August Chatter: AI and Robots Are Changing the World—Are You Paying for It? — 马克汤|AI Finance Buddy
Pioneers
马克汤
Summary
- Nvidia has brought in BlackRock, Brookfield, Goldman and KKR to form a lending syndicate and package data centers into REIT-like assets for financing; the market is also debating whether Nvidia is providing guarantees, but Raymond says that claim is far from certain. 马克汤认为老黄的“卡有残值”叙事在单点违约时可以自圆其说,但真正的违约可能发生在全行业需求被证伪之时,届时“芯片残值本身也会下降得非常猛烈”,是系统性风险而非单点风险。Raymond 补充:供过于求只要 3-5 个百分点,就可能足以让市场价格腰斩;存储缺口 5-10 个百分点,则可能把价格推高两三倍。
- Michael Burry’s fund was tiny, with roughly $150M in AUM before it was shut down; his Substack, Cassandra Unchained, has more than 300,000 paid subscribers paying over $300 a year and nearly $400, implying roughly $100M in annual revenue. 他的核心控诉是循环借贷会崩,以及科技公司折旧造假(芯片应像 iPhone 一样快速 write off)。两位主播的反驳很具体:8 年前的 A100 现在仍能租出 5 年长期租约,价格坚挺、流动性好,因为推理需求分层、客户多元化,本地部署、蛋白质研究、材料科学等场景都可能不需要顶级算力——“他是坐在办公室里做预测的”,应该去硅谷和数据中心转一转。
- Early warning indicators for a bubble include the marginal moves in CDS premiums for Nvidia, CoreWeave and Oracle, along with A100 prices and trading volumes—“I’m certain GB300 is hard currency; I’ll buy as many as I can,” while A100 is a more marginal asset and potentially a better leading signal. 马克汤则强调看“链外”的真实用户需求:产业链内部互相刷单能把双方 ARR 都做上去,“我们赚的钱没有多,甚至还要额外交税”;自由现金流、存货和应收账款都值得看,但公开财务数据可能滞后且不透明。
- China’s open-source-model “kill line” has become a life-or-death threshold for AI application companies: a model must either outperform DeepSeek V4 or cost less than V4, otherwise “most models have no reason to exist”—and switching software takes only 10 minutes, with memory and skills carrying over. 马克汤指出开源模型也开始分梯队:Kimi K3 是 2.8T 参数模型,部署成本较高,需要约 64 张卡才能达到较好的部署效果;DeepSeek 可能只需 4 张或 6 张,具体数字两人并不确定。K3 定价对齐 GPT-5.6 Terra 档,甚至比最便宜的 Luna 还贵;GLM-5.3 Flash 这类小模型才守在斩杀线附近。
- Distillation is the episode’s biggest fault line: 张一鸣 defines it as negative—students who copy the teacher’s homework can never catch up, while organizations fill up with “hackers” instead of people who actually build data and pre-training; Raymond flatly disagrees: “If a frontier-model company is willing to distill others and does it well, I see that as a positive for the company and would be willing to buy it.” 他用拼多多类比:在二选一的环境下砍一刀“蒸馏”了微信社交关系,Temu 则被马克汤描述为抽取供应链精华做白牌,先模仿后超越;马克汤认为中美都在互相蒸馏,中国模型还会公开技术报告,DeepSeek 的 KV Cache、Kimi 的 KDA Attention 等架构优化也被海外模型参考。
- The token-quota era has arrived: companies in China and abroad are limiting employees’ use of closed-source models—“closed-source models have become unaffordable, while open-source models are usable”—and more quota now requires a productivity case; the age of “token maxing” is over, and “real results have to be delivered.” 马克汤判断国产推理卡普及正在加快,但从他看到的报道来看,尚未到完全供应不上的阶段;Raymond 用科威特类比提出疑问:开采成本最低但 Q 不够大,所以 Anthropic 的 P 并没有被逼降一百倍。
- On robots, 马克汤 sees the endgame as UBI leading to UHI and agrees that current use cases, capabilities, BOM costs and end demand still do not work—robots are staging sports meets, performing martial arts, cutting fruit and throwing out their backs while running, while factory deployments are mostly ASIC-like single-task machines rather than GPU-like general-purpose systems; half or more of China’s 1,000-plus robotics startups may die. Raymond 的爆帖直指要害:连洗碗机都不用的中国人,怎么会接受家务机器人进家门?两人都看好行业长期发展,但不等于看好单一企业;马克汤认为宇树估值高或短期目标难达成并不奇怪。
- Macro-wise, a chart cited by 马克汤 is worth remembering: long-end yields could fall under either AI outcome—successful AI drives productivity and deflation, while a burst bubble sends money into Treasuries—possibly by the end of 2027, once the AI outcome becomes clearer. 眼下贝森特的 Treasury Twist(发短债、买长债)“第一天有反应,一个星期之后全部回来”,长端利率处于 2007 年以来最高;30 年期收益率从 5.3% 降到 4.7%-4.8% 时,股价可能不同,而利率走高会直接抬升 Raymond 所说的自由现金流已转负的大厂的融资成本、压低 ROIC。
Deep dive
1. August Was a Sedative, Not a Stimulant
- The opening diagnosis: after a sharp selloff across the hard-tech chain in July, “overall market sentiment recovered in August, but people were still scared, hesitant, worried and anxious.” The market bounced a little; sentiment did not fully recover. Nvidia delivered a good quarter and guidance, but “it rose after earnings and then immediately fell back”—not a durable rally.
- 马克汤 saw something off at the top: CEOs of high-flying companies were “obviously very anxious,” constantly manufacturing new narratives, raising guidance and insisting that “everything’s good”—as more investors concluded this was a rebound, not a reversal.
- His core concern is that Nvidia’s circular lending is “trying to pull forward real AI demand through credit”—“whipping the horse to keep it running.” The market is no longer watching 50% or 80% growth; it is watching the second derivative. “Once the second derivative starts flattening, market sentiment comes down immediately.”
2. Syndicated Financing: Turning Data Centers into REITs
- Raymond separates Nvidia’s two actions. First, it brought in major Wall Street institutions including BlackRock, Brookfield, Goldman and KKR to form a lending syndicate and turn data centers “into assetized commodities like houses”—the equivalent of packaging Shanghai apartments into REITs and using their cash flows to raise financing. Second, the market is debating whether Nvidia is providing guarantees. Raymond says that claim is not particularly certain; in the end, it seems to come down to one sentence from Jensen Huang.
- Huang’s narrative has two legs: demand is still surging, as shown by OpenRouter token usage, and “the cards have residual value; they are highly liquid assets.” An A100 from 8 years ago can still secure a 5-year lease today. 马克汤 repeats the argument but remains cautious: higher token usage “doesn’t necessarily mean higher economic value at the end,” because the tokens could be free or cheaper.
3. The Weakness in the Residual-Value Story: Defaults Are Never Isolated
- 马克汤’s sharpest argument in the episode: selling cards to recover value works perfectly in an isolated default. H100 and GB300 are highly liquid, “no problem there. But the fear is that this won’t be an isolated default.” A single failure may instead mean that industry-wide demand has been disproven, at which point “the residual value of the chips themselves could fall extremely violently.” That is a systemic industry risk, not a single-name risk. It is like a stock-bond portfolio that appears diversified until a systemic crisis sends everything lower.
- Raymond adds a quantitative version of the argument. He cites a claim that a 3-5 percentage-point oversupply could be enough to cut market prices in half, while acknowledging that he does not know whether the underlying data or calculation is solid. Conversely, a 5-10 percentage-point memory shortage could double or even triple prices. Add one more marginal seller, and prices keep sliding.
- His analogy for Chinese listeners: if someone desperate to emigrate cuts the price of a home by 20%-30%, you may treat it as an isolated problem, only to discover a year later that it was part of a collective event. One developer failing is manageable if someone takes over the buildings and land; a sector-wide property collapse like the one of recent years means “the dam breaks directly, and the whole flood comes out.”
4. Michael Burry: A Small Fund and a Newsletter Business
- Raymond’s profile: Burry is one of the few living fund managers to have been portrayed in a movie, The Big Short. “Across his several-decade history, he has spent far more time bearish than bullish.” Apart from getting 2008 right in a major way, he has missed many other calls. Raymond thinks Burry probably began shorting Nvidia and Palantir at least a year earlier, though he cannot remember the precise timeline.
- Raymond says Burry’s fund should now be very small, with roughly $150M in AUM before it was shut down. Cassandra Unchained has more than 300,000 paid Substack subscribers paying over $300 a year and close to $400. At that scale, Burry could be earning roughly $100M a year from the newsletter.
- His two major accusations are that circular lending must eventually break—he has the experience of buying CDS in 2007 and 2008—and that tech companies are broadly misrepresenting depreciation. Chips are electronic products, not productive assets; once the next generation arrives, the previous one should depreciate as quickly as an iPhone on a second-hand marketplace. “Anyone saying four years, five years or six years is committing accounting fraud. This stuff should be written off.”
5. Why the A100 Undercuts the Depreciation Thesis: Layered Demand and Diverse Customers
- The rebuttal runs through a clear mechanism: token consumption is shifting from training to inference. Inference does not require the very best cards for every task, but older cards can still “handle a lot of ordinary work.” Codex and Claude’s standard, fast and ultra-fast modes can use different cards underneath to optimize performance and concurrency, creating differentiated services and differentiated pricing. Firm rental and resale prices “essentially represent its actual usage in data centers.”
- Raymond relays an investor’s observation: CoreWeave’s earlier-generation cards are still in use because its customers are diverse. Local deployments, latency-insensitive workloads, protein research and materials science may all “not need that much compute.” There are enough users across enough workloads that A100 demand could still be expanding.
- Both hosts largely agree on the methodological criticism of Burry: he is “approaching the issue from a financial perspective” and does not appear to understand AI use cases and demand shifts particularly well. “He is making these forecasts sitting in an office.” The most important way to study AI is to use it yourself or visit data centers and semiconductor production lines. Burry is now in California; he could also spend some time in Silicon Valley.
6. Two Layers of Routing: The Task Chooses the Model, the Model Chooses the Chip
- Raymond says the first layer is a model-level router. An agent routes subtasks of different difficulty to models of different capability: simple tasks go to a small, fast model, which “may produce a better result than a smarter large model because it is faster and the answer is no different.” Complex, multistep reasoning goes to a frontier model.
- The second layer is chip adaptation. 马克汤 adds that the TCO of a given model varies across chips, so companies need to use the systems themselves and run their own calculations. Electricity is “the raw material for producing tokens,” making power costs part of the cost structure.
- The same logic extends to the US-China divide. ByteDance and Volcano Engine serve Chinese agent use cases, while Amazon uses Claude to serve US customers. The top-level model may be the same, but “once you go down to the chips, the best choice changes.” The same model can have entirely different deployment plans and cost structures across regions and users.
7. If the Bubble Really Breaks, What Should You Watch?
- 马克汤’s first-principles view is that the AI supply chain—power, chips, cloud and models—may be circularly buying from, investing in and financing one another. “I lend you money to buy my stuff,” causing both sides’ ARR to rise. “In reality, we haven’t made more money, and may even have to pay additional taxes.” The question is whether real demand from outside the chain is genuinely increasing. If not, changes inside the chain “don’t matter; it is just an internal loop.”
- The limits of financial metrics are clear to both hosts. 马克汤 says to watch free cash flow, defined as operating cash flow minus capex, as well as inventory and accounts receivable. But startups such as OpenAI may not disclose those figures, hyperscalers report quarterly with another 2-month lag, and Raymond notes that many numbers are not transparent: “You can make them say whatever you want.”
- Raymond offers two leading indicators. First, CDS for Nvidia, CoreWeave and Oracle, which broadly move together; the key is whether the marginal change in premiums becomes pronounced. Second, A100 prices and trading volumes. “GB300 is definitely hard currency; Americans can’t absorb it, but Chinese buyers can,” while A100 is a more marginal asset and “the more marginal it is, the more sensitive it will be.” Raymond thinks it could be a useful leading indicator. 马克汤 replies, “That makes sense. I might take a look.”
- Another confidence thermometer is the syndicate itself. Its basic function is to spread risk—“if it dies, everyone dies together; they are grasshoppers tied to one rope.” If an institution eventually steps forward and says, “I’ll finance this company directly on my own and guarantee it myself,” that could signal confidence is rising again.
8. David Sacks: A Star Wars-Scale Arms Race That the Pentagon Should Fund
- Raymond first lays out the defense of circular lending. CNBC reportedly asked Huang, “If you don’t give them the money, will they stop buying your cards?” His answer was that even without Nvidia’s financing, they would “take money from someone else, borrow it, or even steal it, and then buy my cards.” Chinese companies cannot buy cards because of geopolitics, but Raymond believes they “need them very, very badly,” so Nvidia investing downstream does not automatically constitute circular financing.
- 马克汤 strongly agrees with Sacks’s framing: AI is a Star Wars-level arms race between China and the US. “Ordinary listeners shouldn’t think that because you use Doubao, we’re no longer at war.” If this is a national-level competition, the Pentagon should be investing in data centers. Another view is that America’s largest defense effort today is lowering interest rates so Hyperscalers do not find borrowing so painful.
- Raymond relays the All-In Podcast’s consistently negative view of Dario Amodei: the hosts believe Dario’s warnings about AI and calls for tighter regulation could slow America’s development. Raymond also says Chinese models have caught up significantly over the past 1-2 months. GPT may refuse a cybersecurity request as too dangerous, while Kimi K3 solves it. China’s support may include backing chip companies and providing cheap, abundant power.
9. The Robotics Endgame: From UBI to UHI
- Building on Elon Musk’s framing, 马克汤’s endgame is for humanoid robots to take humanity first toward Universal Basic Income and then, as productivity rises, toward Universal High Income—“a very good, almost utopian” state. He admits that “this sounds like science fiction”; he is using an end-state framework to understand the industry.
- He is unsparing about the present: use cases, capabilities, BOM costs and end demand do not yet work. “None of that is a problem,” because the industry is still early. Robots are currently staging sports meets, performing martial arts, cutting fruit, throwing out their backs while running and boxing. The machines actually entering factories are mostly robotic-arm-style single-purpose systems—“like ASICs, not a GPU”—unable to move boxes, operate modules, perform maintenance and fix bugs all at once.
- The valuation path does not bother him. From an end-state perspective, Unitree commanding a high valuation, missing short-term targets or 王兴兴 not smiling are all normal. Capital markets often lift valuations first, valuations fall when progress disappoints, and the technology, use cases or other catalysts only arrive later to make the business real.
10. China Is Doing PR for Robots; America Is Manufacturing Fear Around AI
- 马克汤 says robot sports meets and martial-arts demonstrations may not be “fake it until you make it,” but state-level marketing: one way to support a pillar industry is to make people feel positively toward it and comfortable with it. Stadium events, free tickets, social media and the Spring Festival Gala increase exposure and bring people closer to a new product category.
- The comparison is stark. 马克汤 thinks Americans are strongly hostile to AI and robots because the US government has done little to help the public accept the technology, while constantly talking about bans, safety reviews and alignment. Europe is similar. “At the level of top-down design, China is friendlier to frontier technology, while the US and Europe are less friendly,” and public sentiment reflects that.
- His philosophical close: everything ultimately has two sides; “it depends on which side you want to amplify.” Reducing the bad side through technology and business optimization is business development. Increasing familiarity through exposure and narrative is PR. They are different things.
11. The Dishwasher Take: If People Won’t Use Dishwashers, Why Would They Accept Household Robots?
- Raymond’s viral post, which drew hundreds of likes and comments, asked: “So many Chinese people don’t use dishwashers. How will they ever accept a household robot coming into their home to do housework?” The negative replies were nearly universal: dishes are not cleaned properly, pre-treatment is required, there is nowhere to install one, older family members are not used to it, young people order delivery, and some users said they would only buy one if it could wash produce. He also learned that some Chinese dishwashers do have a produce-washing function. His question was: “Don’t these same shortcomings also apply to household robots?”
- The deeper problem is that a humanoid form invites an “auntie”-level expectation of generalization—knowing what you want to eat, covering you with a blanket while you sleep—which makes adoption harder. A robotic arm doing one clear job, such as sorting packages, may feel more useful. 马克汤’s response: “It has to land in a specific use case that people can see. Once a specific scenario is cracked and there is an obvious use case, talk about generalization later.”
- Both hosts agree that China’s primary robotics market is clearly in a bubble. There are more than 1,000 startups, some buying from one another to support delivery figures, while most companies are already close to their limit after producing 1,000-2,000 units a year. Half or more may die. 马克汤 compares it with NetEase during the internet bubble: after most competitors die, companies with enough capital to survive are the ones that remain.
- 马克汤 is highly bullish on the robotics industry over the long term, much as he is bullish on China’s electric-vehicle industry. But “that does not necessarily mean I am bullish on any one company.” The ultimate question is which companies can secure enough capital to survive until the technology or use-case inflection point.
12. The Kill Line: DeepSeek V4 Defines Model Survival
- The rule is simple: “Either your capabilities are better than V4, or your price is lower than V4—the latter is basically impossible, so most models have no reason to exist.” The key difference from electric vehicles is that a bad car purchase can still be driven for 10 years. Software “can be switched in 10 minutes, with all your memory and skills coming with you.” Switching is nearly frictionless, making competition exceptionally brutal.
- From his frontline perspective—菜搭子 selects models every day and bears the model costs—the actual choice set is narrow. Either pay for the best closed-source model to maximize completion rates, accuracy and low hallucination, or choose the cheaper “good enough” option and “avoid basic mistakes.” His team currently relies mainly on DeepSeek and has recently been testing GLM-5.3 Flash.
- The evidence that open-source models are splitting into tiers is clear. Kimi K3 has 2.8T parameters, carries a high deployment cost and needs roughly 64 cards for a good deployment. DeepSeek may need 4 or 6 cards, though neither host is certain of the exact number. K3’s pricing is broadly aligned with the GPT-5.6 Terra tier and is even more expensive than GPT-5.6’s cheapest Luna tier, so it is not a high-value open-source model. The old definition of open source as “cheap, useful, abundant and all-you-can-eat” no longer applies.
13. The Token-Quota Era: Closed-Source Models Have Become Unaffordable
- The new normal, according to 马克汤, is that companies in China and abroad are limiting employee token usage. Some large companies put a self-built harness on top, route requests across models and deduct credits at different multiples. “Once the monthly quota is gone, it is gone.” The only way to get more is to argue: “My output is 5x higher; isn’t it fair to give me 3x more quota?” In the past, “it was the era of token maxing—using twice as many tokens as last month meant I was incredibly productive. Now you have to deliver tangible results.”
- The main restriction is on closed-source models such as GPT and Claude. “Closed-source models have become unaffordable; open-source models are usable.” If Chinese users have access to models with enough capacity at low prices, they may not need to request more quota simply because usage of domestic models rises. 马克汤 says most people still follow the rule: “I use whatever is best first.”
- Raymond’s Kuwait analogy and question: Kuwait has the world’s lowest extraction costs but not enough production volume, so its influence is limited. “In theory, if Americans could use open-source models, Anthropic’s P should have to fall 100x to match them. But clearly it hasn’t. Is that because Q is insufficient?” 马克汤 thinks domestic inference cards are spreading faster. On training cards, he relays media and financial reports of training clusters with 100,000 domestic cards, while stressing that he does not know the actual situation. He does not think supply has yet become completely inadequate.
14. Distillation, Part I: No One Gets to Pretend Others Do Not Distill
- 马克汤’s position is blunt: “China and the US are distilling from each other. No one should say the other side does not distill, and no one should say they themselves do not distill.” Chinese open-source models publish not only weights but also technical reports and architectural improvements. DeepSeek’s KV Cache optimizations and Kimi’s KDA Attention work may be studied by overseas models. Public interviews suggest DeepSeek has disclosed only part of its work, not everything, so it may still be better in practice and potentially cheaper.
- Raymond adds background: according to public information, the main targets of US companies’ accusations over the past few months have been Qwen, MiniMax, TipSick and Kimi, not Zhipu. Relay platforms are directly related to distillation. Raymond also says that virtually every major domestic-model company has its own relay-platform business, while emphasizing that this is what 马克汤 said, not what he himself said.
15. Distillation, Part II: 张一鸣’s Objection and Raymond’s Pinduoduo Defense
- 张一鸣’s two objections are preserved in full. First, a student who copies the teacher’s homework every day can never surpass the teacher; it can only approach indefinitely and is a shortcut. Second, organizationally, it creates a group of “hackers” who know how to distill, while the organization may never build real capabilities in data cleaning, pre-training preparation and training.
- Raymond says this turns distillation into a negative term when it is also a normal technology practice: converting a model like Opus into lite or flash, compressing a large model’s capabilities into a smaller one. DeepSeek has done it too. His position is unequivocal: “If a frontier-model company is willing to distill others and does it well, I see that as a positive from an investment perspective. I would be willing to buy that company.”
- His Pinduoduo analogy is that, in China’s e-commerce environment of choosing one platform over another, Pinduoduo’s “Bargain” feature distilled WeChat’s social relationships, giving Pinduoduo a chance to get onto the playing field. 马克汤 adds that Pinduoduo, including Temu, can also be understood as extracting the essence of merchants’ supply chains and turning it into white-label and fully managed products: “Distillation is essentially extracting the essence.” Raymond then clarifies that Pinduoduo is not distillation in the technical sense. His point is how a smaller company facing a much larger opponent imposing constraints beyond its competitive ability can survive through shortcuts, grow large enough to eventually surpass it—imitate first, then overtake.
- In the latest development, Raymond says Dario’s recently released Claude 5.1 added anti-distillation mechanisms, which may make distillation harder but not impossible. Much of it may take place through coding plans rather than official paid APIs, making it difficult to block completely. Even when GLM-5.3 and Kimi K3 approach closed-source models in capability, relay businesses remain healthy because employees are accustomed to using the best model: “It is hard to go from luxury back to austerity.”
16. The Office-Software War: Chinese Workers Are Trapped in Office Work
- The numbers come first. Raymond cites QuestMobile data showing that, as of July, WorkBuddy had roughly 6 million monthly active users, Qwen Code about 2 million, Trae about 2 million and Qoder more than 200,000. Raymond also heard from another source that WorkBuddy may have daily active users in the millions, but he is unsure whether the figure matches management’s KPI or expectations, given the scale of its advertising spend.
- Raymond’s airport observation: departing China, he saw ads for “Qwen Office, Qwen Office for Professionals,” WorkBuddy, Doubao, Alibaba Cloud and Volcano Engine. In Geneva, the ads were mainly for sports brands. “It was completely two different worlds,” leading him to say that “Chinese people are trapped in office work.” The related contrast is that students at Stanford’s commencement were hostile toward Sundar Pichai, while the US broadly shows AI anxiety and protests against data centers. China is embracing AI, with some of its highest-earning bloggers teaching people to use AI, start side businesses and increase income. Europeans are more likely to think, “This has nothing to do with me.”
- The war is only beginning, in 马克汤’s view. The path runs through three stages—assistance, replacement and digital employees—and “we are nowhere near completing even the first stage.” A future A2A agent network could emerge: “Wukong talks to Wukong, Bajie talks to Bajie.” Agent-to-agent communication across teams, suppliers and customers, along with payments between agents, will become increasingly complex.
- His concern with products like WorkBuddy is that they try to use technology to adapt to an existing world that may itself be dysfunctional. Users may not buy in, then feel that the tool is trying to replace them, leaving them in a persistently “awkward” position. Digital employees are harder still: convincing the people who pay requires group strategy, workforce planning and organizational redesign. At the beginning, several marquee companies may need to serve as case studies, and some may not pay at all—they may participate simply to experiment.
- 马克汤 also mentions Michael’s blog. Michael chose marketing from among the various digital-employee directions because workflows such as ad buying and YouTube advertising are mainly digital and do not require dealing with real people or physical environments. He suggests watching a company whose name sounds like “Hell Yeah,” though the program transcript marks it as “Hell Yeah[?],” so the name is uncertain.
17. Macro: Long-End Yields Could Fall Under Either AI Outcome
- A chart 马克汤 saw 2 days earlier provides the episode’s macro framework. If AI succeeds, penetration and productivity surge, creating some degree of deflation and pushing long-end yields lower. If the bubble bursts, Nasdaq and the AI supply chain fall, investors buy Treasuries and yields also decline. “Either way, long-end yields could inevitably fall by the end of 2027, after the outcome of AI becomes clearer.” Fiscal and inflation pressures could still keep short rates high during the process.
- The current picture is ugly. Bessent’s Treasury Twist—issuing short debt to buy long debt—“worked on the first day, and a week later it had all come back.” US long-end yields are at their highest level since 2007. Repeated official messaging has entered a “boy who cried wolf” phase, with diminishing returns: “By the fifteenth time, nobody pays attention anymore.” 马克汤 thinks rate hikes are relatively unlikely, but feels investors are pricing in a growing US Treasury crisis. Ray Dalio has been discussing the issue for more than a decade, while it remains in a state of “about to explode but not actually exploding.”
- The transmission to tech stocks is straightforward. Raymond says that with free cash flow already negative, major companies such as Google may need equity or debt financing. Higher rates raise financing costs and reduce ROI and ROIC, so “the whole story moves in a negative direction.” His scenario is that current stock prices were formed with the 30-year Treasury yield at 5.3%; if the yield falls to 4.7%-4.8%, stock prices could look different. Both hosts caution against looking only at the long end: an inversion is also bad. Bessent wants yields to move toward inversion, but the market is not cooperating.
18. Closing: Anthropic’s IPO and the “Pursuit Line”
- The closing joke carries a view. At the time of recording, both hosts expected Anthropic could go public soon. Raymond asks how Anthropic will explain itself to investors under pressure from China’s model kill line. 马克汤 corrects him: “This isn’t a kill line. It’s a pursuit line—they’re chasing you down. The one being chased is Anthropic.”
- The imagined answer is: “I’ve been holding back something big; I just haven’t released it yet,” followed by a GPT-6 launch—“because if it isn’t explosive, it won’t be released.” The episode ends in speculation and joking; it is not confirmation that GPT-6 has launched.
- Looking to September, Raymond mentions the rate decision, what officials might say if they do not hike and the signals from the midterm elections. They will also keep watching Treasuries, the yen and oil—Raymond says oil should now be back above $90—and how those macro variables affect hard tech. Fund managers and everyone else are returning from summer vacations and the mountains, so September should be livelier.