This is an engineering roadmap post, and the substance is four concrete data-plane changes, mostly Linux/Android. One: wireguard-go used to copy every decrypted packet into its own 64 KiB buffer even though most packets are ~1 KiB; now packets are sliced in place out of the single large GRO read, for a claimed ~5% speedup, plus shortened queues since most queue depth went unused. Two: subnet routers, app connectors, and exit nodes move from a single ordered single-threaded pipeline to parallel lanes scaled to machine resources, with ordering preserved by keeping each stream pinned to one lane. Three: writev, passing packet fragments to the kernel in one call without copy-and-combine. Four: netmap caching — clients persist the network map to disk and can establish peer connections before reaching the control plane, with warm starts claimed 1–2 orders of magnitude faster on poorly connected tailnets; caveat that it needs a prior successful connection and persistent disk, so SD-card devices and very large tailnets may want it off. Buffer changes and default-on netmap caching land in v1.104, multi-queue and remaining throughput work after. They close by soliciting input on Tailscale-native performance tooling (DERP-vs-direct visibility, QUIC/HTTP/3 support). Open question the post leaves: what happens to sessions established from a stale cached netmap once control is finally reachable and policy has changed?