Scaling AI Agent Populations

New projections from Epoch AI suggest a significant expansion in the capacity for autonomous AI agents. According to data released by the research organization, the volume of AI chips scheduled for shipment through 2027 is sufficient to support an estimated 1.9 billion AI agents running simultaneously. This capacity is reportedly equivalent to the total working hours of the entire human population.

The projected growth is being driven by a dual trend of decreasing costs and increasing computational efficiency. Epoch AI estimates that the cost required to achieve a specific level of AI performance is declining by approximately 13 times per year. For example, the cost to achieve a 75% score on the GPQA Diamond benchmark was roughly $0.30 per question using the o3 model in January 2025; however, Epoch AI projects that GPT-5.6 Luna could achieve the same score for approximately $0.0004 per question in 2026—a reduction of roughly 725x in less than 18 months.

Efficiency Gains in Reasoning

Recent research published in Scientific Reports via Nature indicates that newer large language models are becoming more efficient in how they utilize reasoning tokens. A comparison between OpenAI’s o1 and o3 models revealed that when both o1-mini and o3-mini correctly solved problems on the Omni-MATH benchmark, o3-mini required fewer reasoning tokens 63% of the time. Similarly, on the GPQA Diamond benchmark, o3 used fewer tokens than o1 on 66% of the questions both solved correctly.

These findings suggest that model improvements are not merely a matter of allowing models to “think” longer, but rather increasing the utility of each token used. This aligns with findings from MIT FutureTech / CSAIL, which reported that achieving the same benchmark performance across math, science, reasoning, and coding tests has become 5 to 10 times cheaper per year. MIT researchers noted that model efficiency improvements account for approximately a 3x annual increase in efficiency, independent of hardware cost reductions.

As the demand for compute scales, significant capital is moving into the semiconductor and cloud sectors. Reports indicate that Broadcom Inc is working to amass $60 billion in fresh financing to fund AI chips for Anthropic and other industry players.

In the cloud sector, regulatory scrutiny is increasing, with the European Union preparing to conduct further examinations of the cloud divisions of Microsoft and Amazon.com Inc. Meanwhile, industrial developments continue in the data center space, with reports suggesting that Alibaba and Solaria are in initial discussions regarding power supply arrangements for a data center located in Spain.