The hook
1/10
AI supercomputers may need 9 GW by 2030
The largest frontier training campuses are projected to require up to two million specialized chips and nearly $200 billion in hardware—power equal to multiple large cities.
The largest frontier training campuses are projected to require up to two million specialized chips and nearly $200 billion in hardware—power equal to multiple large cities.
Generative AI has left ordinary cloud data centers behind. Training compute requirements have grown 4.1 times per year since 2010, producing a new class of heavy infrastructure with extreme power, cooling and networking demands.
A traditional enterprise rack draws roughly 10 kW. AI-focused data centers routinely demand 60 kW per rack, and next-generation designs are pushing toward 120 kW to 140 kW, making power and thermal management central constraints.
OpenAI, Microsoft, SoftBank, Oracle and MGX aim to invest up to $500 billion over four years. The Abilene, Texas campus currently draws 421 MW and is projected to reach 1.2 GW with over 450,000 GB200 GPUs at a $15.9 billion hardware cost.
xAI converted a former Electrolux factory into a 100,000-GPU cluster at record speed. Colossus 2 now draws 946 MW with about 1.1 million H100-equivalents and a target of 2 GW, supported by NVIDIA, Dell and Supermicro.
In 2024, U.S. data centers drew from natural gas, renewables, nuclear and coal. Grid strain triggered Virginia electricity price hikes up to 76%, while Microsoft, Google and Amazon pivot toward nuclear and small modular reactors.
Roughly 70% of global computer memory production in fiscal 2026 was purchased exclusively for AI data centers. The result is a severe High Bandwidth Memory shortage and intensifying competition across the semiconductor supply chain.
Extreme rack densities make air cooling obsolete. Direct-to-chip cooling holds about 47% of the liquid cooling segment, while silicon photonics is projected to jump from $2.49 billion in 2025 to $34.3 billion by 2035.
Local opposition has blocked about $130 billion in AI data center projects. xAI faces Clean Air Act lawsuits over gas turbines, and silent data corruption now occurs in roughly one per thousand devices—so huge clusters can harbor hundreds of silent math errors.
India’s IndiaAI Mission is spending ₹10,372 crore to build domestic GPU capacity, targeting 18,000–65,000 GPUs. The entities that control energy, optical manufacturing and sovereign silicon may shape the digital economy. Read the full analysis.
Read the full analysis →