1:03:13
Julien Garran | AI: The Biggest Capital Misallocation in History | No Bull George Noble.
Market Talk with George Noble
Overview
This video discusses the current AI boom, arguing it represents a significant misallocation of capital, potentially the largest in history, far exceeding the dot-com bubble. The speaker posits that AI's commercial viability is limited due to its fundamental design flaws, leading to massive losses for most companies involved. The discussion contrasts the current AI hype with historical market bubbles and explores potential macroeconomic consequences, including the need for significant intervention, and suggests a shift towards resources and emerging markets as a more promising investment strategy.
How was this?
Save this permanently with flashcards, quizzes, and AI chat
Chapters
- The economy thrives on efficient capital allocation; activities that destroy capital lead to trouble.
- Current AI capital misallocation is estimated to be 17 times larger than during the dot-com crisis.
- This situation has serious macroeconomic implications and will be difficult to resolve.
- The AI sector's current GDP impact is estimated at 3%, with potential for a 6% decline if it reverses.
Understanding the scale of capital misallocation is crucial for grasping the potential economic fallout and the challenges in unwinding this situation.
The speaker cites a 'Wixle spread' as a tool to measure capital misallocation, indicating the current situation is 17 times worse than the dot-com bubble.
- Most AI companies, except perhaps Nvidia, are making significant losses.
- The AI model is 'built to fail' commercially due to inherent limitations.
- Companies have shifted from capital-light to capital-heavy, relying heavily on expensive hardware like Nvidia's Blackwell GPUs.
- The reliance on vendor financing, similar to Cisco in 2000, poses a significant risk as receivables balloon.
This chapter explains the fundamental reasons why AI, as currently implemented, is unlikely to achieve widespread commercial success and profitability.
Nvidia's receivables have increased by 770% in 33 months, a stark warning sign analogous to Cisco's situation during the dot-com bust.
- LLMs are primarily correlation engines, not true understanding systems.
- They struggle with tasks requiring genuine construction or complex reasoning, often producing nonsensical outputs.
- The 'Operator' AI by OpenAI only worked 34% of the time on self-set metrics, and real-world tests on platforms like Upwork show success rates as low as 2.5%.
- LLMs are trained on vast datasets, but their ability to apply knowledge is limited to what has been explicitly seen, not generalized understanding.
Understanding these limitations is key to assessing the realistic capabilities and applications of current AI technology.
An AI's inability to construct Pascal's triangle correctly after explaining how to do it demonstrates its lack of true comprehension, relying only on pattern matching from its training data.
- Improving LLM accuracy requires massive increases in computational power and cost, with diminishing returns.
- The historical scaling laws like Denard scaling (power efficiency) and Moore's Law (cost of compute) have ended, making AI development increasingly expensive.
- Significant investments in newer models (e.g., from $50M for GPT-3 to $500M for GPT-4) have yielded underwhelming improvements, suggesting a scaling wall has been hit.
- The immense power consumption and cost of data centers, with individual GPU racks using power equivalent to hundreds of homes, highlight the economic unsustainability.
The end of scaling laws and the escalating costs of AI development and deployment present a significant barrier to profitability and widespread adoption.
The progression of OpenAI's spending from $50 million for GPT-3 to a planned $5 billion for GPT-5, which was delayed because improvements were not significant enough, illustrates the diminishing returns on investment.
- AI models are based on correlation, not causation, leading to spurious findings (e.g., cheese consumption vs. bedsheet entanglement deaths).
- So-called 'rote learning' means AI can recall information but cannot reason or apply it to novel situations.
- AI's inability to ensure code robustness, security, or integration with existing systems limits its practical use in software development.
- Safety overlays and guardrails, while necessary, can unpredictably degrade model performance in other areas (neuroplasticity analogy).
These fundamental issues demonstrate that AI lacks true understanding and causal reasoning, severely limiting its potential for creating genuinely valuable and reliable commercial products.
The correlation between cheese consumption and deaths from bedsheet entanglement, while statistically linked in data, has no causal relationship, highlighting the flaw in relying solely on correlation.
- The market is shifting from rewarding any mention of AI to a more discerning phase, indicating a potential downturn.
- Companies like Oracle and Blue Owl, initially hyped for AI involvement, are now seeing their stock prices fall significantly.
- The 'picks and shovels' trade (like Nvidia) is unsustainable if the gold rush (AI) yields no gold.
- There's a growing conviction that the market is at a career-defining turning point, moving away from big tech/AI towards resources and select emerging markets.
This chapter explains how market sentiment is changing and why investors should consider shifting capital away from overvalued AI stocks towards more fundamentally sound assets.
Oracle's stock price fell below its pre-announcement level after revealing its massive AI data center investments, signaling a loss of market confidence.
- The current AI bubble is unsustainable and likely to reverse, potentially causing a significant economic downturn.
- Governments may need to intervene with macroeconomic policies like lower rates and QE to counteract the downturn.
- This environment favors a 'real picks and shovels' trade in commodities like copper and gold, which have been underinvested.
- Emerging markets, particularly India, are well-positioned for growth due to demographic trends and increasing middle-class demand.
This section provides a forward-looking perspective, identifying sectors and regions likely to outperform as the AI narrative fades and economic conditions shift.
Low inventories across supply chains combined with underinvestment in resources create a strong setup for commodity outperformance when demand increases.
Key takeaways
- The current AI boom is characterized by massive capital misallocation, potentially 17 times larger than the dot-com bubble, with significant macroeconomic risks.
- AI's commercial viability is fundamentally limited by its design as a correlation engine rather than a true understanding system, leading to widespread unprofitability.
- The end of historical scaling laws (Moore's Law, Denard scaling) makes AI development and deployment increasingly expensive with diminishing returns.
- AI models struggle with complex reasoning, real-world application, and robustness, making them unreliable for critical commercial tasks.
- Market sentiment is shifting away from AI hype, with discerning investors beginning to favor tangible assets like commodities and promising emerging markets.
- The 'picks and shovels' trade in AI is unsustainable if the underlying 'gold rush' fails to materialize, suggesting Nvidia and similar companies may face challenges.
- A potential economic downturn driven by the AI bubble's collapse could necessitate significant government intervention, creating opportunities in resources and emerging markets.
Key terms
Capital MisallocationDot-com BubbleLarge Language Models (LLMs)Correlation vs. CausationRote LearningScaling Laws (Moore's Law, Denard Scaling)Vendor FinancingData CentersCommoditiesEmerging Markets
Test your understanding
- How does the current AI capital misallocation compare in scale to the dot-com bubble, and what are the potential macroeconomic consequences?
- What are the fundamental limitations of Large Language Models that prevent widespread commercial success, and why is AI considered 'built to fail' commercially?
- Why have historical scaling laws like Moore's Law ended, and how does this impact the cost and feasibility of developing advanced AI?
- What is the difference between correlation and causation in the context of AI, and why is this distinction critical for assessing AI's reliability?
- Given the potential downturn in AI, what asset classes and geographical regions does the speaker suggest are more promising for investment, and why?