Interview, Fireside Chat, Other
AI Exchanges: Will falling costs drive new opportunities?
- The "AI Exchanges" podcast series hosted by Allison Nathan and George Lee aims to analyze current AI impacts and future trajectory over the next several years.
- Advances in DeepSeek development suggest potential pre-training efficiencies that could significantly reduce capital allocation, causing per-token costs to trend toward marginal zero.
- Reduced capital costs may stimulate expanded pre-training activity by a broader demographic, while infrastructure capital sunk in the past three years is projected to be well utilized.
- Long-term capital necessity is questioned over three, five, seven, and ten-year horizons, prompting Goldman Sachs to reevaluate medium and long-term expenditure requirements.
- Increased model efficiency is expected to trigger a Jevons paradox effect, driving accelerated adoption, new consumption patterns, and unforeseen use cases across legal, financial, and scientific sectors.
- Specific applications include immune system and brain modeling for healthcare, alongside the emergence of ubiquitous conversational AI and consumer-oriented AI agents capable of multi-step complex tasks.
- While AI agents are currently in very early development with slow, deliberate progress, their evolution is anticipated to mirror the early web era of 1997 before becoming ubiquitous.
- Data economies are evolving through new partnerships and licensing, with anticipated markets for synthetic data mimicking medical records and personal data marketplaces allowing individuals to sell data for model training.
- CEO optimism regarding monetary policy and regulation is expected to fuel strategic activity, M&A, IPOs, and investment in the coming year, with investors remaining bullish despite emerging questions.
- Power has emerged as the primary constraint over data, with AI servers requiring ten times the energy of traditional servers and hyperscaler data centers reaching multi-gigawatt scales capable of powering entire cities.
- Massive power demand is projected to drive innovation in green sources, battery storage, small modular nuclear, and nuclear fusion to support scaling needs.
- 2025 is identified as a pivotal year for true enterprise adoption and scaling, with internal tools like the GS AI Assistant facilitating safer, compliant access to leading-edge models for the workforce.
- The near vertical advancement of current models justifies high capital expenditure today, though uncertainty remains regarding future capex requirements as models become more efficient over the next three to five years.