The essence of this round of AI is actually very simple:
The true internet giants that can train large models exchange money for NVDA graphics cards, exchange graphics cards for computing power, and then slice that computing power into tokens for external sale, essentially running a 'computing power retail business.'
Following this underlying logic, several key conclusions can be drawn.
⸻
1. Computing power is advancing, tokens will become cheaper, but demand will only grow larger.
The performance of graphics cards is continuously improving, and the marginal cost of a single token must be decreasing.
But the problem lies in the fact that pricing power is not in the hands of users, but in the hands of model vendors.
As long as:
• Programmers need to write code
• Finance professionals need to do analysis
• Internet professionals need to improve efficiency
• Companies need to automate
Therefore, the consumption of tokens cannot stop, and will only become more intense.
Let alone the fact that there is still a lot:
• API shell projects
• AI intermediaries
• Micro-innovation SaaS
• Various 'AI +' tools
The brokerage software, office SaaS, video platforms, and design tools you see,
Almost all are embedding AI as a standard feature.
The token consumption at the software level will eventually feed back to the hardware level's computing power demand.
⸻
Two, the shortage of computing power in the U.S. is a 'structural problem', not a short-term sentiment
Many people underestimate the difficulty of building data centers in the U.S.
Land approval, electricity installation, rack setup, cooling systems, union costs…
In the U.S., even opening a milk tea shop can take months for renovation.
So, the shortage of computing power is common, not an exception.
This is also why:
• Mining companies suddenly collectively 'transformed into computing power centers'
• With land, electricity, and infrastructure
• Can be used with slight modifications
It is not surprising that mining company stock prices have soared in recent months.
⸻
Three, real estate is 'artificially created scarcity', while AI is driven by real demand
The scarcity of houses is essentially created by policies and financial tools.
When the buying power is insufficient, so-called 'scarcity' will backfire.
City investment builds plates, controls chips, raises premiums,
Logically, it is no different from launching imitation coins —
Houses that are not lived in are treated as financial chips.
Whereas AI is completely the opposite:
• There is real demand
• Efficiency improvement
• There is a willingness to pay continuously
This is also why the long-term fate of the two is destined to be different.
⸻
Four, tokens are becoming cheaper, but 'good models' are becoming more valuable
Technology is advancing, and token costs are being continuously compressed.
If you don't lower the price, others will definitely lower it.
But at the same time:
• The user base is expanding
• Usage frequency is rising
• Usage scenarios are penetrating
This is the typical early expansion stage of the internet.
What is truly scarce is not tokens,
but rather —
Whose model is stronger, more useful, and more irreplaceable.
In a word:
You must use my model to get the job done well.
⸻
Five, data centers are heavy asset hard work, and money will spill over to the entire industrial chain
Building a data center is not 'sexy',
Oracle's stock price trend has already explained everything.
Therefore, large model vendors will definitely:
• Outsourcing
• Subcontract
• Pass the cost of expansion to the market
And this money will flow down the industrial chain:
• Builders
• Graphics card manufacturers
• Power equipment
• Transport logistics
The flow of money is very real and very traditional.
⸻
Six, the final form: AI ≈ Internet + telecommunications operators
In summary, AI is very similar to the early internet:
• token unit price continues to decline
• Usage scale continues to expand
• Model vendors are gradually becoming 'infrastructure-oriented'
The final pattern is likely to be:
• Large model companies → Like telecommunications operators, sell computing power and token premiums
• Ordinary users → Pay for AI-driven applications
Those who can truly enjoy the long-term dividends are still the top platform-level companies.
The conclusion is simple and not very sexy:
👉 Companies like AAPL and GOOG are the long-term beneficiaries.
AI is not a wave of speculation,
It is more like a reconstruction at the infrastructure level.
