Frontier training runs are gated by physical supply, not ideas. Leading-edge capacity is booked years out, high-bandwidth memory is sold before it is made, and grid interconnection queues run for years. Anything on that critical path collects the toll.
The case against
Every capex supercycle in semiconductors has ended in a glut. Efficiency gains, smaller models and custom silicon all cut the accelerators needed per unit of intelligence, and buyers are a handful of firms who can pause in one quarter.
What the weights are worth
Percentages, converted. Nothing is sent anywhere.
Enter an amount to see what each weight is worth. It stays in this browser.
This book has an address
Anyone who opens it rebuilds the same book, weight for weight.