The development of advanced artificial intelligence is likely to continue at its current pace until 2030, requiring unprecedented levels of investment and energy. According to a new report, training a single top-tier AI model will demand investments of hundreds of billions of dollars and consume gigawatts of electrical power. David Owen reports for Epoch AI in an analysis commissioned by Google DeepMind. The author argues that while these challenges are significant, they are likely surmountable.
If current trends persist, a training cluster for a frontier AI model in 2030 could cost over $100 billion. Such a system would use thousands of times more computing power than the model GPT-4. The report examines potential obstacles to this growth, such as data shortages, power supply limitations, and prohibitive costs. However, Owen concludes that these issues are probably solvable. He notes that sufficient public data exists for several years of development, while synthetically generated data can fill future gaps. The massive energy demand could be met through renewable sources like solar, or by distributing training across multiple data centers. According to the report, continued revenue growth for leading AI labs could justify the immense investments.
This continued scaling is expected to unlock transformative capabilities, particularly in scientific research and development. The report predicts that by 2030, AI will function as a powerful assistant in many scientific fields, similar to how coding tools currently assist software engineers. This could lead to productivity improvements of 10 to 20 percent in specific tasks. Examples include AI systems that can autonomously fix software issues, help mathematicians formalize proofs, or answer complex questions about biology lab procedures.
Despite these rapid advances in capability, the report cautions that their real-world impact may take longer to materialize. Fields with long development cycles, such as drug discovery, will likely see effects in early-stage research but are unlikely to have AI-developed drugs approved for sale by 2030. In contrast, the field of software engineering is expected to change dramatically. The author concludes that AI is on track to become a key technology across the economy, urging decision-makers to prioritize AI-related issues.