Keep pulling the thread on Leopold Aschenbrenner.
By the end of the decade, individual AI training clusters are projected to cost over $1 trillion and require power equivalent to more than 20% of US electricity production.
Nvidia's datacenter sales increased from approximately $14 billion annualized to about $90 billion annualized in the last year.
Microsoft and OpenAI are rumored to be developing a $100 billion AI cluster planned for 2028.
Power availability is currently the primary binding constraint for building the largest AI training clusters, more so than capital.
By 2028, total AI compute demand will require approximately 20% of US electricity production and consume nearly 100% of current leading-edge TSMC wafer production capacity.
OpenAI's revenue run rate doubled approximately every 6 months, growing from $1 billion in August 2023 to $2 billion in February 2024.
Chip-on-wafer-on-substrate (CoWoS) advanced packaging and HBM memory are the key current bottlenecks for scaling AI GPU production.
Building AGI data centers in the Middle East creates an irreversible national security risk for the United States.
Nvidia's revenue is projected to exceed $200 billion in calendar year 2025.
A sufficient level of compute for achieving AGI will likely only require a cluster costing around $100 billion or less.
Many trillions of dollars will be invested in GPU, datacenter, and power infrastructure for AI before the end of the decade.
The industrial mobilization for AI will require increasing US electricity production by tens of percent.