It seems I have been preparing for this topic for a long time, and friends have constantly been asking about it. In fact, I prepared to write several times over the past month, but I always felt something was missing.
I used to think the missing piece was data, research, or a structured analysis. I perhaps still believed that by researching and thinking more, I could explain things more clearly. If I failed to explain it well, I saw it not just as a lack of ability, but as a lack of effort.
However, every time I felt the logic and data were complete, the situation changed. So, was it truly complete? Or was I perhaps not even on the "right" path?
Maybe it's simply because, after a year, my way of working has undergone a massive transformation. Adhering to old standards now causes significant discomfort. So, let me try my latest workflow: forget all other tools and windows, stare only at the Google Docs window. I still can't define what "AI Native" is, but focusing, letting go, and setting no limits—that might be "Human Native."
Therefore, the following content consists only of text, no charts. My memories regarding data might be wrong, and it is almost certainly unstructured.
This AI-driven revolution we are experiencing is, to be precise, driven by a small group of people. Everyone involved might be a bit crazy, yet perhaps not entirely so: Sam Altman must be quite mad, which is why he can form a "dream team" with Larry Ellison and Masayoshi Son. By contrast, Jensen Huang seems much more rational. Demis, Ilya, and Dario Amodei are "mad" in different fields—fields they consider "scientific." With these figures at the forefront, others seem less prominent. Of course, let's not forget Brin, the key hero behind Gemini's counterattack; the willful Mark Zuckerberg; and Elon, who finally admitted he couldn't play the political game...
As for whether AI is a bubble and when it might burst, is that just a pure financial question? When there is no longer enough capital to support the dreams of these madmen, the bubble will burst, leaving a mess behind. This event seems imminent because, aside from a few fanatics, most people aren't willing to "bet their lives" to stay in the race. "Steady happiness" is likely the peak of imagination for the masses...
At least, that’s how I am.
Because of this, I am more inclined to believe that market expectations for investment amounts are significantly overestimated. My experience in starting and managing companies—though not entirely pleasant—helps me view problems from a decision-maker's perspective and understand the mindset of most boards: they want everything. Especially when decision amounts start reaching hundreds of billions or trillions of dollars, everyone wants to lead in this era, but most still hope for a "margin of safety."
However, the big players we see don't have their voting power entirely in the hands of the "madmen"; it's held by financial statements and rapidly shrinking free cash flow.
2025 is a magical, crazy year: at the beginning, the ghost story brought by DeepSeek—that "computing power isn't needed"—somehow gained many followers. Yet by the end of the year, a consensus seems to have emerged that giants will continue to expand Capex regardless of the cost.
Under the "energy shortage consensus," people are more willing to look for power investment opportunities, while few do the simple math: what happens if the Earth simply cannot fit that many GPUs and TPUs? Right, we can go to space, or to the bottom of the ocean.
Actually, the decision-making factors are not that complex: when NVIDIA's future technology roadmap is very clear, it's a matter of how much money and how many data centers will exist within a relatively certain timeframe (such as before the first AGI timeline of 2030) and how to allocate each generation of products.
These things can be roughly calculated, and every player surely has detailed data, calculation models, and results. However, we all know that such results are either extremely conservative or extremely optimistic. In a background filled with noise at every level, a rational result doesn't exist.
If we discuss only the financial part: how can a company with a total annual revenue growth rate that rarely exceeds 20%, and whose AI and cloud revenue growth struggles to cross the 40% line, continue to support a 70% growth in Capex? If this company's net free cash flow is rapidly declining and may even turn negative by 2026 according to current trends—and worse, if an increasing portion of new revenue is being offset by rapidly rising depreciation.
There isn't just one such company among the MAG7; there are several.
How will the market and shareholders react?
And what if those companies aiming to be challengers or disruptors are frantically burning every cent they can find, investing in businesses where basic resource constraints lead to rising prices, falling gross margins, and where depreciation and interest expenses push net profit into the negative? How will the market and shareholders react then?
For the latter, the market and shareholders have already voted with their feet.
The bursting of a bubble is always brought about by a liquidity crisis. The starting point of a liquidity crisis begins with accounting and changes in financing rates.
There is only one way to break these concerns: generate more revenue to prove the game can continue.
So, regarding the bubble, perhaps in 2026, we only need to look at the growth rate of AI-driven revenue—not for a single company, but as a whole. In a highly competitive environment, a company can always gain more revenue by releasing a new model or cutting prices. But models have no secrets; it's hard for anyone to lead for more than three months. The way to more revenue can only be through lowering costs to gain a price advantage or providing users with a more complete ecosystem to create more possibilities.
However, even more competitive than model manufacturers are the "users."
The way to lower costs is straightforward: improvements in hardware efficiency or progress in models. If the so-called tenfold performance boost brought by Blackwell can lead to at least a 50% drop in model service prices, and if Gemini-3-Flash can roughly match the capabilities of Gemini-2.5-Pro, then even if the pricing of Gemini-2.5-Flash is maintained, the actual price has dropped significantly. Many tasks running smoothly on Gemini-2.5-Pro can switch to Gemini-3-Flash without any pressure.
Thus, assuming model prices drop by two-thirds or more by 2026 is quite reasonable.
For the "Jevons Paradox" to hold, usage needs to triple, and token usage must increase at a rate of no less than doubling every eight months. If we want to see AI-generated revenue grow by more than 50%, the token growth rate cannot be lower than doubling every six months. Let's wait and see.
Token usage growth might be my most optimistic metric; I still believe a doubling every four months is achievable.
But what I am less optimistic about is the payment rate. In truth, a model is a luxury; it might one day become a free necessity, but it's hard for it to be a reasonably priced common commodity.