A skeptical look at AI investment

Episode Summary

Executive Summary: The episode examines the gap between bullish AI expectations and skeptical assessments of near-term economics. Goldman Sachs researchers outline huge potential productivity gains, but MIT's Daron Acemoglu and GS's Jim Covello argue current AI is expensive, limited in real-world usefulness, and unlikely to justify trillion-dollar investment quickly. The discussion centers on task automation limits, cost-effectiveness, infrastructure spending, and whether AI will eventually create new jobs and scientific breakthroughs.

Main Topics: Bull case vs. skeptical valuation of AI (Priority: 5/5): The episode contrasts market optimism about generative AI's transformative potential with doubts that current economics can support the scale of investment already underway. Task automation and productivity estimates (Priority: 5/5): Daron Acemoglu presents a restrained model in which only a small fraction of tasks will be cost-effectively automated over the next decade, implying modest GDP and productivity gains. Limits of scaling laws and current architecture (Priority: 4/5): Acemoglu argues that more data and more GPUs will not automatically produce major leaps in capability, especially given unclear metrics, data quality constraints, and architectural limits of LLMs. Trillion-dollar AI spending and infrastructure risk (Priority: 5/5): Jim Covello argues AI is unusual because it is a very expensive technology trying to replace relatively cheap labor, raising concern that the infrastructure build-out may overshoot demand. Historical parallels and bubble dynamics (Priority: 4/5): Covello compares AI to prior hype cycles such as VR, blockchain, metaverse, and the internet buildout, warning that bubbles can last a long time before excess capacity becomes visible. Potential for new occupations and scientific discovery (Priority: 3/5): Both skepticism and cautious optimism acknowledge that AI could eventually create new tasks, new jobs, and better scientific workflows, but only if humans remain central to directing and validating outputs. Investment strategy and monitoring signals (Priority: 4/5): Covello advises investors to keep owning infrastructure names for now but watch for the emergence of obvious applications and weakening corporate profits as signs that AI ROI is deteriorating.

Key Arguments: Acemoglu argues that productivity gains depend on both the share of tasks affected and the cost-effectiveness of automating them; he thinks both are limited in the near term. He estimates AI may ultimately affect about 20% of value-added tasks in theory, but only about 4.5%-4.6% of tasks within a 10-year horizon after cost-effectiveness is considered. He rejects the idea that simply adding more compute and data will rapidly deliver dramatically better AI, noting unclear capability metrics, low-quality data problems, and possible architectural limits. He is skeptical that large language models can achieve superintelligence without human involvement in hypothesis generation, testing, and real-world validation. Covello says AI is unlike prior technologies because it is expensive from the start and is replacing low-cost labor, not expensive incumbents, making the economic case unusual. He argues AI cost reductions are not guaranteed because NVIDIA's GPU dominance creates a bottleneck; lower costs would likely require real competition from Intel, AMD, or hyperscalers' custom chips. Covello believes current enterprise use cases are weak, with very few companies actually saving meaningful money from AI today. He still expects continued investment because firms fear being left behind if AI does become important, creating a powerful FOMO dynamic. Both skeptics acknowledge that AI could eventually enable new tasks and scientific workflows, but they believe humans must remain in control for this to work well.

Data Points: Expected AI-related spending: over $1 trillion - Estimated corporate spending on AI-related infrastructure, chips, data centers, and applications over the coming years. Potential task automation (bull case): a quarter of all work tasks - Goldman Sachs research estimate cited by the host as the optimistic scenario for generative AI. U.S. productivity boost (bull case): 9% - Goldman Sachs estimate of long-run productivity improvement from generative AI. U.S. GDP growth boost (bull case): 6.1% cumulatively over the next decade - Goldman Sachs estimate of cumulative GDP impact under the bull case. AI-exposed tasks cost-effectively automated (Acemoglu estimate): only a quarter of AI-exposed tasks - Acemoglu's paper suggests only a limited share of exposed tasks will be worth automating economically. Impact on all work tasks (Acemoglu estimate): less than 5% - His estimate after accounting for cost-effectiveness over the next 10 years. U.S. productivity boost (Acemoglu estimate): 0.5% - Cumulative productivity impact over the next decade in his baseline scenario. U.S. GDP boost (Acemoglu estimate): about 1% cumulatively over the next decade - His estimate of macroeconomic gain from generative AI over 10 years. Value-added share of tasks potentially transformed: about 20% - Acemoglu's reading of task-coding studies on what AI could ultimately transform. Cost-effectively automatable share within 10 years: 20%-25% of what is ultimately doable - Acemoglu's use of computer vision research as a proxy for timing. Combined short-run automatable share: about 4.5%-4.6% - Result of combining task exposure and cost-effectiveness assumptions for the 10-year horizon. Workers in occupations that did not exist 80 years ago: 60% - Statistic cited to show that technology can create new occupations over time. AI build-out horizon concern: 5-10 years - Covello says this is the period over which AI still has not shown a clear, cost-effective killer application. Current enterprise savings: very limited / very few companies - Covello's assessment of early corporate adoption results. Applications window to watch: 6 to 18 months - Eric Sheridan's warning that lack of a killer application in this period would increase concern about ROI.

Pivotal Quotes: "there's not a single thing that this is being used for that's cost-effective at this point." — Jim Covello: He argues current AI deployments have not yet demonstrated meaningful economic value. "What trillion dollar problem is AI going to solve?" — Jim Covello: Core challenge behind his skepticism about the scale of current AI spending. "So the way that I go about doing that is I rely on one of the most comprehensive studies..." — Daron Acemoglu: He explains the method behind his task-automation estimate, using prior research to bound expectations.

Implications: Listeners should expect continued AI capex and hype, but near-term returns may be uneven. Investors may favor infrastructure names for now, while watching for a real killer app, broader enterprise savings, and signs that corporate profits are no longer funding experiments.