AI 李时珍 Takes You “Tasting a Hundred Herbs” — Talk at Xiaoyuzhou’s Xiansheng Event — Solo, 91-Page PPT

Summary

• CapEx is the most frequently recurring keyword across these 76 charts and the AI throughline 庄明浩 returned to repeatedly. Microsoft, Google, Meta and Amazon together spent about $73.3B after their Q2 2022 earnings; that figure reached $331.5B in 2025, versus $222.4B in 2024. Add Oracle and SpaceX, and the six companies are expected to surpass $1.132T in 2027. The CapEx of the 5 biggest spenders has risen from about 1.7% of US GDP to nearly 2.8%. His refrain: “Every time we set our CapEx expectations, we have basically underestimated them over the past 3 years; every earnings release sends the number higher.”
• The comparison with the 1999 internet bubble has become impossible to ignore: internet CapEx was only 1% of GDP then, versus 2.4% for AI today and a projected 3.1% in 2027. The warning charts include Oracle’s CapEx/Sales at 85%, some vendors’ free cash flow “at the breaking point,” Bloomberg’s August 13 update of the circular-financing chart, and the block meme in which Oracle, JPMorgan, Nvidia, SpaceX and Microsoft all sit on “two extremely frail legs,” OpenAI and Anthropic. Whether this produces a 2000-style correction, “nobody knows; this is only a rough timing comparison.”
• The industry’s gross-margin structure runs counter to consensus expectations: chips are at about 41%, energy at 24%, compute and cloud at 11%, while model and application companies are at negative 59%. Everyone expected the layer closest to the user—models and applications—to have the highest margins, “but that now looks a little wishful.” The inversion lines up with the sharp H2 2025 split between a software ETF being sold and an AI ETF being bought.
• The “kill line” has become a new pricing coordinate system: DeepSeek-V3-Flash-0731 drew a rectangle with extremely low pricing and extremely high scores, and “all models theoretically inside it seem to have no meaning anymore.” The leading vendors’ job became “escaping the kill line”; after V3-Pro-0813 launched in mid-August, “the line got even harsher.” On Bloomberg’s pricing chart, DeepSeek was so cheap it looked “as if nothing had been drawn,” earning it the “梁胜” chart moniker—but he warned that “these charts have an extremely limited shelf life”: by September, “良胜 sometimes becomes 梁子.”
• The time it takes a local model to catch up with SOTA has compressed from 33 months to under 9 months, which explains the surge in discussion around edge and local deployment. GPT-3 to Llama 33B took 33 months; GPT-4 to Qwen 2.5-32B took 18 months; Opus 4.5, released in November 2025, to Qwen 3.5-27B took under 9 months. His extrapolation: in another 7 to 8 months, a local model under 30B could match the capabilities of the current leader, Grok 5.
• The OpenAI-Anthropic ARR narrative reversed in Q3: in H1, the consensus was that “Anthropic was growing far faster than OpenAI,” but Ramp data through August 17 showed OpenAI’s Q3 ARR growth moving back ahead of Anthropic’s. Anthropic’s valuation marched from $61.5B to $183B, $380B and $965B, with a reported IPO target of $2T. The ARR data he clipped showed $47B in May, with annual expectations of $100B-$120B. On the OpenAI side, agentic tokens already represented 64%; Codex active users rose from 15M on August 12 to roughly 25M at recording. “You can’t not update—wait a month and every number has to be updated again.”
• His story for the quarter is Sequoia’s “Own Your Intelligence”: vendors providing training data, RL environments and post-training services are seeing revenue and valuations rise together, and the trend has “carried over to China”—“you realize this thing can connect everything.” About 14 of Brex’s 25 fastest-growing software companies are AI infrastructure vendors and about 6 are applications; vertical AI companies generally trade at ARR/valuation multiples in the tens and sometimes above 100x. VC exit concentration shows that from 2023 through 2026, the top 1% of companies captured 80% of exit value—miss the leaders, and you have no exposure to 80% of the industry’s returns.
• The closing meta-question is the one most worth hearing: if you view his information work as a collect-organize-process-output loop, today’s AI, even with all his context, “probably still can’t do it—or can only get 60% to 70% of the way there.” Taken to the extreme, “everyone will have their own model, their own agent and their own harness.” He used the circular clothing rack at a Nike factory store to explain why the process must be reset periodically, and closed with 张玮玮’s “Silver Hotel”: “I can’t wait to put this album out, because only then can I say goodbye to it. Next up, a long road over mountains and rivers.”

Deep dive

1. Experiment design: no organizing, just run through the first 20 days of August as-is

• The setup was the August 22 information session at Xiaoyuzhou’s Shanghai Xiansheng event, focused on how to distinguish real from fake in an age of information overload. 庄明浩’s plan was to “walk through every chart I saw over roughly the first 20 days, without organizing or structuring anything—just list them,” and let the audience see whether a framework or narrative would emerge on its own. The deck ran to nearly 90 pages, 76 of them charts.
• The raw material came from his daily “high-speed surfing” across Substack, Twitter, Feedly RSS, Weibo, Jike and Zhihu: “AI is responsible for shocking the world every day; I’m responsible for saving those shocks to my database.” His August 2026 folder alone held 113 visual charts—5 per business day—and about 30 PDF reports.
• The hidden thread was the question 任鑫 posed on the podcast: each slide in 庄’s deck had one theme paired with an apt chart—“When will AI be able to do this?” The question gets answered directly at the end.

2. The CapEx throughline: from $73.3B to more than $1T in 2027, expectations are always too low

• The first chart was CapEx: Microsoft, Google, Meta and Amazon together spent about $73.3B after their Q2 2022 earnings, rising to $331.5B in 2025 from $222.4B in 2024. His longstanding view: “Every time we set our CapEx expectations, we have basically underestimated them over the past 3 years; every earnings release sends the number higher.”
• The basket kept expanding: after first adding Meta and Oracle, the August 10 version added SpaceX, bringing the total to 6 companies and projected combined 2027 CapEx of $1.132T. The CapEx of the 5 biggest spenders rose from about 1.7% of US GDP to nearly 2.8%.
• The funding question is already surfacing: some vendors’ free cash flow is no longer enough to cover CapEx, leaving them reliant on expectations for “even bigger revenue” in 2028 and 2029. “Will it be as ideal as expected? We’ll see.”

3. Cloud is the clearest monetization path, but warning indicators are flashing

• The 3 major clouds should account for 63% of the market—AWS at 28%, Microsoft Cloud at 20% and Google Cloud at 15%—with Q2 revenue of $143B and 43% growth. The second tier consists of CoreWeave, Oracle and Nebius. His repeated point: “The most direct manifestation of this AI wave is the revenue growth of the cloud vendors.”
• But CapEx/Sales has entered dangerous territory: 85% at Oracle, 60% at Meta, 53% at Microsoft, 40% at Google and 28% at AWS. Software ETFs and AI ETFs “split dramatically” in H2 2025: the market bought AI software and sold traditional software. The S&P’s top 10 gainers through August 14 included 3M, Dell, Seagate, Micron, Intel, Western Digital, Marvell, Lumen, HP and AMD—all hardware companies he classifies as AI-related.
• Gross margins have inverted: roughly 41% for chips, 24% for energy, 11% for compute and cloud, and negative 59% for model and application companies. Everyone expected the layer closest to the user to have the highest margins, “but that now looks a little wishful.”

4. The kill line and “梁胜”: DeepSeek redraws the industry’s coordinates

• The kill line uses price on the horizontal axis and leaderboard score on the vertical axis. DeepSeek-V3-Flash-0731 was both cheap and highly rated; drawing a rectangle around it meant that “all models theoretically inside it seem to have no meaning anymore.” For leading vendors including GPT, Claude, Kimi and GLM, the job became “escaping the kill line.” After V3-Pro-0813 launched in mid-August, the new line was “even harsher”; fewer and fewer models cleared it, and “almost all of them were bought up.”
• He called Bloomberg’s Token-pricing chart the “梁胜” chart. One of the most expensive models charged $50 for input and $50 for output; another charged $10 for input and $50 for output. DeepSeek’s line was so thin it looked “like someone had barely marked a bit of black ink on the axis.” But the chart’s lifespan is limited: “after a while, a new narrative will cover it up,” and by September, “良胜 sometimes becomes 梁子.”
• Across the 3 lines tracking average Token prices, closed-source model pricing is “falling crazily,” open-source model pricing is “edging up slightly,” and the average price continues to decline.

5. SpaceX enters the field and the giants play “StarCraft”

• SpaceX’s CapEx reached $28.5B in H1 2026: $23.6B went to xAI, versus just $2.2B for the rocket business and $2.7B for Starlink. On the revenue side, Starlink generated $4.29B and was profitable; xAI generated $2.56B and was loss-making; Space generated $0.962B and was also loss-making. “One is the here and now, the other is poetry and distant horizons.” With a $1.7T valuation, SpaceX entered the top 10 global companies; in the August 8 ranking, Nvidia at $5.4T, Apple at $4.5T and Google at $4.3T were ahead.
• Among the listed-company holdings disclosed after Nvidia’s Q2 earnings, Intel ranked first, SpaceX second and CoreWeave third. Earlier, he had recorded SpaceX as number one, then corrected himself.
• Sequoia’s David Cahn, author of “The $200 Billion AI Problem,” “The $600 Billion AI Problem,” “The $800 Billion AI Problem” and “The $1 Trillion AI Problem,” compared OpenAI, Google, Microsoft, Amazon, Nvidia, Apple and SpaceX to “StarCraft”: “Every race has its own talents.” Jeff Dean’s 4-page fundraising deck, built in Google Docs when he left Google to raise capital, stated the vision “Automate ML, Science and Engineering for Impactful Problems.” 庄 regarded it as “one of the best ways for a top founder to articulate a fundraising pitch.”

6. Bubble comparison: 2026 AI is being measured against the 1999 internet

• The historical CapEx/GDP comparison is stark: AI in 2024 and the internet in 1997 were both at 0.8%; AI reached 1.4% in 2025, versus 0.9% for the internet in the corresponding year; and in the 2026-versus-1999 comparison, the internet was only 1% while AI is already at 2.4%, with 2027 projected at 3.1%. His wording remained restrained: “Nobody knows whether there will be a major correction like the internet bubble in 2000.”
• Structural fragility shows up in Bloomberg’s circular-financing chart updated August 13 and in the block meme: Oracle, JPMorgan, Nvidia, SpaceX and Microsoft—possibly AWS as well—are all stacked on “two extremely frail legs,” OpenAI and Anthropic. “The revenue of those 2 companies supports this entire AI economy.”
• The real-world comparison is the investment curve for data centers versus offices: after the lines cross, the gap keeps widening. An aerial view of Tesla’s new chip factory is set against the Pentagon and Apple’s headquarters—“possibly dozens of times larger than them. My God.”

7. Model supply is accelerating at breakneck speed; local models catch SOTA in 9 months

• OpenRouter’s monthly count of new models shows just 28 in the peak month of 2024, August—less than 1 per day. The count hit 50 in April 2025, then 80 in April 2026 and 70 in July, or more than 2 per day. “The release frequency of star models is accelerating crazily.” A China-US open-source comparison shows that Chinese models “remain genuinely very strong” on both release frequency and parameter count.
• The chart he valued most tracked the time for locally deployable models to catch SOTA: GPT-3 to Llama 33B took 33 months; GPT-4 to Qwen 2.5-32B took 18 months; and Opus 4.5, released in November 2025, to Qwen 3.5-27B, released this month or last month, took under 9 months. His extrapolation: in another 7 to 8 months, a local model under 30B could reach the capabilities of the current leader, Grok 5. That is also why discussion of edge models has suddenly heated up.
• The hardware knock-on effect is visible in the storage-cost share curves of the 4 biggest CapEx spenders: every line is moving higher. After its IPO, ChangXin became China’s largest listed company by market capitalization; in his recollection, as of August 13, Tencent was second. The AI storage-share chart includes Samsung, Micron, ChangXin and SanDisk.

8. OpenAI vs. Anthropic: valuation sprint and a Q3 growth reversal

• Anthropic’s valuation path was $61.5B in 2025, $183B in September 2025, $380B in February 2026 and $965B in May 2026; recent reports put its IPO target at $2T. ARR estimates vary widely, with versions above $70B and above $60B, but the chart he clipped ran only through May at $47B, against annual expectations of $100B-$120B. On an API-call revenue basis, Anthropic’s share “may exceed 60%.”
• But the narrative has reversed. The consensus in H1 2026 was that “Anthropic’s revenue was growing faster than OpenAI’s—and by a lot.” Ramp data through August 17 showed OpenAI’s Q3 ARR growth moving back ahead of Anthropic’s.
• OpenAI’s internal mix has shifted sharply: after Codex’s rise, agentic tokens accounted for 64%, leaving pure Chat at 36%. Codex active users stood at 15M on August 12, reached 20M on the day of the talk and were around 25M at recording. “You can’t not update—wait a month and every number has to be updated again.”
• The public-markets-style windfall in the private market is CoreWeave’s confirmed $60B acquisition by SpaceX. Its angel investors could see returns of more than 3,000x; OpenAI participated in its angel and seed rounds, with a return of roughly 500x to 800x. Nat Friedman’s angel investments returned 400x to 1,000x, with potentially more than 100x on Series A and several times even in the Growth round—all over very short periods.

9. His quarterly story: Own Your Intelligence and the private-market Matthew effect

• Asked what story he would tell if he were doing a quarterly roundup now, he chose Sequoia’s “Own Your Intelligence.” Every application company can be seen as part of the so-called New SaaS; vendors providing training data, RL environments and post-training services are seeing revenue and valuations rise together, and the trend has “carried over to China.” “You realize this thing can connect everything.” Supporting evidence: about 14 of Brex’s 25 fastest-growing software companies are AI infrastructure vendors and about 6 are applications, closely matching Ramp’s August rankings.
• The CSET 30 heat ranking, as read aloud, put Anduril first and OpenAI third; second place was described only as a US “DJI for defense.” He later said OpenAI had been basically number one throughout last year, while Anthropic was basically number one this year. The list also included Databricks, Stripe, Polymarket, Lambda Labs, Cluely and Neuralink. For vertical leaders such as Ramp, Legora and Harvey, ARR/valuation multiples “could average several dozen times, with some even above 100x.”
• The cold shower comes at exit: from 2023 through 2026, the top 1% of companies captured 80% of total exit value, while the remaining 90% received only 8%. “If you didn’t invest in a top-1% company, you have nothing to do with 80% of the potential returns in this industry.” Hugging Face’s 4-quadrant Loop shows why code and math are the most representative cases: they are easy to edit and easy to verify.

10. The closing meta-question: if I were an agent, what would my loop be?

• After going through all 76 charts, he asked several questions in succession: which keywords did you retain—CapEx, ARR, Human Intelligence, valuation? Which charts do you want to save? Which are already obsolete, and which contradict one another? By September 2027, 1 year later, which will still be worth keeping? “Everyone will have their own answer.” Some people remembered only the memes; “I don’t think there’s anything wrong with that.”
• Returning to 任鑫’s question: can today’s AI, using Skills and even having all his context, produce this deck? “I think it probably still can’t—or can only get 60% to 70% of the way there.” He abstracts his own work into a loop—collect, organize, process, output on a schedule, then start over—because “taken to the extreme, everyone will have their own model, their own agent and their own harness,” and everyone will need to think through what their own loop should look like.
• Why the process must be repeated: he compares himself to the circular clothing rack at a Nike factory store—“I keep hanging clothes on it, but once it reaches a certain point I can’t hang any more. I have to take everything down before I can start hanging again.” He closes with 张玮玮’s “Silver Hotel”: “I can’t wait to put this album out, because only then can I say goodbye to it. Next up, a long road over mountains and rivers.”

Verification Notes

• The reference text is internally inconsistent on who topped the CSET 30: earlier it lists Anduril first and OpenAI third, while later it says Anthropic was basically first this year. The discrepancy has not been forcibly reconciled.