The article highlights a looming "compute crunch" in the AI sector, driven by skyrocketing token consumption and user growth. As AI models improve, users are consuming exponentially more tokens—some, like the author, have seen their daily usage increase by 50 times over three years. The demand for processing this data is fueling a massive infrastructure rollout, with datacenters emerging globally to meet the needs of over one billion active LLM users. While companies like AWS, Azure, and GCP have ramped up their capital expenditures, concerns about how effectively this capital can be deployed remain. Power and memory constraints are key challenges. Many countries already face significant limitations on grid capacity, which impacts the ability to power new datacenters. The author notes that while some datacenters are using gas turbines to mitigate immediate power issues, the bigger problem lies in the supply of DRAM. OpenAI has reportedly purchased a substantial portion of the global DRAM supply. Current estimates suggest the available DRAM can only support the rollout of 15 gigawatts of AI infrastructure, which restricts the number of high-demand users that can be effectively served. The implications of these constraints are serious. If demand for AI services continues to outstrip supply, prices for compute resources are likely to rise. This scenario could hinder the growth of AI applications, especially for users who require large amounts of memory for their workflows. The author suggests that without adequate memory and power, AI companies may struggle to keep pace with the increasing demand for computational resources, complicating the landscape for future AI development.