All that has form is illusion. If you see all forms as formless, then you see the Tathagata.

As a die-hard AI bull, I know that amidst everyone's revelry, any hint of pessimism is bound to be met with scorn and ridicule.

However, I really want to write this piece, which is actually a review and reflection: not just for my own writing on X and trading over the past two and a half years, but also for this frenzied market.

Fortunately, as a leisure hobby, writing on X has allowed me to meet many new friends.

Thanks to mutual appreciation and inspiration, over the past 2 years, I went from liquidating BTC in early April '24 to switching to AVGO and NVDA, buying AAOI, PLTR, TSLA, and INTC before the election, then INTC and GOOG after the tariff war, adding to INTC and AAOI again at the end of '25, to fleeing this Spring Festival and then going all-in on AMD and NOK (recently also trading ORCL and LITE).

Although I had a lottery ticket SUP go to zero, and MRVL and MDB had their ups and downs, I have always firmly believed that AI's demand for tokens would ensure "every dog has its time" for semiconductors, and I have achieved decent returns. During this period, many of my friends have also earned returns of no less than 10x.

However, we must soberly realize at this moment that the returns are not due to our brilliant stock picks or unique vision, but because the market is bullish! I reviewed my own trades in the last week of March this year, where I went all-in 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 at that time.

Therefore, at this point in time, it is even more necessary to calm down, concentrate, and seriously reflect on what the market will do next.

A drunk person does not know they are drunk; someone lost in a beautiful dream does not realize it's a fleeting dream. This process can last a long time, and we can continue to be crazy, but we must be soberly aware in a higher state of consciousness that there is a risk of being jolted awake at any moment.

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

Unlike many on Wall Street, I "firmly believe" in AI.

For the first time in 30 years, AI models use parallel computing to replace serial computing, directly turning semiconductor advancement into productivity. (Previously, technological progress only served as a platform, allowing the software that uses the technology, and the people using the software, to create value.)

The outcome of AI is an exponential increase in token demand vs. semiconductor supply that can only increase linearly.

Coupled with the fact that the entire semiconductor industry's "old guards" did not expand capacity in the past few golden years due to lack of "faith," and now that demand is truly exploding, the entire supply chain is facing shortages.

This has created this great bull market.

However, the lifeline of this great bull market is that the market has priced in:

The acceleration of model revenue, justifying the Capex of Hyperscalers in the not-too-distant future.

It's important to know that anything that shakes the expectations of this lifeline will throw the market into panic.

1. The Impregnable Fortress: One of the horsemen is that the micro-level, company-level picture is truly flawless.

Looking from model architecture, from architecture to demand, examining every semiconductor sector and every company at the micro level, each one is nearly perfect. The more detailed you look, the more you understand deeply, the more confidence you have — this is a typical low-PE bubble.

Rule of Thumb: When you look at dozens or hundreds of companies, each with solid, real demand and profits, you must question whether there's a problem with this market.

A low-PE bubble is still a bubble, it's just different from 2000. A high-PE bubble is an expectations bubble; the market keeps setting higher earnings expectations. Once the core company's expectations shatter, overall expectations adjust immediately, falling like dominoes. This is like last year's IGV, when the market suddenly stopped giving EV/EBITDA valuations.

But a low-PE bubble doesn't have this problem; earnings growth gives you enough confidence to keep rushing in to buy the dip. After all, who can resist single-digit PE memory stocks, with prices and profits rising every year, and 80% gross margins? In this process, buy small dips heavily, buy big dips heavily.

As long as the "lifeline" is not cut, the market will be very resilient.

Sustained positive feedback, like a beautiful dream, makes people unwilling to wake up and not believe it's a dream. (Those who bought P2P products know what I'm talking about.)

Asymmetric feedback: A clear experience in the past year is that the market's reaction to macro-negative risks has been significantly weaker than to positive news. At this time, 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, the Fed — all can be kept at bay.

Conversely, once the lifeline is breached, all the barbarians will come charging back, making up for all the declines that didn't happen before in one go.

2. Deregulation and Abundant Macro Liquidity

The entire financial system underwent a long period of regulation and deleveraging after the 2008 financial tsunami. To compensate for systemic liquidity loss, the Fed correspondingly launched a sustained round of QE. Bank leverage averaged x30~x40 back then, later reduced to x15~20. Last year, the Fed lowered the eSLR requirement, effectively allowing banks to theoretically expand their balance sheets by 4-5 trillion dollars.

In fact, the liquidity flooding the interbank market after April also fueled the frenzy in risk assets, with leverage entering the market under low PE.

Liquidity is king of all, rampant liquidity conquers everything. If the $120B/month during the pandemic was a steady trickle, now it's opening the floodgates. Corresponding to the previously balanced QE without turning to QT, directly allowing bank deregulation creates massive liquidity.

But liquidity is also like dopamine — not a happiness molecule, but an anticipation molecule.

After letting it run wild, now inflation is sticky, and the Fed actually has no room for short-term rate cuts. Even if a crisis occurs, it will be difficult for the Fed to engage in QE; a Fed with only rate cut tools facing a possible crisis is a paper tiger. The other side of short-term liquidity is that currently, leverage in semiconductors, especially memory, is already too high.

3. Rise of the Challengers: He went forth conquering, and to conquer

Perhaps because TSMC took the lead in not believing in AI, the semiconductor industry as a whole did not expand capacity on a large scale to meet AI demand over the past 2 years.

The consequence is that when the entire industry faces the flood of silicon-based demand, it's suddenly dumbfounded. In the past, the semiconductor supply chain was primarily designed and priced for consumer electronics. Now, not only are orders suddenly maxed out, but architecture also needs to iterate continuously.

NVDA has progressed from being a GPU fabless company to a rack solution provider, and is now moving towards being a token factory. The technology and suppliers needed for each step are different. None of the suppliers have gone through such mass production design. It's like forcibly pulling a tractor onto a high-speed rail and making it run at 200 miles per hour — every part will be buzzing. Every sub-sector of semiconductors is facing shortages, lacking capacity, and insufficient capacity ramp-up.

An accompanying problem is: originally, the seller's market only had TSMC as the "gatekeeper" controlling supply, margins, and prices. The original industry structure had meat for everyone, demand-side had expectations for prices and capacity, leading to healthy development.

Today, suddenly the "Five Million Dollar Detective "Reynolds"" can't control it anymore, every "small-time hoodlum" can come to raise prices and collect protection money. The real result is not that all semiconductors are amazing, but loss of control: loss of control will affect the cost calculation for each GW of investment needed, casting a shadow over the lifeline expectation of Capex meeting model revenue. It will cloud the outlook for the convergence of capex and model revenue.

4. Dancing on a Tightrope

AI model revenue expectations are too full.

To recap, the market firmly believes that the pricing of Anthropic + OpenAI + Gemini revenue will increase rapidly, making Capex reasonable. The faith in this lifeline cannot be shaken in the slightest. Even a slight tremor can cause violent fluctuations in the secondary market.

So, in the past, when GPT 5.0 also had models that fell short of expectations, and the market questioned scaling law, why didn't semiconductors fall sharply?

First, semiconductors weren't as broadly flourishing as they are today, nor did they have such high leverage; second, hyperscalers had more cash flow redundancy at the time. As long as they expressed belief and were willing to support with capex money, it was easy to overcome the difficulty.

After all, once they had money and continued capex, NVDA's earnings certainty for the next 2 years was high, and no one could dare to short NVDA against the wind.

Later, there was also an incident where GPT was caught up by Gemini, leading to Orcl being shorted because it held a large backlog of orders. However, Big President's good buddy old Larry dared to say F you to Wall Street, vowing to continue betting even if it meant equity financing. Soon, Opus appeared out of nowhere, making people realize the AGI era had arrived.

However, this year's capex has already reached 770B, next year it will be 1Tril. Wall Street won't look at the Hyperscalers' total revenue (the logic is the same as with ORCL back then: OAI and Anthropic have both received their investment funds). They must see that Anthropic and OAI's growth rates continue; this is the core that sustains the entire chain's self-rotation.

At the same time, in this game, the big tech companies have always played the role of AI's guardian. But their free cash flow has turned negative. "The parents are old, their strength is exhausted," the rest of the road must be walked on their own. Moreover, under the market's full expectations, there is not the slightest tolerance for error on this "main path."

(You can recall NV's rack liquid cooling issues, switch CPO yield problems — these are minor issues that can definitely be solved given time, yet the market had zero tolerance for error.)

The Lifeline Lies in Anthropic: Great excess, meaning the great is in excess. The ridgepole is bent, weak at both ends.

Supply delays, and even technical bottlenecks, are superficial ailments; any disturbance on the model side is the real internal crisis.

The common problem for the three leading models now: insufficient compute, hence, loss of intelligence.

In my extensive use of models for programming, deploying trading programs, and interacting with multiple AWS services, I've found that Opus4.8's actual capabilities have fallen far behind the Chinese model kimi. Although GPT is still barely usable, it is also gradually losing intelligence. The market is thinking about AI with inherent software logic: high system development cost, low usage cost.

If it's good during the trial, it should guarantee stable system quality in the future. But AI is a factory; model output has a cost. If too many people eat at the restaurant, the kitchen can't keep up, and the quality naturally declines.

The second inherent market thinking: Token demand is index-like, without fluctuations, and doesn't change due to quality. In fact, although I've used many more tokens because models have become dumber, I question whether the bizarre scissor gap of declining token quality and increasing token consumption can persist.

Many companies now indeed have KPI assessments for token usage, and many inertia subscription users will also accelerate usage. But if the compute bottleneck cannot be resolved, I am very worried that Anthropic's growth curve will slow. And the compute problem is not solved overnight, nor can a new model solve it.

But when the market begins to realize this problem, investors will ask:

1. If you can't outrun those distilling you, why spend so much money training models?

2. Are models becoming commoditized?

When the model "dream" carrying the entire AI era is also questioned, the entire revenue logic's future ability to self-sustain and run through will be questioned. All semiconductors valued with PE today will similarly be questioned — are you cyclical??

Some will say, if loss of intelligence is due to insufficient compute, then we should increase investment in semiconductors to solve the problem.

Correct! But where's the money?

If you exclude the funding that OAI and Anthropic received from cloud providers, which is fed back to them, how much net cash flow do they have left to continue supporting this investment? This account is very easy for Wall Street to calculate. Of course, it's possible they will, like Orcl back then, continue investing even if it means equity financing, to tough it out.

By then, the market's reaction might far exceed last year's. It's uncertain whether AMZN or Meta, if facing soaring CDS and negative cash flow, can go all-in like old Larry did.

Others might say, can OAI's revenue increase alleviate investor concerns? I believe under such full expectations, being replaced because you're "less bad" is hard to reassure people. (When two armies clash and Guan Yu is beheaded on the battlefield, you say there's still Zhang Fei? It's useless.)

At the same time, we must also realize. This Low-PE bubble is also very resilient; almost invincible as long as it's not directly hit in the lifeline, wounding the core expectation. However, once the fatal spot is hit, it will collapse with a bang.

Since I started working, I've seen several instances of extravagant revelry, tides rising and falling. Not once did a frenzy involving everyone end well;

And not once did those dreaming within it figure out how this sector "could lose." The most recent one familiar to crypto friends is DeFi summer.

Uniswap rolled up tens of billions of USD in a very short time. Back then, it seemed replacing part of traditional finance was just a matter of time and degree; people simply couldn't think, and couldn't figure out, "how could we lose?" However, if a drunk person could figure out how to lose, they wouldn't lose.

Regardless of how Anthropic's model performs, what problems it faces, it will not shake the fact that AI is humanity's third industrial revolution. The AI process will not be affected by any crisis; it will continue moving forward, even over the corpses of revelers, until it replaces hundreds of millions of knowledge workers.

Final comment on the Low-PE bubble: Not to mention shorting it, even missing out on the rally, the pain index is very high.

Look back at those funds that shorted subprime in 2007 — how many survived until the final狂欢?

In an environment where missing a bull market makes you a laughingstock, I will need a lot of Aura power and focus in the future to both watch the downside and stay long.

Please forgive me for reducing the consumption of my cultivation on Twitter for debate. A parting word:

You can drink, but don't get drunk;

You can tell stories, but don't truly believe them;

You can jump on the party table to dance, but your eyes must be fixed on the DJ; if he runs away playing music, you must follow and run too.