Summary

Frontier models are diverging; there may soon be only 2 general-purpose players. Guangmi thinks the only companies still building truly general-purpose models may eventually be Google Gemini and OpenAI; Anthropic has nearly abandoned multimodality and to C, going all in on coding and agentic products—ARR rose from under $100M at the end of 2023 to roughly $950M in 2024, with some guessing annualized ARR could exceed $12B in 2025 and reach $30B-$40B in 2026. At a $170B valuation, “I don’t actually think it’s expensive, because there’s still more than 10x growth in a year.” Kimi and DeepSeek are moving toward the Anthropic playbook: “I think that’s very smart.”
The leaders on the consumer side are converging, and ChatGPT’s moat is nontechnical. As the brand moves downmarket, retention “may be second only to WeChat.” Model-training spend converts into brand the way Netflix’s content spend does: “The money you pay for the model is also money you pay for the brand.” His $100 portfolio has changed from the prior quarter to $40 OpenAI, $40 ByteDance, $10 Anthropic, and $10 Google, because “the moat around coding and the moat around the model are nowhere near as high as ChatGPT’s moat.”
The miners are moving downstream to harvest the value. Model companies spend tens of billions mining intelligence, then see “the front-end application Cursor making money hand over fist… I’m not stupid either.” Claude Code launched 8-9 months after Cursor and overtook it in 3-4 months; it could soon reach $1B ARR and command a standalone valuation of $20B-$30B. The mining window has compressed from nearly 2 years for Perplexity to 9 months for Cursor and 3-4 months for Manus, while startups facing token loss rates of 1:3 to 1:10 “will be dragged to death.”
L4-level experience is the product North Star. The 2 best agents today are ChatGPT Deep Research and Claude Code, both delivering end-to-end aha moments and prompting users to upgrade from $20 subscriptions to $200. The next mine is Office/Workspace: PPT is exactly where coding and agentic capabilities matter most, expanding the user base from 30M coders to 500M-600M office workers.
Google is back in favor. Product forms will ultimately converge into all-in-one bundles, and ChatGPT will inevitably run ads—its new commercialization CEO came from Instacart and previously built Meta’s ad system—while “Google is still the best advertising platform in the world.” Google can claw back ground through the scale effects of end-to-end vertical integration, from chips to the OS; the biggest uncertainty is whether it can change user habits.
Founders may be desperate, but they still need to find “today’s Huang Zheng.” The giants “understand it, take it seriously, and can keep up,” so they are difficult to beat head-on. Yet in 2015 many said Taobao was unassailable, while Pinduoduo was collectively passed over in its A and B rounds. In venture investing, “diversification has failed; concentration is what works”—many Silicon Valley VCs missed OpenAI, Anthropic, and Mira’s $10B angel round because gross margins were not high enough.
The bubble’s watchpoint is the grand narrative. “Just watch what the smartest cohort is using”—quant traders and top programmers overwhelmingly use Claude. The biggest fear is that roughly $300B of AI capex this year becomes “doll-within-a-doll revenue” with no value created at the front end. From the Bay Area, America’s 2 core advantages are financial seigniorage and Silicon Valley’s technology lead: “If China breaks through advanced lithography within 2 years, I don’t know whether global capital will flow back to China.” One veteran’s summary: “Jewish finance, Chinese AGI”（犹太人的金融，华人的AGI）.

Deep dive

1. Divergence and Convergence: Only Google and OpenAI May Still Be Building General-Purpose Models

Guangmi’s overall read on the quarter: “It was absolutely not flat; if anything, it felt more explosive, with stormier seas.” The clearest takeaway is model divergence: going forward, perhaps only Google Gemini and OpenAI will build truly general-purpose models—covering multimodality, reasoning, coding, agentic capabilities, and to C. Google even wants to build a world model, while Anthropic is specializing in coding and agentic capabilities and Mira’s Thinking Machines Lab wants to pursue multimodality and next-generation interaction.
The convergent side is that the top 3—GPT, Gemini, and Claude—“can’t shake one another, but it’s not easy for anyone else to break into the top 3.” The paradigm is already clear: there have been only 2 major paradigms over the past 2 years, pre-train scaling and L scaling, both originating at OpenAI. The competition resembles an F1 race: models used to arrive once every 6-12 months, but this year the cadence is once every 2-3 months. “You also have to make sure you don’t make mistakes, or it’s very easy to fall behind.”

2. Anthropic’s Contrarian Bet: Coding as Another “Virtual Womb” for AGI

The inflection point came after Claude 3.5 in summer 2024, when external developers delivered massive positive feedback and the team went further all in on coding and agentic capabilities. Guangmi jokes that “the management team got reward-hacked by coding too.” ARR went from under $100M at the end of 2023 to about $950M in 2024; some guess annualized ARR could exceed $12B in 2025 and reach $30B-$40B in 2026. At the latest round’s $170B valuation, “I don’t actually think it’s expensive.”
The contrarian thesis: Demis describes Google’s world model Genie 3 as an “infinitely scalable virtual womb” for AGI. Guangmi says coding is another virtual womb: “World models and coding are actually the same thing, just different paths.” The difference between humans and animals is language and the ability to create and use tools; code plus agents lets an agent perform everything humans do on computers and phones in the digital world. “I think that’s already AGI.”
The risk is symmetric. Anthropic would be “in serious danger” if it failed to focus and spread itself thin like the 2 giants. But if OpenAI and Google put their entire organizations behind coding, “Anthropic would also be in danger.” Asked whether the 2 companies should focus only on coding, both management teams said no, because coding alone cannot support a grand narrative. That leaves coding as a major area of non-consensus.
The China mapping is straightforward: Kimi moved from wanting to be “China’s OpenAI” to wanting to be “China’s Anthropic.” Guangmi says, “I think that’s very smart.” DeepSeek and Qwen are also no longer following OpenAI on some capabilities, but moving toward Claude’s path.

3. Thinking Machines Lab: The Most Expensive Angel Round in History and a Product-First Split

The round was done at a $10B valuation with $2B raised—the most expensive angel round in history. The first key point is the team: “the strongest startup team Silicon Valley has produced in the past few years,” perhaps even stronger than xAI’s today. It resembles a split of OpenAI’s post-training and core infrastructure teams. Zuckerberg reportedly offered $1.5B to poach its infrastructure head and was rejected outright; “not a single person was poached.” Meta could move neither Thinking Machines nor Anthropic.
As OpenAI’s former CTO, Mira knows what worked, what did not, and what should come next. In chat, ChatGPT has already won: “The lead has escaped beyond Earth.” Mira may therefore differentiate around natively multimodal systems, build a Her-style consumer product, and explore next-generation interaction. She may not build new hardware, because most human context is already on phones; phones should grow stronger, not weaker, in the AI era.
The first 3 labs are “intelligence-first” cultures; Mira may be “product-first,” believing current intelligence is sufficient to build excellent products. One additional thought: Mira may be globally suited to become Apple’s CEO. It will be worth watching whether Apple’s board invites her over in the next 1-2 years. Without a team that truly understands AI, Apple’s management may not be able to plan the next generation of phones.

4. xAI and Meta: Still Searching for Their Ecosystem Slots

Musk’s biggest previous bet—that a single enormous compute cluster would produce a fundamental model leap—“has not paid back” so far. Massive compute is merely the baseline for staying in the race. Grok is still searching for its ecosystem slot: it cannot beat ChatGPT in chat, Google and Perplexity are ahead in search, and Anthropic leads in coding. It has spent “maybe 10x or 20x” the resources of others without delivering 10x the benefits. “To some extent, Doubao is even more usable than Grok; its end-to-end voice experience is better.” xAI may be folded into Tesla over the next 6-12 months. “I’m worried xAI could fall behind,” unless tens of thousands of GPUs and GB200s eventually force a qualitative model breakthrough.
Meta’s Superintelligence team will still struggle to reach the global top 3 in the short term; even Musk’s relentless push has not gotten him there. If intelligence comes first, Meta may still rank behind Grok; if product comes first, its zero-to-one capabilities are unknown. Zuckerberg needs to ensure he remains at the table, and the scale of the poaching campaign reflects the importance of using financial firepower at this moment—other companies cannot afford to play the same game.

5. The Endgame Map and the $100 Portfolio

The lazy analogy is that ChatGPT could be the next Google, while Anthropic could become the next Windows. Coding provides an API and supports the unlimited expansion of applications: globally, only Windows, iOS, Android, and today’s large-model coding capabilities can support that kind of application expansion. Google, meanwhile, should win through scale effects.
A $100 stock portfolio—“the answer for this particular snapshot in time”—would be $40 OpenAI, $40 ByteDance, $10 Anthropic, and $10 Google, a change from the prior quarter. Why is OpenAI weighted so much more heavily than Anthropic? “ChatGPT has many nontechnical moats… the moat around coding and the moat around the model are nowhere near as high as ChatGPT’s moat.”
He also stresses how difficult it is to call the endgame: “It’s hard to say which company will definitely win, and nobody gets to lie flat and win.” He was extremely bullish on Anthropic last quarter, but everyone is now chasing coding. A SOTA model becomes market average in 3-6 months.

6. Consumer Leaders Converge: ChatGPT’s Coca-Cola Moment

The biggest takeaway from the past quarter is that the consumer leaders are visibly converging. ChatGPT may be the fastest product in history to reach 1B active users. In 2025, the brand moved downmarket while growth continued to accelerate; “its slope is steeper than everyone else’s.” Mass-market users no longer compare options. ChatGPT is the default AI tool, just as Google is the default search engine, placed in the most convenient spot on the phone’s first screen.
Model training resembles Netflix producing content: money burned on training becomes product, mindshare, and brand, rather than simply disappearing into user acquisition. ChatGPT has spent almost nothing on acquisition. Retention “may be astonishingly good, perhaps second only to WeChat”; as the model improves, former unsubscribers return. Like carbonated drinks, Coca-Cola is the first brand people think of even if someone else has a better formula. The quarterly reports’ long-running conclusion is intact: the best model captures the most traffic.

7. Revenue Chasm and Founder Despair

On revenue, OpenAI has disclosed $12B of ARR; Anthropic may be at $5B-$6B today. “Below that, there may be a chasm—you can’t find a $1B AI product today.” The top 2 may account for 70%-80% or more of AI product revenue, with no sign of growth slowing. Their token consumption is rising 20%-30% or more month over month, comparable to the early days of Douyin and Pinduoduo. New ARR from AI labs may already exceed new ARR at listed SaaS companies, “eating into the increment” of many SaaS businesses.
The emotional response is double-edged: the technology is visibly improving every month, which is exciting, but also “a little despairing.” “The top 2 are running too fast and keep eating up opportunities for VCs and founders. If I were the CEO of an AI startup today… I have no strong cards in my hand; I can only grit my teeth and keep moving.” OpenAI may soon join the Mag7, or a new tech giant may be named shortly.

8. The Miners Move Downstream: The Claude Code Lesson

The simplest logic is that model companies spend tens of billions training models, mining intelligence, then discover “the front-end application Cursor making money hand over fist… I’m not stupid either; I’m obviously going to move forward and harvest the value.” Claude Code launched 8-9 months after Cursor and surpassed its ARR within 3-4 months; Cursor has publicly disclosed roughly $500M of ARR. Claude Code could soon reach $1B, perhaps $1.5B, and its growth may even be faster than ChatGPT’s.
The structural conclusion is that every foundation-model company will need to build agents end to end, rather than settle for API economics, because “the API moat and retention are still not good enough.” Even Anthropic, once the company most committed to building an ecosystem, is now making a major push into first-party products. This playbook will surely be replicated across other domains.
The counterintuitively positive point is that, viewed as an independent startup team, Claude Code is also purely API-based. Boris started with 1-2 people and still has only around 10 people. “That shows there is a lot of room for small product teams.” Guangmi’s standalone valuation: “It could already be a $20B-$30B company.”

9. The Mining Window Is Shrinking: From 2 Years to 3 Months

The division of labor is simple: model companies mine the ore; product companies “open the mining window first and process the ore.” The first product to deliver a magic moment that makes users gasp may have captured $500M-$1B of advertising value. Perplexity, Cursor, and Manus all mined capability overflow from models—information retrieval, coding, or agentic capabilities. Guangmi had a colleague map the windows: nearly 2 years for Perplexity, 9 months for Cursor, and 3-4 months for Manus. This does not mean startups are impossible to beat; “it just means competition is more intense.”
The real unfairness is cost. Assuming the agent products perform equally, “the startup’s token costs will drag it to death.” A foundation-model company earns a margin on API sales and can optimize the stack end to end over time. A product priced at $200 per month may see users burn through thousands of dollars of tokens in a few days, making a 1:3 or even 1:10 loss rate unsustainable.
The sensor analogy is useful: the core capabilities of today’s AI models correspond to every sensor on a phone. “Every sensor gave rise to major internet companies”—GPS produced ride-hailing and food delivery, while large screens produced Douyin. The success of AI products depends heavily on the research dividend and the dividend from mining core model capabilities.

10. Horizontal Bundles, Vertical Integration

The all-in-one bundle is powerful: for $20 or $200, a ChatGPT account already includes chat, search, coding, agents, and workspace. “You bought more than a dozen suites.” That is bad news for products built around a single vertical capability; the market for specialized agents is constrained, much like buying the 3-piece Office suite for $1. It may even be that “PPT is no longer needed in the future,” with web coding generating interactive front ends directly.
Google is the vertical-integration example: from TPU chips to Gemini models to agent applications, plus Docs, Chrome, Android, and YouTube in one super-integrated stack. “Being a few months behind on the model is fine, especially since it isn’t behind today.” Google may have the strongest momentum in the second half.
What can startups do? The old PC incumbents “didn’t understand it, looked down on it, and couldn’t keep up.” Today’s giants understand it, take it seriously, and can keep pace—Zuckerberg’s group is operating like an army. The window is narrow, and the logic of a head-on fight says startups are unlikely to beat the giants. Yet in 2015 many thought Taobao’s scale effects were unassailable, while many passed on Pinduoduo’s A and B rounds and on Shein. “The task today is to find the Huang Zheng who can break through.” Have we found today’s Huang Zheng? “We’re working on it.”

11. From Intelligence-First to Product-First: Guangmi Changes His Mind

The first 6 quarterly reports were “extremely intoxicated by exploring the ceiling of intelligence, with a consistently coherent logic.” Over the past 2-3 months, product has received more weight. There were 2 triggers: ChatGPT’s traffic success and the height of its moat; and his confidence last quarter that Claude had a decisive lead in coding, only to see “Gemini 2.5 Pro quickly catch up in coding.” The dividend from model or research leadership has to become a product moat, because the commercial moat is higher.
Sam said in a public interview last year that ChatGPT seemed to have 2 visions: reaching 1B DAU and achieving AGI, in parallel. “It’s possible every leading AI lab in the world can achieve AGI, but not everyone can reach 1B DAU.” OpenAI has the best balance between intelligence and product. The gap among the top models is only 3-6 months, so if the shelf life of “magic technology” is 3-6 months, product becomes more important.
But the floor remains: “If the model is not good, retention will not be good.” Brand investment does not necessarily come through advertising: “The money you pay for the model is also money you pay for the brand.” 张小珺 recalls that Red（音）said at their first meeting that large-model products would resemble consumer products and that brand value would matter. Guangmi strongly agrees: “ChatGPT’s biggest moat today is not a technology moat; it’s a brand moat.”

12. L4-Level Experience: The Next Mine Is Office

The North Star for a good product is to deliver L4-level experience at scale in a particular domain—an end-to-end aha moment or magic moment that genuinely lands. The 2 best agents today have achieved it: ChatGPT Deep Research in information search and Claude Code in software development. L4 delivery can trigger a business-model shift: users who previously paid $20 are now willing to pay $200.
Why these 2 domains? The biggest dividends today still come from language and code: decades of web data and decades of GitHub code represent “the highest-value abstraction of human activity that humanity has produced.” Programmers are the highest-paid white-collar workers, and software development is easiest to close the loop on. “If you can’t deliver L4 experience at scale in coding, where programmers work, other fields will definitely be slower.” Anything involving the physical atomic world iterates more slowly.
The next replication ground is Office/Workspace: PPT and Excel data analysis. These workflows are highly structured and easier to verify end to end, although the reward system is weaker because “you don’t know what makes a PPT good.” But PPT is exactly where coding and agentic capabilities matter, much like generating front-end HTML. The user base can expand from 30M coders to 500M-600M office workers.
The next generation of AI product managers will likely come from algorithms or models and also possess strong product sense. The prior generation was mostly made up of coders. Without model fluency, they cannot capture the model dividend or anticipate how models will change over the next 6-12 months.

13. Google Back in Favor: Advertising and Scale Effects Reverse the Momentum

The basis for the change of view is that ChatGPT and Google will “certainly converge toward the same eventual product form,” by different paths. Xiaohongshu already combines search, short video, and social feed; it has become the first search destination in the minds of young users. ChatGPT will inevitably run ads: its newly hired commercialization CEO came from Instacart and previously built Meta’s ad system. In the future, top users may pay $200-$2,000 per month, paid users may account for 3%-5%, and more than 95% may view ads. “Google is still the best advertising platform in the world.”
This is about scale effects, not technology. ChatGPT wins through brand mindshare; Google’s hard card is that it owns the entire stack end to end—chips, models, voice, and OS—and can “claw back a round” through scale. Its recently launched AMO, likely AI Mode, could also stimulate long-tail demand. The biggest uncertainty is whether Google can change user habits: search users enter 2-3 keywords, while ChatGPT users enter 30-50 words of description, and more context improves matching and reasoning efficiency.
The form of the next dominant entry point is still hard to imagine. It may sit somewhere between a phone screen and an app. Agents will break the end-to-end advantage of apps by calling these services directly.

14. The Bubble, Seigniorage, and “Chinese AGI”

The bubble’s watchpoint is simple: “Just watch what the smartest cohort is using.” Quant traders and elite programmers almost all use Claude. “As long as the smartest people are still using it, things are fine.” The biggest fear is circular revenue: Silicon Valley’s major tech companies may spend a combined $300B on AI capex this year, “I invest it, Nvidia’s revenue rises, the model companies’ revenue rises, but what worries me most is that the front end creates no value.” If the bubble bursts, capital will concentrate in the top few companies. ByteDance is the most robust because it makes money, followed by OpenAI; the financing capacity of companies like xAI would contract.
The Bay Area’s bigger picture is that America’s 2 core assets are financial seigniorage and Silicon Valley’s technology lead. “Apart from those 2 things, almost everything else in America may be getting eaten away by China’s industrialization.” The extreme scenario is: “If Silicon Valley is no longer technologically ahead, I don’t know whether that financial seigniorage still exists. If China breaks through advanced lithography within 2 years, I don’t know whether global capital will flow back to China.” A side note of disbelief: “It’s incredible that the Trump family can issue tokens and extract wealth so openly and aggressively.”
On the role of Chinese people, one veteran’s summary is: “Jewish finance, Chinese AGI.” China has abundant supply and strong manufacturing but insufficient domestic demand, making it prone to excessive internal competition. North America has less abundant supply but very strong paying demand, so Chinese founders should globalize more aggressively, like the East India Company after Britain’s first Industrial Revolution. “The core of doing North America well is doing Silicon Valley well; the core of doing Silicon Valley well is first taking the VC circle or San Francisco’s AI circle.” “The next 3-5 years are extremely promising for Chinese founders.”

15. Investment Methodology: Diversification Fails, Concentration Works, and Find Today’s Huang Zheng

Traditional VC’s spray-and-pray strategy has partly failed: “Diversification has failed; concentration is what works today.” If you did not invest in OpenAI, Anthropic, or xAI and also missed Cursor, “you worked for nothing in this cycle.” Many classic Silicon Valley VCs passed on OpenAI and Anthropic because AI capex was too high and margins too low. Anthropic raised at a $4B-$5B valuation when nobody wanted in; many VCs also passed on Mira’s $10B angel round. “That’s where the non-consensus is.” OpenAI itself resembles a VC: Sam raises huge sums and allocates them to dozens of internal teams, “eating a meaningful part of Silicon Valley VC’s business.” Many have forgotten the pain of Jasper.
The fertilizer thesis is that $300B of AI capex in 1 year and more than $1T over 3 years is “future explosive fuel.” “WeChat was fertilizing Pinduoduo back then.” To achieve 30x DPI, a fund needs several major winners: Gao Rong invested in Pinduoduo, 5Y invested in Xiaomi and Kuaishou, and Sequoia invested in ByteDance. A $300M fund needs $10B of returns to achieve 30x. The remaining work is to spend enough time understanding the next generation of founders. 一鸣 in 2025 has deeply thought through information distribution; 一鸣 in 2015 and 黄征 in 2016 had equally deep insight into social commerce. ByteDance itself survived in the gap between giant platforms under WeChat’s blockade. “The landscape may look settled, but there are always new opportunities.”
The closing list of non-consensus views: the value of exceptional teams is still underestimated—“I still think Anthropic is undervalued, and Mira’s team is undervalued.” Coding remains undervalued; “coding may contribute enormously to this technological revolution.” The next breakthrough in robotics may still come from the language people rather than today’s vision-trained robotics teams, because “only language has achieved generalization,” and robotics may still need to clear several more GPT-4-level milestones of technical maturity.

The CEO Who Cheated Death, Slept an Hour a Night, and Built a Healthcare Empire One More Push Higher Before the Crash?