Toward Flash-Native LLM Appliances: Measured Feasibility of Serving a Hybrid-Attention MoE at 100 tok/s on a Sub-500 GB/s Memory System

Consumer access to large language models is gated by the price of high-bandwidth GPU memory. I investigate whether a purpose-built appliance — a "player" holding a modest fast-memory tier and an accelerator, plus a cheap flashable NAND "disk" holding a mastered model image — can serve a ~120B-parameter mixture-of-experts model at 100 aggregate tokens/s across a 512K-token aggregate context budget. Measurements on Qwen3.6-35B-A3B (256 experts top-8 plus one shared, 3:1 linear-to-full hybrid attention, 1-layer MTP head) as an architectural proxy show: (1) expert routing is measurably concentrated (mean decode entropy 7.44 vs 8.00 uniform bits) and transition-predictable (+17–20 points over a static-popularity baseline), with an LRU expert cache reaching 93.7% hits at 75% residency (~22 MB of flash traffic per output token); (2) multi-token-prediction speculative acceptance is 90/82/66/45% at draft widths 1/2/4/8 and is invariant to context length, concurrency, GPU generation, and bf16-vs-FP8 weight format; (3) a same-format GPU ladder spanning 3.49× memory bandwidth shows per-request decode scaling at 90% of the bandwidth ratio in low-concurrency long-context modes. Calibrating an analytical bandwidth model with these measurements, the target contract fits within 350–500 GB/s of fast-memory bandwidth and 48–64 GB of capacity — LPDDR-class, not HBM-class, hardware. Total measurement cost was $8.01 of rented GPU time. All harnesses, routing traces, and raw results are bundled for reproduction.