AI Investment Review

Summary

• The core trade conclusion: this is the time to reassess your positions, but absolutely not to short. “You could very well end up shorting 2007”—most shorts never made it to 08 or 09, when the problems hit; most investors who stayed in the trade for more than 2 years are up more than 10x. Once an account is up 10x, the right move is to step back and reassess, not bet on a collapse.
• The market has one lifeline: OpenAI and Anthropic’s revenue must grow fast enough to justify hyperscaler capex. That thesis is an impenetrable shield: SIVB, the tariff war, the Iran crisis, Blackwell liquid cooling, and the latest concern over overheating switches are all beside the point. “Any branch problem that is not on the main line is simply not a problem”; negative news has nowhere near the market impact of positive news.
• A low-P/E bubble is still a bubble—even when every shortage and every order is real. The reality of the supply constraints makes investors willing to lever up and buy every dip, like “a P2P platform taking high-yield deposits and giving you positive feedback every day.” Former crypto groups in Korea have bought SK Hynix with 5x leverage, with some pledging their homes; the 2x leveraged SK Hynix ETF (7709) is already one of Hong Kong’s most popular products. Huaqiangbei memory prices rose 5% in 3 days, after another 5% gain the previous week.
• Liquidity is a dopamine-driven expectation factor, not a reward factor. ESLR has opened up roughly $4.5T of assets eligible for bank purchases in a single shot—the pandemic QE was a $120B-a-month river, while this time the dam has been opened directly—and SOFR has traded below OIS for several days. But capital requirements “cannot be cut any further,” so the rush of liquidity will deliver diminishing pleasure.
• The real pressure point is Anthropic. Revenue expectations went from what still appeared to be $9B last December, to $30B by year-end in February-March, to a reported $55B by mid-May (month 5)—the fastest growth in the market. But model quality has deteriorated under call overload, to the point that it is “worse than Kimi at coding,” while the errors themselves drive higher token consumption. If falling quality slows demand growth, “a lower acceleration is enough to give the market a serious shock”—and the market has priced in none of it.
• Model commoditization is the path to a broken dream. Larry, Oracle’s CEO, used the term in early March. Chinese model companies can distill your models, run inference at one-tenth the cost, and deliver roughly comparable results. “The model itself is a dream”; once that dream breaks, “everything breaks,” taking semiconductors in both the US and China down with it.
• Next year’s tightrope is thinner. Hyperscaler capex is roughly $750B-$770B this year and free cash flow has already turned negative; next year’s $1T will need to be funded with debt, assuming Wall Street continues to believe the story of explosive revenue growth at the two companies. The key actionable indicator is to watch OpenAI and Anthropic revenue growth around the middle of each month: “Wall Street cannot tell whether a model is good, and they do not use it to code.”

Deep dive

1. The Golden Shield: A Low-P/E Bubble Is Still a Bubble

• The speaker’s definition of the market’s current state is straightforward: everything rests on OpenAI and Anthropic growing fast enough to support the hyperscaler capex layer in the middle of the stack. “If you touch that main line, you are touching the market’s lifeline.” SIVB, the tariff war, the Iran crisis, CTA selling, Blackwell liquid cooling in 24, possible overheating in switches, and potential delays in optical equipment all miss the artery. “None of these things matter.”
• The most dangerous feature is that everything at the micro level is real: “The more granular you get, the more real it looks.” The shortages are real, the orders are real, and the lack of capacity still brings price increases. That makes investors willing to lever up and buy every dip: “To use an inapt analogy, it is a P2P platform taking high-yield deposits and giving you positive feedback every day.” Former crypto groups in Korea have bought SK Hynix with 5x leverage, some by pledging their homes; 7709, a 2x leveraged SK Hynix ETF, is already one of Hong Kong’s most popular products.
• The on-the-ground evidence is everywhere. Huaqiangbei memory—the original audio sounded like “Net,” likely NAND—rose 5% over the past 3 days after gaining another 5% the previous week. Overseas fabs, including TSMC and GlobalFoundries, are pushing out low-end process customers; after those customers return to China, the price of a single exposure has risen from $25 to $50, a 2x move. “You have seen how P2P schemes blow up, right? When something goes wrong, there is no time to get out, because your P is an illusion. All appearances are empty; once the illusion breaks, it is simply gone.”

2. Liquidity Is Dopamine: An Expectation Factor, Not a Reward Factor

• The form of this liquidity injection is different. Pandemic QE was “printing $120B every month,” a river that slowly filled the pond; this time, ESLR alongside deregulation has opened the dam and released roughly $4.5T of assets eligible for bank purchases in one shot. SOFR has been below OIS for several days, bank reserves are high, and liquidity is extremely ample.
• But the speaker’s analogy is dopamine: “It is not a reward factor; it is an expectation factor. After one cigarette feels great, you need more to get the same high.” Keeping the party going requires a constant stream of new expectations, while “capital requirements cannot be cut any further.” The market’s liquidity high will therefore fade over time.

3. Semiconductors: From Organized-Crime Order to a Field of Rivals

• The old order looked like this: semiconductors were, “to put it bluntly, an organized-crime industry,” operating under a system of near-total planning. TSMC was like the “Five-Hundred-Million-Dollar Detective” Lei Luo, collecting protection money and then allocating the slices; Apple, as the top boss, got first access to 3nm, with Qualcomm and MediaTek taking their turn 6 months later. The volume requirements of mass-produced AI parallel computing forced an old car to run flat out: “It is not just the engine that breaks down—the wheels and brake pads break too.” Bottlenecks appeared everywhere.
• That created an opportunity for “every organized-crime group to come around and collect a fee”: memory prices are rising, MLCCs are rising, ceramic components are rising, CPUs are rising, TSMC is raising prices, one item heard in the original as “维粉” is up 10%, GFS is preparing to raise prices, and Lumentum has not raised them yet. The call made at the start of the year has proved right: “Every dog has its time.” Memory pricing follows the logic of iron ore: spot-market volume is tiny, but it drives contract prices higher round after round.
• But price increases are now eating into the lifeline. If the cost of a 1GW deployment becomes uncontrollable, the expected relationship between capex and revenue breaks down. “Once you break that expectation, the market is actually going to collapse.” The same applies to a popular narrative: “All consumption is carbon-based consumption.” Silicon-based end demand is a false premise; AI will ultimately displace hundreds of millions of jobs. “If that does not happen, the entire Lego-block bubble has to collapse.”

4. The Core Thesis Has Almost No Margin for Error: Dancing on a Tightrope

• The bullish logic has not changed: token demand is rising exponentially while semiconductor capacity can expand only linearly, so “you are always going to be short of supply.” That was the speaker’s call 2 years ago, and it has played out. When entering the trade at the end of March, “it did not matter which stock you bought.” The speaker held a large AMD position, but Qualcomm or ARM would have worked just as well; within optics, AAOI rose 10x and Lumentum rose 10x, and “they still had to rise 10x” because demand for optical components had no visible end.
• The market’s narrow tolerance was demonstrated late last year, when OpenAI was overtaken by Gemini. “It was still left hand replacing right hand; it was not Kimi overtaking it.” The market immediately stopped crediting OpenAI’s backlog with Oracle, Oracle CDS widened by 500-plus points, and the stock fell without recovering through today. Now that OpenAI is back on top, the followers have recovered, but Oracle still has not.
• Next year’s tightrope is thinner still. Capex has reached $750B-$770B this year and free cash flow is already negative. The foundation for “taking it to $1T next year” is that the two companies’ revenue continues to surge and the market continues to accept the story. “The margin for error in your story will become even smaller.”

5. The Pressure Point Is Anthropic: The Market Has Not Priced in the Quality Decline

• Why Anthropic? “If you want to make an adjustment, you have to take out the commander-in-chief.” The strongest rally in the market this year has been Anthropic’s: from what still appeared to be $9B in revenue last December, to an expected $30B by year-end in February-March, to a reported $55B by mid-May (month 5).
• The firsthand evidence is stark. In February and March, using GPT to audit code written by the speaker was “almost flawless.” Now “every module makes mistakes, and several of them are very serious.” “Without exaggerating, Anthropic’s model is now worse than Kimi at coding”; OpenAI is certainly much better. The reason token demand is still rising is almost perverse: “Because I make mistakes, I keep calling the model back and forth, using more tokens while getting worse results.” The unpriced chain is straightforward: demand increases, supply tightens, quality falls, and demand growth may slow. “It does not need to flatten; merely a lower acceleration is enough to give the market a serious shock.”
• The closing analogy is a baby: AI is a baby, hyperscalers are the ones feeding it, and “the wet nurse has run out of milk—whose money are you using?” A hedge-fund friend argued that every problem could be solved with more semiconductor investment. The speaker’s response: “Yes, but where is the money? Who pays? Will Wall Street accept that bill?” Invest another $1.5T and “you necessarily get a crash.”
• The deeper risk is model commoditization. Larry, Oracle’s CEO, used the term in an interview in early March. Chinese model companies can distill your models, run inference at one-tenth the cost, and “deliver roughly the same results.” Why, then, should anyone pay Anthropic or OpenAI for that much training compute? “The model itself is a dream… If that dream is shattered, the impact is not limited to the US. Chinese semiconductors are exposed too. Once it breaks, everything breaks.”

6. The Trade and the Endgame: Intelligence Converges on Concentrated Power

• The key actionable indicator is whether OpenAI and Anthropic’s revenue growth slows, observed around the middle of each month—but do not short. “The closer you get to the underlying technology and the company level, the more bullish the picture looks.” To short, you need a product with extreme leverage; “we will discuss that when the time comes.”
• The long-term conviction remains, with one hedge: “I believe it will all eventually rise back, and no matter how far it falls, it will rise back.” The model gap is real, while capacity investment is only getting underway now; over the past several years, “the semiconductor people simply did not believe in AI,” leaving the industry with years of deferred investment to work through.
• The endgame is deferred to the next bull market. AI concentrates power; it does not distribute it. “The Cultural Revolution’s Little Red Book told you to rebel not to give you power, but to concentrate power further.” The speaker used AI to build a trading program operating “at the same level as Jump and Jane Street.” Going forward, large companies may be willing to pay 5 million to use AI to eliminate staff, and “the final intelligence will ultimately fall into the hands of those with even more concentrated power.”

Verification Notes

• The original audio was heard as “维粉”; its specific material cannot be confirmed from the subtitles alone.