THE BOTTLENECK IS ELECTRICITY
Frontier AI training is no longer GPU-limited — it is power-limited. A single H100 GPU draws ~700 watts; a 100,000-GPU cluster needs ~70 megawatts continuous, plus near-equal cooling load. The constraint that decides who trains the next model is which company can secure firm power on a multi-year horizon.
WHY MEMPHIS
The Tennessee Valley Authority operates one of the cheapest, most reliable grids in North America, with surplus baseload from coal and nuclear plants built for industries that have since shrunk. Memphis specifically sits near a TVA substation network rated for hundreds of megawatts — rare in a country where most new data-center sites wait 4–7 years for grid interconnection.
THE INTERCONNECTION QUEUE
PJM, the largest US grid operator, has roughly 250 GW of projects waiting for grid connection — more capacity than the entire installed base of many countries. Average wait is now over five years. Buying an existing site with live power, as SpaceX did in Memphis, skips the queue entirely. This is why data-center M&A increasingly trades on megawatts-available, not square footage.
THE HYPERSCALER RANK
Frontier AI compute is concentrating in a handful of campuses. Microsoft, Google, Amazon, Meta, and now xAI/SpaceX and Anthropic-aligned facilities are racing for gigawatt-scale sites. Compute capacity has become the new oil reserve — a strategic asset disclosed in earnings calls.
SPACE-BASED COMPUTE
Orbital data centers solve two problems at once: continuous solar power without atmospheric attenuation, and radiative cooling into the 3-kelvin background of space. The constraint is launch cost — historically prohibitive, but Starship's projected cost-per-kilogram makes the math plausible for the first time. The thermal engineering remains unsolved at gigawatt scale.
THE VALUATION ARC
Anthropic's reported $900B raise would put it ahead of every public company except a handful of mega-caps, despite negative cash flow and revenue still measured in single-digit billions. The bet investors are making is not on current earnings but on compute access — capital today buys power contracts and chip allocations that competitors cannot replicate later.