The 10-year Treasury yield sits at 4.2%. The S&P 500 forward P/E is 21.5. The equity risk premium—the excess return stocks offer over risk-free government debt—has collapsed to roughly 40 basis points. That is not a margin of safety. That is a dare.
Yet BlackRock's chief investment strategist for the Americas, Wei Li, recently told clients to favor US equities over government bonds, citing "AI-driven earnings growth" as the primary vector reshaping the investment landscape. The logic is simple: AI is no longer a concept. It is a profit engine. And that engine, Li argues, is powerful enough to offset the gravitational pull of higher-for-longer rates.
I have heard this song before. In 2017, I heard it from three ICO whitepapers that promised to decentralize identity verification. I deployed $150,000 of personal capital based on their tokenomics models. The result was a 92% loss. The lesson was not that blockchain was worthless. The lesson was that narrative density is inversely correlated with data integrity. Hype dies. Data breathes.
So when a strategist at the world's largest asset manager tells me AI earnings will reshape my allocation framework, I do not ask whether AI is real. It is. I ask a different question: whose earnings, exactly, and at what concentration?
Let me decode the signal from the noise.
The Concentration Problem
The phrase "AI-driven earnings growth" is a statistical mirage. It implies a broad-based uplift across the market. The data suggests otherwise. In 2024, the so-called Magnificent Seven—Microsoft, Apple, NVIDIA, Alphabet, Amazon, Meta, and Tesla—accounted for roughly 60% of the S&P 500's total earnings growth. NVIDIA alone contributed more to index-level earnings growth than the bottom 300 companies combined.
This is not a market rally. This is a single-node network effect. And networks, as any systems engineer will tell you, are only as robust as their most critical node. When I audited the BAYC NFT market in early 2021, I found that 60% of early sales were wash trades—the same wallet clusters trading among themselves to fabricate floor price momentum. I shorted leveraged NFT loans six weeks before the peak. The subsequent 70% collapse validated my holder integrity scoring model.
The S&P 500 today is not a diversified index. It is a leveraged bet on roughly seven companies, three of which derive their AI revenue from the same underlying chip supplier. That is not diversification. That is a correlated options position with extra steps.
The Earnings Quality Question
Li's thesis rests on an implicit assumption: AI-driven earnings are high-quality, recurring, and sustainable. My forensic skepticism demands I test that assumption against the actual composition of those earnings.
Microsoft's intelligent cloud revenue grew 20% year-over-year in fiscal 2024. NVIDIA's data center revenue has beaten expectations for six consecutive quarters. These are real numbers. But they are not homogeneous. A significant portion of this "AI earnings" is capital expenditure from hyperscalers—Microsoft, Google, Amazon—buying NVIDIA chips to build out AI infrastructure. In other words, a large chunk of AI revenue is one company's opex becoming another company's top line.
This is not a criticism of NVIDIA. It is a statement about circularity. When I ran my DeFi yield farming algorithms in 2020, I learned that protocols generating yield from other protocols' incentives were not creating value. They were creating leverage. The same dynamic applies here. If the hyperscalers' AI capex does not translate into end-user revenue growth—if the ROI validation cycle extends beyond 18 months—the entire earnings stack unwinds from the top down.
Gartner projects enterprise AI spending will reach 10% of IT budgets by 2025, up from 2% in 2023. That is a massive increase. But it also means 90% of IT budgets are still not AI. The market is pricing AI as if it will consume the entire enterprise software stack within a decade. That is a possible outcome. It is not a probable one. Simplicity scales. Complexity collapses.
The Valuation Trap
Let me put the valuation question in terms any trader understands. The S&P 500 forward P/E of 21.5 implies an earnings yield of approximately 4.65%. The 10-year Treasury yields 4.2%. The equity risk premium is therefore roughly 45 basis points. Historically, that premium has averaged 300-400 basis points. We are at the lowest level since the dot-com peak.
Li's argument is that AI earnings growth will expand the denominator—earnings—fast enough to justify the current multiple. That requires annualized earnings growth of 20-25% for the next three to five years. The consensus estimate for 2025 S&P 500 earnings growth is around 13%. Even the most optimistic AI bulls project 18-20%. The gap between the required growth and the projected growth is the risk premium the market is not paying you to bear.
I have seen this exact setup before. In 2021, I analyzed NFT floor prices using holder distribution entropy. The metric that mattered was not the average price but the concentration of holdings among top wallets. When 10% of wallets controlled 60% of supply, the floor was fragile. The same logic applies to index-level earnings. When seven companies drive 60% of growth, the index is fragile. Your emotion is not my edge. The data is.
The Contrarian Angle: What the Consensus Misses
The consensus view, as articulated by Li, is that AI earnings growth will outpace the drag from higher rates. The contrarian view is not that AI is a bubble. The contrarian view is that the market has already priced in the most favorable AI outcome, leaving no room for execution error.
Consider the following blind spots:
First, the regulatory overhang. The EU AI Act came into force in August 2024. The US has seen 40+ state-level AI bills proposed in 2024 alone. Compliance costs are not zero. They are a tax on AI margins that the current earnings models do not fully incorporate. When I audited stablecoin reserves in 2022, I found critical discrepancies in three major protocols—not because the teams were malicious, but because their models assumed no regulatory intervention. The same assumption is embedded in current AI earnings projections.
Second, the competitive deflationary spiral. API prices for frontier models have fallen 90%+ since GPT-4 launched. OpenAI's GPT-4o mini is dramatically cheaper than the original GPT-4. This is great for consumers. It is terrible for model providers' margins. The AI application layer is becoming commoditized faster than the infrastructure layer can maintain pricing power. The earnings growth is accruing to the picks-and-shovels players—NVIDIA, TSMC—not necessarily to the application companies that the market is also bidding up.
Third, the China factor. DeepSeek, Alibaba's Qwen, and ByteDance's models are closing the capability gap at a fraction of the training cost. If open-source models continue to erode the closed-source moat, the pricing power of US AI leaders erodes with it. The market is not pricing this risk. It is pricing a US-centric AI monopoly that persists indefinitely.
The Takeaway: What I Am Watching
I am not telling you to sell equities and buy bonds. That would be a prediction, and I do not make predictions. I make observations and set triggers.
Here is what I am watching over the next 6-18 months:
- Hyperscaler AI revenue quality: Are Microsoft, Google, and Amazon reporting AI revenue that is incremental and recurring, or is it one-time commitments and internal transfers? I want to see the ratio of AI revenue to AI capex. If that ratio does not improve within four quarters, the circularity problem is real.
- NVIDIA's order book vs. actual deployments: Blackwell shipments are ramping, but I want to see utilization rates at the data centers buying those chips. If utilization lags, the capex cycle peaks sooner than consensus expects.
- Enterprise AI renewal rates: The 2024 cohort of enterprise AI pilots will hit renewal decisions in 2025. If renewal rates fall below 80%, the ROI validation problem is confirmed. I will be tracking Gartner and IDC surveys for this data.
- The 10-year yield: If the 10-year breaks above 4.5% while the S&P 500 forward P/E stays above 21, the equity risk premium goes negative. That is a structural sell signal, regardless of AI narrative.
- Concentration metrics: I am tracking the percentage of S&P 500 earnings growth attributable to the top 7 names. If that number stays above 50% for another year, the index is not a diversified asset. It is a single-factor bet.
The Final Word
Wei Li is not wrong that AI is reshaping the investment landscape. She is wrong to imply that this reshaping is broad-based or that the current valuation has priced in the execution risk. The AI trade is a concentration trade dressed in index clothing. It is a bet on NVIDIA's execution, Microsoft's enterprise distribution, and the absence of regulatory intervention—all at once.
I have been on the wrong side of this type of trade before. In 2022, I lost $200,000 in Terra-Luna despite my risk models, because I underestimated the fragility of uncollateralized debt. The lesson was not to avoid risk. The lesson was to demand a premium for bearing it. The current equity risk premium does not pay you for the concentration risk embedded in the index. It pays you nothing.
Don't buy the noise. Buy the node. And right now, the node is not the S&P 500. It is a handful of companies with verifiable, recurring AI revenue—and even those deserve a margin of safety that the current multiple does not provide.
The market will tell you when the AI earnings thesis breaks. It will show up in renewal rates, in capex-to-revenue ratios, and in the yield curve. Until then, I am not shorting the narrative. I am just refusing to pay a premium for it.
Your emotion is not my edge. The data is. And the data says: the AI trade is real, but it is narrow, it is priced for perfection, and it is one bad earnings call away from a repricing that the index will not absorb without significant drawdown.
Position accordingly. Or don't. The market does not care either way.