Introduction

Can we build AI like a factory, 1 gigawatt a week? This piece maps the costs, power, labor, and who might win the race.

What's the problem it solves?

It explains what it would take to build huge AI data centers fast and at scale. It shows the real limits like money, parts, power, and people.

Quick Summary

The authors ask if Sam Altman’s dream of a gigawatt per week of AI capacity can happen. They find chip demand is so strong that even the giant cost of new fabs looks small next to what buyers will pay. The bigger choke points may be upstream parts for data centers, like turbines, transformers, copper, and switchgear, plus the skilled workers to build and run everything.

Power and time beat everything. Chips are pricey and get replaced every three years, so you cannot afford empty buildings. Energy sources with short lead times win, which makes natural gas attractive now. Solar is cheap per panel but needs lots of land, extra capacity, and batteries to keep chips on 24 or 7. Some data centers may go off grid to skip long grid hookups. The future might split between many 100 MW sites that soak up spare grid power or a few mega sites at 1 to 10 GW that stamp out modular halls like factory parts.

The authors also look at money and geopolitics. If growth keeps up, $400B a year in US AI spending could be matched by revenue later, and $400B plus ARR by decade end seems possible. China could gain in a long timeline world since it leads in many non chip parts and builds power fast. The piece closes with two paths: AI winter with a slowdown, or AI boom with factory scale buildouts and possibly 1 GW per week by the mid 2030s.

Key Takeaways

• Chips are not the only limit. Parts, factories for those parts, and skilled labor may be bigger bottlenecks.
• Speed matters more than perfect efficiency. Short lead time power beats low cost power that arrives late.
• Natural gas is the fastest bridge for new power. Nuclear is clean but too slow right now.
• Off grid data centers can dodge years long grid delays.
• The build style may shift to prefab modules to cut on site time.
• China has an edge on long projects that need lots of hardware and power.
• Two futures are plausible. A cool down, or an acceleration where 1 GW per week arrives by the 2030s.

What to do

• Plan for parts and labor early. Secure turbines, transformers, cables, and crews before breaking ground.
• Favor fast power. Put gas turbines on site while longer term sources mature.
• Go modular. Use factory built racks, cooling, and power blocks to cut build time.
• Hedge the grid risk. Explore off grid designs or flexible curtailment deals to speed interconnects.
• Design for churn. Expect to swap chips every three years and keep shells useful for decades.
• Run scenarios. Model AI winter vs AI boom so you do not overbuild or fall behind.
• Watch China supply chains. Map where you rely on Chinese made power gear and plan backups.

‍