Bill Qian Dissects the AI Bubble’s Endgame: If Chinese Models Become “Too Successful,” Will AI Follow the EV Playbook?

Summary

• Bill Qian’s core framework is that speed may digest a cyclical bubble. The first 2 Industrial Revolutions took 50-100 years to reach adoption, the PC took 10-20 years to reach 1B users, while AI has reached 1B-2B global users in just 3 years—“growth can digest any problem”（有增长可以消化一切问题）. But he is not saying this cycle must be a fast rise and fall: the key question is whether it becomes a Flash Bear or an L-shaped bear market like 2008 or 2000. He sees an L-shaped bear as a scenario to expect, while warning that the cycle could still play out differently. His long-term analogy is Amazon: even someone who bought at the 1999 bubble peak would have earned more than 40x by today.
• The key downside indicator is China’s cheap Token. When Chinese vendors compete in large models, multimodal AI and related fields at one-twentieth or even one-hundredth the price, the AI market could follow the EV or solar playbook: social welfare and adoption surge, but value creation is high while value capture is limited. US vendors’ valuations could be sharply compressed, while rising valuations for Chinese vendors may not fully close the gap. Bill sees Chinese vendors’ global market-share gains as almost certain; OpenAI has already started cutting prices.
• This cycle’s valuation and funding structure is relatively undramatic. Cisco traded at roughly 200x P/E, while storage stocks after the selloff may trade at just 3-4x and TSMC and Nvidia at 20-plus times; capital spending has historically been funded mainly with equity, though the M7 may also be starting to raise debt. By 2028, combined cloud capex, AI capex and AI-related debt could become the second-largest debt market in the US after mortgages.
• The payback period is the gating variable. GPUs typically have a 3-4 year economic window, which could extend to 5-6 years if older units are redeployed for inference. OpenAI’s CFO reportedly said that 1 GW could generate about $10B in ARR per year, implying a 3-5 year payback period. That period determines equity returns for CSPs and new clouds, as well as whether the debt can ultimately be repaid. The key questions are whether downstream Token sales and ARR can keep growing, and how Chinese vendors reshape the market.
• The fundamental chain runs from the bottom up. Application revenue drives cloud vendors’ debt and capex, which in turn drives chip and memory shipments. Coding “can, in some sense, replace almost every white-collar job in the world”; if software and hardware AI replace $10T of a global labor market worth tens of trillions, applying a 10x-20x PS multiple could theoretically produce a $100T-$200T market. Memory experts believe shortages may persist before 2028, while HBM shifts memory’s logic from commodity to TSMC-like advanced manufacturing; supply could still eventually overshoot demand and bring valuations back down.
• At the national level, Bill sees East Asia’s top-down, whole-of-nation model working repeatedly—from steel and autos to solar, EVs and chip manufacturing—challenging pure market orthodoxy. A global AI value-for-money war is unlikely to become a major trade-negotiation issue for now; the real constraints may be sovereign AI and security concerns. “The next Asia is still Asia”; the host called India “the world’s first major country to be shorted by AI,” while Bill explained that its 3M-4M IT workforce could be hit by Anthropic’s ARR growth.
• Robotics may replay the smartphone and EV playbooks. Bill sees robotics as software plus hardware, intelligence plus hardware, with China potentially supplying the world at extremely high value for money over the long term. The US could retain an Apple-like premium position, with national security and trade barriers also in play. An EV with FSD can be viewed as a wheeled robot, but his suggestion that China could ultimately take almost all non-Apple users remains an analogy, not a firm conclusion.
• UBI is a political-structure design problem, not a linear function of GDP. Some European and Nordic countries already have arrangements that resemble it in certain respects, but UBI may also never become universal. Because AI could create enormous value while replacing human intelligence, sovereign AI will persist; governments will intervene in the import and export of AI technology, data and capabilities, and the strongest models may be sold first to allies, much like the F-35.

Deep dive

1. Speed Is the Biggest Variable in This Cycle—Growth May Digest a Cyclical Bubble

• Bill’s opening framework is that the biggest difference between this AI wave and prior cycles is speed: the first 2 Industrial Revolutions took 50-100 years to reach adoption, the PC took 10-20 years to reach 1B users, while AI may reach 1B-2B users in just 3 years. “Growth can digest any problem”（有增长可以消化一切问题）: growth may absorb a cyclical bubble, with user and industry expansion pulling the market back to the table after a selloff.
• That does not mean Bill is predicting a fast rise and fast fall. The question is whether the final outcome is a relatively quick decline and rebound—a Flash Bear—or an L-shaped bear market like the 2008 financial crisis or the 2000 internet bubble.
• Another important variable is a value disruptor from China: cheap Token could make the AI market look like EVs or solar, generating substantial social welfare while potentially compressing US vendors’ valuations for a long time.

2. Memory May Remain Scarce Before 2028 as Valuation Shifts from Commodity to Advanced Manufacturing

• Bill’s research indicates that industry experts broadly believe memory could remain supply-constrained before 2028. The market structure is also changing—from commoditized DRAM to customized HBM. With a long capex ramp, memory should be revalued from a commodity business toward advanced manufacturing more akin to TSMC.
• He retains a long-term caveat: supply will definitely catch up with demand eventually, and the market could then move into oversupply. Better engineering and optimization by Chinese companies could help rebalance supply and demand, at which point upstream manufacturers’ valuations and asset prices could fall.
• Victor added that HBM’s share of high-ASP, high-gross-margin products has risen significantly. AI demand and the capacity squeeze created by HBM production are also affecting DRAM and NAND supply, potentially making prices more stable. Micron has signed 10-plus strategic long-term customer agreements. Memory has not fully escaped its cyclical-stock profile, but this cycle has already moved a considerable distance toward something different.

3. Flash Bear or L-Shaped Bear? Amazon Shows the Potential for Long-Term Returns

• The host described the market as undergoing repeated deleveraging and asked whether that was the central scenario in Bill’s framework. Bill said the process involves both deleveraging and a reset in valuation expectations; the key is whether the end state is a Flash Bear or an L-shaped bear. An L-shaped bear means expectations, fundamentals and valuation recovery all take longer.
• Bill’s current observation is that R&D staff at Anthropic and many other model companies believe development has exceeded expectations, while the market is still trying to identify the inflection point and determine when AGI arrives. User counts, time spent, and downstream willingness to pay and service procurement around Anthropic’s Claude all indicate that adoption is still growing.
• He therefore believes an L-shaped bear market is a scenario to expect, but cautions that the bear-market logic may be right even if the way it plays out is different.
• The host cited several cooling signals: SpaceX falling to $106-$107 per share, OpenAI’s IPO being pushed from this year to next year, and the 24-year-old “stock god” in Germany whose Situational Awareness thesis was later taken wholesale by Citadel. Bill responded that even Coca-Cola is making new highs and sector rotation will persist for a long time; he is more focused on whether model companies can continue to beat revenue expectations next year and the year after.
• Bill’s long-term example is Amazon: someone who started buying in 2001 could have made money most of the time, while even a buyer at the 1999 bubble peak would have earned more than 40x by today. If downstream Token sales are strong, the investment burden carried by cloud providers and the upstream supply chain can ultimately be paid back.

4. Investors Trying to Catch Sector Rotation May Become Exit Liquidity

• Asked about the short-term rotation from semiconductors into software and big tech, Bill said he is not a short-term rotation specialist and that rotation itself is chaotic: hot-money traders and large numbers of small- and mid-cap investors in Hong Kong and Lujiazui exchange stock ideas, buy mid-cap names together, then circulate market rumors and push the same stocks higher.
• The dynamic resembles capital chasing sectors in crypto. Bill believes trying to catch sector rotation is exhausting; an investor focused on catching the next sector often is not the person shaping the narrative but the person providing exit liquidity. That is not a particularly strong investment thesis.
• The exception is identifying an industry bottleneck from fundamentals ahead of time—for example, recognizing an early bottleneck in memory, storage or another part of the supply chain. That is driven by an industrial constraint, not by narrative.
• Bill is not good at calling short-term sector hot spots. Unless an investor is already inside the game and shaping it, he recommends holding assets in which they have strong conviction, or buying broad-based ETFs covering memory, semiconductors and technology more generally. Staying within one’s circle of competence matters more.

5. What Is Familiar: Capex Outruns Revenue; What Is Different: Equity, Speed and Valuation

• As in past cycles, this is real innovation and should be distinguished from the tulip mania; like the railway boom and electrification, innovation-driven capex growth will far outpace revenue. Cloud vendors may spend $700B-plus on capex this year, and by 2028, combined cloud capex, AI capex and AI-related debt could become the second-largest debt market in the US after mortgages. Whether downstream revenue can resolve the upstream investment burden remains the central question.
• One difference is the funding structure: for a long period, investment was funded mainly with equity, while debt remained relatively limited, although the M7 may also be starting to raise debt.
• Another is speed: it took roughly 80 years from the steam engine’s arrival to the point when UK industrial labor exceeded agricultural labor; electrification took nearly half a century from the first generator to full US electrification; and the internet took more than 10 years to reach 1B users. AI took roughly 36 months from GPT’s birth in November 2022 to exceed 1B users. “In martial arts, speed is unbeatable”（天下武功，唯快不破）: sufficiently rapid growth may smooth out some cyclical problems.
• The third difference is valuation. Cisco traded at roughly 200x P/E in the prior cycle; storage stocks after the selloff may trade at just 3-4x, while TSMC and Nvidia trade at 20-plus times. Bill sees no especially extreme fantasy multiples across the industry; the bigger concern is that prices have risen too quickly and too much capital has rushed in, rather than that absolute valuations are universally excessive.

6. A 3-4 Year GPU Economic Window Makes Payback Critical for Equity and Debt

• The host’s pushback was that railways can be used for 50 years and fiber for 25, while GPUs have only a 3-4 year economic window. Even if terminal demand exists, revenue must arrive before the GPUs depreciate.
• Bill said the industry disagrees on GPU depreciation periods. Some believe retired GPUs can be redeployed for inference, extending their useful life to 5-6 years.
• He said everyone is now calculating the payback period. OpenAI’s CFO reportedly said 1 GW could generate about $10B in ARR per year, which could still imply a 3-5 year payback.
• The biggest concern is the overall payback period. It determines the equity returns of traditional cloud service providers, or CSPs, and new clouds—and more importantly, whether the debt can ultimately be repaid.
• The 2 key observations are, first, how well downstream “SRP,” OpenAI and other application vendors monetize Tokens and whether ARR can continue to rise; and second, how Chinese vendors disrupt the entire market.

7. The EV-Style Endgame: High Value Creation, Limited Value Capture

• When Chinese vendors compete in large models, multimodal AI and related fields at one-twentieth or even one-hundredth the Token price, the social welfare and scale of global AI adoption could rise, while US vendors’ revenue falls.
• Chinese vendors may not capture all of the revenue lost by US vendors because they are also engaged in a price war. Developers and users may ultimately adopt a blended model of expensive and inexpensive systems.
• The result could be substantial value creation but limited commercial value capture. US vendors’ valuations could be compressed sharply, while Chinese vendors’ valuations rise without necessarily recovering the full amount of US valuation lost.
• If the industry ultimately follows the EV playbook, consumers and developers will be happy. Whether investors in these companies can earn strong ROE over the next several years requires a separate analysis.

8. Token Budgeting Is Taking Hold; China’s Share Gain Is “Almost Certain”

• Asked whether this scenario would be visible 6 months from now, Bill said an increase in Chinese vendors’ global market share is almost certain. Online benchmarks show that open-weight—what are often called Chinese open-source—models are catching up with closed-source models faster and that the performance gap is narrowing.
• Silicon Valley is moving from Token Maxing, with an unlimited budget, toward granular Token Budgeting: high-end models for high-end tasks, and lower-end or Chinese models for lower-end tasks. That usage pattern is taking shape.
• Bill noted that OpenAI has already started cutting prices. He is uncertain whether Anthropic can list on schedule; if it does, the stock price will reflect competition from Chinese vendors. Here, the price war means comparable products priced at a few-dozenths of the rival’s price, or even 1/100.

9. The Global Value-for-Money War Is Not Yet on the Trade Table; Sovereign AI Is the Obstacle

• Bill does not expect AI to become a major bargaining chip in trade negotiations for now, because those negotiations usually center on industries with large voter and employment bases: Australian beef, US corn and European cars.
• AI remains a relatively elite, highly competitive industry, so whether it enters trade negotiations is uncertain. Intervention is possible if data compliance becomes involved.
• This looks more like a global AI value-for-money war than China’s Hundred-Regiment price war or a regional e-commerce price war.
• The force most likely to stop the price war may not be trade negotiations but governments’ security concerns. Some countries may ultimately care less about cost efficiency than about building their own branded AI.

10. The “Chopsticks Cultural Sphere’s” Whole-of-Nation Model Keeps Working

• From Taiwan’s chips and Japanese and Korean industries to China’s solar and EV sectors, Bill believes the top-down, whole-of-nation model has worked repeatedly. When an issue is important enough, a top-down system can guarantee resources and concentrate them on one objective; waste is inevitable, but the result may still be effective.
• The US and India are also trying to emulate it, but Bill believes smaller government systems have fewer resources to mobilize and will not match the impact of the East Asian chopsticks cultural sphere.
• He sees this as a challenge to pure market orthodoxy: East Asia’s model has worked at times in steel, autos, solar, EVs and now chip manufacturing.
• At the cultural level, East Asian bureaucrats and businesspeople in some sense share a scholar-official ethos, allowing them to share language and resources. That logic may be harder to reproduce in the West, although the US under Roosevelt once had a much larger government.

11. Don’t Guess the Macro; Watch the Fundamentals at the Bottom of the Stack

• Bill said he is not good at calling market trends. The midterm elections and US rate hikes are ultimately macro factors, and the macro backdrop is clearly less optimistic than before.
• His positioning advice is that investors already holding positions should think seriously about the next rebalance; those who are not fully positioned can wait patiently for an opportunity to deploy capital.
• On fundamentals, the key is downstream application growth: whether Anthropic and OpenAI can continue to grow revenue and whether Chinese competition affects them. That could materially change their valuations.
• Another possibility is that Coding grows too quickly and continues to beat expectations. Bill believes Coding can, in some sense, replace almost every white-collar job in the world, making it a key area to watch.
• The chain is straightforward: downstream application revenue determines the debt and capex problem for cloud providers, which then determines upstream chip and memory shipments. Software revenue today still comes mainly from large models; multimodal video and image applications have not caught up, while embodied intelligence and robotics revenue remains small.
• If software and hardware AI replace $10T of a global labor market worth tens of trillions, applying a 10x-20x PS multiple could theoretically create a $100T-$200T market.

12. Robotics May Replay the Smartphone and EV Playbooks

• The host compared robotics with EVs, noting that Apple spent years developing a car without releasing an Apple Car, Xiaomi reached mass production in 3 years, and China has BYD and Xpeng. He also mentioned that China’s 语速科技 reportedly plans to IPO on August 10. Bill agreed that robotics can be compared with the EV industry.
• Bill believes robotics ultimately comes down to software plus hardware, intelligence plus hardware. China is a manufacturing power, and for a long time it may supply the world at extremely high value for money.
• The US may resemble the smartphone industry: China could take almost all non-Apple users globally, while the US retains an Apple-like premium position. The outcome may reflect both market positioning and national-security and trade barriers.
• He believes embodied intelligence and robotics may follow the same logic: a smartphone is a handheld computing tool, an EV with FSD is a wheeled robot, and robots will ultimately adopt the manufacturing logic used to build smartphones and EVs.

13. The Next Asia Is Still Asia; India and Europe Face Structural Problems

• Bill said the rise of sovereign nations depends on a certain population scale, good education and a culture of delayed gratification—the willingness to study, save and invest—along with integration into the global order, geography and external conditions.
• The Confucian cultural sphere is one of the few, perhaps almost the only, postwar regions to move 1B-2B people from agrarian societies into developed economies. That is why Bill believes the next East Asia and the next Asia in the 21st century will still be Asia. Technology gives every country and individual an equal opportunity, but opportunity and development ultimately remain unequal.
• The host quoted Bill’s article calling India “the world’s first major country to be shorted by AI.” Bill explained that agrarian societies can fall into a Malthusian trap; once an industrial society starts a virtuous cycle, population, labor, education and consumption reinforce one another.
• AI may create massive output without creating a comparable number of jobs, which could weaken people’s consumption capacity. This is also why Musk believes future monetary phenomena could be deflationary: if supply is no longer scarce, basic economics and assumptions about population may have to be rewritten.
• India has roughly 3M-4M IT workers, and the sector is part of the country’s middle-class dream. But what defeats India may not be IT engineers from another country—it may be the growth of Anthropic’s ARR. The demographic dividend needs to be assessed case by case.
• Bill said he knows less about Europe, but the acquisition of DeepMind by Google is instructive: Europe can still produce talent, yet appears to lack the financial resources for sustained investment. It may be able to take companies from zero to one, only to become a talent-training base outside the US.
• He expects Europe to face Asian competition in AI, EVs and other frontier industries, and some formerly developed countries could become “rust-belt countries.” Japan’s GDP per capita is currently about $40K, less than half Singapore’s; part of the reason is that one high-value industry after another was taken by competitors over the past 10-20 years, while the auto industry now faces competition from China.

14. UBI Has Arrived and May Never Arrive; Sovereign AI Will Be Controlled Like the F-35

• On “AI communism” and UBI, Bill said the world’s food supply, if averaged across the population, has long been sufficient to prevent hunger, yet people still starve in Africa and Haiti. The US wealth gap and healthcare reform also cannot be solved by aggregate resources alone. UBI is therefore first a problem of political-structure design, not a linear function of total GDP.
• In some well-governed European and Nordic countries, unemployment insurance and benefits are relatively generous, creating arrangements that resemble UBI in some respects. UBI is therefore not purely a future issue, but it may also never become universal.
• Bill believes governments will inevitably intervene because the value AI can create and the human intelligence it can replace are too powerful. Regardless of how private-sector AI develops, sovereign AI will exist.
• Governments will intervene heavily in the import and export of AI technology, AI data and AI capabilities. AI could eventually be treated like arms: the best systems sold only to allies, much as the US sells F-35 fighter jets.