Original video in Chinese.
Key Takeaway
- AI market stages: from the software era (model size/algorithm leadership) to the hardware era (GPU supply/data centers), and now entering the energy era (power supply), testing full-chain system capability; Microsoft’s idle GPUs confirm the power bottleneck.
- Energy demand opportunities: AI training/inference electricity use is expected to account for 2% to 3% of global power consumption, requiring nuclear power, grid upgrades, and copper for transmission and cooling; copper miners and grid stocks have performed strongly, and the main upswing will come once the energy issue is solved, so this is not a bubble.
- All in AI strategy: I make money in multiple ways (content/products/investments), and the capital market validates my judgment; beyond being a self-media creator, I encourage full participation in the AI wave.
Microsoft bought a pile of GPUs, but in the end couldn’t use them. That’s the most awkward and also the most typical thing I’ve heard recently.
Their CEO, Nadella, said on a show that the power supply and physical space in data centers are nearing their limits, which means a pile of Nvidia chips were bought but can’t be used and can only sit in storage.
I’m bringing this up because it happens to confirm a point I made earlier inside Knowledge Planet: the market’s understanding of AI has already entered the third stage. And this stage is more difficult than the previous ones, because it tests “full-chain system capability.”
The first stage of the AI market, I call the “software era.”
Back then, everyone was focused on the model itself. In other words, whoever could train a model with more parameters and whoever had a more efficient algorithm would take the lead.
Look, early OpenAI was so damn impressive precisely because it was big enough.
The reason DeepSeek could shake the industry so much at the beginning of the year was that they had a unique algorithm that allowed them to complete training with limited computing power.
Before, everyone thought that as long as you pushed model parameters from the hundred-million level to the trillion level, breakthroughs would come. But the market quickly realized:
The prerequisite for brute force creating miracles is that you must have brute force, meaning lots and lots of computing power.
Training a large model may require tens of thousands of cards and several months of operation. And to get that many cards, you have to place orders with Nvidia two years in advance, otherwise you won’t even be able to spend the money.
This is the second stage, which I call the “hardware era.”
The bottleneck in this stage is chip supply and data center scale. Nvidia became the center of global attention. Whether it was GPUs or the stock, if you could get it, you made money.
But today, the market has a new understanding and a new consensus. This time, the target is energy.
Actually, the logic is very simple:
If you want more powerful AI, you need bigger models.
If you want bigger models, you need more GPUs to provide computing power.
If you want more GPUs to run, you need more electricity — and isn’t that energy?
AI training and inference are both massive power consumers. AI’s power consumption is expected to account for 2% to 3% of global electricity by 2030. Without electricity, no matter how many cards you have, they’re useless and can only be decorative.
This is the third stage. I call it the “energy era.”
In this stage, whoever can guarantee sufficient and inexpensive electricity supply can keep developing and can keep the lead.
There are huge investment opportunities here.
For example, where is all this electricity going to come from? Shouldn’t it rely on nuclear power?
U.S. power grids are so outdated they can’t handle it at all. Doesn’t that mean they need large-scale upgrades?
And when building new infrastructure, not to mention anything else, copper as the basic material for power transmission — shouldn’t demand for it rise?
Besides power transmission, cooling also needs copper. So shouldn’t copper demand surge too?
If you don’t believe it, you can go look at U.S. stocks and commodities. Copper miner stocks and grid stocks have performed very impressively. In fact, I believe those who had this awareness have already made a killing over the past half year.
So I think that only after the energy issue is fully resolved will AI’s main upward wave arrive. Right now, it’s still early. There’s no need to rush, and there’s no need to shout about a bubble. I previously even wrote a detailed article specifically about whether AI is a bubble; if you’re interested, you can go read it.
As a tech blogger who talks about AI, the reason I pay so much attention to AI investing is that since I’ve decided to go all in on AI, I have to take the whole package:
Making AI content can make money, making AI products can make money, and investing in AI-related targets can also make money.
Content, products, and investments are my current sources of income.
And making real money from the market, especially from the capital markets, is the best way to validate my understanding of AI. Just making content and being a self-media creator is too elementary; that only fools beginners.
OK, that’s all for this episode. If you want to understand AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you next time!