Two new system-on-chip for the top of the Android range, and a message that goes beyond the benchmark race. Qualcomm has introduced the Snapdragon 8 Elite Gen 6 and Elite Extreme Gen 6, featuring Oryon CPU cores, GPU matrix cores and LPDDR6 support. The company has not yet released complete data on memory and power consumption, but the combination already says a lot: the matrix multiplication work typical of LLMs is being moved into the SoC GPU, while the new-generation memory is meant to support local inference workloads without going through the cloud.

The thesis to follow is not "more power". It is that the boundary of data sovereignty is shifting into users' pockets. A phone capable of running quantized LLMs locally reduces dependence on round trips to remote servers. For anyone evaluating on-premise or edge deployments, the point is not just speed: it is the ability to keep sensitive data on the device, with smaller exposure surfaces and, at least in theory, a different TCO from a cloud API for repetitive tasks. The analytical frameworks that AI-RADAR proposes on /llm-onpremise are there precisely to distinguish these trade-offs, avoiding reducing everything to the spec sheet.

The real sticking point remains memory. LPDDR6 improves bandwidth, but the announcement does not resolve the maximum capacity constraint: how large can a model be if it must stay resident without clogging the system? In the absence of numbers, the critical variable becomes quantization. A local FP16 model requires far more memory than one compressed to 4 or 8 bits; quality degrades, but the privacy and latency advantage can remain. The Oryon cores and GPU matrix cores are not enough on their own: drivers and runtimes are also needed to avoid bottlenecks between CPU, GPU and NPU.

Structurally, the message for the market is clear. Apple and MediaTek can no longer treat on-device AI as a marginal feature. Developers designing apps for privacy, air-gapped environments or enterprise devices may gain more room to run LLMs locally. The losers are cloud services for short, repetitive inference jobs, which risk becoming a superfluous cost when the device can handle the task itself. The game is no longer played only on smartphones: it is played at the point where data stops leaving the device.