Gpu

Frontier model training has quietly shifted from a chip problem to a power-and-capital problem. A grounded look at the four scarce inputs (compute, power, data, talent) and who should (and shouldn't) build from scratch.

A systems-level deep dive into vLLM: PagedAttention, the scheduler, KV cache management, continuous batching, tensor parallelism, speculative decoding, and the V1 architecture.