A skeptical look at AI investment

Market Context and Spending

• Generative AI is projected to attract over $1 trillion in investment in the coming years, driving over 60% of the S&P 500's year-to-date returns.
• Despite the spending, no "killer application" has yet been identified that is cost-effective for business use.

Productivity and Economic Impact Estimates

• Goldman Sachs Bull Case: Estimates suggest generative AI could automate 25% of work tasks, boost U.S. productivity by 9%, and increase GDP by 6.1% cumulatively over the next decade.
• Darren Acemoglu (MIT) Skeptical Case: Projects only 4.5%–4.6% of tasks will be cost-effective to automate within 10 years, implying a 0.5% productivity boost and a 1% GDP increase.
  • Acemoglu attributes this limitation to AI's inability to handle multifaceted real-world interactions (transport, manufacturing) compared to pure mental tasks.
  • He argues that scaling laws (doubling data/compute) do not linearly double capabilities due to undefined metrics for "improvement" and a lack of high-quality data sources.
• Jim Covello (Goldman Sachs) Structural Critique: Highlights that AI replaces low-cost labor with high-cost infrastructure (GPUs), unlike e-commerce or the internet which started as cheaper alternatives from day one.
  • Covello notes a lack of competition in the GPU market (dominated by NVIDIA), preventing immediate cost reductions comparable to historical tech cycles.

Investment Risks and Corporate Behavior

• FOMO-Driven Spend: Major tech companies are compelled to invest in an "AI arms race" to avoid falling behind, even without immediate ROI, similar to past hype cycles like the Metaverse.
• Wasted Capital Risk: Acemoglu suggests a portion of the investment boom will be wasted on attempts to automate tasks that are not yet technically or economically viable.
• Covello's Historical Parallel: Compares the current build-out to the 2000–2001 internet bubble, warning that excess infrastructure capacity will eventually require a long period of market absorption.
• Early Results: Current enterprise adoption shows limited applications and very few companies are realizing actual cost savings.

Forward-Looking Indicators and Timelines

• 12–18 Month Horizon: Covello and Eric Sheridan (Goldman Sachs) identify the next 12 to 18 months as a critical window for the emergence of visible, tangible "killer applications."
• Inflection Point: If significant, cost-effective applications do not materialize within this timeframe, skepticism regarding returns will likely intensify.
• Key Metric to Watch: Corporate profit trends are cited as the primary indicator; negative ROI experiments are the first to be cut when profits slow.

Counter-Perspective on Valuation

• Cash Rangan and Eric Sheridan: Argue that current AI capital expenditure (CapEx) as a share of revenue aligns with prior tech cycles and that incumbent companies with low capital costs and massive distribution networks are better positioned for returns.
• Investment Strategy: Both Cohen and Rangan suggest maintaining exposure to infrastructure providers, noting that valuations are driven by fundamental growth expectations rather than immediate profitability.

Long-Term Technological Trajectory

• Cognitive Limits: Acemoglu questions the feasibility of superintelligence within 20–30 years, arguing current Large Language Models lack the multi-modal reasoning and sensory inputs required for human-like cognition.
• Human-in-the-Loop Scenario: The likely path to scientific advancement involves AI providing inputs while humans remain in control of decision-making, testing, and real-world validation.