
Fu Peng: AI infrastructure stocks are exhibiting characteristics of older assets; during the deleveraging phase, prices will still fluctuate violently even with fundamentals.
Author: Fu Peng

A couple of days ago, at the Xinhuo Monthly VIP Private Meeting, I exchanged views with some old friends. What everyone cared about most was still the next phase of the AI sector. In the financial market, the same thing placed in different stages may have completely opposite answers. In the past two years, as long as big companies were willing to increase capital expenditure (CapEx) to invest in infrastructure, the market would give high valuations; but by the second quarter of this year, the market began to strictly question "Free Cash Flow" (FCF).
It is clear from the earnings reports of big companies that when the free cash flow of giants tends toward zero, the era of making capital expenditures with cash on hand at no cost is over. Next, if they want to continue expanding CapEx, they must borrow or raise funds externally, and the current interest cost is 6%-7%. The capital market no longer blindly recognizes the "left hand to right hand" capital cycle, but begins to demand: after investing so much in infrastructure, is there actually business traffic and real profit downstream to recoup the costs?
This is like the traditional large-scale infrastructure of the past: in 2003, building roads and bridges, upstream orders for cement, steel, and construction machinery exploded, and performance surged; but after the roads were built, there must be "cars" running on them. The current AI sector is at this critical window period: the first phase climax of the infrastructure layer (GPU/HBM/Computing Power) is over, but true terminal applications (such as Medical AI, Autonomous Driving, Enterprise Vertical Models, etc.) have not yet been fully realized.
The biggest mistake many people make in trading is "treating possibility as certainty," and "blindly adding leverage within extreme certainty." Storage (DRAM/HBM) is ultimately a commodity and cannot escape the commodity cycle law. When industry fundamentals are extremely certain and volatility drops to extremely low levels, the excessive accumulation of off-exchange financial leverage is often the invisible killer that triggers plummeting and flash crashes. When deleveraging occurs, even if company fundamentals are fine, prices will fluctuate violently.
In the macro environment of tightening liquidity and credit differentiation, the market will inevitably follow a "shrinking circle" logic: capital will first clear out peripheral, high-elasticity, speculative assets lacking cash flow support, and continue to concentrate on the most core, highest certainty targets. This is also why assets like Bitcoin, which are pure "denominator-side" (pure liquidity assets), often digest valuation pressure before the traditional stock market.
The AI industry is currently in a "transition period from midstream to downstream"—the infrastructure layer represented by Nvidia has entered a mature stage (characteristics of "mature stocks"), but the downstream application layer has not yet seen a ChatGPT-like milestone breakthrough. The next 10-18 months are a transition period for the industry cycle. The second half of AI is absolutely not about repeating the story of "burning money on infrastructure" again, but rather seeing whether AI, as an efficiency tool entering real industries, can generate stable "payments, profits, and diffusion." General large models will eventually become universal tools; what can truly build a moat must be applications and commercial closed loops deeply cultivated in vertical fields.













