Summary

奔波儿 is currently most bullish on China’s domestic compute stack, putting the odds at “probably above 50%” with attractive upside relative to downside. The thesis is that domestic models will cross a usability inflection point—similar to Claude 4.6, which went from basic coding to coding, Agents, and long-running tasks—and that most users remain constrained by account bans and network-access barriers. If China’s major platforms bring the product threshold down far enough, demand could explode, but that explosion will take time. The stocks have already sold off, making domestic compute at least worth watching.
He believes the commercial model for domestic inference has already been proven, making continued capex increases by Chinese vendors the likely outcome. He cites 梁文锋’s comments that DeepSeek recovered its costs in just 10 months despite very low pricing: “I buy GPUs, make money, then use the money to buy more GPUs and train again.” Supply-side evidence includes 智谱’s Coding Plan selling out “like train tickets,” while Kimi’s Coding Plan also became unavailable almost immediately. Prices for domestic models are rising and demand is strong, although the ultimate scale of that demand remains to be seen.
奔波儿 sees no clear slowdown yet in AI ARR growth and says the binding constraint is GPUs, not demand; 星辰 flatly disagrees. On his figures, Anthropic is at more than $70B of ARR and OpenAI at $40-50B, with each still adding more than $10B a month; renting 1GW of data-center capacity costs roughly $50B, and rental prices for older GPUs have yet to fall. 星辰’s response: “The US is already spending $800-900B a year buying GPUs, and you’re still telling me there aren’t enough? I don’t buy it.” 奔波儿 does not expect the issue to be settled today and suggests revisiting it in 1-2 months.
AI may be at the electrification stage, with capex at roughly 2% of GDP—perhaps only at 1919, or not even there yet. The penetration rate depends on the denominator. On coding alone, with global programmer compensation at roughly $1.2-1.3T, penetration is already around 20% and approaching the late stage. Against the roughly $50T global white-collar wage pool for all tasks that can be done on a computer, it is still very low—about 2% using the US white-collar pool. Anthropic’s usage data shows coding at roughly 35%, with other use cases already showing penetration ranging from 5% down to 1%.
He is no longer focused primarily on second derivatives, instead watching narrative, valuation, and stock-price reaction. US-listed GV has an order-to-revenue ratio above 3x, orders growing 100%-200% year over year, and prices up 50% year over year, yet the stock fell 8%-9% after the company said it would expand capacity by 25% over 3 years after 2028. The market may already be pricing it at 25-28x 2028 EV/EBITDA. Memory stocks are being hit by the short-term, hard-to-falsify narrative that algorithms will “take a bite out” of memory demand; 星辰 warns that memory names trading at high-single-digit PEs could become value traps.
The other line worth watching, in his view, is open-source models paired with applications. Closed-source models once captured most of the economics at 80%-90% gross margins. Once open-source models drive prices down, profits could be redistributed to open-source models and applications. When Cursor uses the best available model, almost all the revenue goes to Anthropic and Cursor’s own gross margin approaches zero; simple translation and full-document summarization do not necessarily require Opus. He is not bullish on every application, but the group was badly punished by the “models will eat everything” thesis, leaving some names with better risk-reward.
The biggest disagreement on the show is whether coding’s success can be extrapolated to a much broader set of use cases. 星辰 worries that the classic bubble mistake is extrapolating success in one direction into a second and third, and notes that AI’s historical breakthroughs may be separated by long desert periods. 奔波儿 counters that Agents appeared only about 6 months ago, Scaling Law is still advancing, and AI for Science, finance, law, and management are progressing in parallel; there is no reason to assume an abrupt stall.
In practice, he breaks investment research into five stages: idea generation, information gathering, rapid and deep research, decision-making, and feedback tracking. AI can lead the information and monitoring work—for example, automatically tracking Nordic data-center companies, New York State legislation affecting data centers above 50MW, and European local governments’ land sales. Humans still make the decisions, but rules such as “don’t chase” and a PE ceiling can be delegated to AI to keep impulses in check. Against minute-level quantitative factors, ordinary investors should rely on longer time horizons and smaller samples, while watching inflection points such as Kimi K3 that could reset pricing and volatility.

Deep dive

1. Opening exchange: 星辰 puts 浩哥’s “conviction” through the wringer

星辰’s cross-examination: HBM is in shortage through 2030; “memory is no longer cyclical, it’s a matter of faith”; every day brings claims that institutions have no position, retail investors have not boarded, and a historic primary uptrend is coming. “So why is it that the moment I build a position, I start studying PE, cyclicality, and risk?” Down 2%, he asks whether to cut; down 5%, he says to respect the market; down 10%, he says he has to rethink cash flow.
浩哥’s confession: “My faith in the industry is still as hard as my mouth, but my account curve no longer supports being that hard.” So he brought in the author behind the public account “奔波儿,” whom he repeatedly reads whenever his own conviction starts to waver, to replenish the faith.

2. Guest profile: an eighth-year researcher who moved from renewables to AI hardware

奔波儿 says he has covered renewables, Chinese manufacturing, power, and cyclical industries, and shifted to AI hardware this year. His public account began with tutorials on using Claude Code for investment research, followed by a retrospective comparing AI with electrification.
浩哥’s endorsement centers on the article about “investment taste”—simple questions versus complex questions—and its attack on traditional and classical value investing: “Value investing is not a fig leaf for refusing to change, and it is not a shield for your losses.” “Everything in it was something I had been thinking.”

3. Trading stocks in the AI era: information is flattened, so identify the phase before picking the trade

奔波儿’s starting point: TSMC’s earnings-call transcript is available almost immediately after the call, and investors can layer on leveraged long ETFs, so price discovery happens extremely fast. The key question is now “what stage is AI in, and which directions should I watch and invest in at that stage?”
He adds that quantitative funds, self-media, and AI are converging in a three-way wave, producing heavy overlap in stock selection. Both bullish and bearish catalysts can be realized almost instantly—“they can hit you with several limit-down sessions like 德明利, leaving you unable to get out”—making investing increasingly difficult for ordinary people.

4. The denominator fight: “Defining the denominator is defining what stage AI is in”

The framework: AI capex at roughly 2% of GDP is comparable with electrification. The most comfortable phase of a technology revolution is 5%-10% penetration: “The total market is already large, you still expect it to double every year and octuple in 3 years… supply is permanently short and profits are permanently exploding.”
On the numerator, Anthropic plus OpenAI will “definitely” exceed $200B of ARR by year-end. If the denominator is only coding—$550B of US programmer compensation and roughly $1.2-1.3T globally—penetration is already around 20% and in the middle-to-late stage. If it is “all tasks that can be done on a computer”—$10T of US white-collar compensation and $50T globally—penetration is around 2% using the US wage pool. If the value created by AI for Science is included, such as improving drug-development success rates, the denominator becomes larger still.

5. No Jevons-paradox debates: the premise is the conclusion; only take the direction that is 80% right

“If someone wants to debate Jevons paradox with me, I simply won’t debate them.” Both bulls and bears can make a case, but the premise—whether latent demand has peaked—is itself the conclusion, so neither side can persuade the other. “You can be 40% right and I can be 60% right; that means nothing. I want to be 80% right.”
The example is whether overseas compute is bullish, which is not worth debating: “You know Kimi’s emergence is definitely bullish for domestic compute.” It is not a one-off event; domestic progress will keep appearing. “You didn’t believe the first one, you didn’t believe the second—what about the third, fourth, and fifth?” Short-term selloffs can instead reflect ETF flows, leverage, and other trading mechanics.

6. Two high-certainty directions: open source and applications take share from closed-source profits; domestic models clear the usability bar

First thesis: in the previous phase, “closed-source models ate everything,” with gross margins of 80%-90%. Open source pushes prices down and releases profits back into the ecosystem. Cursor is the example: users want the best model, so almost all the money goes to Anthropic and Cursor’s own gross margin approaches zero, forcing it to train its own coding model. “For simple translation and full-document summaries, I really don’t need Opus.” Applications were devastated by the “models will eat everything” thesis; for some names, even if the bet is wrong, “it looks like you won’t lose money, or at least won’t lose much.”
Second thesis: once domestic models reach 智谱 5.2’s level, “that’s good enough; if it gets much better, people like us can’t even make use of that intelligence.” “Before, it was an elementary-school student. Now it’s a college student. If you’re a PhD, I can’t afford to hire you; I’ll hire the college student.”
星辰 unexpectedly became an ally. He has long questioned the sustainability of the overseas chain but agrees with the direction of domestic models and applications, adding that cybersecurity stocks have already broken out. More capable models make hacker attacks more dangerous, potentially forcing companies to raise security budgets. 浩哥 offers the opposite warning: AI applications remain a “super-hard problem,” so investors should be careful.

7. Domestic inference has a working business model: against a 10-month payback, capex is likely to rise

奔波儿 thinks a turning point may emerge from Q3 this year. Domestic players previously lagged overseas operators badly in both investment and monetization, but Anthropic has already established the To B API model. 梁文锋’s comments on DeepSeek said it recovered its costs in “just 10 months” despite very low pricing: “I buy GPUs, make money, then use the money to buy more GPUs and train again, improving intelligence. That model already works.”
His conclusion is that domestic vendors will probably increase capex continuously, as overseas players did over the past 2 years. Evidence includes the persistent strength of the supernode-related sector and the heavy attention attracted by Huawei’s 950 supernode at the Shanghai WEC conference.

8. 星辰 questions end-user ROI as supply-side evidence points to a buying frenzy

星辰’s concern is that DeepSeek is not the end user, and B-side research may not show especially high ROI. Demand is still “carbon-based”: the benefit comes from replacing carbon-based employees with silicon-based GPUs. Internet giants could theoretically cut 60%-70% of their workforce without much trouble, so how much of the growth is trial demand or performative demand?
One response comes from a ByteDance contact: people who once built ad-delivery systems are now building “AI ad-delivery systems” and embedding their workflows into them. The redundancy may fundamentally reflect friction costs in large organizations. “Two AI Agents go one-to-one, context one-to-one”; eliminating an existing organization in one stroke does not mean nothing changes.
The second response is from the supply side. 智谱’s Coding Plan sells “like train tickets—you have to fight for it at noon every day,” while Kimi’s Coding Plan also became unavailable almost immediately, in roughly 2 days. “If end demand weren’t there, who would pay 5x the price for a new model?”

9. Is this just extrapolating one success to every direction? Usage data says coding is not the whole story

星辰 says many historical bubbles broke because investors linearly extrapolated success in one direction across multiple others: “Problems emerge with the second and the third.”
奔波儿 cites Anthropic’s published data from around January 2026: coding accounts for only 35%, while the remaining 65% consists of science and education, creative work, middle management, and other use cases. Back-solving from the wage pool, a $200B year-end ARR assumption implies coding penetration in the low double digits, while science and education and creative work are already around 5%. “It’s not one straight line for coding and zero everywhere else; it’s a descending gradient curve.” Many use cases still have substantial room to reach 5%-10%.

10. Has AI slowed? 奔波儿 says the constraint is GPUs, not demand

星辰 cites 李贝’s view that AI could slow materially. 奔波儿’s figures put Anthropic’s ARR at more than $70B and OpenAI’s at $40-50B, with monthly additions still above $10B. “The reason it hasn’t reached that ARR is not a lack of demand. It’s because GPUs are constrained.”
His evidence is pricing: renting a data center costs “$50B for 1GW,” while 1GW of old GPUs used for inference generates only $50-60B of revenue—meaning GPUs may be getting even scarcer. “If old-GPU prices were falling, that would show supply and demand were starting to reverse, and then you might need to worry. But we haven’t seen that.” China’s GPU shortage is compared with lithium carbonate at RMB200K-600K.
星辰 does not buy it: the US spends $800-900B a year on GPUs and still claims there is not enough supply? “The things you can’t get are always the most attractive.” He even suspects a marketing component—volumes are down, but prices are up. 奔波儿 does not press the point: the issue cannot be resolved in debate, so set it aside and watch for another 1-2 months.

11. No longer focused mainly on second derivatives: narrative, pricing, and the stock reaction matter more

奔波儿 says he used to focus on sequential and year-over-year growth, but now prefers narrative, valuation, and the stock’s reaction. “How would I know whether the market expects 20, 30, or 10? But if it delivers 20 and falls, that tells me expectations were above 20.” Grinding through weekly data is “like studying baijiu stocks used to be.”
US-listed GV is the example: orders are more than 3x revenue, orders are up 100%-200% year over year, prices are up 50% year over year and another 10%-20% sequentially, yet the stock fell 8%-9% after the company announced a 25% capacity expansion over 3 years after 2028. It may already be up 10-20x since 2023, with the valuation at 25-28x 2028 EV/EBITDA. “At some point, no matter how good it is, the case becomes impossible to explain.”

12. Memory: the short-term, unfalsifiable “algorithm takes a bite” is a complex question

The supply problem in memory has always been there. After rising more than 10x, the sector may have absorbed more than $1T—or $600-700B—of capex, while the industry chain captured most of the profits. When prices were rising, everyone treated the thesis as a joke. After the selloff, the short thesis came back under review: “It sounds kind of reasonable. Maybe I should sell too.”
There are 2 blades. The supply-expansion argument lands on 长鑫存储: with ROIC at 200%-300%, “you invest 1 and make 2-3; only an idiot wouldn’t invest,” but they still do not invest. The sharper blade is algorithmic optimization: it costs almost nothing, is completely unfalsifiable in the short term, and could become reality. But memory could also get hit after optimization, only for demand to prove strong and Jevons paradox to be vindicated. There is no conclusion yet, and no reason to fight the market.
星辰 offers the bear warning: people around him who never used to follow AI are now buying memory, which is “genuinely frightening.” 杨哥, a classic value investor, likes Korean memory stocks at single-digit PEs: “A lot of value traps I’ve seen were high-single-digit PEs at the top. The industry may be disrupted. At least if you buy 海狗, the technology roadmap won’t change.” 浩哥 wails: “I brought you here to replenish my faith. Why does it feel like I’m the one being converted?”

13. Why compare AI with electrification rather than the internet: both the scale and the stage line up

星辰 sets up the distinction: the internet connects things and primarily improves efficiency; AI replaces intellectual labor. 奔波儿’s scale test is AI capex at 2% of GDP, versus less than 1% for fiber investment during the internet era. Excluding real estate, “history has no technology revolution with infrastructure investment on this scale.” He spent roughly 10 hours studying AI and the relevant history.
There are 3 stages. Electric lighting appeared in the 1870s, followed by roughly 20 years of debate over technical routes, analogous to the pre-GPT-3 arguments. The infrastructure-building period for electricity may correspond to 2023-2025: several years of capex at 2% of GDP, while manufacturing productivity growth may have been only 1%. “Infrastructure-phase stocks got beaten like dogs; General Electric only started rising in the application phase.” Electrification took off after 1919. “We may only be at 1919 now—perhaps we haven’t even reached it.”
With only electric lighting, no one could have imagined radios or washing machines. “Before, tokens were question-and-answer; the addressable space was like lighting. Once you have Agents, the space gets much bigger.” AI is the foundation for more creation in the future; “you just don’t know what its future form will look like.”

14. 星辰’s Bayesian cold shower: desert periods may be the norm; 2 points do not imply a third wave

The counterargument is that electrification was a century-long transformation, not a one-way acceleration. This cycle has already produced 2 acceleration waves—chatbots and coding. If the third wave takes 5 or 10 years, “the entire industry chain’s stocks will definitely suffer a major drawdown.” Computers appeared more than 100 years after electricity was invented.
He has recently seen extensive disagreement over world models and physical AI. Replacing all blue-collar labor would be no smaller a market than replacing white-collar labor, but experts are polarized: some think everyone building world models is a fraud, while others think the GPT-3 moment arrives within the next 2-3 years. Once the sector is already at 2% of GDP, “is it getting a little difficult” to keep trading it toward 4% or 5%?
He also notes that the second and third AI waves were separated by 30 years historically. After the Dartmouth conference, “the desert period was the mainstream.” Two breakthroughs in 3 years do shorten expectations for a third, but they cannot prove that another breakthrough must arrive within the next 2 years. “You are too optimistic; maybe I look very bearish from your perspective.”

15. 奔波儿’s rebuttal: forecasting a discontinuity is betting on a low-probability event

On physical constraints: wires and grids did not exist before, and electricity infrastructure took roughly 20 years to build. AI already has the internet and the interconnections among providers as a foundation. “With AI, you saw coding emerge within 3 years.” Agents appeared only 6 months ago; if the field has been accelerating all along, it is difficult to forecast an abrupt stall. Capex and the influx of the smartest people in the world also create industry momentum.
On stages: the previous AI wave was still debating rules-based versus heuristic systems, perhaps even earlier than the AC/DC fight. Today, major use cases have already worked, while AI for Science, finance, law, and management are advancing in parallel; one use case does not have to finish before the next begins. He also avoids strong calls: “I like to place bets when something is probably going to become reality, but some people haven’t reacted to it yet.”

16. Stockpicking philosophy: only change creates mispricing; non-AI ideas face a higher bar

The core framework is that “stocks only offer pricing-dislocation opportunities when something is changing.” From power to AI hardware and then to domestic compute or applications, there is always somewhere that has changed but has not yet been priced. Keep positions larger when the wind is at your back and smaller when it is against you. Disagreement appears every year; after the Agent use case was validated between the end of last year and the start of this year, the debate faded and the rally resumed.
The bar is higher outside AI. Oil may be an undervalued direction. Cyclicals are simple questions: price is low, valuation is low, expectations are low; answer whether demand will peak and supply will surge, then judge when the cycle turns. “I can wait 6 months for you. There is no need to wait 2 years.” 浩哥 jokes that 6 months is already patient capital for a hedge fund, while AI can run through an old-timer’s 6 months of emotion and price action in 1 day.

17. The AI research workflow: hand collection and monitoring to AI, then let AI restrain you at the decision point

The 5 stages are idea generation, information gathering, rapid and deep research, decision-making, and tracking feedback. For information gathering, “connect the data sources and ask whatever you want; it spits out the answer.” At the decision stage, feed AI your rules—don’t chase, set a PE ceiling—and when you get carried away, “discuss it with AI for a couple of lines and you immediately calm down.”
The most striking case is a Nordic data-center company for which roughly half of the existing coverage carries sell ratings. He only told AI to “track this stock and theme,” and it automatically broke down the hypothesis that the US might be unable to build enough data centers, sending demand north to the Nordics. It tracked New York State’s law barring data centers above 50MW, which European players were acquiring land from city governments, and the project owners and builders. “Who told it to track the agreements signed with city governments, how much land was bought, how much was spent, and how many megawatts that represented?”
The current reality is: “Without AI, I can’t work.” AI leads information gathering and monitoring; humans lead the other stages with AI as an assistant, and humans ultimately connect the workflow end to end.

18. Tools and getting started: “Let AI do everything first”

He chose Claude Code for its stronger creativity; Opus “thinks one step further” and breaks tasks down automatically. Domestic models can also work, but the user has to decompose the task more finely. “Our business deals with money. Using the best model obviously offers a higher return.” But even the best model may not be necessary, and 奔波儿 mentions having his own model account banned. After Kimi K3 launched, he used it to write a PPT.
For beginners, download a domestic app such as Work Body: no VPN and no model selection are required because the cloud deployment is already in place. His closed-loop workflow, AI Workspace Hub, and information-scraping tool, AI Signal, are on GitHub and can be installed with one click. The simplest starting point is daily monitoring: give AI a stock pool, have it search for relevant information when something moves, and receive a daily push.
The core principle is: “Let AI do everything first. Don’t do it yourself first.” It is smarter than you in most areas, faster, and capable of more parallel work. Step in to assist when it gets stuck, then make the final call yourself. Outsource the most annoying and time-consuming tasks and preserve your own time for thinking about stocks and the next market theme.

19. Most bullish today: domestic compute, “odds above 50%, and attractive upside”

The analogy is Claude 4.6’s inflection point: “Before 4.6 it was basically a small coding tool; after 4.6 it generalized.” If domestic models reach a similar usability inflection point and domestic platforms lower the product threshold far enough, demand could explode—but “the explosion still needs time.” The stocks have also sold off significantly, so domestic compute is at least worth watching now.
Whether to buy large or small caps, and whether to focus upstream, midstream, or downstream, is a matter of taste. 奔波儿 says he does not know whether the probability is above 50%, but “it probably is, and the odds are attractive.” Applications remain a specialist problem: study them if you have the time, skip them if you do not. 星辰 adds that supply chains dependent on Japan and South Korea, as well as domestic substitution, deserve further work. 浩哥 says he is already trapped in the trade—just a little early.
There is no need to compete on high-altitude conviction in names such as memory and 闪迪 that have risen 80x in 1 year. “However you look at them, they’re extreme small-sample events.” This is no longer a contest in research; it is a contest in nerve and trading. While traveling in Shanghai, 奔波儿 even heard passersby discussing 德明利 and the earnings surprise gap in memory, prompting the question: “Everyone tells me the surprise is something other people don’t know. So who exactly is the person who doesn’t know?”

20. Living alongside quants, old solar scars, and “firm conviction, flexible posture”

Quantitative trading has reached intraday and minute-level factors: “Daily factors are already too low-frequency to have any Alpha.” Quant index-enhancement funds have had no Alpha this year and are beginning to capture and harvest one another. “If you put money in, you’re just handing it to them.” Ordinary investors should extend their time horizon, shrink their sample size, and watch major inflection points such as Kimi K3 that could reset the pricing mean and volatility. Anyone trying to trade short-term or intraday will “be beaten to a pulp; mathematics guarantees it.” The market’s insanity is visible in the 6 straight limit-down sessions after a reusable commercial-space rocket appeared, the ChiNext’s roughly 11% year-to-date gain through July 23 despite multiple days with swings above 10%, and 星辰’s observation that South Korea saw more than 20 circuit breakers that month.
His old scar came from solar. Early in his career, he calculated that a sector leader could earn more than RMB10B after 5 years and, at a 10x valuation, be worth RMB150B; against a RMB90B market cap, that was only 60% upside, “not very interesting.” It went on to reach RMB500B, then fell back to RMB100B. “You collected half the abuse and made not a penny.” The lesson: “A stock is a stock, a company is a company, fundamentals are fundamentals… if you play this game, you have to follow its rules.”
浩哥 closes by describing the guest as having “firm conviction and a flexible posture,” and says his own faith has been “topped up a little.” 奔波儿 no longer answers 5-year and 10-year questions either; looking back 5 years, many of the judgments that worked then no longer work today. The final takeaway: “No matter how hard anyone talks, none of us can withstand the iron fist of A-shares.” Keep walking and see what happens.

游凯超 on vLLM, Open-Source Infra and Model Co-design SAP's 原欣 on Whether AI Is Coming for Enterprise Software Giants