47. After July: Putting the Last AI Bull Through the Wringer! — 唐奕波 | 奔波儿r

Summary

• 唐奕波’s core bull case is to benchmark this AI cycle against electrification, not the internet: electrification capex ran at 2% of GDP for 15 years with no productivity gains in the first 10, while AI is already at roughly 2% in the US this year—“it has actually already surpassed the internet.” The key analogy is the radio: RCA rose more than 300x after 1920 on a “wrapped-electricity” application that would have been impossible to imagine in an era with only electric lights and streetcars. “Replacing humans is only the starting point… If you had a pile of very cheap geniuses, what would that world look like? At this point, you simply have no way to imagine it.” Raymond’s conclusion after repeated bear challenges: disagreeing with this framework is essentially “selling an AI call option.”
• Unit economics are the anchor for the entire debate: before GB300, a gigawatt cost roughly $40B—now the latest estimate is $60B—while a gigawatt deployed for Anthropic inference could generate about $60B in annual revenue at an assumed 60% operating margin; the same capacity for open-source models generates roughly $30B at a far lower margin. Global capex of about $1.2T corresponds to 30 GW; Tang’s relatively optimistic view is for 80%-100% growth to $2.5T next year, with hyperscalers accounting for half. He rejects using free cash flow to judge whether capex is justified: “Every industry starts with capex above operating cash flow, and only later recovers that capital.”
• Open-source models have merely blown up the incumbents’ excess profits, leaving frontier providers with 2 choices: cut prices to target medium and low intelligence, or “keep racing at the frontier”—and they “cannot win” a cost war with Chinese models at medium intelligence. Unlike solar, whose ceiling was the price of thermal power, AI is creating value: if drug-development success rates rise from 5% to 15%, “the frontier model may account for only 5% of usage, or even 2%, but ultimately capture 50% of the value.”
• After July, the market’s central question is not whether CSPs have enough money, but whether AI has a second use case beyond coding: coding’s TAM is only $1T, and Anthropic plus OpenAI could reach $130B-$140B in revenue by year-end—already 30% penetration. Dario’s answer is to redefine the denominator: the target is the $40T global white-collar economy, where penetration is only 0.5%. Raymond’s bear rebuttal: the Magnificent Seven generate less than $2T in combined revenue, so why should 2 new companies generate more than $1T? Tang’s response: “Your benchmark is still the internet, still the information economy.” The right comparison is the several-trillion-dollar global electricity market.
• Annual depreciation of $1T would require at least $3T-$5T in revenue under an asset-heavy model; if the model is less asset-heavy, the requirement could be $10T. Tang sees a credible path: the 2 companies should generate more than $200B by year-end; signed commitments imply roughly 22 GW in 2028 and close to $1T in revenue by the end of that year, with revenue reaching $3T in 2029-30—“that part is imagination.” He asks Raymond: “Do you simply think the number is big?”
• Organization design is an underappreciated variable: a 10x increase in individual efficiency may lift company output by only 20%-30% because communication friction is the bottleneck; a company-wide Agent like Claude Tag, paired with organizational redesign, could lift output by 8x-10x. Tang points to Block, which cut its workforce 40% this February, from 1,000 to 600, while maintaining strong operating results. His analogy: Token Maxing merely replaces the steam engine’s central shaft with an electric motor; real redesign looks like Ford decoupling small motors across every stage and cutting vehicle production time from 12 hours to 1.5 hours.
• The risk profile is asymmetric in 2 ways: back-solving from capex as a share of GDP puts the peak at roughly $4T, while the memory supply-demand imbalance “could be resolved by algorithmic optimization in the short term”—DeepSeek and Kimi have sharply reduced KV cache requirements, pushing memory’s share from below 10% to 20%-30%; “your ROIC is simply too high—everyone wants to take a swing at you.” The lesson from debt history is that leverage, not debt itself, is fatal: railways and power ultimately ended with equity-to-debt ratios of roughly 1:3, and “a 5% or 10% move can take you out.” AI today could withstand a 30%-40% move.
• AI has already changed his own research production function: instead of covering only 300 companies, he extracts the traits of winning stocks from a universe of 3,000 and then has AI screen for them. In March, using those criteria, he screened 800 names down to 20 in 20-30 minutes—No. 1 Taiyo Yuden, No. 6 Murata, No. 10 Yageo—and completed the MLCC research in 2 days. The key is not the scan itself, but the review and synthesis beforehand; Raymond estimates the process could generate $100K in returns for a $200 payment. Tang’s demand-side footnote: “I’m not paying enough right now. It’s extracting too little from me.”

Deep dive

1. “I don’t understand why free cash flow has to be used to judge whether capex is reasonable”

• Raymond set the backdrop first: several large US cloud companies are spending $700B-$800B on capex in 2026, with next year potentially exceeding $1T. Some are directing almost all of their free cash flow toward data-center construction. His question was where the money comes from and whether there is enough of it.
• Tang’s opening thesis was that in the early growth phase of a new industry, capex must exceed operating cash flow. In solar’s early years, no one argued the industry was unsustainable simply because capex exceeded operating cash flow; during the First Industrial Revolution, railways could not have covered their initial investment with operating cash flow from day one. From the perspective of CSP shareholders expecting a stable business model, stable cash flow, buybacks, dividends, and 15%-20% growth, the need for free cash flow is obvious. Change the lens, and the spending is entirely normal.
• Raymond’s bear case is that the market had previously treated the Magnificent Seven as mature companies. Deep pockets can absorb the spending, but once free cash flow is no longer sufficient and companies have to turn to capital markets, the scale of the funding need—combined with the fact that “they’re starting to borrow”—creates a generalized fear.
• On the math, before GB300 a gigawatt cost roughly $40B, so $1.2T of capex corresponded to 30 GW. Tang’s relatively optimistic view is for 80%-100% growth next year to $2.5T, with hyperscalers accounting for half. Of the $40B per gigawatt, roughly $10B goes to construction, power, and electrical equipment; the remaining $30B goes into servers. GPU’s share is declining but remains the largest component, with storage in second place.

2. The payback math on 1 GW: $60B for Anthropic versus $30B for open source

• Raymond was startled by the revised benchmark: “Sorry, it’s already at $60B?” GB300 consumes less power and fits more cards into each gigawatt, taking capex per gigawatt to $60B but also generating more revenue. Under the old $40B benchmark, one gigawatt used for Anthropic inference—API accounts for roughly 90% of its revenue—could generate about $60B in annual revenue at an inference gross margin above 80%. Assuming a 60% operating margin, that translates into more than $30B of annual operating profit: “The return is extremely high.”
• The same gigawatt deployed for open-source models generates 40%-50% less revenue, or roughly $30B. Tang admitted the estimate was imprecise: “I’ve discussed it with AI many times and still haven’t arrived at a precise number.” Operating margins are also far lower.
• Raymond reconciled the figures with third-party data: since June, closed-source token prices have fallen from just above $3 per million to $2.7-$2.8—“the experience is that Anthropic gives you Fable, while Codex keeps giving you reset, reset, reset”—while Chinese open-source pricing has risen from $0.7 toward $1. The gap is still more than 2x, consistent with the $60B-versus-$30B revenue split.

3. Open source destroys excess profits; the frontier has only 2 options

• Token prices are clearly falling: last week, OpenAI made its lowest tier free and cut the mid-tier price sharply. News reports said Anthropic would launch Sonnet 5 at the same $15-$25 Sonnet-tier price, with capability only 10%-20% below Opus. Tang does not see that as a problem. Anthropic could theoretically avoid building data centers altogether, rent cloud compute at $20B per gigawatt, and resell it: “My ROIC—I feel like it’s in the thousands or even tens of thousands.” In any industry, making that much money is “simply unreasonable.”
• Tang sees only 2 possible paths: cut prices for medium and low intelligence, or “keep racing at the frontier and make the frontier achieve things you cannot imagine—that is the only way out.” A cost battle with Chinese models at medium intelligence is “impossible to win”; the outcome would be the same as in every other manufacturing industry.
• Raymond worried about a solar-style outcome in which price falls, volume rises, and total revenue stays flat. Tang’s response is that the elasticity depends on what AI is replacing. Solar competes with thermal power, a total revenue pool of only a little over RMB200B. AI intelligence has already surpassed human capability: if the frontier “turns on the lights” and lifts drug-development success rates from 5% to 15%, the frontier model might account for only 5% or even 2% of usage while capturing 50% of the value. It is therefore too early to reach the same conclusion about AI as about solar.

4. Anthropic reaches SOTA with a fraction of the compute because OpenAI went first

• The mechanism behind the latecomer advantage is straightforward: training a model requires experimenting with 10 different paths before selecting one. Over the past 2 years, OpenAI tried multimodality, 2C, and other directions, then cut them one by one and found that “coding is the only broad highway to the peak of intelligence.” OpenAI “went down that road first for everyone.” The side effect was that OpenAI said multimodality was the direction of travel and “led Google into the ditch.” Google was still first-tier, and perhaps even No. 1, last October; it no longer is.
• China’s lower costs stem from architecture innovation under constraints. Kimi and DeepSeek’s papers contain substantial innovation—not in training a stronger model, but in testing whether roughly 90% of the performance can be achieved with 1/10 of the cards and 1/10 of the memory. In inference, the question is how to use 1/10 of the KV cache to achieve 90% of Anthropic’s capability. “You definitely cannot reach 100%, because all compression involves information loss.” The frontier providers have already done much of that exploration for them.
• Raymond’s extrapolation received Tang’s agreement: US and Chinese intelligence may run in parallel, but “whoever wants to take one more step forward has to spend another $100B.” China may spend first on video—Seedance 2.0 and MiniMax’s H3—after which the next generation of Gemini could be cheaper once the path has been mapped. Tang’s view: “Waste is sometimes necessary; it is the cost of exploration.” Eight or 9 out of 10 R&D bets are wrong.

5. The central problem is not money but “no second use case”

• Whether CSPs have enough money is only the surface issue. The market’s deeply rooted concern is the same as in the second half of last year: demand. Coding is an unusually favorable use case—the users are programmers, and penetration can rise quickly—but its TAM is only $1T. Anthropic and OpenAI could reach $130B-$140B in combined revenue by year-end, implying 30% penetration. “It should already become a mature industry.” What happens if there is no second use case next year?
• Anthropic and OpenAI could IPO in the second half of this year. Dario’s response is to redefine the denominator: “Our target is not coding, and it’s not software. Our target is every white-collar worker in the world”—a $40T market, which implies only 0.5% penetration today. Tang says Dario’s job is to convince the market that the denominator is not a single use case. 梁文锋 has likewise told investors that AI could eventually account for 20%-30% of global GDP.
• Raymond’s rebuttal is worth preserving in full: global SaaS revenue is more than $300B, including $180B in the US; the Magnificent Seven generate less than $2T in combined revenue—so why should 2 new companies create value equal to that of all 7? Tang acknowledged that revenue “cannot really be counted twice,” but said the benchmark was wrong: “You’re still benchmarking against the internet, still against the information economy.” The internet improves the flow of information; the proper comparison is the several-trillion-dollar market for global electricity consumption multiplied by the price of power.

6. A 10x increase in employee productivity may lift company output by only 20%-30%

• Tang’s diagnosis is that Claude Code is fundamentally a personal assistant. Once embedded in enterprise workflows, the bottleneck becomes communication friction between people: a 10x increase in each individual’s efficiency may produce only 20%-30% more company output. Claude Tag is the turning point. Anyone in a Slack group can @ it with a task; it reads the group chat and learns the company’s entire workflow, “turning a personal-level Agent into a company-level Agent.” If the organization is rebuilt around that structure, Tang believes output could improve by 8x-10%, rather than 20%-30%.
• Raymond remains skeptical. Silicon Valley’s Token Maxing experiments have not produced visible productivity gains in earnings reports, and layoffs have not been large: “Someone invited a Pokémon into the group chat that knows your company’s expense policy—so what?” Shopify’s CEO allows Tag only in public channels and prohibits direct messages to preserve the full context.
• Tang’s case study is Block, which cut its workforce 40% in February, from 1,000 to 600, while operating and maintaining the business with “very good” results. The dividing line is whether a company is “actually rebuilding its organizational structure around the Agent.” His analogy: Token Maxing is like replacing the central shaft that drove a factory with an electric motor, nothing more. The real breakthrough is Ford decoupling small motors across every stage and taking the output time for a car from 12 hours to 1.5 hours.

7. Finding an overlooked MLCC name in a universe of 3,000

• Over the past 6 months, with 4.6 as the dividing line, the real change in Tang’s work was not faster screening—cutting the process from 2 or 3 days to 10 minutes—but the radius of coverage. An individual analyst can cover at most 300 companies. Tang distilled the traits of last year’s winners, including Delta Electronics, optical-fiber names, and Mitsui Kinzoku, which rose 10x-15x after June 2025: AI exposure below 10%, an industry that had been bad for years, low valuation, a 3-year gain of no more than 150%, and a non-Tech profile. He had AI screen roughly 800 global names and identify 20 in 20-30 minutes: No. 1 Taiyo Yuden, No. 6 Murata, and No. 10 Yageo. In March, the MLCC project took only 2 days from “what exactly is an MLCC?” to a completed report.
• The details provided a reverse confirmation. An electronics investor told him, “What a garbage industry MLCC is,” that price increases were impossible, and that Chinese manufacturers were leaving equipment in warehouses because they were afraid to take it out and use it in case it had to be capitalized. “I was even more excited after hearing that.”
• Raymond argued that even by August, diffusion remained insufficient: if a skill is already used by many people—a 1M-follower audience and 250,000 GitHub stars—it should theoretically have been fully mined and priced. Tang’s answer is that the difference between scanning with AI and without AI is enormous. But if everyone has AI, “the review and synthesis beforehand become the most important part.” He asked himself whether that earlier success was pure luck or whether there was “some element of it.”
• Raymond also ran the demand-side math: the process could generate $100K in returns while requiring only a $200 payment. Tang’s demand-side footnote was blunt: “I’m not paying enough right now. It’s extracting too little from me.” He is willing to pay more for a better product.

8. Building 2 Three Gorges projects a year: the bet is on TAM

• Raymond estimates that capex could reach $1.2T-$1.3T in 2027 and $1.3T-$1.45T in 2028. That means more than $1T a year going forward—roughly 1.5 to 2 Three Gorges projects annually, or a 22 GW buildout. With short depreciation lives, annual depreciation could reach $1T in 2029-30. If this is an asset-heavy industry, revenue would need to reach at least $3T-$5T; if it is less asset-heavy, the requirement could be $10T.
• Tang’s response: “It’s very simple.” The 2 companies should generate more than $200B by year-end; a 10x increase gets to $3T. This year’s 2-3 GW, and roughly 22 GW in 2028 based on signed commitments, are still conservative. “By the end of 2028, revenue will be close to $1T.” Growth will then slow, and $3T in 2030—“that part is imagination.” He asks again: “Do you simply think the number is big?”
• Raymond worries about a scenario in which everyone uses AI, Anthropic has 3B daily active users, but its market cap and revenue do not change. Tang says, “I cannot imagine it, but I think it’s possible.” There will eventually be a wall; at that point, “you won’t care whose token it is,” because the product will have been commoditized. “But even if it is commoditized, I don’t think that gets in the way”—the product of electricity prices and electricity consumption would still be a several-trillion-dollar market, comparable to oil and coal. He also warns that electricity is not frictionless: rising US power prices have already triggered protests, Trump introduced a dual-track power system in PJM, and New York State drove out another data center.

9. “Debt is not bad; high leverage is”; algorithms may resolve the memory squeeze

• The debt review is consistent across railways, canals, and power: equity ultimately accounted for roughly 1 part to 3 parts of debt, while canals were closer to 1:1. The bubbles ended in runs, with equity wiped out first. Debt itself was initially healthy—“that is why capitalism, stock markets, and financial markets exist.” The fatal problem was leverage: “A 5% or 10% move and you’re out.” AI can currently absorb a 30%-40% move. If stocks fall 30%-40% this time, only the levered players go bankrupt; “the people who need to buy servers are still buying.”
• Raymond also back-solves the capex peak from its GDP ratio to roughly $4T. Companies can use free cash flow or equity as collateral to borrow more, roll the debt forward, scale up, and generate more revenue—a case of “the left foot stepping on the right foot.”
• The paradox is memory. Tang is extremely bullish on the broader cycle but uncertain about memory. DeepSeek and Kimi pursued linear attention and other architecture innovations precisely because they lacked enough memory. “Foreign companies don’t want to do this—they’re still exploring the frontier, and they’re not short of money, so they don’t need to economize.” Memory’s share has risen from below 10% to 20%-30%: “Your ROIC is simply too high; everyone wants to take a swing at you.” But the next paradigm, such as Continual Learning, could lift demand again. “It’s too difficult to judge.”
• Raymond pointed out that the same logic applies to Nvidia. Tang conceded the point: Nvidia’s 50% share makes the mix especially exposed, while GPU’s share is being squeezed by memory, optimization, and TPU. “That’s because you’re too good. If you weren’t that good, you might not be getting squeezed.” That, in itself, is the cycle.

10. Electrification and the radio: replacing humans is only the starting point

• Raymond summarized the core bet this way: if AI is only narrow coding and the market is already fully penetrated, “we might as well buy hogs.” If it is a vast new frontier with penetration still at only 0.5%, investors should keep betting on compute and memory.
• The reason to use electrification as the benchmark is that its capex ran at 2% of GDP for 15 years, with no productivity gains in the first 10. At night there was only lighting; during the day there were only streetcars. Total-factor productivity contributed just 1% to GDP. Internet fiber accounted for only 1.3%, and investment lasted 4 years before the market realized that much of what had been built was not being used. AI is already roughly 2% of US GDP this year—“it has actually already surpassed the internet.” The internet improved information efficiency; electricity was infrastructure that enabled assembly lines and an endless range of appliances. “AI is intelligence,” so its scale should be comparable to power.
• The radio example is the cleanest illustration. RCA rose more than 300x after 1920. “Radio was really a package for electricity, electricity in a wrapper.” If all you could see from 1900 to 1910 were electric lights and streetcars, the radio would have been impossible to imagine. “The idea of replacing humans is too simple. Replacing humans is not the limit of AI… If you had a pile of very cheap geniuses, what would that world look like? At this point, you simply have no way to imagine it.” Tang also acknowledged that he may have “presupposed the conclusion” by choosing electrification as the comparison.
• Raymond’s closing formulation was that disagreeing with Tang essentially means denying the possibility of a major AI breakthrough in the next 6 months: “I’m really selling an AI call option… It may be worthless, and it may be valuable. Thinking about it this way, I find it a little frightening.” Tang cited 汪天凡’s acceleration thesis: civilization cycles of 3,000 years, 300 years, and 30 years, with the fourth potentially lasting only 3 years—“it is possible that the cycle we are facing will be completed in 5 to 10 years.” Raymond’s response: “Then let’s cherish today’s opportunity.”