Qualcomm sees AI experimentation ahead

Qualcomm's pivot to agentic AI feels like the right bet, but I'm skeptical about the timeline. Amon's framing suggests on-device reasoning will explode within a generation, yet current NPU architectures still struggle with basic context windows. We saw this with the Snapdragon X Elite-great benchmarks, but real-world LLM inference lagged behind cloud-based models by a noticeable margin. The real question isn't whether chips get faster; it's whether the energy-per-token ratio can drop below a threshold where users stop noticing battery drain. I've been running local models on a mid-range ARM laptop, and the thermal throttling kicks in after 90 seconds of sustained generation. That's a hard wall for agentic loops that need to persist across hours. Amon's experimentation comment echoes what we saw in the edge AI boom of 2023-lots of prototypes, few mass-market hits. The difference now is that multimodal agents demand simultaneous vision, audio, and text processing, which forces a complete rethinking of memory hierarchy. Is anyone else watching the SRAM vs. HBM trade-off for mobile? That'll determine if 2027 actually delivers on the promise.