the part about US financial seigniorage and Chinese entrepreneurship is lower information density.