November 16, 2026

After Jensen Huang: Will Everything Stay the Same?

The only spark · Two shortcuts overseas · AI bills of 200 versus 10,000 · When the old guard steps in
Contents

    “Jensen is the only spark.” — Trader Jiu, on NVIDIA’s succession problem

    Guests: TC (Tan Cong), founder of Bedrock (本拾投资); Trader Jiu (Trader韭), financial blogger

    “I think, out of that $4 trillion, Jensen alone accounts for $2 trillion.”

    In this conversation, recorded on November 15, 2026, I used a very rough number to express something I could not quite resolve: how much of a company’s valuation belongs to its products, customers and supply chain, and how much belongs to the person who can organize all of them?

    Three days earlier, NVIDIA announced Jensen Huang’s resignation, with finance chief Colette Kress serving as interim CEO. The company is still there, and so is its product roadmap. But the person who could discuss technology, production capacity and hundreds of billions of dollars in financing at the same time has left. The market needs to reprice the company, and the entire AI industry is doing its sums again.

    This episode of Mossfire is a crossover with Bedrock’s BEDROCK podcast. TC focuses more on continuity in technology and operations: slower growth may also mean lower risk. Trader Jiu cares more about where the money comes from and who is still willing to pay, and he is more pessimistic about overseas demand for Chinese AI and political resistance. My repeated question is: could the change extend beyond each company’s share of the market to the size of the market itself?

    We start with NVIDIA’s successor, move to how chipmakers and cloud providers divide profits, and then discuss the paths open to Chinese models, capital flows and valuations in 2027. We disagree throughout, including over whether better AI will improve corporate revenue and ordinary people’s jobs at the same time.

    The conversation follows, edited and abridged from the recording.

    1. “Jensen is the only spark”

    Raymond: Let’s start with your reactions. From seeing the news to Huang leaving the company, what did you go through?

    TC: Initially, I was shocked. I think Huang’s influence on NVIDIA is no less than Steve Jobs’s influence on Apple. Without Jensen, there would be no NVIDIA, and even this AI revolution might not have gone so smoothly.

    If you only had large models but no GPUs—graphics processing units—or nobody who could use GPUs well, AI would not have developed this quickly. I cannot say this revolution definitely would not have happened without Jensen, but at least it would not have gone so smoothly.

    He understands technology, has foresight, places bets when others dare not, and can lead so many people in doing it together, across supply chains in both China and the US. Building the world’s largest company by market capitalization from nothing over more than thirty years takes something almost superhuman. Someone with such extraordinary achievements and capabilities leaves in such a short space of time. I am still processing the shock.

    Trader Jiu: Initially, I looked at it more as a trader and put it on my list of risks to watch the following week. But actually seeing him leave stirred some emotion and empathy. Such an important person, someone who has contributed so much, simply gets pushed out. So abruptly, so carelessly.

    When you are standing that high and are suddenly rejected outright, your psychological defenses can collapse too. I think I can empathize with that feeling.

    Raymond: I also think there may be a lot we do not know. An outsider would ask: how can you push out a founder who has been there for more than thirty years? Some things may only emerge afterward. But the immediate impact is already enormous.

    To make succession more concrete: apart from yourselves, who would be suited to being NVIDIA’s CEO? Names being discussed include Arm’s Rene Haas and Ian Buck, who helped create CUDA, NVIDIA’s parallel-computing platform and programming framework.

    TC: I do not think anyone can do what Jensen did. People bring up those two because they have influence within NVIDIA’s world, have accomplished things, and there are not that many better candidates. But neither can take on everything Jensen held in his hands.

    Jensen understands technology and financing. When there is a deal worth hundreds of billions of dollars to discuss on Wall Street, he goes himself. He also knows the Asian supply chain well. He can find someone at TSMC, or have skewers with the Koreans, and get straight into the discussion. Finding a strong technologist does not fill all those gaps.

    A successor may therefore be able to do only a third or a quarter of the original job and delegate the rest. Technology, supply chain and finance will each have their own people in charge. Authority will disperse, and the company will become more professionally managed. Whoever takes over, I think that process will happen.

    Raymond: But when the board chooses someone, it first has to answer what the company really stands on. When Jobs left, Apple ultimately chose Tim Cook, who ran the supply chain. That was effectively saying this was the flame they most wanted to preserve.

    What is that flame at NVIDIA? Products, technology, or the supply chain? Different candidates represent different judgments about the company’s strengths.

    TC: NVIDIA faces a difficult situation. B2B competition—selling to business customers—is fierce. Design the product badly and customers leave immediately. Design it well but fail to manufacture it and they still leave.

    With consumer products, “I just like this brand” can sustain some continuity. NVIDIA cannot easily say that as long as the product is well designed, the supply chain does not matter. No single metric guarantees everything.

    Trader Jiu: Jensen is the only spark. Jensen is that one remaining spark, and the only spark is gone. Among those left, it comes down to who gets along with the board.

    Raymond: I think a lot of the big banks’ valuation adjustments are taking a deduction for exactly that. Earnings forecasts for the next two or three years can stay unchanged for now, because the products and roadmap he left behind remain. But the valuation multiple beyond that needs a discount.

    Sometimes the valuations people give Tesla and SpaceX today are strangely high, or almost impossible to calculate. Aren’t they also paying for Musk himself?

    TC: I would add a corporate-governance point. NVIDIA was founded very early and does not have an A/B share arrangement with different voting rights. Jensen owns a little over 3%.

    I think the same thing happening at Meta or Tesla might not produce the same result. When you are achieving great things, it may not matter. But the moment something happens, there is nobody who can stand behind him. I find that rather sad. Corporate governance and the board’s choices have a great deal to do with it.

    2. Becoming ordinary could also mean less risk

    Raymond: It sounds a little like the shift from an entrepreneurial company under Jobs to a public company under Cook. Has NVIDIA reached that turning point too?

    TC: I am not that pessimistic about the practical effect on the industry. Post-Jobs Apple may have been less creative and moved more slowly, but smartphones had reached a stage where plenty of people still knew how to keep things moving.

    AI is similar. The earlier stage was an enormous rush forward, with money, policy and personnel all providing a tailwind. Cloud providers’ cash flow had not yet been squeezed into a particularly difficult position either, so capital-expenditure expectations kept rising: one trillion, two trillion, three trillion, four trillion.

    Those expectations may now gradually flatten. Interest rates are high, policy faces resistance, and Huang has left. All these signals say we cannot keep charging ahead in the most aggressive way we used to.

    But slowing down also means the risk of overexpansion is getting smaller. You may no longer need to worry quite so much about the industry racing to an extreme peak and then falling off a cliff, as in the dot-com bubble of 2000.

    That signal reaches demand and supply together. Buyers of computing capacity will be cautious, and TSMC and Korean memory manufacturers will be cautious when planning capacity too. If both sides slow a little, you may avoid a situation where one side has severe excess capacity while the other suddenly collapses.

    After all, this is already a $4 trillion company. Expecting it to double every year, or grow at a high double-digit rate, was never particularly realistic.

    Less circular financing and less risk from overextending financing support to customers, slower growth but greater shareholder returns—that could be a choice. Healthier finances and less excessive risk-taking do not necessarily mean a worse valuation.

    Apple went through this too. People once worried Chinese Android manufacturers would beat it, and its valuation fell below ten times earnings before later reaching more than thirty times. NVIDIA could even go through a similar phase: slower growth, healthier finances, but giving up some market share. I think it could become more ordinary while the risk of overexpansion declines.

    Raymond: Let me push this a step further. What you just described is the same pie being divided differently. Could the pie be 100 this year and have been expected to reach 200 next year, but now become 150, or even less, because Jensen has left? The AI hardware market itself could become smaller.

    TC: That is possible, but I remain relatively optimistic. Next year’s budgets are largely determined by the leading CSPs—cloud service providers. Those big companies account for most of the spending.

    Neoclouds, the newer cloud providers specializing in AI computing capacity, may face greater pressure, and their contribution to NVIDIA’s incremental growth is important. So I understand going from an expected 200 to 180. But whether it drops straight to 150, substantially below earlier expectations, I am not that pessimistic yet.

    Computing capacity is still in short supply. A slower pace does not mean demand immediately disappears.

    3. A cost of one and a selling price of four: how long can that premium last?

    Raymond: Which companies benefit, and which lose out? This is not investment advice; let’s discuss broad industry judgments.

    Trader Jiu: In the short term, the most direct beneficiaries are competitors that can take NVIDIA’s share. There are direct GPU competitors, and ASICs—chips designed for specific uses—will take some demand too. Chip designers and cloud providers developing their own chips may benefit, such as Google’s TPU and Amazon’s Trainium.

    Further out, I care more about how profits are divided. I believe AI applications will keep growing. But hardware used to take the largest share. Will it gradually have to yield profits to cloud services, models and applications?

    Previously, Jensen held the whole thing together and let everyone along the supply chain eat well. If professional managers run it like Apple in the future, will they still give suppliers such good gross margins? Even optical-module assemblers can get 40% gross margins. Could it eventually resemble Apple’s supply chain, gradually squeezing hardware prices and capital spending?

    Would cloud providers, large-model companies and companies building vertical agents then make more money? I do not know exactly how it will develop, but I think the division of the pie will change.

    TC: Let me add something about hardware. NVIDIA used to have advantages in software and working across platforms. Its general-purpose third-party solution was easy to use. But as coding capabilities improve, some of those advantages are weakening.

    When Jensen was there, he reserved supply-chain capacity on one side and helped customers finance purchases on the other. He had influence over both sides and took more risk on both. If his successor does not dare reserve as much capacity or support customers as aggressively, others have an opening.

    AMD, Google TPU and AWS, for example, could reserve more CoWoS capacity—TSMC’s advanced packaging process—and support new customers for computing capacity. That would divide up the pie.

    The most aggressive forecast I have seen says NVIDIA’s share could fall from 80% to 40% over a long period. Everyone is guessing about how long that would take, and I do not know whether it would actually reach 40%. But I think a downward trend is hard to avoid.

    NVIDIA is a company with a 75% gross margin. At that margin, something costing one has to sell for four to maintain the same level. That includes HBM—high-bandwidth memory—interconnects and chips, not just the piece NVIDIA designs itself.

    Other companies do not necessarily need that margin. If you make your own ASIC, you mainly pay design and foundry costs. Plenty of people might be willing to sell something costing one for two.

    NVIDIA could charge that premium because its software was easy to use and its platform was versatile. If more people buy computing capacity through the cloud without caring which card sits behind it, its cross-platform advantage weakens too. But if it is unwilling to accept a collapse in gross margins, it may choose to give up some share.

    Raymond: So this is essentially a “one whale falls, and all things flourish” story, right?

    TC: First, not that fast. Second, its ecosystem is still there. Enterprise customers deploying systems themselves may not be able to replace NVIDIA easily with AMD or another ASIC. They may still buy NVIDIA.

    Trader Jiu: I have a question. When Jobs left, Apple already had iOS and iCloud. It just had to maintain its moat. But models are evolving so quickly now, including the possibility of RSI—recursive self-improvement. Will CUDA’s moat remain that strong?

    TC: I would push back a little. NVIDIA’s competitive advantage already extends far beyond CUDA. It has gone from a chip supplier to a systems supplier. Software, architecture, memory, interconnects and how the entire system is assembled—it can deliver all of that together. That is a powerful capability, and it does not all disappear because Huang leaves.

    Raymond: Could that capability leave with Jensen too?

    TC: He built it, but the architecture already exists, and solutions such as Rubin and Rubin Ultra are already there. Continuing along that roadmap still leaves it competitive. I agree it has weakened at the margin, but you cannot say it only has CUDA and suddenly has nothing left.

    The big cloud providers already have strong systems-design capabilities, so NVIDIA’s competitive position against them will weaken more visibly. For enterprises without those capabilities that want a complete solution, NVIDIA may still be the best choice.

    4. If both shortcuts disappear, what can Chinese AI do?

    Raymond: When NVIDIA fell sharply, Alibaba fell even more, while domestic chip companies such as Cambricon rose. What is happening to Chinese companies here?

    Trader Jiu: First, your cloud business cannot buy the cards, and access may even be cut off later. Second, competing in large models becomes harder. Others are putting more computing capacity into inference and cutting prices too.

    I think the most hurt are Chinese cloud providers, data centers in China and Southeast Asia, and Chinese open-source models.

    TC: Beyond domestic demand, Chinese models previously had another big story: train overseas, use relatively low costs and more readily available data to build open-source models, then expand abroad to serve overseas customers.

    If restrictions extend beyond chip exports to China and also constrain Chinese companies’ use of those chips overseas, that obviously benefits their competitors. With fewer vendors able to provide the service, customers are more likely to stay on existing platforms.

    I also think model companies that charge more and target the higher end have a strong commercial incentive not to let others catch up quickly through distillation.

    Raymond: Suppose you went to Hangzhou or Shenzhen today to advise these companies on strategy. What could you suggest?

    TC: One option is self-reliance: sort out the chips yourself. But the pressure is not just in design. It extends to process technology and other areas, and none of that can be done overnight. The previous solution was to go overseas. If that path also faces greater risk, it really is hard to think of particularly good advice.

    Trader Jiu: China really needs its own NVIDIA. Otherwise, it cannot make this work.

    Previously, you could skip making chips and then use distillation to skip part of pretraining. If both shortcuts are gone, you have to start from scratch and go back to where things were a few years ago. And even after grinding away at it, you may not make any money.

    If the main revenue source for large models in their first ten years is still a paid-software model like SaaS—software as a service—you will find you do not have that many paying customers. In China, apart from people making short dramas, who else is willing to pay?

    TC: The conditions in the two markets really are different. AI replaces labor with algorithms and computing capacity. The price gap between them matters enormously.

    A programmer in the US might earn $200,000 a year. Replacing half their functions could be worth $100,000, and replacing more could be worth more. That $200,000 may only be what some junior employees earn.

    In China, you can hire someone for RMB 200,000, with people lining up. If the model is less capable too, the value generated by replacing labor is smaller. That is a difference between the domestic and international markets.

    If the domestic market is small, can you go overseas—effectively export? That depends on how creatively companies work within the restrictions. Serving overseas customers with domestic chips and domestic models is not impossible. If you run DeepSeek on Chinese chips in Southeast Asia and serve local customers, others may not necessarily be able to interfere.

    Trader Jiu: Damn, that is hard. You grind away, invest hundreds of billions, and end up serving people in Southeast Asia.

    Raymond: Chinese companies certainly face challenges, but there is no need to belittle ourselves either. Maybe DeepSeek will unveil something big in a couple of days. Anyway, we are counting on Mr. Liang for ideas.

    5. Can computing infrastructure still be built through a concentrated push?

    Raymond: AI is no longer just a technology-sector issue. Jobs, data centers, electricity prices and taxes all affect votes. How will the tug-of-war between the White House and Congress affect AI investment over the next two years?

    TC: It will definitely slow relative to the most optimistic earlier expectations, but I am not that pessimistic. Local demands are often about getting a share of the benefits.

    You build a data center here, disrupt local life, perhaps drive up electricity and water prices, and all the benefits go to big companies. Of course I want jobs, investment or compensation. This is not just about what a few senators say. Many decisions ultimately come down to the county level. Whether local people support you and whether you give back to the community matter a great deal.

    Approving every project and building extremely quickly is unrealistic. Banning everything is unrealistic too. If you do not want it, someone else will consider investment and jobs a good thing and welcome it.

    Ultimately, it depends on whether end users need that much computing capacity, whether GPU rental prices hold up and whether the data-center economics work. After accounting for some social costs, if the project’s return still exceeds financing and resource costs, it can move forward.

    Trader Jiu: My concern is that when partisan conflict clashes with the national economic interest, many people put partisan conflict first.

    Once Trump loses his grip, Republicans will no longer all be in the same boat. The establishment never liked the MAGA approach; it previously thought it could win elections. If it cannot win anymore, then I do not want that magic ring.

    Add high financing costs and resistance from local voters, and it becomes hard to concentrate resources behind a big push as before. There will be a lot of back-and-forth and checks and balances. It will not be easy.

    TC: We can foresee it being harder than before, and the market is pricing that in. The more important question is how much is already reflected in expectations.

    Some people previously forecast 50 or 60 GW of computing infrastructure being built annually over the next two years. I do not think that is achievable. But if you lower it to 30 GW, is that necessarily impossible? If the market has already pushed expectations very low, the question becomes whether that lower expectation can be met.

    Raymond: Looking specifically at Anthropic’s IPO—its initial public offering—what do you both think?

    Trader Jiu: I think it still has to go public. If it waits any longer, there will be no window. This AI bubble is developing so quickly. It is not necessarily completely over, but it feels as though it is running at triple speed. The financing window is very short.

    SpaceX may have caught a very good opportunity. The window left for Anthropic and OpenAI could keep narrowing. They need to seize the time available.

    6. Stocks are cheaper. Why is nobody setting a price?

    Raymond: I have been confused all week. The market has not fully digested the midterm elections or expectations of delayed data-center construction, and Huang has added another huge variable. Whether manufacturers keep expanding, what happens to orders, how much Anthropic can raise in its listing—all of it is changing.

    What exactly is the market expecting now? Where are we in this process?

    TC: My rough guess is that 2026 is nearly over, and most 2027 orders already involve commitments. There are advance payments and non-cancellable orders. Whether construction or procurement, those are more certain than things further out.

    On expected 2027 earnings, NVIDIA trades at a little over ten times. Some interconnect-related companies are more expensive, at roughly twenty times, while memory could be two to five times. The market seems to be treating the 2027 numbers as a peak, perhaps even assuming a decline in 2028.

    If 2028 does not decline, and there is no significant decline in 2029 or 2030 either—perhaps even gentle growth—then current valuations are already very low. But if you think 2028 earnings will be only half of 2027’s, forecasts need to be cut sharply, and today’s prices may not be cheap.

    The key is whether you think the aggressive 100, 200, 300, 400 growth path merely becomes gentler growth, or whether things really collapse afterward. Those two judgments mean completely different things for valuation.

    Trader Jiu: Expectations are very low. Nobody can see 2028 clearly, so nobody dares assign a valuation. On near-term earnings, you could give something a P/E—price-to-earnings ratio—of two. One or ten could work too. There is no leader setting the price of these assets now.

    Previously, someone was doing that. China had some young mutual-fund managers, and the US had “young gun” fund managers such as Leopold. That force is weaker now. Many long-only institutions—investors that mainly take long positions—are underweight. Perhaps they held NVIDIA at benchmark weight and later cut it too.

    I have spoken with friends at several large institutions. They are trading swings as well: buy after a sharp selloff and sell the next day. When funds of that size do it, it feels absurd.

    TC: I think 2027 could be a fairly important year. Capital-spending expectations may be slowing, but spending is still out of balance with revenue.

    Take cloud businesses. Investment still exceeds business revenue, making people worry about overinvestment and demand failing to materialize. But if revenue sustains another year of high growth while capital expenditure gradually flattens, the gap will narrow.

    At some stage, the market may become more confident that free cash flow has a strong chance of turning positive in 2028. Only then might money from traditional large institutions—the “old guard,” as we call them—be willing to come in. Right now, they cannot prove this definitely is not a bubble.

    Cloud computing went through something similar. Investment initially far exceeded revenue, and several companies’ cloud businesses lost money for a long time. As profits improved, people stopped arguing every day over whether cloud was a bubble or capital expenditure would fall off a cliff.

    It is perfectly normal for large asset managers to buy Microsoft and Amazon now. Cloud once looked like something only the young guns would buy. But at a certain point, cash flow and revenue made it intelligible to more investors, and the mix of capital changed.

    AI could go through the same process. I am not saying we are there today. But at some point, the gap between revenue and investment expectations could gradually close, reducing the market’s fear of a bubble.

    7. People with monthly AI bills of 200 are pricing AI

    Raymond: Could I watch another indicator? If Anthropic goes public and then reports very strong quarterly revenue, with its annualized revenue continuing to rise, could that also give the market a lot of confidence?

    Earlier, people worried about competition from Chinese models, chips and OpenAI. If annualized revenue keeps accelerating, the market might think reaching 150 billion or 200 billion was easier than expected, and then guidance gets raised again. I do not know—those numbers are just guesses—but could that kind of result change investors’ attitudes?

    Trader Jiu: Of course. That would be a game changer. Proven demand is the strongest signal. It could even give the market more confidence than a company projecting higher capital expenditure for the next year.

    TC: I agree, but I would add that we cannot use just two model companies to represent all AI demand.

    Many applications consume computing capacity through open-source or proprietary models. They do not use Anthropic or OpenAI, but they still pay cloud bills. Some usage happens internally at companies, so you do not see an application’s user count suddenly surge. It is difficult to observe directly.

    It is a bit like smartphones. Apple is very important, but Apple’s sales and global smartphone sales are different things. Apple can be an important signal without being equivalent to the entire industry.

    I do not think Anthropic equals everything. AI applications already extend far beyond Anthropic.

    Raymond: I see it a little differently. It is now emphasizing that it can use all kinds of chips and mix them together. From its perspective, they form a pool of computing capacity it can adjust at any time. In effect, model companies do not want to be constrained by a single hardware supplier.

    And its ARR—here meaning an annualized revenue metric—could become one of the most important monthly indicators for the market. It is a more timely signal than cloud revenue reported only quarterly, and easier to obtain.

    TC: From the perspective of someone spending money to build data centers, I still think cloud bills matter more.

    Cloud revenue could accelerate while one model company’s revenue slows. Both can happen at once. You watch the model company because there is no better high-frequency indicator in between. But ultimately, validating the whole investment in computing capacity requires broader revenue and usage data.

    In many technological revolutions, Wall Street people are late adopters themselves. They use the technology less and need more evidence to understand it. People in Silicon Valley are genuinely willing to pay for computing capacity. You cannot simply assume they are all fools buying like crazy despite knowing there is no demand.

    The two groups disagree sharply about the prospects for the underlying applications.

    Raymond: Last time we discussed this, you said something I thought was memorable: AI is being priced on Wall Street by people whose monthly AI bills are only 200, while the actual frontier users have monthly bills of at least 10,000.

    The people spending 10,000 find it extremely useful and need more computing capacity. The people spending only 200 think there is excess capacity and the bubble is about to burst.

    TC: Someone who only chats with a chatbot each month cannot understand why you would be willing to spend $10,000.

    Raymond: Companies such as Jane Street that are already making money from this may even build their own data centers. They spend so much on computing capacity that building their own becomes more economical.

    8. After high-growth stocks peak, do they become high-dividend stocks?

    Raymond: So much has happened in 2026. I have learned a lot this year about the Middle East, oil in different countries, tariffs and US elections. I suspect A-share investors have learned even more than I have.

    But this year is largely behind us. Looking ahead, what are your important calls for 2027? We do not have to stay with AI. Chinese stocks listed overseas and energy are fair game too.

    Trader Jiu: I am an A-share investor, so I will go first.

    First, I still believe AI capabilities will continue growing exponentially. They will move beyond coding into more fields and change how people work.

    But in investing, hardware will diverge. Companies with high valuations that need constant short-term excitement may never return to their peaks. Other companies, cyclical stocks in the traditional sense, will keep falling until they simply cannot fall much further and become high-dividend stocks.

    Like oil-tanker stocks back then, high dividend yields and shareholder returns can support the share price, followed by a gradual bull run. But this only applies to a few areas where demand is relatively certain and suppliers stop expanding capacity aggressively. Some memory manufacturers beginning to halt expansion is the kind of signal I would watch.

    Cloud providers may emerge from the mire first. They could take a larger share of the industry’s profits while winning recognition from both Silicon Valley and Wall Street, potentially supporting better valuations. If interest rates fall too, lower financing costs will also make their financial statements look better.

    I think the AI trade in A-shares is pretty much over, especially the overseas-demand part. With Jensen gone, there may still be pockets of opportunity in the domestic supply chain. A-shares can only hope for a new round of fiscal stimulus for domestic demand and see what new direction emerges. But the previous approach is basically finished.

    I think gold and Bitcoin can establish a medium- to long-term bottom around the end of this year or the beginning of next year.

    Raymond: TC, anything to add?

    TC: Strategically, we remain bullish on AI. This is the first time technology is replacing human intelligence. Humans have never had enough computing capacity of their own. Now there is a technology that can accelerate that process, but many people are unprepared and do not know how to use it.

    It is a little as though we are still living in a world of handicrafts, except everyone has been given a few electric drills. The real change comes when an entirely new assembly line appears, with AI doing a great deal of the work and people supporting the line.

    Buying a few power tools for each person does not cost much. But when you rebuild the entire production line, spending far more on machines than on labor is perfectly reasonable, because the machines generate far more productive power than people do.

    The intelligence revolution could be similar. You cannot look only at whether large companies are willing to spend 200 per employee today and assume that is the ceiling on demand. New companies will spend in completely different ways.

    Model companies themselves are the clearest example. Their spending on computing capacity can far exceed employee salaries. There may be more and more companies like that.

    I think you once told me about a law firm. How much can you expect an existing firm to spend giving each lawyer a chatbot? But new firms could emerge where AI does most of the work and the lawyer adds the final stamp. Such a company would view computing capacity completely differently.

    So we think AI remains the most strategically important opportunity for our generation over the next five or ten years. If we can go to the Moon or Mars later, that is another discussion. For now, this is where we place that strategic priority.

    Tactically, we still need to watch areas of excessive expansion. If demand falls while suppliers keep aggressively adding capacity, fundamentals could suffer a stampede. I do not think next year will be a particularly big beta year (with returns driven by the market’s overall rise or fall). Alpha (excess returns) will matter more.

    9. Does cutting costs necessarily increase revenue?

    Raymond: I have a concern. AI may keep evolving and productivity may rise, but that does not necessarily mean revenue and GDP across the economy rise with it.

    The change may come through cost-cutting first. Microsoft, for example, introduced a voluntary-exit plan covering roughly 60,000 employees. Microsoft should be one of the companies best at using AI worldwide, yet it is still making this kind of workforce adjustment. I think that is a signal.

    More technology companies may take similar steps, followed by non-technology companies, putting pressure on the labor market. Right now, the US job market still looks strong, with multiple jobs competing for one person and little reason to cut rates. But a change in that relationship could also change interest-rate and market expectations.

    TC: I agree about displacement in the labor market. We may only be seeing the first signs now, and it could intensify next year. But cutting costs also improves efficiency. If my costs fall by half, can I cut prices, gain a larger market and turn that into revenue?

    Raymond: I understand what you mean. Efficiency can improve. But suppose you do not cut prices: does your revenue necessarily increase? Your revenue is someone else’s cost. It does not come from outer space. Others also have a strong incentive to replace your service with AI. SaaS has already shown us some of these changes.

    This is a dynamic game. But what seems more certain to me is that employees get laid off, go home, and expenditure immediately drops on the financial statements. That is my biggest concern.

    TC: The Industrial Revolution went through something similar. More money went into machines, machines took some people’s jobs, and society rebalanced. Some people bore the frictional costs, while others benefited more.

    But there will also be companies that previously could not afford to hire so many people and can now grow faster through computing capacity. I am more inclined to see it as a redistribution.

    Looking at it from a very broad macroeconomic perspective, global GDP is roughly $120 trillion, and US GDP is more than $30 trillion. Suppose global nominal growth runs at 4% to 5% annually in the coming years. By 2030, global GDP might reach around $140 trillion. The US might also reach about $38 trillion by then. Compared with current levels, that is approximately $20 trillion more globally and $8 trillion more in the US.

    Optimistically, how much of the additional GDP over the next few years could benefit from AI? I think a substantial share might be related to it. At that scale, there is less reason to worry that AI has no ability to generate revenue at all.

    Raymond: Thanks, TC, for giving me some confidence. We can come back and check in a few years.

    But personally, I really do worry about being replaced and not earning a penny. Especially the more I use AI, the more I feel I have little value. It feels as though that could happen any minute.

    10. Will Jensen come back?

    Raymond: One last speculative question. Sam Altman was pushed out by his board and later returned to OpenAI. As things stand today, could Huang come back? None of us has much basis for this; let’s just take a guess.

    Trader Jiu: If the share price falls enough, major shareholders such as BlackRock will feel the pain. Could they put pressure on the board? OpenAI had a strategic shareholder like Microsoft that could actively exert influence. NVIDIA depends on its public-market shareholders, and institutions such as BlackRock do not necessarily look that proactive. Maybe they will just let it go.

    TC: OpenAI was smaller at the time, and a very high proportion of employees signed in support. NVIDIA has tens of thousands of employees. You would need a substantial proportion to say, “If Jensen does not come back, I leave.” I do not think that proportion would be very high.

    And with institutional shareholders, it is hard to say whether they would actively stand up and back Jensen directly.

    But if they really cannot find the right person, keep going around in circles, and calls for his return grow loud, you cannot rule it out completely. Those people do not naturally sit on the board forever. Their seats are not welded in place.

    Raymond: I see two variables. One is that finding the next CEO really is difficult; people struggle to imagine what that person should look like. The other is a continually falling share price. At some point, could major shareholders change their minds?

    At the time of recording, I saw Polymarket probabilities of 4% for a return by the end of November, 10% by the end of this year and 30% by the end of next year. I will keep watching. If trading behaves unusually, someone may know more than we do.

    Trader Jiu: In terms of probability pricing, next week’s earnings report is important. Who will speak then? People are used to Jensen’s expansive vision and big promises. With someone else speaking, they may immediately feel something is missing.

    TC: There is also a practical issue. Jensen used to confidently commit when asking for capacity and arranging financing. Will new management dare make equally strong commitments? If those commitments weaken, market expectations change too.

    Raymond: Analysts will definitely ask about the outlook for the coming year. If management promises too much, people will ask: has TSMC agreed?

    Trader Jiu: I do not think Jensen is likely to return during that round, but at least the Polymarket probability of his return will surge.

    I would actually like to ask you something. You went to Columbia as an undergraduate. Ethnic Chinese people have never occupied so many positions in America’s most mainstream industries as they do in this wave. Do you think that can continue?

    Raymond: I think it can. But first, being ethnically Chinese and taking a pro-China position are different things. I do not expect all ethnic Chinese people to have the same feelings about China. Their environments and nationalities differ.

    The change in the talent base, however, is obvious. On my last visit to Silicon Valley, I attended a HYSTA—the Hua Yuan Science and Technology Association—event and saw people such as Ruoming Pang and Yuandong Tian onstage. I also spoke with a VC who mainly invests in ethnic Chinese founders. He may see more than a thousand such projects each year.

    So many people work at companies such as Meta, NVIDIA and Qualcomm, understand US culture and its ecosystem, and then start their own businesses. Ten years ago, Eric Yuan was still a particularly prominent example of someone who grew up in China and started a business in the US. There are more and more such people now, and the foundation has changed. Lin Qiao at Fireworks, building an AI inference platform, is one example of the growing number of these founders. I think we will see more people who attended middle or high school in China and later built successful companies in the US.

    TC: I think forces produced by a technological revolution are difficult to stop through politics alone. Whether in algorithms or the supply chain, AI cannot do without ethnic Chinese people—or, more broadly, Asia. Unless you give up the industry, you have to rely on these people and capabilities.

    Raymond: Software and hardware also differ. Software companies may see China as a competitor and take a harder political stance. Hardware companies depend on the Asian supply chain, so their attitude may differ. Who’s your daddy? TSMC.

    I once made a rough count based on names at Anthropic and OpenAI. Around 30% of people appeared to be Asian. Half of those, or roughly 15% of the total, had names written in pinyin. The other half had names using Korean, Hong Kong, Taiwanese and other spelling systems.

    People with pinyin names are very likely to speak Chinese. Some were born and raised in China; others left in middle or high school. If you go to Silicon Valley and look at the people outside company offices, you can feel that change.

    Trader Jiu: Ethnic Chinese people really do have power—real power. How everyone gets along in the future will have to be worked out along the way.

    TC: Maybe the next Jensen Huang is Lisa Su. I do not think there is any need to worry about ethnic Chinese people. Whatever happens, it is very hard to change the fact that the supply chain is in Asia. If you want to make these things, you have to rely on the Asian supply chain.

    If you're working on this too — or you think we've got it wrong — write to us at [email protected]; if you'd rather not write, just leave your email below.

    Topics AI Infrastructure & PowerPublic MarketsChina & the WorldWork & Organizations
    Disclaimer. This content is for general informational purposes only. It does not constitute investment, legal, tax, or accounting advice, nor an offer to sell or a solicitation of an offer to buy any security or interest in any fund managed by Mossfire Capital. The Firm and its affiliates may hold positions in the instruments discussed; actual positions may differ from — and even be contrary to — the views expressed, and are not disclosed. See our full Disclosures.