The market does not care about your feelings. It cares about structure. And structure is telling us something critical about the AI investment narrative.
Over the past six months, a subtle but powerful shift has occurred in institutional discourse. JPMorgan strategist Gabriela Santos publicly recommended diversifying AI investments across regions and sectors. This is not a casual suggestion. It is a structural signal. When a top-tier global financial institution—one that manages trillions in assets—begins to advocate for dispersion over concentration, it means the easy alpha is gone. The phase of buying the picks and shovels (NVIDIA, Microsoft, the hyperscalers) and watching them double is maturing. The next phase requires a different toolkit.
Context: The Historical Narrative Cycle
Every technological revolution follows a predictable narrative arc. The 1990s internet boom: first, infrastructure (Cisco, Sun Microsystems), then applications (Amazon, eBay). The 2010s mobile revolution: first, hardware (Apple, Samsung), then software (Uber, Instagram). The AI revolution is no different. From 2022 to 2024, capital flowed aggressively into the foundational layer—GPU manufacturers, cloud providers, and foundational model builders. The narrative was simple: "AI is the new electricity, and everyone needs the wires." That narrative worked. NVIDIA's market cap exploded from $300 billion to over $2 trillion. The hyperscalers reported AI revenue at annualized rates of tens of billions.
But now, the narrative is fracturing. The law of large numbers is catching up. The growth rates of the infrastructure giants are decelerating, not because AI is slowing, but because the base is too large. The market is searching for the next leg. This is where Santos's recommendation lands. It is not a radical idea—diversification is basic portfolio theory. But its timing is everything.
Core: The Mechanical Logic of Diversification in AI
Let me break down the mechanics. Santos's argument rests on three pillars: regional dispersion, sectoral dispersion, and the recognition that AI value is diffusing from a monopolistic infrastructure layer to a multi-polar application layer. This is not an opinion. It is a reflection of on-chain data—or in this case, market data.
First, the regional dispersion logic. The U.S. dominates foundational model research and cutting-edge chip design. China leads in application scale and engineering efficiency. Europe has carved out a niche in regulatory technology and ethical AI frameworks. These regions have different drivers—U.S. AI companies are R&D-intensive with long profit cycles; Chinese AI firms are revenue-driven with rapid deployment; European AI players focus on compliance and vertical integration. The correlation between these regional AI baskets is lower than the correlation between any two U.S. tech giants. This is the mathematical basis for diversification. Yield is the lie; liquidity is the truth. The liquidity of AI investment flows is now being redirected across borders, not just within the U.S. tech ecosystem.
Second, the sectoral dispersion logic. The adoption curve of AI across industries is not uniform. Financial services and technology are at 50%+ penetration. Manufacturing, construction, and healthcare are still in single digits. This gradient creates a massive value dislocation. The market is pricing all AI-related stocks with a similar growth premium, but the actual revenue realization varies dramatically. A medical imaging AI startup may have a 10-year adoption runway, while a customer service chatbot AI may hit saturation in 3 years. Diversification across sectors is not just risk management—it is a bet on the timing of the S-curve. Arbitrage exposes the cracks in consensus. The consensus is that AI is a monolithic growth story. The reality is that each vertical has its own capital cycle, regulatory speed, and user adoption curve. The arbitrage lies in identifying which verticals are undervalued relative to their actual adoption path.
Third, the infrastructure-to-application transition. The cost of inference has dropped by 80-90% since 2023. This is a seismic shift. When inference is cheap, the barrier to building AI applications collapses. The market is still pricing AI companies as if the cost structure is static. It is not. The marginal cost of AI reasoning is approaching zero. This means the value capture shifts from the chip seller (one-time hardware sale) to the application provider (recurring software revenue). The diversification recommendation is a tacit acknowledgment that the infrastructure phase has peaked in terms of marginal returns. The next wave of growth will come from thousands of niche applications, not from a handful of infrastructure giants.
Contrarian Angle: The Hidden Trap of Pseudo-Diversification
Here is where the narrative gets twisted. Santos's recommendation is rational, but it may be a defense mechanism for overvalued assets. The market is quietly signaling that the top-tier AI stocks are running out of room to re-rate. The P/E multiples of NVIDIA and Microsoft are already pricing in years of future growth. Diversification, in this context, is a way to spread the risk of a single point of failure. But there is a catch: the diversification may be a mirage.
Most AI companies—whether in the U.S., China, or Europe—share a common risk factor: they are all dependent on the same underlying technology stack. Transformers, GPUs, and cloud infrastructure. If a paradigm shift occurs (e.g., a non-Transformer architecture emerges that requires different hardware), the entire basket of AI stocks could suffer. Regional and sectoral diversification does not hedge against a fundamental paradigm shift. It only hedges against idiosyncratic risks. Auditing the code, not the charisma. The charisma of the diversification narrative is appealing. But the code—the underlying correlation structure—reveals that the diversification is largely superficial. The real risk is systemic, not idiosyncratic.
Moreover, the diversification argument assumes that the market is efficient enough to price the second-tier AI stocks correctly. It is not. Most investors cannot distinguish between a company that has genuine AI revenue and one that is simply rebranding itself as an AI company. The "AI washing" phenomenon is rampant. In 2024, over 40% of companies that mentioned AI in their earnings calls had no material AI revenue. Diversifying into such stocks does not reduce risk; it increases exposure to noise. The true alpha comes from deep fundamental analysis—auditing the revenue streams, the unit economics, and the defensibility of the moat. Not from spreading capital across a broad index.
Finally, there is the geopolitical dimension. The diversification recommendation implicitly assumes that cross-border capital flows remain frictionless. They will not. The U.S. has imposed export controls on advanced AI chips. China is building its own semiconductor ecosystem. Europe is introducing stringent AI regulations. The result is a fragmented landscape where diversification across regions introduces regulatory risk, not just market risk. The correlation between U.S. and Chinese AI stocks may be low, but that is because they are both being hit by different shocks. A Taiwanese chipmaker's supply chain disruption affects both markets. A U.S. export ban on GPU sales to China affects Chinese AI companies but also affects U.S. GPU manufacturers. The diversification is not clean.
Takeaway: What the Data Reveals About the Next Narrative
The signal from JPMorgan is clear: the AI narrative is entering a new chapter. The infrastructure phase is over. The application phase has begun. But the path forward is not a simple buy-and-hold diversification strategy. It is a game of precision. The market is transitioning from a phase where you could buy the entire sector and win to a phase where you must pick the winners within each vertical. This is a higher-difficulty environment.
Based on my experience analyzing tokenomics during the ICO boom—where 80% of whitepapers had no viable utility—I see a parallel. The AI boom is now entering its "application proof" stage. Just as many DeFi projects failed to deliver on their promises, many AI companies will fail to convert hype into revenue. The diversification recommendation is a risk management tool, but it is also a signal that the easy money is gone. Narrative follows logic, never precedes it. The logic now points to a selective, research-intensive approach. The diversification that Santos recommends may be a starting point, but it is not the destination. The destination is identifying the specific AI applications that have cracked the unit economics and are scaling profitably.
Watch for the following signals: the earnings reports of AI application companies, the growth in paid user accounts for AI tools, and the capital expenditure guidance of hyperscalers. If the hyperscalers start lowering their capex forecasts, it means they see diminishing returns from infrastructure investment. That would be the final confirmation that the next phase—application-led growth—is upon us. Until then, treat the diversification narrative as a proxy for the market's uncertainty. And in uncertain markets, structure prevails. Auditing the code, not the charisma.
The floor prices of AI stocks may bleed, but the structural opportunity in undervalued application layers remains. The next narrative is not about diversification. It is about discrimination. About separating the AI companies that have real revenue from those that only have real buzz. The yield is in the details.
Pivot not panic. The data reveals the path.