Frontier labs don't use most AI compute(yet) (epoch.ai)

OpenAI's 2022 launch of ChatGPT initiated a massive boom in AI compute, with global AI computing power now estimated at the capacity of around 20 million Nvidia H100 GPUs, significantly fueled by nearly $1 trillion in annual capital expenditures. Despite their influential role, OpenAI and other leading AI labs like Anthropic are currently utilizing only 10% to 15% of the world's operational AI compute resources, a figure that could grow but remains below half of the total global capacity. This underutilization highlights a substantial gap in deployable compute, signaling that while frontier AI models have sparked extensive growth, they haven't monopolized the infrastructure. The implications for the AI and ML community are profound. As top labs continue to grow their resource demands at an impressive pace—OpenAI foresees its data center capacity reaching up to 12 GW by 2027—there is an urgent need for an acceleration in overall compute production to sustain advancements in model capabilities. If computing power growth at these labs continues to outpace the global supply, we may see a consolidation of AI resources among a small number of players, potentially leading to heightened competition and increased prices for GPU hours. This situation prompts key questions about the sustainability of compute growth and the broader economic impacts if the frontier AI developers consume available resources at a faster rate than they can be produced, suggesting a pivotal shift in how AI development evolves in the coming years.