Author: Fu Peng

A couple of days ago, at the monthly VIP private meeting of New Fire, I exchanged views with some old friends, and everyone's main concern remains the next phase of the AI track. In the financial market, the same issue may yield completely opposite answers at different stages. In the past two years, as long as big companies were willing to increase capital expenditures (CapEx) to invest in infrastructure, the market offered high valuations; however, by the second quarter of this year, the market began to rigorously question "free cash flow" (FCF).

From the financial reports of large companies, it is very clear that when the free cash flow of giants tends towards zero, the era of relying on cash on hand for costless capital expenditure is over. Moving forward, if they want to continue to expand CapEx, they must borrow money or seek financing externally, and the current interest rate cost is 6%-7%. The capital market is no longer blindly accepting the "left hand to right hand" capital cycle; instead, it is starting to demand: after investing so much in infrastructure, is there actually business flow and real profit downstream to recoup the costs?

This is similar to traditional large-scale infrastructure: in 2003, when roads and bridges were built, upstream orders for cement, steel bars, and construction machinery skyrocketed, and performance surged; but once the roads are completed, there must be “vehicles” to run on them. The current AI track is in this critical window period: the climax of the infrastructure layer (GPU/HBM/computing power) has passed, but the truly terminal applications (such as medical AI, autonomous driving, enterprise-level vertical models, etc.) have not yet been fully realized.

One of the biggest misconceptions that many traders have is "treating possibilities as certainties" and "blindly leveraging in extreme certainty." Storage (DRAM/HBM) ultimately is also a commodity, and it cannot escape the cyclical nature of commodities. When the industry fundamentals are extremely certain and volatility has dropped to a very low level, the excessive accumulation of out-of-market financial leverage often becomes the invisible killer that triggers crashes and flash collapses. When deleveraging occurs, even if the company's fundamentals are sound, prices can still fluctuate violently.

In the context of tightening liquidity and credit differentiation, the market is bound to follow a "contracting circle" logic: funds will first clear out speculative assets that are on the periphery, have high elasticity, and lack cash flow support, continuing to concentrate on the most core and highest-certainty targets. This is also why assets such as Bitcoin, which are purely "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 ("old Deng stock" characteristics), but the downstream application layer has not yet seen milestone breakthroughs similar to ChatGPT. The next 10-18 months will be a transitional period for the industry cycle. The second half of AI is certainly not about repeating the story of "burning cash on infrastructure" again, but rather seeing whether AI, as an efficiency tool, can produce stable "payment, profit, and diffusion" after entering real industries. The universal large model will ultimately become a universal tool, and what can truly build a moat will inevitably be applications and commercial closed loops deeply cultivated in vertical fields.