Dylan Patel says OpenAI and Anthropic are taking a rapidly growing share of new AI infrastructureAI compute infrastructure is the hardware, facilities, networks, power, cooling, storage, and software used to train and run AI models.. He expects the two laboratories to consume roughly half of incremental computeAI compute capacity is the available ability of hardware and supporting systems to perform AI training or inference work over time. by the end of 2027, with their newer hardware delivering more useful processing per wattEnergy-efficient AI inference produces useful model outputs while minimizing the electricity consumed by computation, memory, networking, and cooling. than the older global stock.
The economic engine is a widening gap between the cost of infrastructure and the revenue frontier models can produce. As inference margins rise, the laboratories can fund more training and internal research, while cloud providers, chipmakers and independent infrastructure owners raise prices to capture part of the value.
Dwarkesh Patel and Dylan Patel examine constraints that could slow this trajectory. Semiconductor equipment, power generation, data centers and credit markets cannot expand instantlyAn AI infrastructure bottleneck is a constrained component that limits how quickly or efficiently AI compute capacity can expand., while higher borrowing costs and political resistance could redirect capital away from AI or limit where new capacity is built.
The discussion ends with a governance problem. Economies of scale, learning from deployment and the ability of advanced models to improve later systems all reward the leading laboratory. Without a credible decentralizing force, a small number of companies could control both the largest compute fleetsAI compute concentration occurs when a small number of organizations control a large share of the infrastructure capable of training or serving advanced models. and a rapidly growing population of capable AI workers.
Watch on YouTube
