100B phone LLM demand hits a RAM wall
The memory wall is the story. A 4-bit quantized 100B LLM needs 50GB of RAM on a phone, and the brief has no DRAM cost, no supply quantity, no thermal measurement, and no inference speed.
| 100B LLM | |
|---|---|
| parameters | 100B |
| target year | 2028 |
| 4-bit RAM | 50GB |
| 2-bit RAM | 25GB |
| current Android memory | 24GB LPDDR5X |
| RAM threshold | 16GB |
| distilled alternative | 20B to 30B |
| 1-bit variants | 1-bit |
Even a 2-bit variant needs 25GB for weights alone, and current Android devices sit at 24GB LPDDR5X while makers steer clear of higher RAM variants.
The brief reads the constraint as memory, not silicon ambition. The 2028 claim is a buyer demand, not a shipping plan; the unnamed AI companies are constrained by DRAM, and Qualcomm is constrained by the absence of a named roadmap.
The realistic path is a distilled 20B to 30B model or MoE offload, not a dense 100B resident in DRAM.
Still unmeasured: the actual phone model, the DRAM price, the heat envelope, and the flash-offload latency.
DRAM is the binding constraint, not compute. The 2028 statement is a demand, not a shipping plan.
After @firesidealpha. We did not report this. The pictures, if any, are theirs.
