Original author: Herman Jin, US stock KOL

All forms are illusions. If you see forms as non-forms, then you see the Tathagata.

As an AI die-hard bull, I know that when everyone is celebrating, any slight expression of pessimism will be mocked and ridiculed.

However, I really want to write this article, mainly as a review and reflection: not only about what I have been doing for the past two and a half years with X and trading; but also as a commentary on this crazy market.

Fortunately, as a leisure hobby, writing X has helped me meet many new friends.

Thanks to mutual appreciation and inspiration, over the past two years, I’ve shifted from clearing out BTC in early April 2024 to AVGO and NVDA, then before the election to AAOI, PLTR, TSLA, INTC, and after the trade war to INTC and GOOG. At the end of 2025, I added more INTC and AAOI, then escaped before the Spring Festival this year, going all-in on AMD and NOK (recently also traded ORCL and LITE).

Although I experienced lottery-like zeroing out with SUP, and some gains and losses with MRVL and MDB, I have always believed that the rising demand for tokens driven by AI will make semiconductors a “dog’s time,” and I’ve gained decent returns. During this period, many friends also earned at least tenfold profits.

However, at this moment, we must soberly realize that profits are not because we are stock pickers with keen insight, but because the market is bullish! I reviewed my own last week of March this year, when I bought full positions before the weekend. Making money wasn’t because I bought AMD correctly at that time, but because I correctly called the market’s capital structure.

Therefore, at this point, it’s even more necessary to calm down and reflect on how the market might develop next.

Drunk people don’t know they’re drunk; those immersed in beautiful dreams don’t realize it’s just a fleeting illusion. This process can last long, and we can continue to go crazy, but we must be sober at a higher level of consciousness, aware of the risk of being awakened at any moment.

One of the four beasts saying, COME AND SEE – Revelation 6:1

Unlike many on Wall Street, I am “firmly believing” in AI.

For the first time in 30 years, AI models replace serial computation with parallel computing, directly turning semiconductor advances into productivity. (Previously, technological progress was just a platform for software and software users to create value.)

The result is an exponential increase in token demand, while semiconductor supply can only grow linearly.

Plus, the entire semiconductor industry’s old timers, during the past few golden years, did not expand capacity out of “belief,” and now that demand has truly exploded, the whole industry chain is out of stock.

This has created this round of bull market.

However, the lifeline of a bull market is market price in:

The model’s revenue acceleration, which in the near future can justify hyperscalers’ Capex.

It’s important to know that any expectation that shakes this lifeline will cause panic in the market.

1. Ironclad: One of the horsemen is that micro and company-level fundamentals are truly impeccable ==================================

From analyzing model architecture and demand from the architecture perspective, looking at each semiconductor track and each company, they are nearly perfect. The more detailed the analysis, the deeper the understanding, the more confident—this is a typical low PE bubble.

Rule of Thumb: After examining dozens or hundreds of companies, each with solid demand and profitability, you must question whether there is a problem with this market.

A low PE bubble is still a bubble, just different from 2000. High PE is an expectation bubble, where the market keeps setting higher earnings expectations. Once core expectations collapse, the entire outlook adjusts immediately, like dominoes falling one after another. It’s like last year’s IGV, where the market suddenly stopped valuing EV/EBITDA.

But a low PE bubble doesn’t have this problem; earnings growth gives you enough confidence to keep adding positions after slight dips. After all, who can resist storage stocks with single-digit PE, rising prices and profits annually, with 80% gross margin? During this process, small dips mean buy, big dips mean buy more.

As long as it doesn’t hit the “vital point,” the market will remain very resilient.

Persistent positive feedback, like a beautiful dream, makes people unwilling to wake up, and they don’t believe it’s just a dream. (Those who have bought P2P know what I mean.)

Asymmetrical feedback: The obvious experience over the past year is that the market’s response to macro negative risks is significantly weaker than to positive news. At this point, it’s not that there are no “barbarians,” but that the fortress of corporate profitability is impregnable. Whether it’s the Iran crisis, oil prices, inflation, interest rates, or the Fed, they can all be kept at bay.

Conversely, once the vital point is broken, all the barbarians will rush back, and the market will quickly fill the gaps from previous declines.

2. Deregulation and ample macro liquidity ==============

Since the 2008 financial crisis, the entire financial system has undergone a long period of regulation and deleveraging. To compensate for systemic liquidity loss, the Fed launched a continuous QE. Banks’ average leverage was initially x30–x40, later reduced to x15–x20. Last year, the Fed lowered the eSLR requirement, effectively enabling banks to expand their balance sheets by 4–5 trillion USD.

In fact, after April, the flood of interbank liquidity also intensified the frenzy in risk assets, with leverage entering the market at low PE levels.

Liquidity is everything; flood it, and it conquers all. If during the pandemic, 120 billion per month was a slow trickle, now it’s a floodgate opening. Compared to past QE that was balanced and then turned to QT, this direct laissez-faire approach by banks has created massive liquidity.

But liquidity is like dopamine, not a happiness molecule, but an expectation molecule.

After the laissez-faire, with inflation sticky, the Fed has little room for short-term rate cuts. Even if a crisis occurs, it’s hard to do QE again; the Fed’s only tool would be rate cuts, which is a paper tiger. On the other hand, the leverage in semiconductors, especially in storage, is already too high.

3. Heroes rising: He went forth conquering, and to conquer ================================================

Perhaps because TSMC led the way in not believing in AI, the semiconductor industry as a whole did not respond with large-scale capacity expansion to AI demand over the past two years.

The consequence is that when the entire industry faces a silicon flood demand, it’s caught off guard. Historically, the semiconductor supply chain was mainly designed for consumer electronics, with pricing based on demand. Now, not only are orders suddenly full, but continuous architecture iteration is required.

NVDA has evolved from GPU fabless to rack solutions, and now to token factories. Each step requires different technologies and suppliers. All suppliers have never experienced such mass production design. It’s like forcing a tractor onto a high-speed train, running at 200 miles per hour, with every part humming. Every segment of semiconductors is out of stock, under capacity, and unable to meet demand.

A related problem is: the seller’s market was once controlled by TSMC as the “gatekeeper,” managing supply, margins, and prices. The industry had some meat to eat, demand expected prices and capacity, and developed healthily.

Today, suddenly, five hundred million “Reynolds” can’t be controlled, and every “little thug” can raise prices and collect protection fees. The real result isn’t that semiconductors are all excellent, but that it’s out of control: out of control can affect the cost calculations per GW, impacting the lifeline expectations. It will cast a shadow over Capex and revenue models.

4. The wire dance =======

AI model revenue expectations are overly optimistic.

Looking back, the market firmly believes that the revenue from Anthropic, OpenAI, and Gemini will grow rapidly, making Capex justifiable. This lifeline’s belief must not waver. Even slight doubts can cause severe volatility in the secondary market.

So, why did GPT 5.0’s model underperform, and why didn’t the market crash due to the scaling law not holding?

First, at that time, semiconductors weren’t as widespread as today, nor was leverage as high; second, hyperscalers had abundant cash flow redundancy, and as long as they expressed confidence, funding Capex was easy to get through tough times.

After all, if they have money and continue Capex, it shows that NVDA’s earnings certainty over two years is high, and no one dares to short NVDA.

Later, GPT was caught up by Gemini, leading Oracle to hold a large backlog of orders and be shorted. But the big Larry dared to tell Wall Street “F you,” and continued to raise funds through equity. Soon, Opus emerged, signaling that the AGI era had arrived.

However, this year, Capex has already reached 770 billion, and next year it will be 1 trillion. Wall Street doesn’t care about hyperscalers’ total revenue (just like Oracle back then: OAI and Anthropic are funded with their investments). They must see continuous growth in Anthropic and OAI, which is the core to keep the entire chain spinning.

Meanwhile, big tech firms have always played the role of guardians of AI. But their free cash flow has turned negative. “Parents are old and exhausted,” and the road ahead depends on themselves. Moreover, under market expectations, their unwavering focus on this path leaves no room for error in the “main line.”

(You can recall issues like NVDA’s liquid cooling failure or switch CPO yield problems—these minor issues are nothing compared to major problems, and given time, they will be solved. But the market currently has no tolerance for errors.)

The vital point is Anthropic: the big one, the big one that surpasses others. The main pillar, the weak foundation.

Supply delays and even technical bottlenecks are like superficial ailments; any fluctuation at the model level is a major threat.

Now, the unified problem of the three leading models: insufficient computing power, leading to “dementia.”

In my extensive use of models for programming and trading system deployment, interacting with multiple AWS services, I found that Opus 4.8’s actual capability has already fallen far behind Chinese models like Kimi. Although GPT is still barely usable, it’s gradually losing intelligence. The market’s software-centric thinking about AI, with high development costs and very low usage costs, is problematic.

Trying out good models can ensure system quality in the long run. But AI is a factory; model output requires cost. As more people eat at the restaurant, the kitchen can’t keep up, and quality naturally declines.

The second inherent market thinking: token demand is non-volatile and unaffected by quality. In fact, although I use more tokens because the model is getting dumber, I doubt the strange divergence where token consumption increases as quality decreases can last.

Many companies now have KPIs for token usage, and many habitual subscribers accelerate usage. But if computing bottlenecks can’t be solved, I worry that Anthropic’s growth curve will slow. And solving computing issues isn’t quick or easy, nor can a new model fix it.

But when the market realizes this, investors will ask:

1. Why run faster than your distillation? Why spend so much on training models?

2. Are models becoming commoditized?

When the “dream” of models supporting the entire AI era is questioned, the entire revenue logic’s sustainability will be doubted. All semiconductor stocks valued with PE today will also be questioned—are you cyclical?

Some say, since dementia is caused by insufficient computing power, we should increase semiconductor investments to solve the problem.

Exactly! But what about the money?

If we exclude the funds from OAI and Anthropic that are reinvested into cloud providers, how much net cash flow do they have to continue supporting this investment? This calculation is very straightforward for Wall Street. Of course, it’s also possible they’ll follow Oracle’s example, raising funds through equity to keep investing and weather this storm.

By then, the market’s reaction could far surpass last year’s. If Amazon and Meta face soaring CDS spreads and negative cash flow, can they gamble like Larry did? It’s uncertain.

Additionally, some say that increased revenue from OAI can offset investor concerns. I believe that in such high expectations, being “replaced by the worst” due to “bad luck” is very unsettling (like a battlefield where Guan Yu was slain—who’s left? Zhang Fei? Useless).

Meanwhile, we must realize that this low PE bubble is very resilient. Until it hits the vital point and damages the lifeline expectations, it’s almost invincible. But once the fatal flaw is struck, it will collapse suddenly.

Since I started working, I’ve seen several times the illusion of paper turning into gold, tides rising and falling. No wave of mass madness has ever ended well;

nor has anyone in the dream ever truly understood “how to lose” in this race. The most recent familiar example is the DeFi summer in crypto.

Uniswap quickly swept through billions of USD in a short time, making us think that replacing parts of traditional finance was just a matter of time, with no real concern about “how to lose.” But if a drunkard wants to understand how to lose, he won’t lose.

No matter how Anthropic’s models perform or what problems they face, it will not shake the belief that AI is the third industrial revolution for humanity. The AI process will continue unaffected by crises, crushing the carcasses of revelers, until it replaces hundreds of millions of mental workers.

Finally, a comment on the low PE bubble: don’t even think about shorting, even missing out on gains, the pain index is very high.

Looking back at the funds that shorted subprime in 2007, how many survived until the final celebration?

In an environment where missing the bull market becomes a joke, I will need a lot of Aura’s strength and focus, to watch the market cautiously while going long.

Please forgive me for reducing my time on Twitter, debating and refining my skills. A parting word:

You can drink alcohol, but don’t get drunk;

You can tell stories, but don’t believe them blindly;

You can jump on the party table and dance, but keep your eyes fixed on the DJ. If he stops the music and runs, you must follow.

Original link

Click to learn about Rhythm BlockBeats’ job openings

Welcome to join the official Rhythm BlockBeats community:

Telegram subscription group: https://t.me/theblockbeats

Telegram discussion group: https://t.me/BlockBeats_App

Twitter official account: https://twitter.com/BlockBeatsAsia