AI Exchanges: Will falling costs drive new opportunities?

Market Context and Trigger Event

• Recent market volatility stems from DeepSeek, a China-based AI firm, launching a low-cost AI tool.
• This development challenges assumptions regarding massive capital expenditure (CapEx) by major U.S. tech companies and questions the economic viability of high-cost pre-training infrastructure.

Efficiency Economics and Jevons Paradox

• DeepSeek's pricing suggests per-token costs are trending toward marginal zero, potentially reducing the capital required for pre-training.
• Host George Lee argues this validates efficiency gains, addressing skepticism about high costs and the scarcity of utility.
• The group applies Jevons Paradox: as the price of computing power declines, demand and total volume of consumption are expected to increase, offsetting lower unit prices.
• Lee posits that lower costs will breed abundant new use cases, rendering current sunk capital and infrastructure planned for the next three years "well used."

Debate on Capital Expenditure (CapEx) Viability

• Kim Possett notes that despite efficiency concerns, current CapEx levels remain appropriate given the "near vertical" advancements in model complexity.
• Historical Data Point: The four major hyperscalers (Amazon, Alphabet, Meta, Microsoft) increased combined annual CapEx from $116 billion in 2022 to nearly $200 billion in 2024.
• Forward-Looking Commitments:
  • Meta announced a $60–$65 billion AI-related CapEx budget for the current year.
  • Microsoft disclosed an $80 billion AI CapEx budget.
  • The "Stargate" AI infrastructure joint venture was recently announced.
• Medium-Term Uncertainty: While current spending is justified, the trajectory of CapEx in 3–10 years remains an open question dependent on future efficiency curves.

Identifying Key Constraints: Power vs. Data

• Consensus Shift: The consensus among guests has shifted from viewing data as the primary bottleneck to identifying power availability as the critical constraint.
• Power Demand Dynamics:
  • AI servers require approximately 10x the power of traditional servers.
  • New AI data center campuses are projected to require multi-gigawatt power, sufficient to power entire cities.
  • This represents a tectonic shift from historical baseload power demand growth of under 3% annually in the U.S.
• Innovation Implications: The power constraint is expected to drive innovation in energy delivery, including green sources, battery storage, small modular nuclear reactors, and fusion.

Data Supply and Emerging Economies

• While human-generated data may be approaching saturation, the ecosystem is evolving toward synthetic data marketplaces.
• New Data Models:
  • AI-generated data mimicking real-time information (e.g., medical records) for model training.
  • Personal data marketplaces where individuals can opt in and sell proprietary data to businesses.
• Licensing Trends: Continued formation of partnerships and licensing deals involving publishers, social media platforms, and stock photography firms.

AI Agents and Operational Capabilities

• Definition: AI agents are systems capable of autonomously executing multi-step, linked tasks (e.g., booking flights, hotels, and cars simultaneously).
• Current State: The technology is in "early days," with current consumer-oriented examples described as "herky-jerky" and slow compared to human interaction.
• Future Potential: Despite current limitations, agents are viewed as a powerful vector for automation, moving beyond repetitive tasks to complex process management.

Enterprise Adoption and Deal Making

• Adoption Curve: 2023 was characterized by testing; 2024 is identified as the inflection point for true enterprise adoption and scaling.
• Corporate Sentiment: CEOs express optimism regarding easier monetary and regulatory environments, predicting increased M&A, IPOs, and investment.
• Strategic M&A: The group anticipates a rise in AI-driven strategic mergers and acquisitions in the coming year, following precedents set in 2023.
• Goldman Sachs Implementation: The firm launched the "GS AI Assistant" to provide employees with secure, compliant access to leading-edge models, aiming to foster creativity and innovation among junior staff.

Macro Outlook

• Despite short-term market volatility, the guests maintain a bullish long-term view on AI's trajectory.
• The rapid decline in costs and rise of agents signal a discontinuous moment in technological history, though it represents a small blip in the broader 120-year curve of declining computing costs (Moore's Law).
• Generational shifts in workforce fluency are expected to accelerate productivity and embed AI into professional and personal workflows.