Why generative AI may not actually cause a compute crunch

The AI computing power landscape: Debunking scarcity concerns: Despite widespread worries about generative AI (genAI) causing a shortage of computing power, a detailed analysis suggests that the long-standing regime of affordable and widely available computing resources is likely to persist.

• Even under aggressive assumptions, genAI workloads are projected to account for only about 34% of global data centre-based AI computing supply by 2028.
• This forecast challenges the notion of an impending scarcity crisis in AI computing resources.

Breaking down genAI computational demands: GenAI computational requirements into three distinct areas, each with varying resource intensities and usage patterns.

• Model Training: This is the most resource-intensive phase, typically undertaken by a small number of companies with substantial computing capabilities.
• Fine-Tuning: Less demanding than initial training, fine-tuning requires approximately 0.1-10% of the resources used in training.
• Inference: The most common form of genAI usage, inference can often be performed on less specialized hardware, making it more accessible and less resource-intensive.

Quantitative modeling and bullish projections: A comprehensive quantitative model was developed to assess the future landscape of AI computing supply and demand, incorporating ambitious growth projections and realistic supply constraints.

• The model assumes continued scaling of AI models, with projections reaching up to 15 trillion parameters per prompt by 2028.
• Aggressive global adoption rates and exponential growth in utilization were factored into the demand projections.
• On the supply side, estimates suggest AI computing power will increase by approximately 60 times by the end of 2025 compared to early 2023 levels.
• Growth is expected to moderate to around 60% annually through 2028, still representing significant expansion.

Potential breaking points beyond hardware: While computing hardware supply appears sufficient to meet demand, other potential constraints could impact the AI computing landscape.

• A sudden explosion in consumer demand for inference could strain available resources.
• Supply chain disruptions could affect the production and distribution of necessary hardware components.
• Energy constraints are identified as the most pressing concern, potentially limiting the expansion of AI computing capabilities.

Implications for businesses: Given the projected abundance of computing power, businesses should focus on leveraging this resource rather than preparing for scarcity.

• Companies are advised to develop strategies to capitalize on the anticipated availability of AI computing resources.
• This approach could lead to more innovative and ambitious AI projects, as concerns about resource limitations are alleviated.

Broader context and future outlook: Future analysis promises to propose a framework for leveraging the expected abundance of computing power.

• This comprehensive approach to analyzing AI computing resources provides valuable insights for both technology strategists and policymakers.
• The findings challenge prevailing narratives about AI resource scarcity, potentially reshaping discussions about the future of AI development and implementation.

Critical analysis and remaining questions: Several important questions remain unanswered.

• How might geopolitical factors or unexpected technological breakthroughs affect the projected supply and demand dynamics?
• What are the environmental implications of the massive expansion in AI computing power, particularly in light of the identified energy constraints?
• How might the democratization of AI computing resources impact innovation and competition in the tech industry?

By addressing these questions in future research, a more complete picture of the AI computing landscape can emerge, helping stakeholders navigate the rapidly evolving world of artificial intelligence.