Trillion-parameter AI stopped being a memory question

Trillion-parameter AI inference has stopped being a memory question. It is now a residency policy: what to keep resident in RAM, what to stream from disk, what to warm into a bounded cache. The engineering constraint is no longer total memory size — it is what stays hot on NVMe, what cycles through the cache, and what stays resident in DRAM. SQLiteAI's WASTE engine and Kimi K3 (2.78 trillion parameters) make that boundary visible. A 64 GB MacBook Pro runs the full model by holding the trunk resident, streaming selected experts from a 982 GiB on-disk container, and using the rest of memory as a bounded expert cache. Final logits match a PyTorch reference to within 3.6e-06, so reachable on a laptop does not mean approximate. It does mean slow: about 0.49 to 0.54 tokens per second, roughly 31 seconds for 16 tokens. The slowness is the signal, not the caveat. Kimi K3 is a 3584-wide latent mixture-of-experts with 896 experts per layer and top-16 routing, so only a sliver of the 2.78 trillion parameters activates per token, and the engineering problem is what to keep hot on NVMe versus what to warm into the cache. SQLiteAI's own framing is honest: no bibliography, no comparison table, and an explicit invitation to send a counter-example. The machine did not get smaller. The unit of account got smaller: from parameter to expert, from fits in RAM to stays resident.