Why is compute infrastructure so central to the AI race, and who currently controls it?

Compute infrastructure is central to the AI race due to its dominance in costs and operational demands, accounting for over 50% of expenses at major AI firms like Anthropic and others, making AI development highly capital-intensive [2]. The shift from software competition to controlling vertically integrated infrastructure stacks by large AI labs underscores this, as scarcity in data centers, chips, and power creates severe bottlenecks, with build times of 18-36 months and power grid waits of 3-5 years limiting scalability [1][6]. Additionally, power constraints, not just compute, are a key issue, with inferencing costs up to 15 times higher than training and expected to consume 75% of AI compute by 2030, while lack of compute makes AI expensive and restricts it to high-value tasks [8][9]. Base resources like training clusters, high-end GPUs, electricity, and cooling are concentrated in a handful of "super-platforms," where big software companies have become hardware players with heavy capital expenditures [10]. Control of compute infrastructure is held by a limited set of hyperscalers and large AI labs, with examples like Anthropic leveraging multi-provider strategies across Google TPUs, AWS Trainium2, and Nvidia GPUs for cost and iteration advantages [12]. These super-platforms manage the core resources amid skyrocketing data center costs and demand from sectors pulling in robotics and materials [6][10][11].