Vol.72 Technology, Applications, and Capital: A 154-Page Review of the AI Industry in September 2025
Summary
As of Q3 2025, AI is still advancing along OpenAI’s L1 Chatbot, L2 Reasoning, L3 Agent, L4 Innovator, and L5 Organizer roadmap, with intelligence itself remaining the main storyline. GPT-5 has drawn sharply divided reviews, but its benchmark performance still places it in the leading tier; as fixed evaluations approach perfect scores, the length of time models can execute tasks continuously has stretched from several dozen seconds for the first ChatGPT to roughly 2 hours 17 minutes. “It actually works” remains the most important fact of this technology wave.
Reinforcement learning has shifted the center of competition from pretraining to post-training, but the real flywheel still depends on synthetic data, self-evaluation, and verifiable rewards working in practice. Mathematics and coding are progressing fastest because their answers are well-defined; other fields remain constrained by environment construction, reward hacking, and the need for high-quality, PhD-level data. Having AI generate training data is the “first perpetual-motion machine”; having AI grade AI is the “second perpetual-motion machine”—simple in theory, but far from solved in engineering.
Chinese models are using open source to claim a seat at the global main table, but the top tier has clearly contracted from the “Six Little Dragons.” The top 15 models on one designer-focused leaderboard were all Chinese, with a GPT-5-series variant not appearing until No. 16; an a16z partner said roughly 80% of the AI startups in which the firm invests are using Chinese open-source models. 庄明浩 summarized the new landscape as “there are no longer six little dragons”: DeepSeek and Qwen are strongest, Zhipu, MiniMax, and Kimi remain at the table, while Baichuan and 01.AI have faded from the core competition.
Coding found PMF first, while multimodality shows the contest is far from over, and data ownership is being converted into model advantage. Cursor, Replit, and Lovable are growing rapidly; even Chinese internet giants building for global developers are creating proprietary tools because of code-security concerns. In images, the war has repeatedly been “finished and restarted,” from Midjourney to GPT-4o and Nano Banana; ByteDance, Kuaishou, and MiniMax occupy much of the video leaderboard. “Who owns the video data” is the key to explaining the capability gap between YouTube, ByteDance, and Kuaishou.
ChatGPT is growing from the leading model gateway into a platform, and product data explains OpenAI’s lead better than any single leaderboard. As of August 2025, the program cited 700M weekly active users and 2.5B prompts per day; average monthly usage was about 13 days, while the tail of paid retention turned upward, producing the “smile curve” associated with mature internet products. Faced with the choice between “AGI or a billion-daily-active-user platform,” Sam Altman has at least temporarily bet correctly on the latter.
An Agent is not simply a model capability, but a combination of model, scaffolding, context engineering, and runtime environment. 杨植麟 argues that third-party products are “reverse-engineering the model-training process,” while model companies can incorporate tools and environments directly into training, potentially giving first-party products such as Claude Code and ChatGPT Agent a higher ceiling. The Manus-side addition is that the third role an Agent adds beyond a Chatbot is precisely the “environment.” That expands the investment opportunity into MCP, memory, browsers, computer use, databases, routing, virtual machines, security, and sandboxes.
Capital has made AI its third major investment theme, but returns are becoming increasingly concentrated in a handful of giants and super-sized deals. Roughly 58% of US VC dollars in 2025 flowed to AI, but those investments represented only 36% of deals; deals above $500M accounted for about 63% of funding, versus just 18% in 2021. Public markets are diverging as well: NVIDIA, Meta, and Microsoft lead, while Apple and Tesla have temporarily fallen behind; even though 91% of surveyed fund managers believe US equities are overvalued, the reality remains that “relative to holding, selling may carry greater risk—you’re in the market, and you have no choice.”
AI’s most dangerous contradiction is not a lack of revenue, but the possibility that ARR, gross margin, and Token consumption are all being misrepresented at once. A “tens-of-millions-of-dollars ARR” figure linearly annualized from nine days of activity cannot account for annual fees, refunds, or retention; leading AI applications may have gross margins of only about 25%, coding products about 20%-30%, and some have reached “$100M ARR and $120M in Token costs.” Lower model prices have not automatically become profits because users will not use old models, while Agent task lengths are growing geometrically: “one money-losing business is built on top of another,” with the trust cost of AI-generated misinformation added on top—“AI is the problem, and AI is also the solution.”
Deep dive
1. The Industry Is Moving So Fast That Every Review Comes With an Expiration Date
庄明浩 set a hard cutoff for the deck: preparation began in August, the manuscript was locked on September 4, and recording took place on September 6. “Every morning when you open your eyes,” a new model, policy, or earnings report from the previous night may require the analysis to be rewritten.
The milestones during production were almost absurdly dense: Google DeepMind updated its world model and video model in early August; Claude 4.1 arrived on August 6, GPT-5 on August 8, and DeepSeek V3.1 on August 19; the State Council issued its “AI+” action guidance on August 27, followed by NVIDIA’s quarterly earnings report on August 28, which the global market was watching closely.
He compressed the entire deck into 3 metaphors: “Reasoning is fighting monsters, scenarios are opening the map, and investing is paying to level up.” Technology, applications, and capital are tightly interlocked, ultimately leading to bubbles, employment, and polarization; as for how to make money today, a friend’s joking answer was: “Make money by trading stocks.”
2. OpenAI’s Five-Level Roadmap Still Governs Frontier Models
The starting point for this cycle remains the November 30, 2022 launch of ChatGPT: GPT-3.5 plus a chat UI gave people their first real experience of what general-purpose foundation models could do. In March 2023, Bill Gates called AI the second major technological revolution he had experienced after the GUI, placing it alongside chips, computers, the internet, and smartphones.
OpenAI’s roadmap has not become obsolete: L1 Chatbot, L2 Reasoning, L3 Agent, L4 Innovator, and L5 Organizer. 庄明浩’s core judgment is not that OpenAI will lead forever, but that as of Q3 2025, the world’s leading companies are still competing for the same stage inside the coordinate system OpenAI defined.
When Ilya Sutskever was asked for “the most beautiful and surprising idea in deep learning,” his answer was one sentence: “It actually works.” Scale up neural networks and feed them enough data, and capability really does keep improving; 庄明浩 compares it to “more force really can produce miracles,” and to the moment after Columbus found the New World, when everyone finally knew the route existed.
3. o1 Shifted the Training Center of Gravity From Pretraining to Post-Training
After Claude 3.5 in June 2024 and OpenAI o1 in September, the industry formally entered L2. From Q3 2024 through Q1 2025, the common mission of leading model companies was almost simply to reproduce o1: make the Chatbot “slow down” and turn a single answer into a reasoning process that could explore, make mistakes, and converge.
The path laid out by Kimi researchers was explicit: use RL to solve problems continuously, require precise rewards, avoid imposing a structured-thinking template, let the model explore its own paradigm, and “allow mistakes.” 庄明浩 initially grouped Kimi 1.5 and DeepSeek R1 together as key implementations, while adding that “Kimi 1.5 probably was not open-sourced”; the important point was that both demonstrated the effectiveness of using reinforcement learning to solve problems.
DeepSeek’s chain was V3 as the base model, reinforced through RL into R1-Zero, then combined with cold-start data and usability adjustments to produce R1; R1 was updated to R1-0528 on May 28, while V3 was upgraded to V3.1 in August. Further training from V3.1 could produce a stronger R-series model, but whether that would be called R2 or wait for V4 is a decision only DeepSeek can make internally.
The leading technical advantage was estimated at roughly 6 to 9 months at the time, and may have since compressed to about 3 months. Reasoning models becoming standard does not mean training is over: base and reasoning models are still being upgraded, and “the evolution of intelligence itself remains the industry’s main storyline.”
4. Reinforcement Learning’s Two “Perpetual-Motion Machines” Have Yet to Work in Practice
The scaling of pretraining data was once thought to be failing, but models generating synthetic data that did not previously exist opened a new possibility. If that data can reliably feed back into the model, it amounts to “AI training AI”—庄明浩’s “first perpetual-motion machine.”
The second is more aggressive: if some models are already considered to have reached a human PhD level, can models grade model answers and construct rewards? If it works, self-training could enter an accelerating loop; but “a perpetual-motion machine is a logic that violates the laws of physics.” It is easy to reason through on paper, while implementation still presents a long list of problems.
Mathematics and coding have verifiable answers, while other domains lack stable reward functions; how to combine environments in software and hardware, avoid reward hacking, and obtain high-quality data are usually engineering problems. Meta’s $14.8B investment in Scale AI and acquisition of its core team reflect the upgrade in data work: in 2018, workers labeled roads, vehicles, and road signs; today, programmers, doctors, lawyers, finance professionals, and even PhDs grade models.
5. Once Leaderboards Fail, Task Duration Becomes the New Measure of Capability
GPT-5’s reception after launch was sharply polarized, with some users even asking for the old model back, but 庄明浩 still placed it among the world’s leading models when looking only at benchmark data. Meanwhile, the unit cost of frontier models continues to fall, with the long-term trend approaching “an order-of-magnitude decline every year.”
The problem is that once a fixed leaderboard becomes a KPI, vendors optimize around the test set and quickly push scores toward 100. As models continue to improve, humans lose an effective measuring tool; capability has not stalled—the ruler has simply reached its limit first.
Once Agents began executing human work, the length of time they could complete tasks continuously became a substitute metric. Early ChatGPT interactions lasted only a dozen or so seconds; the GPT-5-related figure cited by the program was already about 2 hours 17 minutes. Moving from dozens of seconds to several hours represents an order-of-magnitude change in task duration.
6. Agents Continue the Shift Toward Greater Generality and Less Control
Even looking only at the first 3 quarters of 2025, the year deserves to be called the “Year of the Agent.” A platform tracking open-source models and open-source databases showed that AI Agents had created 4x as many databases as humans in recent months; if more of the internet is operated by AI Agents in the future, the ecosystems designed for humans today may all have to be rebuilt.
The common direction from L0 through L3 is greater generality, less control, and stronger learning. Generality means a system cannot outperform specialized solutions in every vertical scenario at the outset, but it also means L3 should not depend indefinitely on preset workflows; it should increasingly solve problems on its own.
The user experience keeps oscillating between “simple—complex—simple.” L1 initially required only conversation, followed by prompt engineering; L2 reasoning models could again understand simple instructions; L3 was once surrounded by workflows and is now moving back toward a general model decomposing tasks autonomously.
7. China Has Used Open Source to Reach the Main Table, but the Seats Have Clearly Contracted
庄明浩 describes China’s model landscape as “there are no longer six little dragons,” and sees DeepSeek and Qwen as the world’s 2 leading open-source AI models. Chinese models updated in waves in July, while global leaders continued to follow in August; open source now appears to have shifted from a differentiating strategy to an industry consensus, although the answer to whether “open source is the goal or the means” differs by company.
On a platform for visual artists and frontend designers, all top 15 open-source models came from China, with a GPT-5-series variant mentioned by the program not appearing until No. 16; an a16z partner wrote on Reddit that roughly 80% of the AI startups in which the firm invests are using Chinese open-source models. 庄明浩 left the key question open: is this a temporary result or the starting point of a trend?
If language models remain the main AGI table, the US seats belong to OpenAI, xAI, Google, Anthropic, and Meta, which remains in the game; China’s seats belong to Alibaba’s Qwen, DeepSeek, Zhipu, MiniMax, and Kimi, with France’s Mistral also present. Baichuan, 01.AI, and the other former “Six Little Dragons” have faded, while ByteDance, Tencent, and Huawei are not on his list of current core seats.
DeepSeek’s delay in releasing R2 is itself evidence that naming does not define the roadmap: reinforcement learning can continue after V3.1, but whether the result is a new R-series model, the rumored R2, or a larger upgrade after V4, the program did not decide for the company and simply advised: “wait and see.”
8. Coding Found PMF First, While Security Created Domestic Demand
“When machines master language, strong AI has arrived” represents the language-centered view; coding itself, however, can also be treated as a language. If AI can write all the code in the world, whether coding is a subset of large language models or another path toward AGI stops being merely a classification question.
Whatever the answer, Cursor, Replit, and Lovable have already demonstrated through rapid revenue growth that AI coding may be the first application category in this cycle to find PMF. Model vendors are competing to release coding capabilities and first-party products, with the contest spanning clients, cloud, databases, security, frontend, backend, and application generation.
庄明浩 initially assumed programmers writing code in English were naturally the least national of all user groups, meaning Chinese vendors might not need to rebuild the stack. He later changed his view: if the internal code of Tencent, ByteDance, Alibaba, Meituan, or Xiaohongshu flows into third-party tools, no CIO will accept it. Security requirements will create a domestic AI-coding market by force, just as enterprises once built their own IM systems.
9. The Image and Video Wars Are Far From Over, and Data Ownership Determines the Rankings
Stable Diffusion was open-sourced in August 2022, before ChatGPT; Midjourney said its AI ARR had reached $200M by 2023, and 庄明浩 estimates it may be at $500M today. Image generation has been declared “finished” multiple times, only to be reopened by GPT-4o’s Ghibli style and Google’s Nano Banana.
Nano Banana’s image editing, consistency, and perspective conversion reminded him of “words becoming reality”: after recombining people, products, and backgrounds, the model still preserves consistency, and can even shift from a bird’s-eye view to eye level or infer the back of an object from the front. Images remain like photography apps in the mobile-internet era: a strategic battleground where new products continue to emerge.
Video has moved from the distorted “Will Smith eating spaghetti” phase to Veo 3 achieving synchronized audio and video; the boundaries between images, image-to-video, text-to-video, and video editing are rapidly blurring. On the leaderboard cited by the program, ByteDance, MiniMax, and Kuaishou’s Kling occupy many of the top positions, while OpenAI’s best model ranked only No. 13 on one text-to-video leaderboard.
庄明浩’s mechanism-level explanation is data. America’s strongest video-data source is YouTube, but even within Google, copyright gray areas limit how fully it can open the data for training; ByteDance and Kuaishou in China both own massive video libraries and operate real production environments, giving them an advantage in earlier video-model progress and leaderboard share.
10. Voice, 3D, and World Models Continue Expanding the AGI Main Table
On the August 27 voice leaderboard, MiniMax ranked No. 1 and No. 3, OpenAI No. 2 and No. 4, ElevenLabs No. 5 through No. 7, and Fish Audio No. 8. 庄明浩 inferred that voice models require lower talent, compute, and data thresholds than text, images, or video, making voice an “excellent value-for-money zone” for the AI Labs of mid-sized companies such as Xiaomi, Bilibili, and Xiaohongshu.
That view was revised the following day: on August 28, OpenAI released the GPT Realtime API, while Microsoft launched 2 proprietary models, 1 of them a voice model. Voice is not merely a side market that mid-sized companies can enter; it has also become a battlefield for the giants.
In 3D, game companies including Tencent, Microsoft, and Roblox possess data and usage scenarios that may give them a natural advantage. One step further, Google Genie 3, 李飞飞’s World Labs, and Tencent Hunyuan, which released the first version of its world model on September 3, are moving from local generation toward infinitely expandable virtual worlds. The old “metaverse” narrative has returned on a new technical foundation.
11. From Models to Data Centers, AI Has Created a Closed Capital Loop for the Giants
On Gartner’s 2025 technology curve, AI Agents sit at the “peak of inflated expectations,” generative AI is beginning to slide into the trough, and cloud computing has entered a stable phase. 庄明浩’s reminder is that technology does not move upward in a straight line; different stages imply completely different product expectations and capital requirements.
The past 2 years have produced a simple and expensive chain: more capital builds data centers, stronger models are trained, better results are produced, and those results support the giants’ stock prices and the next round of CapEx. In 2025, Amazon, Meta, Microsoft, and Google together spent more than $300B on related investment, making AI the shared narrative of the “Magnificent Seven.”
US data-center construction investment has rapidly approached office construction investment: the former is rising as the latter falls. 庄明浩 describes it as the “offices” of the silicon-based world about to exceed those of the carbon-based world. The image is partly a joke, but it points directly to compute infrastructure moving into the center of macro investment.
Microsoft’s cloud-revenue additions have exceeded AWS’s in recent quarters, which he attributes to Microsoft’s deep tie-up with OpenAI; Amazon invested in Anthropic, but Anthropic had also previously used Google Cloud extensively. The ultimate beneficiary remains NVIDIA, which has completed a “three-wave combo”—from gaming GPUs to Web3 mining to AI data centers—and reached a market capitalization of more than $4T.
12. ChatGPT Is Growing From a Model Gateway Into a Platform
As of August 2025, the ChatGPT figures cited by the program were 700M weekly active users and 2.5B prompts per day; traffic was still growing about 135% in July. Faced with the choice between “AGI or a billion-daily-active-user platform,” Sam Altman naturally wants both, but the actual priority is clearly the platform.
More important is paid retention. For a normal product, retention should continue falling as the months pass, but ChatGPT’s curve has begun to rise at the tail. 庄明浩 calls it a “smile curve”: once brand strength, competitive barriers, or monopoly power crosses a threshold, older users become more stable. The phenomenon once seen with Douyin and Pinduoduo is now repeating with ChatGPT.
On Sensor Tower’s basis, ChatGPT is used about 13 days per month on average, steadily up from roughly 4 days early after launch; the comparison figures are about 12 days for Claude, 13.7 days for xAI and Twitter, and 18.6 days for Google. Looking only at external traffic, he believes Chatbot traffic competition has essentially already been won by ChatGPT.
13. An Agent’s Ceiling Depends on How Model, Product, and Environment Come Together
In a Chatbot, the user gives the model one sentence and the model returns a result; the relationship contains only the user and the model. Peak, one of Manus’s founders, frames the Agent as adding a third key element: the environment. What product teams are really building is often not a chat interface, but a runtime space where a model can work over an extended period.
杨植麟 explains it from the model-company side: many third-party Agents build “scaffolding” and tools on top of base models, essentially reverse-engineering the model-training process and benefiting from capability spillover. When model companies build first-party products, they can design the tools, context engineering, and environment first, then train the model directly inside them, giving them a potentially higher theoretical ceiling.
He specifically named Claude Code and ChatGPT Agent as first-party products. 杨植麟 said Kimi’s investment in first-party products was still limited at the time, with most resources going to model training; the key question going forward is not whether first-party products must inevitably eliminate third-party products, but how the 2 sides divide the work and whether the ecosystem continues to allow third parties to capture model spillover.
Agent infrastructure’s “environment” is far broader than an interface: protocols such as MCP, memory and context, browsers and computer use, databases, model routing, result presentation, virtual machines, sandboxes, and security all belong to it. The Agent users see is only the tip of the iceberg; reliable operation depends on the engineering beneath the surface.
14. Chinese Applications Are Strong on the Leaderboards, While Social Companions Hide in Borderline Traffic
In a16z’s fifth Top 50 ranking, 12 Chinese companies or products appeared on the web list and 22 on the app list, close to half the total. Its categories were domestic-serving giants, exporters aimed at global markets, and hybrid players such as Meitu that operate on both sides. China’s advantage exists not only in open-source models, but also in application productization.
Data overlooked by the media: 10 of the top 50 web products were social or companion products, a meaningful share. Most remain stuck in the “1.0 experience” of chatting with virtual characters, however, and much of the content is NSFW or mildly suggestive. 庄明浩 asks the question directly: “Apart from doing borderline content, what else can you do?”
More design-led attempts include the alien-companion product Tolan, a woodland therapy room for communicating with small animals, 独响, which resembles an AI social feed and extends into wristband hardware, and the AI boyfriend-and-girlfriend product Eve. Viewed calmly from an investment perspective, some of these projects have broken through mainly at the interaction and design layers; monetization, growth, and model moats remain unproven.
For companies that can no longer compete for a seat at the foundation-model table, social is a side table. ByteDance, Meituan, Kunlun, MiniMax, Zuoyebang, Yuanxiang, Westlake Omics, and even game companies are experimenting. Game studios use voice-interactive NPCs to solve tasks, reconnecting games, companionship, and world models in a single chain.
15. As Model Capabilities Spill Over, Application Boundaries Expand Faster Than Categories
For the first time, a16z’s ranking included 4 Google projects at once: Gemini, NotebookLM, Google Labs, and Google AI Studio. The benefit is progress across every business line; the other side is that departmental walls remain clear. Even so, the performance of both models and products has revived the market’s discussion of “Google is back.”
Nano Banana best demonstrates capability spillover. The developer 花生 gave the model an indoor photo of his girlfriend and a beach location from a map, and it generated an almost flawless beach-travel photo. 庄明浩’s reaction was: “Package this function, and it seems like a small product already. You could call it virtual tourism.”
三色塔 tracks not the absolute number of AI applications, but the incremental categories created by AI. The top growth category was nutrition and food, while religion and spirituality also ranked highly. His explanation is that these fields had weaker capabilities to begin with, so AI’s marginal enhancement is more direct; they may therefore generate new products more easily than crowded general-purpose tools.
Web Coding “squeezed out” roughly $1B of incremental market from 2024 to 2025. Anthropic went all in on coding, and Claude Code’s growth even began to affect Cursor; but OpenAI and Google, as leaders, cannot bet on only 1 direction. After coding, marketing, HR, entrepreneurship, and every other kind of work may be rebuilt for the web or turned into an Agent.
16. ARR Has Become the Most Popular—and Most Easily Abused—Number
BVP defines AI superstars as reaching $100M ARR within 18 months, while rankings and financing news also routinely use ARR to measure companies. But whether ARR means annual revenue, annual subscriptions, or a linear extrapolation, there is no consistent treatment of refunds, discounts, annual-fee recognition, or cyclical fluctuations.
庄明浩 quotes his friend 兰溪’s satire: “Products that have been live for less than a year are not allowed to use ARR as a metric. It has already been legislated; violators face the death penalty.” A product goes live for 9 days, multiplies collections by 40, and claims tens of millions of dollars in ARR; if the collections include annual fees paid upfront, the same revenue has effectively been projected 40 times when it should not be repeated.
When Manus disclosed roughly $90M of “real ARR,” its method was to look directly at the payment backend: divide annual fees by 12, count monthly fees in the month received, and multiply monthly recurring revenue by 12. 庄明浩’s summary was: “A real man opens the payment backend for you. Don’t talk to me about those run rates.”
17. Public Markets Are Re-Ranking Giants by Their Ability to Convert AI Into Results
The top 10 companies in the S&P 500 now account for close to 40% of total market capitalization, and NVIDIA’s quarterly earnings report has become a barometer for the global economy. But the law of large numbers is taking effect: even when revenue sets another record and meets expectations, the market no longer gets excited automatically; decelerating growth alone can trigger a correction.
The AI paths of the 4 giants differ. Amazon relies mainly on AWS; Google spans applications, models, cloud, and chips; Microsoft’s full-stack story is constrained by its new relationship with OpenAI; and Meta is using AI to repair ad-targeting capabilities. Its former disadvantage of lacking end-user hardware has been partly offset by AI-driven advertising efficiency.
As of roughly August 20, the year-to-date gains cited by the program were about 30% for NVIDIA, 28% for Meta, and 21% for Microsoft, versus an average of about 10% for the 7 giants; Amazon and Google were each up about 5%, Apple was down about 10%, and Tesla was down about 18%. A weak model did not stop Meta from benefiting: product performance and advertising revenue are more important support for the stock than model rankings.
Other beneficiaries are more extreme. Palantir trades at roughly 90x PE and 250x PS, “riding alone at the front” on the S&P 500 valuation chart; AppLovin also benefited directly from AI recommendation efficiency after selling its gaming business and focusing on advertising. 庄明浩 closes with Graham’s line: “In the short run, the stock market is a voting machine; in the long run, it is a weighing machine.”
18. More Primary-Market Capital Is Piling Into a Handful of Super-Deals
Roughly 58% of US VC investment dollars in 2025 flowed to AI, but those deals represented only about 36% of total projects; the program cited roughly $440B in financing over the previous 3 quarters. The difference comes from pricing: AI companies may raise about 40% more than ordinary companies even in seed rounds, while later-stage deals run into the billions or even hundreds of billions of dollars.
米拉·穆拉蒂’s new company raised $2B in its angel round at a $12B valuation based solely on its team; Anthropic then raised $13B at a $183B valuation. 庄明浩 summarized this pricing with the phrase “姜太公钓鱼，愿者上钩”—the fisherman waits, and those willing to bite come to him.
Deals above $500M now account for about 63% of the market; even at the most exuberant point in 2021, they represented only 18%. Funds are concentrating at the same time: in the first half of 2025, the top 10 funds received 50% of market capital and the top 30 received 74%. Concentration in AI startups and the Matthew effect in asset management are reinforcing each other.
Anthropic has roughly 86 investors, OpenAI about 83, and Hugging Face about 53; SPVs conceal the actual sources of capital behind still more layers. OpenAI and Anthropic have publicly opposed investors stacking another SPV on top; 庄明浩 also warns that people claiming to be investors in OpenAI or SpaceX without a legitimate institutional background should be asked whether they are direct shareholders or merely placed a small amount of money into some investment channel.
19. Super-Unicorns Do Not Need to Rush to IPO, and Acquisitions Have Become the Main Exit Route
Companies such as ByteDance and Xiaohongshu continue to enjoy smooth business growth, fundraising, and secondary trading. The market valuations mentioned by the program were about $330B for ByteDance and a disclosed $31B for Xiaohongshu, with actual transaction pricing possibly close to $38B. As long as there are buyers, and companies can even repurchase employee equity, “why go public?” becomes a real question.
庄明浩 calls 2025 “an M&A year in the absolute sense.” By midyear, transaction volume had already exceeded the full-year totals for 2023 and 2024, while AI-related M&A was about 2x the 5-year average. Many transactions are “hollowing-out acquisitions”: they provide liquidity to investors and buy core talent at high prices, while the original company and business shell remain in place.
OpenAI has also shifted from investment target to acquirer. After appointing Fidji Simo as CEO of applications, it acquired application data-monitoring company Statsig for about $1.1B. A super-private company can also act as an industry consolidator; it does not need to wait for an IPO before beginning acquisitions.
The IPO window itself has not truly opened, but Figma created an extremely powerful example: its angel-round price was about $0.09 per share, versus $110 per share on the first trading day. Based on prices during the first few days after listing, Index’s return was about 1,850x, Series A about 480x, Series B about 240x, and Series C close to 100x. Even after the stock later halved from its peak, the returns were still enough to reignite risk appetite.
20. The Evidence of a Bubble Is Ample, but the Market Fears Leaving Early Even More
The S&P 500’s PE and PS ratios have exceeded their levels during the 1999-2000 internet bubble, while AI data-center and infrastructure investment as a share of GDP has also surpassed the internet-infrastructure peak. Adjusted for inflation, the cumulative valuations of today’s leading AI companies likewise exceed those of that era, and many of the largest companies have not even gone public.
Sam Altman has publicly acknowledged that AI investment contains a bubble. An MIT report claiming that “95% of enterprise AI attempts have no effect” triggered several consecutive days of 2%-3% declines in US equities. 庄明浩’s assessment after reading it was that the report’s data sources, survey method, and reasoning quality were weak; the market was still struck by a frightening number, which itself showed that sentiment was “too tightly wound.”
A Bank of America survey of roughly 3,000 fund managers found that 91% considered US equity valuations too high; the same report said the risk of selling those stocks could be greater than the risk of continuing to hold them. Investors know prices are high, but benchmarks, career risk, and the fear of missing further upside leave them “in the market, with no choice.”
Employment requires a time dimension. The decline in jobs over the past few years cannot simply be blamed on AI, because unemployment would also have occurred without AI; the future impact remains uncertain, but current signs suggest that entry-level jobs may be taking a more direct hit than higher-level roles.
21. Geometric Token Growth Is Eating AI-Application Gross Margins
BVP data shows that fast-growing AI applications may have gross margins of only about 25%, far below the 50%-60% common in traditional SaaS; model companies may still have 50%-60%, while AI coding applications commonly run at about 20%-30% and can even be negative at certain stages. “Is negative gross margin acceptable?” has shifted from a joke to a question valuations must answer.
庄明浩 draws the industry chain as a game of hot potato: users pay applications, applications use revenue and VC subsidies to buy model APIs, and model companies use revenue and financing to pay for cloud and chips. A company may therefore have $100M ARR while carrying $120M in Token costs—“one money-losing business built on top of another money-losing business.”
“Model costs fall 10x every year” does not automatically rescue profits for 2 reasons. Users will not use old models—“just as nobody reads yesterday’s newspaper”—and the price of the most advanced models has not fallen 10x in parallel; meanwhile, Agent tasks have stretched from dozens of seconds to several hours, with the program citing a doubling roughly every 7 months. Token consumption is geometric, not linear.
As a result, both models and applications are redesigning pricing. Claude Code once produced a top user who paid $200 per month while consuming tens of thousands of dollars in Tokens; that plan cannot continue indefinitely. At the same time, QuestMobile found that only about one-third of AI apps and web products were growing, while roughly three-quarters of plugins were growing. The independent founder’s problem is: “Who the fuck is willing to just build a plugin?”
22. Narrative Inflation and AI-Generated Misinformation Are Simultaneously Draining Trust
The A-share market’s “falling in love with the new and abandoning the old” has become a running joke. Cambricon once traded at about 91x PS and 555x PE, with PE exceeding 4,000x before profits improved, bringing its valuation close to Intel’s. The valuation joke at the start of the year was “there is a 1% chance it becomes NVIDIA, so it is worth 1% of NVIDIA”; when that probability approached 2%, the joke became “isn’t NVIDIA just America’s small Cambricon?”
After Cambricon’s share price surpassed Moutai’s, 庄明浩 warned that this resembles a “curse” in A-share history: many companies that overtook Moutai eventually lost more than 99% of their value. The joke itself is not a trading signal, but it is a thermometer. When “being anxious about not owning Cambricon” becomes “being chilled to the bone,” market heat has spilled beyond fundamental analysis.
Product marketing has also developed a fixed SOP: add enough qualifiers to “the world’s first,” fill the launch video with jargon such as “end-to-end,” show the CEO’s face, include an invitation code, and coordinate same-day media placement. A toy was described as “the world’s first end-to-end AI interactive toy”; when ordinary users asked what that meant, the official explanation was simply “a smooth interactive experience comparable to a real person.” Even a16z has begun teaching founders how to shoot product videos and run social media, showing that formulaic marketing has penetrated investment firms themselves.
More difficult is information pollution. Parody images of military parades and Trump, a fake Zhihu answer from 梁文锋, and fabricated DeepSeek financing news may all be repeatedly summarized, cited, and re-ingested by multiple AIs until their original sources disappear. The AI-generated-content labeling rules implemented from September 1 require platforms to apply visible or invisible labels. 庄明浩 believes they may not cure the problem, but “something is better than nothing.” The episode ends with a single paradox: “AI is the problem, and AI is also the solution.”