Authors: Victor, Mr. Z, 168X
A technology can be extremely successful yet still be a failed investment.
In early August 2026, the AI and semiconductor sectors experienced a significant downturn following declines in June and July, reigniting market debates over whether a bubble was about to burst: SpaceX’s stock price fell back to the low hundreds of dollars, OpenAI’s IPO was postponed from this year to next, and the prominent AI-themed fund Situational Awareness was forced to shut down. Meanwhile, Chinese open-source models emerged densely in July, offering performance close to that of proprietary flagship models at extremely low costs, thereby reshaping the global pricing structure for intelligent services.
This episode of 168X welcomes back a familiar face: Bill Qian (@billqian_uae), a top investor in AI and Crypto, and the creator of the media platform "Bill It Up." Recently, Bill has been examining AI within a broader economic and social framework, exploring whether the demographic dividend has run its course, why governments are re-entering the arena, and how the AI bubble might ultimately resolve. In this nearly two-hour conversation, he presents a counterintuitive core insight: the greatest fundamental risk to the AI bubble may not be technological failure, but rather "the Chinese model is too successful." When China turns intelligence into a low-margin commodity, global social welfare rises dramatically—but industry revenues and profits may not keep pace, creating a unique Jevons Paradox: a technology can be wildly successful, yet still be a failed investment.
I. Why This AI Bubble Is Different: Speed, Valuation, and "Growth Absorbs Everything"
Mr. Z: Bill has recently devoted a lot of energy to AI and often seems deeply concerned about the broader implications, viewing AI within the larger context of economic and social systems. Your recent articles frequently raise fundamental structural questions: Is the demographic dividend no longer effective? Why are national governments now actively intervening in the AI landscape? And how will the AI bubble eventually burst? Could we start by discussing the AI bubble?
Bill Qian: Alright. I think the biggest characteristic of this wave of AI, compared to previous cycles, is speed. This speed is unprecedented: the First and Second Industrial Revolutions may have taken 50 to 100 years to achieve widespread adoption, PCs may have taken a decade or two to reach a billion users, but AI has achieved 1 to 2 billion users worldwide in just three years.
As the pace of adoption accelerates, many cyclical bubbles may be absorbed by the speed of growth. There’s a saying: “Growth can absorb all problems,” and I believe this is a key point to consider when evaluating AI today. Every technological cycle tends to bring bubbles, followed by bursts—but this time, we need to understand that speed itself may alter the entire cycle’s evolution, potentially turning it into a very brief market correction. Because once the market dips, a sudden surge in user adoption and industry growth can quickly bring everyone back to the table—an observation we’ve seen multiple times before.
Another point is that during this process, we must remain vigilant toward these “value disruptors” from China. In the future, inexpensive tokens from China could make the AI industry resemble the electric vehicle or photovoltaic industries: generating massive societal welfare while, over an extended period, suppressing valuations on the U.S. side. These are two phenomena I’ve recently observed and discussed with friends.
Mr. Z: Reading your article, you are a bull on AI fundamentals. Your reasoning includes: AI will reach one billion users within three years, with adoption rates far surpassing those of the steam engine, electricity, PCs, and the internet; model companies are growing revenue rapidly; current public market valuations have not yet reached the levels of past bubbles; and most are backed by cash flows from large tech companies, not high-leverage financing. Therefore, a dynamic emerges: each time the market corrects or marks down valuations, new breakthroughs in models, revenue growth, and user expansion expand the opportunity, allowing valuations to be reabsorbed. So are you suggesting that the market is merely undergoing repeated deleveraging cycles, accompanied by some V-shaped pullbacks?
Bill Qian: On one hand, it’s about deleveraging; on the other, it’s about adjusting valuation expectations. But throughout this process, we need to closely monitor one key point: will it ultimately be a “Flash Bear”—a relatively quick rise and fall—or will it resemble the 2008 financial crisis or the 2000 dot-com bubble burst, with a more L-shaped recovery?
What is an L-shaped rebound? Essentially, it’s equivalent to an expected recovery where fundamentals catch up to valuations, requiring more time. But looking at it now, I feel that—even among many researchers at Anthropic and other large model companies—the pace of development has actually exceeded their expectations. Everyone is now asking: When will the singularity arrive? When will AGI arrive? From the perspective of adoption metrics—such as user numbers, user engagement time, and the willingness of downstream services like Anthropic and Claude to purchase—while one can still anticipate an L-shaped bear market, one must also be cautious: the underlying logic of the bear market may ultimately hold true, but its manifestation this time might be subtly adjusted. This is precisely what I’ve been closely observing.
Mr. Z: I myself can feel that the AI wave may indeed be losing momentum in the short term (within the next six to twelve months). First, SpaceX’s stock has dropped to $106; second, OpenAI, which had consistently planned to go public this year, has now delayed its IPO until next year; third, situational awareness-style thematic funds have been completely liquidated, leveraged down, and collapsed. So, is the market’s next move likely to see tension and capital shift away from AI toward other sectors as enthusiasm for AI wanes?
Bill Qian: Yes, we can see that even Coca-Cola’s stock has recently hit new highs—sector rotation will certainly continue. During this rotation, investors naturally seek new themes, and these market sentiments will persist over the long term. But I’m more focused on this: returning to AI, if AI revenues ultimately exceed expectations—for example, if Anthropic or China’s large models continue to beat forecasts next year and the year after—then investment in AI will remain strong. At that point, questions about whether CAPEX can be recovered and payback periods may also be resolved. As long as downstream tokens sell well, and the cloud services in the middle do too, the entire value chain can ultimately be absorbed and sustained.
From this perspective, it’s like buying Amazon starting in 2001: for most of that time, it was profitable. Over the past couple of decades, Amazon has grown tremendously as an e-commerce stock, and later as e-commerce plus cloud, and buying it at almost any point has essentially been the right decision—even if you bought it at the peak of the 1999 bubble, you’d now have more than a 40x return.
II. Storage: Reevaluating Advanced Manufacturing from Commodities to "TSMC-like" Entities
Mr. Z: I’ve heard that storage might account for 13% to 18% of costs, but after next year and into the year after, both storage itself and even the chips could experience price collapses due to oversupply, thereby squeezing the revenue and profits of these large companies. Will this actually happen, or is it inevitably going to be undersupply?
Bill Qian: I believe supply will ultimately catch up with demand. However, based on research I’ve conducted regarding the growing computational needs downstream, industry experts generally agree that, for storage in particular, there will likely remain a significant supply shortage until at least 2028.
At the same time, the market structure of memory itself is changing: from the previously highly homogeneous DRAM to today’s customized, advanced manufacturing HBM (High Bandwidth Memory). Within this framework, the CAPEX ramp-up cycle for memory itself will be significantly longer in the future; our overall perspective on memory should no longer treat it as a commodity, but rather adopt an advanced manufacturing logic akin to TSMC. Under this new framework, future valuations will also need to be reassessed.
But the point you just mentioned—whether future demand will be adequately met by supply—I believe this question will persist over the long term. There may be significant oversupply in the end; however, on the demand side, due to China’s improving engineering optimization capabilities, equilibrium may ultimately be achieved. In that case, the valuation or asset prices of upstream manufacturing would also need to adjust downward.
Victor: Storage has been the most standout sector in this cycle. It’s inherently a strong cyclical industry—factories take a long time to build, require massive capital investment, and have historically shown clear cyclical patterns: booming profits during upswings, and severe losses when supply surges. In this AI cycle, storage has some similarities and some differences: high-margin, high-unit-price products like HBM now make up a significantly larger share of sales; HBM capacity has crowded out production capacity for conventional storage like DRAM and NAND, ironically helping stabilize prices for traditional memory; companies like Micron are also signing longer-term contracts, entering into over a dozen strategic customer agreements. Overall, in terms of the AI bubble, what do you think is different this time—and what remains the same?
Bill Qian: Let’s start with the similarities. First, it is a genuine innovation—just like the previous generations of the internet, electrification, and industrialization—we will distinguish this innovation from speculative bubbles like the tulip mania; it is a real advancement. Second, it is true that capital expenditures (CAPEX) for all innovations have historically outpaced business revenue—whether during the railroad boom or electrification, CAPEX growth far exceeded income. This is why many are concerned today: cloud providers are spending over $700 billion in CAPEX this year, and in two years, by 2028, cloud CAPEX, AI CAPEX, and AI-related debt could become the second-largest debt market in the U.S., behind only mortgages. Under this logic, the key question is whether downstream revenues can keep up with upstream investments.
But there are several key differences this time. First, for a long time, investment was primarily equity-based, and debt levels were not particularly high—only recently have companies like the Magnificent Seven (M7) begun raising debt capital. Second, and even more significant, is the speed. The industrial revolution took roughly 80 years for steam power to shift the UK’s workforce from agriculture to industry; electrification took nearly half a century from the invention of the first generator until the entire U.S. was electrified; the internet took over a decade to reach one billion users. Yet today, just 36 months after GPT’s emergence in November 2022, AI has already surpassed one billion users worldwide. Of course, you could argue that AI innovation builds upon the foundation of global digitization, PCs, the internet, and cloud computing—but the speed itself... in martial arts, speed is unbeatable. When adoption happens this fast, doesn’t velocity itself smooth out some of the growth challenges we face today?
Another point is that the overall valuation is not excessive. During the previous cycle, Cisco had a P/E ratio of around 200; but today, after the storage sector declined, P/E ratios are only three to four times, while TSMC and NVIDIA are both around 20 times. You’ll find that the entire industry doesn’t have particularly inflated “dream valuations”—most companies are within a very reasonable P/E range. The main concern people have is that prices have risen too quickly and too much capital has flowed in, which naturally raises red flags; but from an absolute valuation perspective, things are still fairly reasonable.
III. GPU Depreciation: Railways can last 50 years, but GPUs cannot wait that long
Mr. Z: There’s another key point here. After reading your article, I believe the main focus is on the depreciation rate of GPUs. GPUs have an economic lifespan of only three to four years, whereas railways can last 50 years and fiber optics can last 25 years. Even if CAPEX eventually generates demand, revenue must still catch up before the GPUs depreciate. Could you elaborate more on this point?
Bill Qian: Yes, there are differing opinions within the industry regarding the depreciation period for GPUs. Some believe that even outdated GPUs can still be used for inference, potentially extending their useful life to five or six years. What I want to emphasize is that everyone is now running these calculations. The other day, it seemed OpenAI’s CFO mentioned that one gigawatt generates approximately $10 billion in annual recurring revenue (ARR) per year, meaning it would take three to five years to break even.
So regarding the payback period, everyone is still in a wait-and-see mode. Currently, the biggest concern stems from the overall payback period, as this will ultimately determine: whether traditional cloud service providers (CSPs) and NeoCloud (new-generation cloud providers) can deliver equity returns—and more importantly—whether they can repay their debt. This is the central question on everyone’s mind.
Below this question, I see two key points. First, on the downstream side: how well can Anthropic, OpenAI, or other application providers sell their tokens? If their ARR continues to rise, the upstream CAPEX will ultimately become less of an issue. Second, we need to observe whether Chinese vendors will ultimately disrupt the entire playing field.
IV. The Value Disruption of Cheap Chinese Tokens: Successful Technology, Failed Investment?
Bill Qian: When Chinese manufacturers eventually engage in price wars using token prices as low as 1/20 or even 1% across various domains, including large models and multimodal technologies, the likely outcome is that global social welfare from AI increases and AI adoption becomes massive—but U.S. companies’ revenues decline; and the revenue lost by U.S. companies is not fully captured by Chinese firms either, since they are also engaged in price competition among themselves.
In the end, the vast majority of developers and users adopted a hybrid model combining high-priced and low-priced models. From this perspective, there was significant value creation, but much less value capture in commercial terms. This manifested as: U.S. companies’ valuations were significantly reduced; while Chinese companies rose in prominence, their gains did not match the amount of valuation lost by the U.S. firms.
So this is a point to watch moving forward. If this industry eventually becomes similar to the electric vehicle sector—where both consumers and developers are satisfied—but whether investing in these companies can still generate strong ROE returns over the next few years is something everyone needs to consider.
Mr. Z: That sounds pretty alarming. Suppose Kimi K3, Qwen, or whatever model starts significantly outperforming Claude or GPT, gains widespread adoption, and begins to eat into OpenAI’s and Anthropic’s user bases—leading to a collapse in their valuations, a market reassessment, and compressed valuations for leading model companies. Would the secondary market also suffer as a result, and would the IPO timelines for these two giants be delayed? Could this scenario play out like that? Could we see this happen in six months?
Bill Qian: I believe it’s almost certain that Chinese manufacturers will continue to increase their global market share. If you look at online reviews, you’ll see that open-weight models—often referred to as Chinese open-source models—are closing the performance gap with proprietary models at an accelerating rate.
Five: Token Budgeting Replaces Token Maxxing: The Era of Hybrid Closed- and Open-Source Approaches
Bill Qian: Under this logic, you’ll now find in Silicon Valley that people’s mindset has shifted from the original “my budget is unlimited”—known as Token Maxxing—to “I need to use my tokens very precisely”: deciding which applications warrant high-end tokens and which can use mid- or low-end ones—essentially, Token Budgeting. This mindset has already taken shape.
Under this logic, globally, high-end users rely on models like Claude, while mid- and low-end users turn to Chinese models—this has already become a widespread usage habit. So I believe this trend is inevitable. That’s also why we recently saw OpenAI begin lowering its prices. I’m not certain whether Anthropic will go public on schedule, but if it does, its stock price will clearly reflect this dynamic: the price competition from Chinese vendors. By “price competition,” I mean offering product performance comparable to yours, but at a fraction—1/10 or even 1%—of the price.
Mr. Z: This sounds like something the Chinese market is very good at—like the price wars between JD.com or Meituan back then, fighting until one side is wiped out, until the government steps in and says, “Stop this nonsense.” How will this end? Because a clash between the U.S. and China feels like it could go on indefinitely.
Bill Qian: I think it depends on whether this issue will enter trade negotiations among countries. Personally, I feel its priority isn’t that high yet, because trade negotiations typically focus on industries with large numbers of domestic voters and employed workers. For example, Australia cares deeply about its beef, the U.S. about its corn, and Europe about its automobiles—because these are industries with substantial voter and workforce bases. But AI is relatively an elite, highly competitive sector, so I’m uncertain whether it will become a bargaining chip in future trade disputes. That said, if the negotiations involve data compliance issues, I think it’s possible.
But in the short term, I feel this is again a global technological competition. It’s different from before—previous battles, like China’s “hundred-group war” in e-commerce or price wars, were regional; this time, in a sense, it’s a global AI value competition.
Six: Will the price war enter trade negotiations? Sovereign AI is the real wildcard.
Bill Qian: The only thing that might prevent everyone from engaging in this battle, I believe, isn't necessarily the trade negotiations, but rather each country's concerns about its own security strategy. This could lead certain regions to ultimately disregard cost-effectiveness and insist on developing their own sovereign AI. That, I think, is a separate topic.
Mr. Z: I’m also curious about your perspective on this. Clearly, in East Asia, Taiwan, Japan, and South Korea were essentially nurtured by the United States and can be considered part of the American-aligned semiconductor bloc. You’ve previously mentioned that to understand the current hype around Asian stock markets, one might need to trace back to Japan’s MITI, South Korea’s economic planning system, and Taiwan’s Ministry of Economic Affairs and ITRI—organizations that mobilize private enterprises. Looking toward China, one must examine the National Development and Reform Commission, local governments, and industrial funds. This logic resembles the government selecting winners. How do you view the roles of Eastern Asian and Western economic systems in mobilizing key national industries, and what are their respective strengths and weaknesses?
Bill Qian: I think during this process, people will realize that the top-down, nationwide system has repeatedly proven effective. Many economists previously believed markets couldn’t be intervened in and should be left 100% to market forces. But we can see that across East Asia’s chopstick cultural sphere—whether in South Korea, Japan, China, or other regions—this top-down approach has enabled many industries to thrive.
There will be some waste in this process, but when something is extremely important, a top-down approach ensures two things: first, absolute allocation of resources, and second, the concentration of all resources on a single task. We see this pattern repeatedly—whether in Taiwan’s semiconductor industry, South Korea, or Japan.
Seven: Government Returns to the Stage: Advantages of East Asia’s National System and the Chopstick Cultural Circle
Bill Qian: Under this logic, you’ll also see the U.S. and India attempting to emulate it. But I believe such emulation may have some significance, yet it certainly won’t be as pronounced as in the chopstick cultural sphere. After all, under a smaller government, the resources available are inherently limited. Today, for those purely market-oriented thinkers, this has been a wake-up call. In the end, it becomes clear that the logic of the chopstick cultural sphere—from steel and automobiles in the past, to China’s photovoltaics and electric vehicles, and now to the entire East Asian dominance in chip manufacturing—has consistently demonstrated that top-down approaches can indeed be highly effective.
Mr. Z: If this top-down government coordination logic is applied again, can it be precisely implemented at the level of leading models? For example, could China precisely coordinate resources toward Alibaba, Tencent, or Moonshot AI? Or, as with Trump in the U.S., could political power be leveraged—such as previously citing national security concerns with Claude’s Fable 5 to restrict its availability to the general public and limit access solely to the U.S. military? Could this top-down coordination still work with leading models? I’m referring to how Taiwan’s Ministry of Economic Affairs and ITRI once declared, “We will direct all resources, talent, and funding toward Hsinchu.”
Bill Qian: I understand. I believe this phenomenon will persist long-term in East Asia, as it’s tied to the region’s broader preference for a larger government role. Moreover, East Asia has long cultivated a cultural mindset—shared by both bureaucrats and businesspeople—of a scholar-official ethos, which enables people to communicate through a common language and pool resources collaboratively to achieve a goal. By the same logic, I think this would be more difficult in today’s West. It may have been possible in the past, such as during Roosevelt’s presidency in the U.S., when the American government was also quite large.
Eight: Sector Rotation and Circle of Competence: Don’t Be the One Who Takes the Bathtub
Victor: On the big tech side, I’m curious about Bill’s perspective on short-term versus long-term outlooks. Over a 20-year horizon, internet companies have indeed multiplied many times over; however, many investors are still focused on identifying the best sectors right now that can deliver solid returns in the near term. Recently, capital has begun shifting from semiconductors toward software and big tech. Could this become our theme for the second half of the year, or should we take it one step at a time? Where is the money starting to flow, and are there any early signs emerging?
Bill Qian: First of all, I am not an expert in short-term rotation strategies. Moreover, as far as I know, short-term rotation itself is inherently very chaotic. There are numerous speculative funds, Korean hedge funds, and investors from Hong Kong and Lujiazui who frequently trade small-cap to mid-cap stocks. They often share ideas and collectively buy into mid-cap companies, after which rumors or informal narratives circulate in the market, prompting everyone to pump the stock together. In a sense, this is quite similar to how capital in the cryptocurrency space has historically chased specific sectors.
In this process, trying to catch sectors is, in my view, a difficult endeavor. When someone attempts to catch a sector, it suggests that you are not shaping the narrative yourself, but rather becoming the one who takes over after others have moved in. This, in itself, is not a particularly sound investment logic. Unless you’ve already identified key bottlenecks from a fundamental perspective—such as recognizing early on that memory is a bottleneck for the industry—then what drives your decision isn’t a trend, but rather an analysis of bottlenecks across the entire industrial chain.
So, to answer your question, I’m not particularly skilled at identifying short-term sector trends. Unless you’re directly involved in the game—or even the one orchestrating it—I’d recommend holding onto the assets you truly believe in; or, in the end, consider investing in broader-based ETFs, such as those focused on storage, semiconductors, or broader technology sectors. I believe staying within your circle of competence is far more important.
Nine: Downstream Decides Everything: The Transmission Chain from Application Layer, Cloud Providers to Chips and Storage
Mr. Z: From today’s perspective, how do you view the future development of the AI market? We’ve seen a significant downturn in the semiconductor and AI sectors during June and July. After the U.S. midterm elections in November, do you expect a correction or crash in the secondary market, or do you anticipate a rally?
Bill Qian: I’m not good at predicting market trends. But I still believe that things like the midterm elections or U.S. interest rate hikes ultimately come down to macro factors. The current macro environment is certainly less optimistic than before. I think those holding positions should carefully consider rebalancing going forward; and for those who haven’t yet established positions, congratulations—you have the patience to wait for the right opportunity to enter.
Regarding fundamentals, I think a key point is to observe the growth of downstream applications—or whether Anthropic and OpenAI can continue to grow their revenues, or whether they’ll face headwinds, such as competition from Chinese companies. This could significantly impact their valuation. Alternatively, if the coding business grows too rapidly and consistently exceeds expectations, it could be significant, since coding, in a sense, has the potential to eventually replace nearly all white-collar jobs worldwide.
I think it’s important to observe this, as it’s the furthest downstream—and downstream demand determines the debt crises and CAPEX issues of midstream cloud providers, which in turn determine the shipment volumes of upstream chips and memory. This creates a complete chain of logic from downstream to midstream to upstream. I recommend everyone monitor the actual underlying revenue of this industry to see whether it can smooth out the sector’s hypergrowth.
Mr. Z: From my perspective, the furthest downstream segment is the AI application layer, but one level above that is actually the large model layer. Regarding the AI application layer, it seems like there aren’t yet many other applications beyond large models—is that correct?
Bill Qian: Yes, I agree with this understanding. We can divide AI into software and hardware. In terms of software, today’s largest revenues still come from large models; revenues from video and image modalities within multimodal AI have not yet caught up with those of large models. As for hardware, whether it’s embodied intelligence or robotics in the U.S. or China, I believe current revenues are still relatively small.
Then, when you talk about buying growth and betting on the future, you’re clearly hoping that software will eventually replace 10% or 20% of white-collar jobs globally; and with hardware, various service industries and physical labor could also be substituted to some extent. From this perspective, this amounts to: the global labor market is worth tens of trillions, and perhaps ten trillion of it could eventually be replaced by software AI and hardware AI. If that ten trillion is multiplied by a price-to-sales ratio of 10x or 20x, that’s a market capitalization of $100 to $200 trillion. In the future, within the global industrial landscape, agriculture will likely be a small portion, industrialization will follow, and the largest segment will be AI-enabled industries—because AI itself will generate even greater value in services and finance.
Ten: Robots and Embodied Intelligence: Another U.S.-China Race, Like the Electric Vehicle Era
Mr. Z: You just mentioned embodied intelligence, and I’m curious if you’ve been following the robotics space? In the U.S., everyone’s talking about leading robotics companies like Figure AI with high valuations, while in China, the most talked-about is Unitree Robotics, which is reportedly set to go public on August 10. The development of robotics in China and the U.S. feels a bit like the early days of electric vehicles: Apple in the U.S. said it wanted to build a car and talked about it for ten years—but still has no Apple Car. Meanwhile, Xiaomi said it wanted to make cars and launched production within three years, alongside strong players like BYD and XPeng. I feel China still excels in manufacturing—what do you think?
Bill Qian: I’d prefer to compare the entire robotics industry to electric vehicles, because in the end, it’s all about software plus hardware—intelligence combined with physical components. China is already a global manufacturing powerhouse, and I believe that for a long time to come, China will continue to supply the world with exceptionally high-value, cost-effective products.
In the U.S., it might be somewhat like the smartphone industry: there will be an iPhone. China may ultimately capture nearly all non-Apple users worldwide, while the U.S. will have its own high-end "Apple." Whether it’s a matter of market positioning, national security concerns, or trade barriers, I believe the same logic applies to embodied intelligence and robotics.
In the end, it all boils down to the same thing: in a sense, a smartphone is also a handheld computing device; once electric vehicles gain FSD (Full Self-Driving), they essentially become wheeled robots; and robots, in turn, will be built using the same logic that was used to create smartphones or electric vehicles. So I anticipate that the future of this industry will closely resemble that of the smartphone or electric vehicle industries.
Eleven: The Rise and Fall of Sovereign Nations: Why the Next Asia Is Still Asia
Victor: You previously posted about attending the World Artificial Intelligence Conference (WAIC). Given the current involvement of governments, I’m curious: East Asian countries like Japan, South Korea, Taiwan, and China have all benefited significantly from this AI wave; meanwhile, the U.S., Europe, and the Middle East are also making massive AI investments, pursuing breakthroughs in different ways. You mentioned that East Asia was nearly the only region in the previous era to achieve leapfrog development through government-led, planned economic strategies—accomplishing something akin to “surpassing Britain and catching up with America.” In the next wave of AI, which countries or regions do you think have similar potential? For example, could Southeast Asian nations like Malaysia and Vietnam, or countries in Latin America or Africa, seize this opportunity by shifting their government models?
Bill Qian: When we observe the trade war, we can see that countries like Mexico have indeed become locations for阶段性 arbitrage benefits. However, overall, from the perspective of the rise of sovereign nations, I believe there are several key points to consider.
First, a certain population size. Then, good education and a culture of "delayed gratification"—in plain terms, hard work, the willingness to invest for tomorrow. This manifests itself not only in diligent studying but also in diligent saving and investing. As a result, you’ll find that these factors have made the Confucian cultural sphere one of the few, perhaps the only, regions in the post-war era capable of transforming a population of one or even two billion people from an agrarian society into a developed economy.
From a systemic perspective, I still believe that in the 21st century, the next Asia will be Asia itself. Some regions are more advanced and progressing faster, such as Singapore and South Korea, while others like mainland China are gradually catching up. Southeast Asian countries also need to meet certain conditions—willingness to work, diligence, and a readiness to invest—which is why we’re seeing growth in countries like Vietnam.
Sometimes I view a country as a corporate team to assess its potential—it’s very similar to evaluating a company: the ability to delay gratification, willingness to invest in the future, and openness to participating in global systems, combined with favorable geographic and external conditions. So I don’t believe the rise of sovereign nations will follow a rotating sectoral pattern; I still think the future’s benefits will largely be captured by Asian countries. Ultimately, development between nations comes down to this: technology has granted all countries and individuals equal opportunities, but opportunities and outcomes remain unequal.
Twelve: The End of the Demographic Dividend: Why India Is the "First Major Country Shorted by AI"
Victor: This also applies to your other article on the demographic dividend. The main advantage of these Asian countries has always been the demographic dividend. You mentioned that India is the first major country to be shorted by AI, primarily because its software industry is highly developed—somewhat like a manufacturing hub for Silicon Valley. But now, AI is primarily replacing white-collar workers and engineers, and India is hit first and hardest. What about other populous countries like China and Japan? Will their future populations be able to turn into an advantage, or will they face the same issues as India?
Bill Qian: I think this question requires some context. In an agrarian society, population isn’t necessarily a benefit due to the Malthusian Trap: you can easily have many children, but if food supplies can’t keep up, many will starve, and the population will return to equilibrium. However, in an industrial society, once you establish a positive economic cycle and provide educational opportunities, people can collectively move from low-value labor to high-value work. In this process, sovereign states, demographic dividends, and the working-age population are all positively correlated. That’s why you’ll find that many industrialized sovereign states actively encourage childbirth.
At its core, this means: people need jobs, which create many employment opportunities and produce a great deal of goods; people then earn income, which they use to consume those goods, completing a positive cycle. Now, in the field of AI, we see something different: AI generates significant output, but it may not create as many jobs, leaving people with less capacity to consume. This is why Musk suggests that, in the future, the entire monetary phenomenon could become deflationary. If supply is no longer scarce, then fundamental economic principles—and even our basic understanding of population dynamics—may need to be rewritten.
Returning to the micro-case of India: India once had about three to four million people in the IT sector, which represented the entire domestic middle-class dream—everyone hoped to grow up and work in IT. But what has now surpassed it is not IT engineers from other countries, but Anthropic’s ARR growth. I find this a fascinating phenomenon. Therefore, how we view the demographic dividend in the future also requires context-specific thinking.
Thirteen: Europe’s Dilemma: Brain Drain Seen Through DeepMind’s Sale to Google
Mr. Z: What about Europe? It seems everyone has overlooked Europe. Last week, Musk had an interview with the editor-in-chief of The Economist, during which he argued that Europe has been disrupted by immigration and has failed to capitalize on the post-internet era, making it lag behind. Are you keeping an eye on AI development in Europe?
Bill Qian: I’m not very familiar with this area, so I’ll just share my thoughts. The story of AI in Europe can be understood through the example of DeepMind’s eventual acquisition by Google. At the time, DeepMind themselves said they wanted to continue developing, but only the U.S. had the capital to fund their next steps, to cash out, and to acquire them. You’ll notice that Europe still produces talent, but its production capacity and financial resources seem limited to achieving things from 0 to 1—effectively turning Europe into a talent training ground outside the U.S.
So I believe that Europe will face significant competition from Asia in the future—not just in AI, but also in electric vehicles and other cutting-edge industries. There was once a saying that JD Vance emerged from America’s Rust Belt; in the future, some former developed countries may become rustbelt nations. For example, Japan’s current GDP per capita is only $40,000—less than half of Singapore’s—primarily because over the past decade and a half, one by one, the advanced industries that once generated profits have been taken over by competitors, including pressure from China’s automotive industry.
Fourteen: UBI, Sovereign AI, and Governance: AI Will Be Sold Only to Allies, Like the F-35
Victor: Finally, I’m curious—earlier we discussed AI, global competition, open-source versus closed-source models, and the U.S.-China AI rivalry. What are your thoughts on how the governance model of AI might evolve between governments and industry in the future? In the long term, as AI could replace a large portion of the workforce and become a public infrastructure—similar to how telecommunications companies eventually became infrastructure—what do you think about government governance of AI? Earlier in July, a former OpenAI executive suggested that if the U.S. doesn’t restrict the export of closed-source models or curb competition from China’s open-source models, AI could evolve into a communist model where the state provides all these resources. What are your views on how governments might govern AI in the future? And if AI truly has a major impact on labor and employment, could models like UBI emerge? Although you mentioned in your article that UBI may be more suitable for small, homogeneous nations.
Bill Qian: I’d like to start by discussing UBI (Universal Basic Income). I’ve had several objections. First: if we globally redistributed food equally today, no one would starve—so why do people still starve? Second: the wealth gap in the U.S. is enormous, yet why is it so difficult to increase income for the poor or implement healthcare reforms?
So what I want to convey is this: UBI is ultimately a question of political design. It may not be strictly linearly correlated with the total GDP of society at that time. But today, in some well-governed European small countries, unemployment insurance and unemployment benefits are quite high—in a sense, they have already achieved UBI. So I think: first, UBI has never been a matter of the future; it is already happening in Nordic countries today; second, I also believe it may never fully materialize, because even today, when we look at the issue of food, we see that even though there is more than enough food for all of humanity, people still starve in Africa and Haiti.
Regarding governance by the state, given the immense value that artificial intelligence can currently create—or the extent to which it can replace human intelligence—it is clear that government intervention will inevitably occur in the future. Regardless of how private-sector AI develops, I believe sovereign AI will always exist. Governments will strongly regulate the import and export of AI technologies, AI data, and AI capabilities. In fact, AI in the future may become like arms: you will only sell your best AI systems to your allies, much like the United States does with its F-35 fighter jets.