Topics Pioneers Insight Method Author
Back

168X Conversation with Investment Talk Jun: How Will AI Be Priced After the Midterms? #2

·168X

Rendering the long image…

Topics AI Data Center Power Bottlenecks Trillion-Dollar AI Capex and Bubble Risks GPU Depreciation and Cloud Market Shakeup Commercial Shifts in Mobile Recommendation Algorithms Enterprise AI ROI and Cost Optimization AI Venture Funding and Bubble Risks

Summary

  • Talk Jun reads the Texas dispute as a negotiation over new approvals and how the benefits are shared, not a halt to AI data center construction in the US. As he understands the policy stance from coverage he heard, already-approved projects can continue, enough to cover 2027 and the first 1–2 quarters of 2028; new approvals could face a 3–4 month pause, with the actual impact depending on how quickly legislation moves. He says the market’s leap from a Democratic election win to a blanket construction freeze is “a very big misreading.”
  • He is willing to add some beta amid indiscriminate panic, but has no idea when the market will stop falling. Wednesday’s semiconductor losses and CSP gains still fit the logic that slower capex could improve cash flow; Thursday’s 4% drop in even the Dow looked more like deleveraging and forced selling. “A bottom is an event; a top is a process.” He is waiting for a clear catalyst and worries that falling confidence could, in turn, hurt financing and AI demand.
  • Under the extreme assumption that everyone stops building, existing compute would become more valuable and cloud providers’ cash flow could improve. That does not mean new clouds are risk-free: high financing costs, orders that are not yet fully funded, customer credit and delivery capability could all become problems. But if new capex is forcibly halted, blaming every compute holder’s decline on financing risk conflicts with the “everybody stop building” assumption.
  • Delayed capex does not mean canceled demand, and the semiconductor supply chain cannot be treated as one short trade. He gives the example of buy-side estimates for 50% capex growth in 2027 falling to 40%, with some spending potentially shifting into 2028. GPU deliveries would be affected first; HBM and CPUs also depend on system specs and pricing. If slower server growth frees up memory capacity, consumer-facing businesses at Intel, Apple and Qualcomm could see cost pressures ease—provided AI demand keeps growing.
  • Intel’s PC misread deserves a price, but he has not abandoned the thesis on servers, packaging and foundry improvement. He thinks a 10%–15% post-earnings drop is understandable, while a 30% drop has emotion mixed in. Lip-Bu Tan’s discipline of “I invest when I see demand” also helps explain why upstream capacity could not keep up once demand arrived. Intel has slipped to his third or fourth-largest holding as its price fell; Microsoft and Amazon are now his top 2.
  • Meta has more direct monetization levers, while Google’s key test is turning AI search traffic into ad revenue. Meta already has traffic, commercial use cases, models and compute, and can manage profits by selling excess compute. Google’s traffic and growth are at records, but its ad products for AI Mode are not keeping up with demand. He personally thinks Google can solve this, while keeping his “who knows?” caveat, and compares the challenge with Reels taking 6–9 months to find its ad model.
  • His 2027 portfolio remains centered on AI applications, with a target allocation of about 20% to semiconductors and hardware. For AI agents, he continues to watch Google, Meta, Apple and Qualcomm; Apple’s valuation is too high for him to buy for now, while he has kept the Microsoft position he built at the start of the year. In healthcare, he is focused on operations, logistics and supply chains, and hospital efficiency—not drug development he does not understand. His read is that “everyone is killing the cycle,” not that AI adoption has slowed.
  • Talk Jun does not plan to participate in an Anthropic IPO, mainly because of its funding needs and potential selling pressure. The host mentioned a listing around November 16 and inferred that the roadshow might start November 9. Talk Jun thinks a weaker market will certainly mean a lower valuation, but does not know whether it will affect the amount raised. For both Anthropic and SpaceX, he would rather wait for financing pressure and early investors’ exit demand to clear: “Buying its stock is just lending it money, isn’t it?”

Deep dive

1. Texas pausing new approvals does not mean approved data centers will stop work

  • Responding to the market selloff after the host described Democrats as having “swept the Senate and both chambers of Congress,” Investment Talk Jun first unpacked the headline. Chinese-language coverage he saw summarized Texas’s position as “data centers must pause,” but his understanding after listening to the interview was that approved data centers would keep going; only new approvals would be paused until lawmakers spell out the rules.
  • The timing gap is his first point: a data center takes at least 1–1.5 years from construction start to delivery, while chips are typically ordered about 6 months in advance. Data center capex therefore comes first, chip capex later. In his view, currently approved projects can cover all of 2027 and the first 1–2 quarters of 2028, so a pause on new approvals cannot be equated with near-term chip orders disappearing.
  • For new projects, he estimates there could be a 3–4 month gap in approvals, but the end result depends on the pace of legislation and negotiation. The requirements did not appear from nowhere: he says Texas had already asked applicants for more documentation and delayed some approvals. The new policy tightens the screws further; the longer the delay, the greater the impact on later deliveries.
  • He uses the backlog to explain why approval figures cannot be treated as real investment demand: applications in Texas exceed 400 GW. At about $50B per GW, that implies $20T in capex. “Who has the money to invest $20T in capex?” In his view, there was already a great deal of speculation in the pipeline that needed to be screened out.

2. State governments want a say in how the benefits are shared, and tech companies will adjust their negotiating stance

  • In describing Texas’s policy demands, Talk Jun focuses on “a seat at the table”: residents want a voice, and the government wants tax revenue while ensuring data centers do not affect residents’ water and power. He notes that some areas offer tax breaks on electricity use, and sees the dispute as residents and government both wanting “a piece of the pie,” rather than rejecting all investment.
  • He thinks the added requirements may not materially change project economics, based on the cost structure he cites: GPUs and memory account for 80%–90% of data center costs, while power, building shells and the rest make up 10%–20%. New taxes would largely fall on that latter portion. He also cites SemiAnalysis as saying that the highest-paid electricians in Texas can earn $500,000 a year, evidence that construction is already creating scarce, well-paid jobs.
  • The host adds that elected politicians often open with extreme statements to reassure voters, then negotiate the terms. Talk Jun agrees: “You don’t show your hand.” The final compromise should land “somewhere in between”; if one state’s terms are too demanding, investment could move to another state willing to take it.
  • He thinks tech companies will adapt quickly to political requirements, citing Sam Altman’s statement that OpenAI’s data center partners must promise not to affect residents’ water and power. He also says about 50% of Anthropic’s spending goes to the big cloud providers, helping explain why frontier model companies want to build their own infrastructure. Building in-house still means accepting new requirements; it does not bypass negotiations over who benefits.

3. The election was a catalyst; leverage and forced selling magnified the decline

  • Talk Jun admits he did not anticipate how aggressively the market would sell off after the election; even looking back, he would not readily have predicted a 5% drop in the S&P 500 and 6% in the Nasdaq. He does not dress up a post hoc explanation as foresight. Instead, he starts from the price action and separates policy impact, positioning and market sentiment.
  • That day, he saw and heard reports of some hedge funds liquidating, returning capital or posting poor returns. He suspects one-way semiconductor bets layered with leverage were liquidated again: after the October decline, investors re-levered, then faced a new catalyst and were forced out. “Leverage is one of the original sins behind a big selloff”; the policy news provided the trigger.
  • He draws a sharp distinction between Wednesday and Thursday. When the election result came out, futures briefly fell about 2%, but CSPs ultimately closed higher on Wednesday, on heavy volume, and sector rotation was still visible. On Thursday even value stocks fell, with the Dow down 4%—more like a macro event than an industry trade focused only on data centers.
  • TSMC and ASML also fell about 10%, which is why he calls it an “indiscriminate selloff.” The market stopped distinguishing alpha within the supply chain and sold the entire AI-related beta; index selling also means selling everything by weight. “Sell first, ask questions later.” Whether the call was right comes afterward.

4. Take a blanket construction halt to its extreme, and existing compute and cash flow could benefit

  • Talk Jun tests the market narrative with an extreme assumption: “everybody stop building.” If all of the roughly 23–25 GW of existing US compute capacity stopped adding new construction, existing compute should become scarcer as long as demand remained strong. Owners would also be pushed to monetize rather than expand. He is not predicting this will happen; he is testing the pricing logic.
  • For the big cloud providers, his question is: the market disliked high capex, so if they are now forced to spend less, why are they all being sold when cash flow and near-term financials should improve? “I’m not spending the money. I’m being forced not to spend it.” The installed compute base should also be easier to monetize. That is why he thinks Wednesday’s CSP gains made more sense than Thursday’s blanket selloff.
  • The host’s challenge is worth preserving: new clouds are heavily indebted, have weaker financial structures, and their orders may not be solid; Elon Musk has even said that perhaps 90% will die. Talk Jun agrees there is a basis for concern: new clouds face higher financing costs than big clouds, and may lag in operations, construction speed, quality and downtime. If they cannot pass financing costs through to contract margins, they will run into trouble.
  • But under the premise of a complete halt to expansion, financing needs would also fall, so the logic should differ from a world of continued expansion. He thinks rising short-term rental prices could benefit new clouds more directly, citing what he thinks may be Nebius as an example. Over the past week, GPU prices rose 5% at the low end and more than 10% at the high end; B-series prices rose about 13%, and short-term rental prices also increased, though they have not clearly risen further over the past 2 days.

5. Musk criticizes new clouds but cannot avoid massive financing needs himself

  • Talk Jun applies the same funding constraint to SpaceX: if it wants to build 10 GW in 2027, at about $50B per GW that means $500B in investment. He estimates that cash on hand, Starlink operating cash flow and credit facilities combined would not be enough, leaving roughly $200B–$300B to raise. Even big tech would struggle to absorb that level of spending all at once.
  • Debt cannot be piled on indefinitely. He says recently issued debt is already carrying interest above 8%; as leverage rises, the next financing will only cost more. An equity raise would also be constrained by a recent listing and shares that are not yet fully unlocked. Musk has questioned whether SoftBank has the money to build AI infrastructure; Talk Jun says the same question can be asked of Musk.
  • He does not understand why Musk is taking shots at new clouds while financing conditions across the industry are fragile. With sentiment deteriorating, “nobody benefits.” On the theory that Musk wants new clouds to go bankrupt and then acquire them, he adds that Musk has said he has no interest in acquisitions; that cannot be treated as an established strategy.
  • He offers another possibility, explicitly as speculation: if new clouds exit, Musk might hope to secure more Nvidia Vera Rubin allocations, take over the short-term rental business himself and become a bigger customer for Jensen Huang. If compute paid back in just 8 months, or even 6, the business would be tempting. But he ends by saying he still does not know why Musk is doing this.

6. 2027 capex may shift into 2028 rather than disappear

  • Talk Jun’s baseline is that Wall Street expects 35% capex growth in 2027, while the buy side expects about 50%. If approval delays cut the buy-side estimate from 50% to 40%, that does not automatically mean demand has disappeared. The key is whether big tech gives up its willingness to build, and at the time he saw no signs of a deliberate capex slowdown.
  • He thinks the deferred spending is more likely to move into 2028 than go toward buybacks. Growth in 2028 might have been 20%–30%; after the delay, it could be 30%–40%, perhaps roughly in line with 2027. That is a change in the timing of allocation. Whether it plays out that way still depends on when legislation takes effect and how quickly projects move once approvals resume.
  • Amazon is his example for why he retains a long-term demand view. He cites Andy’s comments that Amazon can grow 40% over the next 2 years and reach $1T in revenue by 2030. The former relies on past investment; the latter requires capex in 2027–2028. He does not deny the impact of delays, but rejects the leap from delay to canceled long-term investment.

7. GPU deliveries are the first-order impact; HBM, CPUs and consumer businesses need another layer of analysis

  • Talk Jun acknowledges that if data center deliveries fall short of expectations, GPU shipments will be affected first. But some Nvidia and AMD GPUs were previously downgraded because of limited HBM capacity. If GPU deliveries fall, original configurations could return: “Put in as much HBM as the system calls for.” HBM therefore cannot simply be cut in proportion to GPU volumes.
  • CPUs require the same distinction. If fewer data centers can be built, each GW will need more token throughput, which could drive adoption of higher-spec companion chips. Unit volumes may fall while unit prices rise; the ultimate impact on HBM and CPUs depends on how much GPU delivery falls and how much configurations improve.
  • He also questions GPU pricing: if buyers can pay back a GPU in 6–8 months, with the remaining 3.5 or 4.5 years mostly producing returns, why wouldn’t Nvidia raise prices? That is why he does not rule out “volumes being forced down somewhat, but prices going up.” This does not deny delivery risk; it highlights that revenue and profit also depend on selling prices.
  • Another link the market may be overlooking runs through consumer markets: server demand is taking memory capacity and pushing prices higher, already weighing on PC and phone demand and margins. If server growth slows, capacity and cost pressures could ease for consumer businesses, potentially benefiting Intel, Apple and Qualcomm. He stresses that the whole thesis depends on AI demand continuing to grow.

8. A bottom needs a catalyst, and falling confidence could feed back into fundamentals

  • Asked when the market might stabilize, Talk Jun answers directly: he does not know. Looking back at selloffs this broad, the last comparable episode was the tariff shock on April 3 and 4, 2025, when the market fell even more over 2 consecutive days. In his experience, when correlations approach 1 and institutions are forced to hand over positions, adding some exposure or beta is worth considering; that does not mean knowing the date of a reversal.
  • He is waiting for a defining event: “A bottom is an event; a top is a process.” A person’s remarks, a news story or a clear policy signal could mark the bottom. If the event appears and triggers a rebound, that gives investors more observable evidence of a bottom than simply saying, “It’s fallen a lot.”
  • The biggest risk is that the decline “feeds on itself”: lost confidence in public markets could hurt private financing; if AI-native companies then fail, model demand could weaken and transmit to frontier labs and CSPs. He rejects the idea that stock prices and fundamentals can be fully separated. Capital markets themselves can force fundamentals to deteriorate.
  • The wealth effect is part of this feedback loop. He says the top 10% of earners in the US account for about 50% of consumption, so falling equity wealth can quickly affect spending. He points to the 2022 tech selloff and the sharp drop in Silicon Valley home prices. The stock market can therefore affect not just financing, but consumption, GDP, employment and inflation.

9. Anthropic’s IPO window has worsened, and political alignment cannot erase its funding needs

  • The host mentions that Anthropic is considering an IPO around November 16 and infers a roadshow start on November 9, worrying that the offering could draw money away from other stocks. That is a different market environment from SpaceX’s IPO, which he says was on June 16. Talk Jun thinks a lower valuation is certain, but does not know whether the amount raised will be affected.
  • The host also offers a counterargument: Anthropic’s focus on safety may make its values more aligned with Democrats. Talk Jun allows that it “may be more aligned in terms of values,” and even imagines the company could get an easier path to building its own data centers. But big tech already builds relationships across political camps; values alignment alone cannot explain commercial outcomes.
  • When deciding whether to buy, he cares more about the obvious funding needs. SpaceX plans to build compute at scale but does not have the corresponding cash. In his metaphor, buying the stock now is “lending it money to build data centers.” Investors who want to act as that source of capital can participate; he would rather wait.
  • Another concern with Anthropic is early shareholder selling. He thinks the company reached enormous scale in a very short time, leaving employees and institutions with large paper gains and an incentive to cash out after listing. He would rather wait for the selling pressure and for “the people who need to leave” to do so; liking AI does not mean having to buy at the IPO.

10. Intel’s PC misread should be held to account, but it does not settle the long-term business outlook

  • Asked about his “largest position,” Talk Jun first corrects the premise: “not anymore.” Intel has slipped to third or fourth place as its price fell; he is not describing an active reduction. On the earnings report, he agrees management clearly underestimated the PC slowdown and that the stock deserved to fall. He thinks 10%–15% is understandable; 30% is harder to justify.
  • He separates the current drag from the sources of growth. PCs are indeed a problem, but the main upside he watches is servers, packaging and improvement in foundry losses. On those fronts, he has not seen enough negative change to materially overturn his forward view, so he has not rejected Lip-Bu Tan and management wholesale on the basis of one report.
  • He compares Intel with AMD in 2022: results and earnings growth were still good, but the market had already priced in a downturn in the PC cycle. As he remembers it, Lisa Su pushed her bottoming estimate from Q1 to Q2, then to Q3, repeatedly underestimating the slowdown. That error did not mean investors should then dismiss AMD management’s overall execution.
  • He also draws a line around what he can know. Without a team, channel access and time spent deep in the supply chain, it is hard to spot in advance when management is too optimistic about PCs. “Personally, I don’t think I have that ability.” He keeps his focus on the bigger picture, leaves some specific judgments to management and accepts that they cannot get every call right.

11. Intel ordered upstream capacity too late, the other side of conservative capital discipline

  • On the question of why Intel did not order more substrate if it knew CPU demand was strong, Talk Jun traces it to Lip-Bu Tan’s cautious approach when he first took over. The company said it would not invest recklessly without clear demand. Revenue growth picked up while capex stayed flat, and the market was pleased; that was part of repairing the earlier loss of control over capital spending.
  • He thinks management only realized how short CPUs were in Q1, “when they sold all the junk CPUs in inventory like they were treasures.” By then, upstream orders could already have lead times of 9 months or even more than a year, creating shortages in Q1 and Q2 and a gradual recovery only in Q3 and Q4.
  • He interprets Lip-Bu Tan’s recent visits to different parts of Asia as an effort to secure capacity, though that is his own read. The core trade-off is: “I invest when I see demand.” That may mean acting too late once demand is confirmed. He is not arguing for a return to heavy investment before demand is visible; he accepts that discipline comes at a cost in delivery speed.

12. Meta has agent use cases and a lever to manage profits by selling compute

  • Talk Jun thinks Meta has the clearest path to AI agent adoption: “It has the traffic.” It already has monetization channels and use cases, alongside models, data centers and its own chip efforts. Compared with companies still looking for an entry point or business model, Meta can roll agents out through its existing businesses or sell excess compute.
  • He says Meta did not sell compute to Anthropic in Q3, but could monetize it that way if Anthropic needed it. He compares this with SpaceX, which he says sold $6B of compute last quarter. The cost comes from past investment and must be depreciated either way, so he thinks incremental compute sales can quickly improve both the top and bottom lines; “basically,” the money earned goes into profit.
  • Management’s real allocation decision is between training, deploying and promoting Muse, and directly selling bare compute. He thinks selling $1B–$2B of compute would have a major impact on the bottom line. Having that lever would make it easier for the CFO to manage operating income. That is why he values Meta’s financial flexibility.

13. Google’s search demand is growing, but its ad products have not caught up with how people use search

  • Talk Jun thinks Google’s latest model has returned to the frontier; the more important question now is how traditional search transitions to AI Mode. He saw AI search pick up noticeably after June, especially in July and August, and thinks agents are part of the reason: one query might search 17 or 18 webpages, directly or through Google, generating far more visits than before.
  • As he describes it, Google’s traffic and traffic growth are both at records. The problem is monetization. New AI usage is driving traffic growth, but the existing ad products have not caught up. He also compares Muse’s roughly 10M users with Gemini’s roughly 1B to show that Google’s challenge cannot be explained simply by a new entry point taking users away.
  • The analogy he cares most about is Meta in 2022. After TikTok’s rise, Reels grew quickly, but Meta could not simply transfer its text and image advertising playbook to short video. It had to figure out where ads should go and what would work; that took about 6–9 months. Google may now be in a phase where demand is ahead and ad products are catching up.
  • He sees monetization in AI Mode as an industry-wide challenge, noting that OpenAI has ads and Claude does not, and that an OpenAI and Walmart partnership was canceled after low conversion. On whether Google can solve it, he keeps his “who knows?” caveat while personally judging, based on its advertising experience, “I think Google can solve this problem.” He gives no timeline.

14. Hong Kong stocks’ relative outperformance starts with a narrative; long-term investors still need policy confidence

  • The host observes that Hong Kong stocks are holding up better during the US market crash and asks whether investors are trading the idea that “slower US AI means China benefits relatively.” Talk Jun thinks this is more “narrative-driven”: money is looking for a haven or pricing in a possibility early. But when it comes to fundamentals—for example, whether Nvidia might therefore sell more chips to China—both he and the host think that remains difficult.
  • He thinks international investors need to regain confidence in policy predictability before holding Chinese assets for the long term. His clearest example is the requirement that private tutoring companies become nonprofits. Even investors who like Tencent, Alibaba and Meituan need time to overcome doubts left by that rules shock. He also says explicitly that he does not follow Hong Kong or mainland companies closely; he already cannot keep up with all of the US market.

15. He still likes 4 AI agent companies, but valuation determines whether he buys

  • Google, Meta, Apple and Qualcomm remain Talk Jun’s preferred AI agent beneficiaries. Google has both an AI tailwind and an advertising monetization headwind, which the company has to solve itself. He does not erase the uncertainty: “There’s risk and reward because there’s uncertainty.” Market concerns are part of the risk-reward equation.
  • On Meta, he notes that it was trading above $600 at the time, with a recent low around $550, after moving sideways for 2 years. He thinks it has relatively little speculative positioning, while its path to agent adoption is clearer and it is already leading. That makes it different from the semiconductor trade that had attracted a great deal of capital.
  • Qualcomm’s decline is harder for him to understand: its valuation is relatively cheap, it is not heavily speculated on, and it did not participate in the earlier semiconductor rally. He expects data center revenue to start coming through gradually next quarter—that is, next year—so he does not think it should simply take the same liquidation as crowded AI hardware names.
  • Apple is a case of liking the business but not accepting the price. He reduced his position around $330–$340 and admits, in hindsight, that this may not have been a good decision, since holding through the selloff would at least have provided some resilience. But he still thinks Apple is too expensive. His view of Apple’s ability to deliver agents is strong: given its distribution, “it will definitely be easy.” He still will not chase the stock.

16. The 2027 portfolio is centered on applications, with about 20% reserved for hardware

  • Asked to look 3–6 months ahead, Talk Jun does not guarantee a short-term market call. Instead, he lays out his 2027 portfolio framework: a target allocation of about 20% to semiconductors and hardware, with greater weight on AI agents, software and healthcare applications. He remains bullish on AI, but is now more focused on how applications translate into results.
  • Microsoft is his largest current position, followed by Amazon, with Intel falling back after its decline. He built and added to Microsoft at the start of the year and has not touched the position since, nor added materially in this selloff. For now, he does not plan to adjust it. That is different from saying he actively switched into Microsoft because the market fell.
  • In healthcare, he avoids drug development that he does not understand and considers highly volatile. He is watching whether AI can improve company efficiency, logistics and supply chains, change operating models and empower hospitals. But recent defensive flows have already pushed many healthcare names to new highs or elevated levels, so he wants to keep watching; he does not name a specific stock to buy now.
  • He is keeping his traditional holdings at their existing allocations, and also includes stablecoins in his thinking about AI applications. He stresses that he has not seen AI adoption slow; application stocks have recently fallen relatively little, while semiconductors have taken the brunt. “Everyone is killing the cycle, but that doesn’t mean the market has lost confidence in AI adoption.” That is why he continues to hold to the AI theme.

17. You do not need to bet on when the war will end to form a view on oil

  • Asked when the war between the US and Iran might end, Talk Jun says he does not know and has not followed it closely for a long time: it is hard to predict, and even getting the timing right would not necessarily mean predicting the market impact. For holdings he owns with some exposure to traditional aviation, he pays more attention to the operating pressure from high oil prices than to betting on when the war ends.
  • His long-term view is that “oil prices can’t stay high forever.” The US is a net exporter but still has to consider the burden on consumers; China, a major importer, does not want expensive oil, and Gulf producers also do not want high prices to suppress demand for a prolonged period. High oil prices squeeze demand. They would prefer a “sweet spot” that balances sales and profits; he is not trying to predict the path back down.