$1 In, $100 Out: Does the AI Math Add Up? (ganeshanmalhotra.substack.com)

In a recent discussion with Dylan Patel, a striking claim emerged: for every dollar spent on fabrication equipment for AI, there's a potential return of about $100 in downstream revenue. This assertion is pivotal for the AI/ML community, highlighting a rapid centralization of computing power within major labs. The numbers indicate that AI infrastructure, particularly data centers and servers, has been instrumental in driving U.S. GDP growth, with reports showing a staggering 92% contribution from information-processing investments in the first half of 2025. However, Patel's figures invite scrutiny, particularly regarding the costs associated with data center infrastructure. While he suggests a cost of $10–15 million per megawatt for compute infrastructure, analyses reveal the true costs to be significantly higher, around $38 million per megawatt when accounting for hardware. This adjustment dramatically alters the revenue-to-cost ratio and indicates that the projected revenue figures are aspirational rather than reflective of current realities. Despite these corrections, the underlying message remains: a concentrated number of companies are poised to dominate AI computing power, with frontier labs expected to control a substantial share of global operational AI compute by 2028, underscoring the urgent need for stakeholders to pay attention to these shifting dynamics.