AI Exchanges: Will falling costs drive new opportunities?

Episode Summary

Executive Summary: Goldman Sachs launches its AI Exchanges series by arguing that DeepSeek’s low-cost model is more a sign of AI efficiency gains than a threat to the build-out thesis. George Lee and Kim Posnett say falling token and compute costs could expand adoption, enable new use cases, and support continued infrastructure investment, while highlighting power, data, and enterprise adoption as the next major constraints.

Main Topics: DeepSeek and the AI capex debate (Priority: 5/5): The hosts frame DeepSeek’s low-cost model as a challenge to assumptions about massive AI spending, but argue it mainly accelerates efficiency and broader utility rather than invalidating the investment thesis. Jevons paradox and cheaper AI tokens (Priority: 5/5): Lee and Posnett argue that lower per-token costs should stimulate more usage, more pre-training, and more inference-driven applications, making AI more abundant rather than less important. Enterprise use cases and adoption (Priority: 4/5): The conversation emphasizes that lower costs could unlock practical applications in legal, finance, research, healthcare, and conversational AI, speeding enterprise uptake in 2025. AI agents and new interfaces (Priority: 4/5): The discussion explains AI agents as systems that execute multi-step tasks autonomously, with examples like travel booking and web-based shopping assistance, though the technology is still early and clunky. Power and data as bottlenecks (Priority: 5/5): Posnett argues the biggest constraint is increasingly power rather than data, citing the energy intensity of AI infrastructure and the need for new power-generation and storage solutions. Dealmaking, financing, and ecosystem growth (Priority: 3/5): The speakers expect AI to spur strategic M&A, financing, IPOs, and growth in adjacent tooling, security, and infrastructure businesses as the broader corporate backdrop improves. Goldman’s internal AI adoption (Priority: 4/5): Goldman Sachs highlights its own GSAI assistant as a firmwide, compliant way to access leading-edge models, signaling how regulated enterprises are beginning to scale AI internally.

Key Arguments: DeepSeek’s lower-cost approach may reduce some long-term capital intensity, but it does not end the need for AI infrastructure; it may actually broaden participation by lowering the barrier to entry. Falling token costs should trigger Jevons paradox: cheaper intelligence leads to more consumption, more use cases, and ultimately more total compute demand. Most of the current AI infrastructure spending is still justified because near-term model scaling remains highly capital- and energy-intensive. AI is moving from experimentation to true enterprise adoption, with 2025 expected to be a breakout year for scaling real business use cases. Agents are an important next step in AI because they can complete linked tasks autonomously, though current products remain early and imperfect. Power has become an increasingly serious bottleneck for AI scaling, potentially surpassing data as the binding constraint. The AI ecosystem will likely create new markets for synthetic data, data licensing, and personal data marketplaces. Lower regulatory and monetary policy headwinds could support more tech dealmaking, AI-related M&A, and financing activity.

Data Points: Goldman Sachs AI Exchanges launch date: January 29 - Episode recorded on Wednesday, January 29th. CapEx spend of Amazon, Alphabet, Meta, and Microsoft in 2022: Over $116 billion - Used by Kim Posnett to illustrate the scale of AI-related infrastructure investment when GPT was launched to the public. CapEx spend of Amazon, Alphabet, Meta, and Microsoft in 2024: Just under $200 billion - Shown as evidence that AI infrastructure investment nearly doubled in two years. Meta AI-related CapEx guidance: $60 to $65 billion - Announcement cited as an example of continued heavy spending on AI infrastructure. Microsoft AI-related CapEx guidance: $80 billion - Referenced to show ongoing large-scale investment in AI data centers and compute. Baseload power demand growth historically: Sub-3% annually for decades - Posnett contrasts historical U.S. power demand growth with the current AI-driven surge. AI server power requirement: About 10x traditional servers - Used to explain why AI is putting unprecedented strain on power infrastructure.

Pivotal Quotes: "I think this answers some of Jim's principal concerns, not all of them, but some of them." — George Lee: On whether DeepSeek’s low-cost model undermines skepticism about AI capex spending. "The cost of compute is coming down dramatically. The price per token is coming down dramatically." — Kim Posnett: On why lower costs are broadly positive for AI adoption and efficiency. "This year is the year of true enterprise adoption and scaling." — Kim Posnett: On the expectation that companies will move from testing AI to operational deployment.

Implications: AI’s economics may be shifting from scarcity to abundance: lower costs could accelerate adoption, expand use cases, and sustain infrastructure demand, but power supply and enterprise execution will determine how far and how fast the market grows.