Cristiano Amon: AI labs want 100B-parameter models on phones by 2028 - The Next Web

A 100B-parameter model on a phone is not an iteration—it is a category collapse. Even at 4-bit quantization, weights demand roughly 50GB plus KV-cache overhead, while current flagships address only 12-16GB. This gap forces HBM-class stacking into consumer silicon, dragging datacenter architecture into the pocket. Chain reaction: Qualcomm is effectively pressuring TSMC (Taiwan, China) to accelerate 2nm advanced packaging for mobile, while Samsung and SK Hynix must pull LPDDR6 mass production forward. NPUs will claim 40%+ of die area; sustained thermal design power jumps from 15W to 40W+. The phone form factor is under structural stress. Supply-chain risk: US export controls on advanced packaging, combined with EU AI Act transparency mandates for on-device inference, create dual regulatory friction. Memory remains a Korean oligopoly; any geopolitical disruption cascades directly into mobile AI timelines. Competitive read: Apple will likely leapfrog via its unified-memory A-series. MediaTek and Exynos are bandwidth-constrained. The wildcard: if NVIDIA licenses mobile AI IP, the SoC value chain fragments. 12-24 month call: 32-64GB RAM flagships ship in 2025-26; the "AI phone" locks into the $1,500+ tier. The 2028 100B target is supply-chain anchoring—realistic on-device deployment lands at 30-50B with speculative decoding. The phone bifurcates into a compute appliance and a communication device.