Shiwen Dialogue 045 | All Cry AI Bubble, Money Flows to NVIDIA

Opening Dialogue

Shiwen: Yongliang, where does this article come from? Who is the author?

Yongliang: The author is personal blogger Matteo Raso. He posts AI-related comments and reviews on his blog. The article was published on Sunday, November 8th, and appeared on HN on Monday, November 9th, Beijing time. When we collected it, there were over 500▲ views and nearly a thousand comments, and the number is still increasing. The author isn’t some big shot; his GitHub profile is almost anonymous, and he has only a dozen followers. So this isn’t an institutional report, but rather an ordinary person’s reverse judgment that hit a nerve.

Shiwen: What is the core argument of the article?

Yongliang: In one sentence: The bubble theory has become a echo chamber, but the data doesn’t support it. According to the data in the article, the AI industry has promised to invest at least 3 trillion dollars, of which about 1.9 trillion have already been spent. The author acknowledges that this number is frightening, but he believes the key lies in the nature of the expenditures: data centers are one-time capital expenditures, not ongoing operating costs. After they’re built, the spending will drop sharply. On the revenue side, he cited Anthropic’s S-1 financial report: for every $1 of computing power, there’s $2.08 in revenue; revenue increased from $41 million in 2023 to 386 million in 2024, 4.585 billion in 2025, and reached $16.3 billion in the first half of 2026. These are all figures cited by the author. I didn’t verify the original documents independently, so I framed these statements with this limitation.

Shiwen: What about the debt issue? That’s the most common argument for the bubble theory.

Yongliang: He solved a problem he called “mathematics on the back of an envelope”: assuming the industry has 5 trillion dollars in debt to be repaid over 20 years, at $250 billion per year, and Alphabet had revenues of $162 billion in 2019 alone, after inflation adjustment, it would be $213 billion. So repayment isn’t impossible. He also mentioned that some of the debt can be converted into shares: according to the author, by the first half of 2026, Anthropic had $72 billion in losses owed in the form of shares. But because its valuation has risen to around 1 trillion dollars, the equity is sufficient to cover these losses. On the market level, he made an observation: publicly listed AI companies like Alphabet, Meta, and NVIDIA have P/E ratios stable at just over 20, unlike during the internet bubble when they were often hundreds of times higher.

Q1: What’s most offensive in the full text is not the data, but the diagnosis.

Shiwen: Among the nearly a thousand comments, which part resonates the most?

Yongliang: Strangely, it’s not the financial data, but the psychological diagnosis at the end. The author says that the prevalence of the bubble theory is not essentially a financial issue: it carries a desire to return the world to its former state. People fear that AI will change work and life, so they tell themselves that it’s all a bubble that will eventually burst, and then everything will return to normal. He calls this mindset “wishful thinking.” And the most widely spread observation in the article is his observation that it’s one thing to say AI is a bubble, but another to say where investors’ money comes from. His original words mean that when people have to stop talking about money, they will spend it on NVIDIA.

Shiwen: Do the comments accept this diagnosis?

Yongliang: It’s divided into two factions. Commenter npilk agrees, saying this explains why pessimistic narratives have a large audience: people want to hear that things will eventually return to normal. Another commenter, GolfPopper, argues that the diagnosis itself is projection: whether AI companies collapse or succeed, the outcome may not be good for ordinary people. So believing in the bubble theory is not a wish, but a judgment. Both interpretations make sense. I think what should be remembered more is the author’s self-restraint: he says this isn’t a major critique of the new bubble theory and even acknowledges that complex financing structures are one of the most significant arguments supporting it. A counterargument that acknowledges its strongest attacks actually increases tolerance in the comments section.

Q2: The refutations in the comments section are more hardcore than the article itself.

Shiwen: On a data level, how does the comment section handle this?

Yongliang: The most severe criticism is on the depreciation period. Commenter wmf says the author assumes 20 years to amortize 5 trillion, but the biggest cost in data centers is GPUs, not buildings. GPUs should be depreciated over 5 to 6 years, which directly quadruples the required annual income. hobofan adds: Computing hardware accounts for about half of data center costs. With a 5-year lifespan for GPUs, the depreciation period is closer to 10 years. This criticism hits the key math in the back of the envelope: changing the denominator changes the entire conclusion.

Shiwen: What about valuation?

Yongliang: Commenter blltprfmnk directly asks: Why does Anthropic have a clear positive cash flow path to value 1.2 trillion? amluto provides a “double-peak distribution” framework, which I think is worth remembering: one future where Anthropic builds a sustainable profit model would be worth 7 trillion in ten years; another where revenue growth stagnates and they are forced into default. Both futures are entirely possible, so any single-point valuation is essentially betting on which peak you fall into. vannevar’s analogy is even more toxic: the author says AI is profitable after ignoring infrastructure costs, but if you ignore car manufacturing costs, car manufacturing is also profitable. He also points out that according to S-1 data, Anthropic’s losses are expanding, not narrowing.

Shiwen: Did the author address this himself?

Yongliang: He did. Commenter jxcole suggests that DeepSeek and open-source weight models can switch at any time, and a 2.08-fold gross profit can’t sustain price competition. The author replied that this point has been discussed in the article. Another comment mentions Goldman Sachs: it is internally restricting the use of Claude Code while almost halving its IPO target within four weeks (mentioned in the comments section, related to yesterday’s topic, not expanded here). hattmall points out another signal: CoreWeave’s bond yields are as high as 16%, indicating that the bond market has a different pricing for risk, which is not the same as optimism in the stock market.

Q3: Should ordinary tech professionals make career decisions based on the bubble theory?

Shiwen: Bringing it back to the workplace. Many readers are asking: The bubble theory sounds reasonable, should I avoid AI-related positions and directions?

Yongliang: My answer is, don’t bet on “the bubble won’t burst” or “the bubble will burst”. First, look at history. Commenter swatcoder listed railways, semiconductors, plastics, and networks: the uses and profitability of these technologies were obvious back then, yet many investors still lost their money. His conclusion is that the bubble doesn’t lie in potential but in timing. Wongarsu added a brighter perspective: the decade after the internet bubble burst actually saw the birth of YouTube, Reddit, and Spotify. I have personally experienced the mobile internet and O2O cycles; this is my experience, not statistics: when the bubble breaks, companies that told stories die, but infrastructure and user habits remain. The next wave of entrepreneurs uses ready-made paths.

Shiwen: So, the AI skills you learn now aren’t wasted?

Yongliang: We need to distinguish two things. Commenter rayiner provided a good framework: the bubble bet is on “which company will profit,” not on “whether the technology is useful.” He compares large models to electric motors: electric motors changed the world, but today’s companies that invented electric motors aren’t valuable because technology has become a commodity. So, betting on “AI is useful” is a high-probability event, while betting on “a certain AI company will win” is a low-probability event. Ordinary tech people should accumulate transferable skills: using AI to solve engineering problems, evaluating AI outputs, and integrating AI into team workflows. At the company level, who wins or loses should follow the tasks, not stock codes.

Shiwen: What if the bubble really breaks?

Yongliang: In the short term, it will be difficult—recruitment cuts, project eliminations, this is cycle experience, not prediction. But note what commenter motionlessveloc pointed out: “AI is a bubble about to burst” and “AI is filling entry-level positions” are actually hard to hold true at the same time. Reality is likely mixed: there’s useful parts and investments that never return. The way to prepare for this mix isn’t to bet on a single script, but to have a place in both scripts.

Q4: Corporate Side: Budget tightening and AI spending increases coexist

Shiwen: From a management perspective, many companies are tightening budgets while increasing AI spending this year—isn’t that contradictory?

Yongliang: It isn’t contradictory, because the nature of the two funds is different. The author says data centers are one-time capital expenditures; commenter AngryData counters: hardware won’t last forever, maintenance isn’t free, electricity costs won’t significantly drop, and if models become more efficient, no such expensive hardware is needed. If models continue to improve, new generations of hardware will directly bury old investments. Tuna-Fish gives a more specific explanation: modern hardware’s useful life is estimated to be only 2 to 3 years. This is the real situation in corporate IT decisions: money for tools must be spent, but each expense should be expected to depreciate or become obsolete in two years. When I led a team, this was experience, not statistics: buy capabilities, not assets; use current money to solve current problems, don’t lock future three years into “AI is the future” budgets.

Shiwen: Where are the risks hidden?

Yongliang: Hidden inInvisible places. Commenter dofm pointed out that a large amount of debt has been moved out of the balance sheet through special purpose entities, with pensions and sovereign wealth funds possibly taking over in the end. This echoes the author’s own admission: complex financing structures are one of the most significant arguments supporting the bubble theory. After risks are packaged layer by layer, no one really knows who will bear them. The same lesson applies to decision-makers in ordinary companies: when evaluating AI suppliers, don’t just look at model demonstrations, but also their financial structure and service continuity. If your core business is built on a supplier that may default, this risk must be written into contracts and contingency plans in advance.

Q5: Three honest advice for ordinary people

Shiwen: At the end of the article, does the author give his own advice?

Yongliang: Don’t bet your life on the AI revolution. If you lose, don’t blame it. I think this is the most honest sentence in the whole article, because after writing a whole reverse article, he finally brought himself back: no one can see the future, but he said there are quite realistic possibilities. My three advice for ordinary people are based on this sentence. First, replace “I think” with “What would I do”: set up falsifiable conditions for yourself, such as what signals appear, and admit that you were wrong in your judgment about how quickly AI will penetrate the industry. Write it down in advance and execute it when the time comes, not relying on stubbornness on the spot. Second, bet on skills, not assets: you can’t decide whether a company’s valuation rises or falls, but you can decide whether you can use AI to do work faster and better. Third, leave a way out for each dependent: tools, suppliers, industry narratives are all the same. Don’t bet all your work assets on a single format or a single company. This is the conclusion of our last conversation, and it still holds true this time.

Shiwen: If you had to end with one sentence?

Yongliang: Whether there is a bubble may only be known afterwards, but deviations between words and actions are visible now: many people who shout the most about bubbles are also seriously using AI and buying AI stocks. For ordinary people, a more stable approach is not to judge whether it will burst, but to ensure that no matter which scenario plays out, you are still at the table. I don’t predict stock prices, just say this sentence.

---

Signed: Yongliang