AI isn't a bubble — it's a war where the giants will burn their last dollar rather than leave the table
What people call an AI bubble is an accounting done on commercial returns; the giants treat it as a new kind of national defense or nuclear weapon, and what they're calculating is the risk of being disrupted — so they'd rather burn their last dollar than leave the table.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Not a bubble, but an AI war you can't afford to lose
The trigger was the $1.4 trillion compute commitment Sam floated in October: 30 gigawatts, GPUs depreciated over six years, which works out to OpenAI burning an average of more than $200 billion a year, while its own public forecast for 2030 is only just over $200 billion in annual revenue and $500-600 billion cumulative over the next five years — on a cash-flow view it doesn't add up in the short term. But Guangmi says OpenAI's cash flow will be fine for the next two to three years, and it recently raised another $100 billion at a valuation of over $800 billion. He thinks this isn't a valuation game but an arms race: developer selection has shifted from "AWS or Azure" to "which model," and cloud vendors are being white-labeled; once the Agent becomes the entry point, Super Apps and information intermediaries like Uber and Ctrip all get short-circuited — hail an Uber in the Bay Area and the platform can take 50%. People who can't afford to lose don't leave the table.
— Guang MiYou can see $200 billion clearly, but not ten T
What's clear is three blocks: subscriptions — assume 4 billion MAU, 2.5 billion WAU and 1.5 billion DAU by 2030, a 10% paid rate means 400 million users at $200 a year each, about $80 billion; advertising and commerce — over 2 billion WAU at $20 to $50 per user, about $50 to $100 billion; API — a rough guess of $50 to $100 billion. That adds up to two or three hundred billion dollars a year, and Guangmi's assessment is "it's just one more internet platform, nothing that interesting really." The part you can't see clearly is the key: if Agents replace labor end to end, with 1 billion white-collar workers worldwide at $1,000 to $10,000 per Agent white-collar worker, that's $1 to $10 trillion in revenue. The reference point is the $150 billion software development market, while AI coding this year is only just over $10 billion.
— Guang MiNvidia is Android, Google is Apple
This war splits into two camps: Nvidia's GPUs and Google's TPUs. For the first time Google has integrated models, chips, cloud and products end to end, and Guangmi says it looks more like the Apple of the AI era; Nvidia stands where Android stood in the AI era, with OpenAI and Anthropic below it, and those two are somewhat better on talent density. He thinks the Nvidia-OpenAI line is currently undervalued: GPUs are better than TPUs overall, the drawback is cost, but TSMC's capacity is only so big, the Ruby generation of cards should pull a full generation ahead of TPU V8, and Nvidia has already locked up TSMC's 1.6-nanometer capacity in advance. Another trajectory judgment: the stronger Google gets, the more an anti-Google alliance forms, and the stronger OpenAI gets, the more an anti-OpenAI alliance forms — OpenAI was founded precisely to challenge Google.
— Guang MiAre you doing vertical e-commerce, or Xiaohongshu
Guangmi's borrowed analogy is: foundation model companies look a lot like the general e-commerce players of a decade ago, where scaling SKUs equals scaling data — add data for whichever capability you want to do well, and it's more efficient. Like Amazon in its early years, which used book sales to pull together logistics, warehousing and users, then expanded horizontally into categories, today's model companies are doing PPT, Excel, data analysis and investment research. The conclusion is that foundation models still capture most of the value in the value chain. Applied to startups, it's his question: are you doing vertical e-commerce, or Xiaohongshu? Vertical e-commerce can make money too, but you have to be able to retreat flexibly, and historically vertical e-commerce has never produced a very large platform company; Xiaohongshu built a layer of content assets on top of a general e-commerce platform. Whether Perplexity and Cursor are one or the other, there's no answer yet.
— Guang MiRobots and world models may all be fake problems
This is the judgment Guangmi himself labels a hot take: robots, world models and even multimodality may largely be fake problems, and Online Learning may be the only truly important real problem — only when it breaks through are the other problems solved in essence, because the core is generalization; without generalization you can only walk the old autonomous driving road, where however much human labor you have is however much intelligence you get. The corollary is heavy: future breakthroughs in robots and world models may not necessarily come from the people working on them today, and today's cohort may end up like the previous generation of NLP. The metaphor for the three paradigms is: pretraining data is like oil, fossil fuel, huge in volume but finite, and 70-80% of it is already used; RL's expert data is like new energy, useful but small in total, which is why outsourcing companies that specifically recruit human experts have sprung up; Online Learning is like nuclear fusion, not yet broken through, and if it breaks through it's invincible.
— Guang MiThe agentic web is the real web3
What he most wants to see in 2026 is the next nuclear-grade paradigm breakthrough, such as Online Learning, because that will bring more proactive Agents and change product forms. Another inspiration comes from the Doubao phone, which he calls the agentic web: this is the real web3, the first time startups and new companies have a chance to flip the table, which was very hard in the past few years. The logic is traffic short-circuiting — once the Agent becomes the traffic entry point, today's Super Apps, Google Search, phone makers, and information intermediaries like Uber and Ctrip all get affected; the essence is seizing the power to allocate traffic. He also sees 2026 as a big year for multimodality, with the biggest opportunity for knowledge workers, especially in coding, and phone makers will be an important theme next year, because there may be more than one entry point.
— Guang MiThe more top-tier the company, the cheaper it is, the less bubble
Counting ARR from top to bottom: OpenAI $20-21 billion, Anthropic $9-10 billion (of which Claude Code has publicly passed $1 billion), Cursor over $1 billion, the two data labeling companies Surge AI and Mercor both over $1 billion, Midjourney $700-800 million, serving company Together AI $300 million; after that basically everything is under $300 million — enterprise search Glean and web coding Replit at $200-300 million, voice company Eleven Labs at nearly $300 million, front-end Lovable at over $200 million, the medical ChatGPT Open Evidence at $150 million, Sierra and Harvey between $150 and $200 million. His conclusion is that the more top-tier the company, the cheaper it is, and the more top-tier the company, the less bubble there is; if there really is a bubble it's probably in some of the companies further down.
— Guang MiDon't package yourself as an American company
Three pieces of advice for Chinese founders: first, firmly go after the global market, especially the high-paying US market — the US can contribute 60% of revenue and 70% of profit; second, make good use of China's talent and engineer dividend, don't give up China's industrial advantages, and don't awkwardly package yourself as an American or Singaporean company — if the team is 100% Chinese, building a good product is the way. Third, getting money from top-tier Silicon Valley VCs is actually not easy, trust chains are hard to build, so take whatever money you can get early on. He also mentions a shortcoming: over the past three or four years there have been more than a hundred AI application companies worldwide that grew into unicorns, and he can only think of three to five Chinese teams — a ratio far lower than in the internet era.
— Guang MiIn their own words · checked verbatim
So a lot of people will probably burn their last dollar and still not easily leave the table.
所以很多人可能就会砸光最后一分钱 也不会轻易下牌桌吧
Guang Mi5:04
You see, the stronger Google gets, the more an anti-Google alliance forms; the stronger OpenEye gets, the more an anti-OpenEye alliance forms.
你看Google越强 越会形成一个反Google联盟 OpenEye越强 也会形成一个反OpenEye联盟
Guang Mi17:11
I think there's a hot take here: the robots, world models and even multimodality everyone talks about may actually be mostly fake problems … Online Learning may be the only truly important real problem.
我觉得这里有一个暴论啊 就是大家提的机器人世界模型甚至多谋态 其实可能很多是假问题 … Online Learning 可能才是唯一重要的真问题
Guang Mi39:29
There's a metaphor for online learning: nuclear fusion. It's nuclear energy, not yet broken through, but if it breaks through, it's invincible — humanity enters the trajectory era.
那online learning有一个比喻 就是叫核聚变 这是核能 还没有突破 但如果突破了 就是无敌的 人类就进入轨迹时代了
Guang Mi41:31
If the model's data distribution doesn't contain this kind of data, this kind of task just doesn't work; only after compressing this kind of data does the task work.
如果模型数据分布里面 没有这类数据 这类任务就是不work 只有压缩过这类数据 这个任务才work
Guang Mi42:32
It's called the agentic web. I think this is the real web3. Finally there's this kind of power that gives startups or new companies a chance to flip the table.
就是叫agentic web 我觉得这才是真正的web3 这个终于有了这样一个power 让创业公司或者新公司 有掀桌子的机会
Guang Mi50:39
The more top-tier the company, the cheaper it is; the more top-tier the company, the less bubble there is. Even if there is a bubble, it's probably in some of the companies further down.
越头部的公司越便宜, 越头部的公司越没有bubble, 就算有bubble的可能还是后面的某些公司
Guang Mi1:04:55
Before it might have been a zero-to-five probability, a 0 to 5 percent probability; now I think it's reached a 20 percent probability.
之前可能是零到五的概率 百分之零到五的概率 现在我觉得到了百分之二十的概率
Guang Mi1:16:07
Figures
| OpenAI's user assumptions before 2030 | 4 billion MAU / 2.5 billion WAU / 1.5 billion DAU | 8:06 |
| Gemini vs ChatGPT DAU/MAU | Gemini 10%, ChatGPT 25%; Gemini MAU is about 20%-25% of ChatGPT's | 32:22 |
| AGI Index suggested allocation | 25% OpenAI + 25% ByteDance + 10% Google + 10% Anthropic + 10% Nvidia and TSMC | 49:37 |
| Generalist real-machine data volume | Claims 270,000 entries | 59:48 |
| Cursor annualized revenue | Over $1 billion; data labeling firms Surge AI and Mercor also over $1 billion | 1:04:55 |
| Doubao DAU | Over 100 million | 1:06:58 |
| Probability of a Chinese-background team building the world's leading AI company | Up from 0-5% to 20% | 1:16:07 |
Glossary
- Online Learning
- A model that keeps learning and updating itself during inference and interaction, no longer relying only on static pretraining.
- Agentic Web
- The network form after Agents become the traffic entry point, which Guangmi calls the real web3.
- Neo Labs
- A new batch of small frontier labs in Silicon Valley founded by top researchers who left big companies, such as SSI and Thinking Machines.
- Credit Assignment
- In reinforcement learning, judging how much each step's action contributed to the final result.
How to listen
Investors who care about the model landscape and how to build a portfolio, and founders looking for a niche in the AI application ecosystem; the technical judgments are concentrated after the 40-minute mark.
The opening pleasantries and the quarterly review (roughly the first 3 minutes) can be skipped.