Article Summary

Rohit Krishnan explores the potential economic and resource implications of widespread Artificial General Intelligence (AGI), grounding the discussion in current trends and known constraints.

He begins by acknowledging AI CEOs' predictions of significant job displacement, potentially affecting 60-80% of jobs within a decade. The author emphasizes the need to prepare for this future by examining concrete data about AI's resource consumption.

Data centers are projected to consume 12% of US electricity by 2030, with AI operations responsible for over 40% of that power. AI model training is rapidly increasing, but data movement bottlenecks may pose limitations. The piece also dives into the costs associated with AI models like OpenAI's, detailing expenses related to model usage and inference.

Krishnan then considers how AGI might evolve, presenting various scenarios. One scenario involves AGI acting as autonomous "employees" capable of long-range planning, costing significantly less than current models.

Another contemplates a future where AGI isn't a perfect human replacement but functions as a productivity tool, requiring human oversight and correction. It is also possible AGI excels in specific domains like math and science, leading to scientific breakthroughs, but potentially posing monetization challenges.

Resource constraints are also a true bottleneck.

Considering electricity consumption, Krishnan estimates that by 2030, the US could run approximately 44 million concurrent H100 GPUs, translating to a substantial number of AGI "employees." Meeting this demand would require significant expansion of the semiconductor, energy, and AI industries.

He ends by suggesting that the economy will likely undergo a reshuffling as AI becomes more integrated, leading to a mix-and-match approach with different types of AI serving various economic needs.