Just chatting casually over the weekend.
First, let's talk about AI. Since GPT-4 in 2023, there have been two major 'negative factors' surrounding AI: first, hardware is not shipping (tariffs, production scheduling), and second, AI intelligence has reached a training bottleneck. This year, in the hot summer of AI in 2025, H20 (hardware) and the report on GPT-5 by The Information over the weekend may not meet expectations, and coupled with the fact that the sector is indeed at a high level, different voices are emerging.
What I want to say is that in the past two years, AI has been disturbed by countless such voices (this time it is really very lightweight), but each time has proven to be interference and noise, which does not affect NV and its partners from continuing to reach new highs; this is the power of the trend. Recently, I watched the latest interview with the two co-founders of Anthropic, Dario and Mann (by the way, Claude code is really an AI application product I am very optimistic about). When the big shots mentioned the scaling law, I felt it is more verifiable than many physical laws. Mann's original words were that there are very few such laws in the universe that still hold when the scale expands by several hundred thousand times (look at quantum mechanics and general relativity
[dog head]
). This year, Claude may reach 4 billion dollars in revenue, and the computing power-application flywheel has essentially been activated (Dario specifically mentioned that it was less than 100 million dollars in 2023, and if it continues like this in 2027-2028, it may reach hundreds of billions of dollars to catch up with Meta.
[can't bear to look]
This means there is a positive correlation between computing power investment and income returns.
There was a lot of interesting information over the weekend, for instance, O3 is considered 'hard to use' because it is too smart; communicating in human language makes it seem very clumsy, similar to Einstein doing elementary grammar exercises.
The North American chain is definitely in short supply, no need to say more. In fact, last week there was another signal worth noting: non-mainstream liquid cooling and power supply also received news in the North American chain. This is precisely a signal of sharply increasing demand. Why do I say this? If you are an emerging supplier looking to enter a major client, what opportunities are there? A. Policy encouragement (domestic alternatives). B. Striving to improve product performance. C. Prices lower than old suppliers. Actually, ABC are not the most direct factors; in my years of industry research, these factors are not as direct as a sudden surge in demand from major clients leading to insufficient upstream capacity, which is the golden opportunity for emerging suppliers to enter the supply chain. The kouzhao period is the most representative; the long-discussed domestic alternative is not as good as the opportunity brought by the surge in medical equipment capacity leading to insufficient overseas component supply. In the A-share market, small suppliers with a market value of less than 10 billion suddenly entering major clients can see a frightening increase.
In fact, I have always said that AI hardware is not just optical modules or PCBs; it is just that we have relative advantages in these two areas. Liquid cooling and server power supplies are also markets with a scale of tens of billions of dollars. So why have I always looked forward to opportunities in domestic AI hardware? Because there are many selectable targets. Moreover, even though I have already made a lot of money in this wave of AI, I still have confidence that I can at least double my account based on the future potential of domestic AI.
Over the weekend, I also reflected on anti-involution. I think it may not be as easy as some people think, because pushing it further will generate more new problems. For example, if upstream prices rise, who will ultimately pay? Should the midstream also limit production? If upstream is standard products, it is still manageable, but who in the midstream complete machinery manufacturers will reduce production (unless you organize a car group purchase)?
[dog head]
). It is even less likely for downstream to pay; doesn't this contradict another national policy (stimulating consumption)? Just like the consumption mentioned last year, I also wrote a long article thinking it would be very difficult to achieve. If anti-involution this year only focuses on single points, it may also be very difficult.
I think we should neither underestimate nor overestimate the government's will and capability. For example, in the real estate market, and in the stock market, I think these are already relatively 'autonomous and controllable' areas. However, after so many years and so many rounds of brutal chip declines, we have finally reached a hard-earned level of control since last September.
That's all~