A fundamental shift from the TCP paradigm
The buffer-sizing decision for a 16,000-port AI fabric came down to a whiteboard argument between two engineers. The first wanted to size for a p99.9 microburst at 256-fan-in incast, plus 30% headroom. The second wanted to size for p99.99, plus 50%. Both were guessing, and both knew it.
The queueing model on the whiteboard had eleven parameters and three inputs nobody could measure precisely. The number that ended up in the BOM was 144 MB per chip. It was right by accident. Most of these decisions are.