AI Compute
Buying the chips is not the same as running them: without grid connections and power, GPUs sit in racks waiting for electricity. The bottleneck is moving from process nodes and VRAM to substations and invoices.
The compute bottleneck has shifted from chips to power: grid-connection queues, transformer lead times and electricity contracts are delaying data centers by years. On the supply side, serverless platforms like Modal bill GPUs by the second, inference, fine-tuning and batch sharing one abstraction. On the demand side, invisible reasoning tokens often eat the budget: a few hundred words can burn tens of thousands of thinking tokens.
