Video: https://www.youtube.com/watch?v=yGKTjvfur4k
Speaker: Shane Greenstein, Martin Marshall Professor of Business Administration, Harvard Business School
Shane Greenstein compared the current AI buildout to the late 1990s internet boom and argued that AI has become an infrastructure story rather than a software one. He traced the path from ImageNet through the scaling wall that pushed both OpenAI and Google DeepMind toward very large data centers, then showed where those facilities are going and why. Siting logic has flipped, from capacity that
tracked city size because compute sat near customers to low density places chosen for electricity and gas access. Prineville, Oregon, population about 10,000, now hosts roughly 180 megawatts of Meta capacity. He closed on the gap that matters most, which is that capital spending is running well ahead of the applications and revenue that would justify it.
Top takeaways
- Roughly half of US GDP growth last year could be attributed to construction tied to data centers, a concentration of investment Greenstein said has only a handful of precedents in US history.
- The economics increasingly favor incumbents, with a 150 megawatt shell at $2 billion to $4 billion, hardware at $5 billion to $15 billion, and annual operating costs near $500 million, with the frontier moving toward gigawatt scale. Chip fabrication, gas generators and data center electricians are all scarce, and suppliers prioritize existing customers, which is a harder barrier to entry than new firms faced in the dot-com period.
- Diffusion is far faster than the internet’s but the payoff is unproven. Netscape reached about 100 million users by 1998, while OpenAI reached about a billion in a little over three years. The applications that would justify today’s spending have mostly not been built yet.
