Summary

As of 8/5, the S&P 500 set new highs after weathering July’s “epic” correction, but the market has split internally: AI assets are rising while traditional software sits 50%-70% below its highs, with some names down more than 90%. 庄明浩’s conclusion is that the selloff itself is not important because demand remains strong across the value chain and no one has dared cut expectations; but the market appears to be repricing SaaS on the expectation that traditional software could be displaced by AI.
AI capex is approaching 3% of US GDP, and AWS, Microsoft, Google, and Meta’s 2026 capex is nearly double 2025 levels—both a confirmation of demand and entry into a danger zone comparable to the fiber-optics cycle. AWS quarterly growth should be back above 10%; Microsoft Cloud and Google Cloud growth figures cited orally are 45%, 50%, and 60%. The real question is “What about 2027?” because the same concern raised last year was answered this year with further upward revisions.
The latest figures cited by 庄明浩 put OpenAI’s July ARR at more than $40B and, according to rumors, xAI’s at more than $70B. OpenAI and Salesforce ARR are also widely used as indicators of AI adoption, but ARR may not capture everything. Their revenue-sharing definitions differ, so the figures cannot be compared directly; model prices still have substantial room to fall. A cost of 1 and price of 10 produces 90% gross margin; halve the price and margin remains 80%; at 25% of the original price, it is still 60%. Models therefore have to “escape the kill zone”—by becoming stronger or cheaper.
The application consensus has shifted from Coding Agent to Work Agent and Enterprise: “coding was first, now everything else.” Codex has become the new ChatGPT; desktop clients now feature ChatGPT Work and Claude’s Chat/Work split, with every major player betting on enterprise. But the strongest hard data on China’s desktop AI productivity market remain WorkBuddy’s 20.97M visits through June and 18M users across all PC clients.
China’s AI office market is colder on payment than usage: Doubao has roughly 300M-400M MAU, but its professional tier had only “a few hundred thousand” paying users nearly 2 months after launch, implying roughly 0.1% conversion. 庄明浩 considers 3% “extremely optimistic,” with the normal rate potentially below 1%; no one among the 100-200 investors at the event had bought it. In the US, only 2.2% of households pay for AI tools. “Everything is just beginning.”
The AI commercialization opportunities easiest to see today are training data, RL environments, inference services, software infrastructure, and compute—not standalone consumer applications. The observation threshold for private AI companies has risen from last year’s $100M ARR club to $500M, but the rankings are packed with data, compute, and model companies, with very few true apps. The list is both a ranking of the leaders and potentially a “death list” of businesses model vendors are watching.
Early-stage funding offers no clearer answer on an application breakout: the Deal Count top 10 is concentrated in humanoid robots, world models, coding, legal, model services, and GPUs, with only a few categories qualifying as true applications. China’s domestic 10M-MAU rankings are almost entirely projects from previous-generation giants. Kling, MiniMax, ACE Studio, and Immersive Translate may qualify as apps, but they still do not answer what the next major app category will be.
Roblox most clearly exposes the gap between understanding the AI narrative and buying the right stock: after rising from the $20s to $150, it fell about 26% the day Google DeepMind released world model GPT 3, then dropped 33% or 37% in a single session on its latest earnings report as DAU declined, returning to roughly $20. 庄明浩 initially judged the earlier decline to be an overreaction because commercial disruption was still far away, but the downtrend did not reverse. The episode closes on RSI’s “everything is loop,” arguing that this may be the biggest consensus in AI today.

Deep dive

1. Index Highs Cannot Hide the AI-Software Rift

庄明浩 framed the roughly 20-plus-page deck as “no views, just a record.” As of 8/5, the S&P 500 had repeatedly set new highs despite war, China-US trade friction, and July’s violent correction; over a 1-year horizon, external shocks appear not to have altered the index’s upward path.
Beneath the index lies an entirely different market: AI-related names have outperformed over the past 6 months, while most traditional software and SaaS names appear 50%, 60%, or 70% below their all-time highs; many are down more than 90%. “AI and software appear to have split into separate paths.”
The demand side offers shorts no easy answer: cloud vendors, chipmakers, semiconductor companies, and SaaS companies have “not one” dared to lower expectations. 庄明浩 therefore said the selloff itself does not matter, then immediately asked: “That sounds dangerous, doesn’t it?”

2. Capex Keeps Beating Expectations While Model Prices Still Have Huge Room to Fall

For the past 3 years, every earnings cycle has brought higher estimates for the following year’s capex, while year-end forecasts have repeatedly proved too conservative. Spending by US megacap tech companies is now approaching 3% of GDP, and AWS, Microsoft, Google, and Meta’s 2026 capex is nearly double 2025 levels. By analogy with the fiber-optics cycle, this is already dangerous territory.
His own counterargument is worth preserving: when so many things approach the limits of human execution, “how much further can they be revised upward?” But the same question was asked at the end of last year, and the market continued to ratchet up 2026 spending, leaving the answer for 2027 simply: “We don’t know.”
AI revenue is showing up first in Cloud: AWS quarterly growth should have returned above 10%; based on verbally cited figures, Microsoft Cloud and Google Cloud are growing 45%, 50%, and even 60% per quarter. For more direct evidence of industry adoption, OpenAI and Salesforce ARR are often used as reference points. OpenAI’s July ARR should have topped $40B, while rumors put xAI above $70B.
But ARR may not capture everything: OpenAI’s figure should deduct Microsoft’s revenue share, while xAI’s does not deduct AWS and Google Cloud’s shares. The same dollar of revenue can therefore be counted multiple times, as in gaming.
The math of model pricing is harsher: a model with a cost of 1 sold at 10 has 90% gross margin; halve the price and margin remains 80%; if DeepSeek V4 reaches 25% of the original price, the same assumptions still produce 60%. Using DeepSeek Flash 0731 as the reference point, models that are weaker or more expensive fall into the “kill-zone” rectangle. Every vendor is therefore trying to “escape the kill zone”—either by becoming stronger or cheaper. Single-axis scoring may be flawed, but the direction is clear; DeepSeek may also cut prices sharply in the near term.

3. Coding Was Only the First Stop; Enterprise Is Becoming the Shared Bet

庄明浩 sees Codex becoming the new ChatGPT as a paradigm shift: download the ChatGPT desktop app and what you actually get is what used to be Codex, while the pure chat version, “ChatGPT Classical,” is harder to find. Agents are expanding from Coding Agent to Work Agent: “coding was first, now everything else.”
The Coatue chart he cited shows Copilot expanding from traditional models into finance, legal, security, biopharma, design, and enterprise services. According to market rumors, Cursor is also quietly roadshowing investors, with coding now accounting for 40% of revenue and enterprise taking its place. Cursor is rumored to target a $300B valuation by the end of next year; breaking $100B this year is likely not difficult.
Google, OpenAI, Tencent WorkBuddy, Qwen Office, and Doubao Enterprise therefore point in the same direction. The Claude client likewise splits its entry points directly into Chat and Work.

4. China’s Office AI Has Reached Million-Level Daily Activity, but Paid Conversion Is Still 0.1%

Analysys data show WorkBuddy at 20.97M visits through June, but that is an interaction count, not MAU or UV. QuestMobile puts China’s PC-client AI-app users at 18M, including legacy Doubao, Qwen, and Yuanbao clients; the current Work Agent push only really began around Q2.
WorkBuddy’s mobile app did not launch until July 18, and usage remains primarily PC-based. 庄明浩’s off-the-cuff estimate puts the upper bound for potential white-collar and enterprise productivity users at 100M-200M. Applying a 10%-20% coverage rate for users not yet reached gives a total of a few tens of millions. He stressed that the estimate was optimistic but relatively reasonable.
Market rumors of 12M-13M DAU were dismissed by his friend Will as “pretty far-fetched”; the source was simply a typo by an investment-bank analyst. The more credible figure is “a little over 1M.” LatePost reported a clear weekday-weekend gap for WorkBuddy, with average usage at the million level—meaning it is not a stable multimillion-DAU product.
Doubao began testing paid access to its professional tier in June. With roughly 300M-400M MAU, a 1% conversion rate would imply about 3M paying users; actual paid users over nearly 2 months numbered only “a few hundred thousand,” or roughly 0.1%. At a 100-200-person investor meeting, 庄明浩 asked who had bought the paid version of Doubao. “Not one hand went up. It was starkly revealing.”

5. AI Revenue Lands First in the Middle Layer, Making App Rankings “Very Boring”

Model vendors and enterprises cannot handle the entire training pipeline themselves, so AI training data, RL environments, inference services, and software infrastructure are becoming important middle layers. A group of US data companies may already have ARR above $1B and multi-billion-dollar valuations. Firms that build RL environments for clients or handle more granular execution for models and Agents are also growing rapidly.
The observation threshold for private AI companies has risen from last year’s $100M ARR club to $500M. Cursor, Scale AI, Surge AI, Mercor, Together AI, Sierra AI, Lambda, Replit, Perplexity, Glean, Lovable, ElevenLabs, Mistral, and Midjourney are all on the list, but most are still data, compute, or model businesses. Under a strict definition, only a handful are true apps, making the ranking “very, very boring.”
The sharper point is that once a sector or direction produces a high-revenue company, model vendors quickly take notice. This Top AI startups list may therefore be “a death list” in disguise: revenue validates demand while exposing the directions model vendors may target.
CB Insights’ Q2 Deal Count top 10 includes humanoid robots, world models, coding, legal, model services, Edge AI, image and video generation, personal AI assistants, AI for science, and GPUs. Under a strict definition, only coding, legal, AI photo and video, and personal AI assistants qualify as applications. China’s 10M-MAU rankings are again almost entirely dominated by major incumbents. ACE Studio’s shift from an independent music app to a music model plus Agent shows that the boundary between applications and models is still retreating.

6. Roblox Shows That Understanding the Narrative Still Doesn’t Mean Buying the Right Stock

During the 2021 metaverse boom, Roblox rose from the tens of dollars to above $100; after the bubble burst, it fell back to the teens and traded sideways for years. In 2025, improving generative capabilities, rising DAU and MAU, and the “AI will change games” narrative matured again, sending the stock from the $20s to around $150.
When Google DeepMind released its new world model GPT 3, the demo included generated scenes featuring Dragon Ball’s Goku and a Subway Surfers-like environment. Fears that “the game industry is over” drove Roblox down about 26% that day. 庄明浩 judged it “definitely an overreaction,” because the world model was still “far, far” from commercially disrupting games.
But the market did not trade on that fundamental reasoning: the stock’s downtrend continued, and the latest earnings report triggered another 33% or 37% one-day drop on declining DAU, sending it back to roughly $20. “You always feel you can see through many so-called narratives,” but a massive gap remains between understanding a narrative and trading it successfully in the secondary market.
Another set of numbers pulls the conclusion back to “only just beginning”: Luna, the next-lower-tier version of ChatGPT 5.6, is now available for free, while an a16z chart shows that only 2.2% of US households pay for AI tools. Jeff Dean’s BP reduces the business to a loop of experimentation environment, running, evaluation, feedback, and redesign. 庄明浩 calls it “everything is loop,” also potentially the industry’s biggest consensus under the logic of RSI-driven autonomous evolution.

Runta Founder 戴冠兰: Model Capability Is Good Enough—Compete on Infra Ep. 024 - SpaceX's 10GW Plan Drives $300B ARR by 2027 (Datacenter, Energy)