The signal is weak; the noise is deafening.
Every major technological inflection point produces the same mirage: a capital formation cycle that pretends to be demand. The latest record in AI infrastructure spending by listed companies is not a testament to genuine adoption. It is a self-referential loop where financing begets procurement, procurement begets supplier earnings, and supplier earnings beget more financing. The underlying asset—AI compute—has become a speculative commodity, detached from its final utility.
Context: The Global Liquidity Map Meets the AI Gold Rush
Over the past 12 months, a growing number of publicly traded companies—from traditional data center REITs to industrial conglomerates—have announced capital raises explicitly earmarked for AI infrastructure. The justifications vary: "securing future compute capacity," "building sovereign AI capability," or "capturing the next wave of cloud growth." But the macroeconomic backdrop reveals a common driver. Global M2 money supply has been expanding, albeit unevenly, as central banks maintain accommodative stances to manage lingering post-pandemic imbalances. Liquidity is searching for yield, and AI infrastructure offers a narrative thick enough to absorb it.
But there is a catch. The inflation-adjusted cost of capital has not fallen in tandem. Real interest rates remain elevated relative to the risk-free rate of the past decade. This means the capital being raised today carries a higher cost burden than the previous cycle's crypto mining mania. And the payoff timeline is longer. Unlike Bitcoin mining rigs, which could be repurposed or resold with reasonable liquidity, AI GPU clusters are tied to specific architectures—H100, B200—that depreciate rapidly with each new generation. The capital cycle is funding assets that will be obsolete before their loans are repaid.
Core Insight: The Capital Cycle as a False Demand Signal
The numbers paint a clear picture. Publicly traded companies raised over $40 billion in debt and equity for AI infrastructure in the first half of 2025 alone. That figure exceeds the entire venture capital flow into AI startups for the same period. The capital is not going into building better models or novel applications. It is going into warehouses of silicon.
I have seen this pattern before. In 2017, I audited tokenomics whitepapers for ICOs that raised millions for “decentralized compute networks.” Almost none of them ever generated the revenue to justify the hardware. The same logic applies today: a company that raises capital to buy GPUs is not creating inherent value unless those GPUs are actually used to solve a problem that a paying customer acknowledges. And right now, the flood of new supply is outpacing the growth in actual AI inference demand. The electricity draw of newly commissioned data centers in Northern Virginia alone would power a small country—yet the utilization rates for many of these clusters sit below 40%, according to industry reports.
This is the classic bubble formation. The act of raising capital and spending it creates a temporary demand spike for upstream suppliers (NVIDIA, AMD, Super Micro, etc.). Their stock prices rise. The rising stock prices make it easier for the same companies to raise more capital. The cycle feeds itself. But the end consumer—the enterprise that buys AI software subscriptions—has not materially changed behavior. They are still waiting for killer applications, not infrastructure.
Contrarian Angle: The Decoupling Thesis
The prevailing narrative says that AI infrastructure spending is a leading indicator of structural economic transformation. I argue the opposite. The capital cycle is decoupling from real-world adoption. It is a temporary equilibrium where suppliers capture all the gains, and the downstream companies that actually use the compute pay a premium that stifles their margins. The only winners are the ones selling the picks and shovels—and even they face an accelerating depreciation curve.
Consider the analogy to the fiber optic bubble of the early 2000s. Between 1996 and 2001, telecom companies laid millions of miles of fiber, raising astronomical sums to build “information superhighways.” The technology was real. The demand was real—eventually. But the timing was off by nearly a decade. When the bubble burst, most of that fiber went dark, and the companies that financed it went bankrupt. The survivors—companies that waited and bought distressed assets—reaped the rewards later. Today, AI infrastructure is being built on a similar premise: build it now, and they will come. But the capital commitment is far larger, and the debt structure is far more fragile.
Another hidden risk: the concentration of supply in a single company, NVIDIA. While AMD and Intel scramble to catch up, NVIDIA’s CUDA ecosystem and interconnect technology create a moat that makes its GPUs effectively mandatory for large-scale AI training. This means the capital being raised is funneled into a single point of failure. Any disruption—a geopolitical trade restriction, a design flaw, or a shift to a non-CUDA architecture—could render billions in hardware obsolete overnight. Systemic risk hides where the charts are too clean.
Institutions smell blood when retail smells profit. Right now, retail investors are piling into AI infrastructure ETFs, chasing the narrative. But the institutions behind the capital raises are hedging. They are issuing convertible bonds with low conversion prices, or selling secured notes that prioritize creditor claims over equity. They know that the underlying asset’s value is volatile. They are placing asymmetric bets: if the AI boom continues, their equity surges; if it collapses, they walk away with the hardware at a discount. The capital raise is not a bet on the future—it is a transfer of risk from insiders to the public markets.
Takeaway: Positioning for the Cycle
Volatility is the price of entry, not the exit. The current phase demands that we ignore the narrative and watch the liquidity. Specifically, monitor the yield on investment-grade corporate bonds tied to AI infrastructure. If spreads widen, it means creditors are pricing in default risk. Also track the secondary market prices for GPUs—if the premium over list price collapses, the “scarcity” fiction is over.
The most prudent move is not to short the infrastructure providers. They are too correlated with the broader tech index. Instead, look for companies that own the application layer—the ones that can convert cheap future compute into sticky user behavior. The bubble will not burst until a meaningful fraction of that new compute sits idle. When the next earnings season reveals that a major listed infrastructure company is cutting CapEx guidance, that will be the signal.
Chasing shadows in the algorithmic dark of AI CAPEX leads to ruin. Wait for the blood in the streets—and by that, I mean the write-downs on stranded assets. That is when the real opportunity emerges.
