March 2025 — BlackRock Investment Strategist Wei Li has made the call: AI-driven earnings growth makes U.S. equities structurally more attractive than government bonds. On the surface, this reads as standard institutional positioning in a bull market. But strip away the asset allocation language, and the underlying assumptions reveal something far more fragile.
I've spent the last four years auditing the gap between AI narratives and actual protocol-level fundamentals. What I'm seeing now in Wei Li's thesis mirrors patterns I've tracked since the 2020 DeFi summer — the same concentrated optimism, the same missing layers of risk assessment, and the same tendency to confuse momentum with structural transformation.
The "AI earnings growth" framing isn't an investment thesis; it's a bet on three specific assumptions that remain unproven.
The Context: Why This Narrative Is Getting Traction
The timing matters. BlackRock's public commentary arrives as the S&P 500 trades at roughly 21-22x forward earnings against a historical average of 16-17x. Meanwhile, the Mag 7 cohort carries forward P/Es in the 30-35x range, with NVIDIA still hovering at 60-70x based on trailing fundamentals.
Traditional valuation metrics are stretched by any historical measure. But the AI narrative provides a convenient justification: this time the growth is real.
There's truth in that. Microsoft's Intelligent Cloud revenue (including Azure AI) posted 20%+ growth in FY2024. NVIDIA's data center segment has blown through analyst estimates for six consecutive quarters. Anthropic and OpenAI are tracking annualized revenues of $1B and $3.4B respectively. Enterprise AI budgets jumped from roughly 2% of total IT spend in 2023 to an estimated 5-8% by end of 2024, with Gartner projecting 10%+ by 2025.
These are meaningful signals. AI commercialization has crossed the proof-of-concept threshold. The infrastructure layer — chips, cloud, data centers — is generating real cash flows.
But here's the problem with the equity-over-bonds thesis: it assumes AI-driven earnings growth will outpace the risk-free rate over the investment horizon. With the 10-year Treasury yielding 4.0-4.5% and the S&P 500 earnings yield sitting around 4.5-4.8%, the equity risk premium has compressed to roughly 0.3-0.5% — near historic lows.
The Core: Three Structural Assumptions That Need Auditing
Let me break down the actual components of the AI equity trade.
Assumption One: AI Earnings Growth Is Broad-Based
Wei Li's framing suggests AI-driven earnings are reshaping the broader market. The data tells a different story. AI revenue accumulation is brutally concentrated. Microsoft, Google, NVIDIA, and Amazon capture the overwhelming majority of AI-related revenue growth in public markets. Second-tier players — IBM, Intel, Oracle — contribute marginal AI revenue at best.
This isn't a market-wide transformation. It's a concentrated bet on roughly six to eight companies trading at premium multiples.
During the 2021 BAYC liquidity crunch, I watched retail investors mistake NFT floor price momentum for a liquid asset class. The same conflation is happening here: market-wide index movement is being interpreted as broad-based AI adoption, when it's actually six companies carrying an entire index.
Assumption Two: AI Revenue Quality Justifies Premium Valuations
This is where I get more skeptical. The "AI-driven earnings growth" in current earnings reports splits into two categories: incremental revenue from new AI products, and efficiency gains from AI-optimized operations. These have very different implications for earnings quality.
Incremental AI revenue — actual customer spend on AI services — is verifiable. You can track Azure AI consumption, NVIDIA data center shipments, and enterprise Copilot seat expansion. I've done this forensic analysis on public filings since the 2020 Yearn.finance yield optimization work, and the signal is clear: the infrastructure and model layers are generating real cash flows.
But efficiency-driven earnings growth is a harder sell. When companies claim "AI-driven operational improvements" as a margin expansion story, I'm reminded of the liquidity mirage I saw in Terra/Luna's collapse — the appearance of stability masking structural fragility. AI-driven efficiency gains in enterprise settings often fail to survive contact with real-world deployment complexity. The ROl verification cycle for enterprise AI is still in early innings, and renewal rates remain unproven.
Assumption Three: AI's Competitive Dynamics Won't Erode Margins
The market's pricing of AI equities implicitly assumes sustainable competitive advantages. But look closer at the competitive landscape:
- Model layer: OpenAI, Anthropic, and Google lead, but Meta's open-source Llama series continues to close the gap. API prices for frontier models have already plummeted — GPT-4o mini costs 90%+ less than GPT-4 at launch. This is textbook margin compression.
- Compute layer: NVIDIA dominates with >80% market share, but AMD's MI300, Google's TPU, and Microsoft's Maia and Amazon's Trainium custom silicon all threaten the monopoly. Semiconductor supply has historically been cyclical, and the 2025-2026 capacity release cycle could hit pricing hard.
- Application layer: The most fragmented and competitive segment. Microsoft Copilot, Salesforce Einstein, ServiceNow Now Assist... no clear winner has emerged, which means marketing spend and price competition will likely compress margins before scale benefits materialize.
The Contrarian Angle: What the Bull Case Ignores
Here's the angle I'm not hearing from BlackRock or other institutional voices: the equity-over-bonds call requires ignoring the regulatory and structural risks that are only beginning to materialize.
The EU AI Act became legally effective in August 2024, and compliance costs are already hitting enterprise AI deployment timelines. Over 40 U.S. states have introduced AI-related legislation — the 2024-2025 state legislative session saw a massive increase in proposed AI bills. Copyright litigation against OpenAI, Anthropic, and Google (New York Times v. OpenAI, Getty Images v. Stability AI) carries potentially existential implications for training-data economics.
None of these risks are priced into the current equity risk premium, because they're not yet visible in earnings reports. But as my 2017 Parity multisig audit taught me — the risk that kills you is the one that doesn't show up in the transaction flow until it's too late.
The market is pricing AI earnings growth as if the regulatory environment will remain static. Historical precedent — from social media regulation to antitrust enforcement — suggests that's an aggressive assumption.
Additionally, the AI energy problem remains unaddressed. Large AI clusters now consume 10-20% of operating costs in power. Energy constraints could cap compute expansion faster than demand forecasts suggest. This isn't a peripheral issue; it's a structural constraint on the very growth narrative underpinning the trade.
The Takeaway: What to Watch Instead
The equity-over-bonds call isn't wrong — it's just incomplete. The AI earnings narrative will likely continue driving index performance in the near term, and the infrastructure investment cycle has legs for at least 6-18 months.
But I've learned that the best trades come from identifying what the consensus narrative omits.
Watch these specific signals instead:
Near-term (0-6 months): - Azure AI and AWS Bedrock consumption growth rates — are they accelerating or plateauing? - NVIDIA's data center revenue and Blackwell shipment pace — any guidance cuts break the entire trade - 10-year Treasury yield trajectory — above 4.5%, and the equity risk premium argument collapses
Medium-term (6-18 months): - Enterprise AI renewals and ROI validation — if Copilot-style deployments fail renewal tests, the application-layer narrative breaks - Regulatory enforcement actions under the EU AI Act — the first fines will reset compliance cost assumptions - Chinese AI competition (DeepSeek, Alibaba's Qwen) — open-source pressure on model pricing accelerates margin compression
The AI trade will continue to work for those who understand its concentrated, infrastructure-heavy composition. But the broader market narrative — that AI is broadly reshaping earnings across industries — remains unproven. Speed without precision is just noise. The precision here means knowing exactly which layer of the AI stack generates real cash flows, and which merely rides the narrative.
The market is pricing AI as a secular transformation. I'm still waiting for the earnings data to confirm that's more than a concentrated infrastructure cycle with a growth-label attached.