Qualcomm's AI Play: Edge Inference, 6G, and the Physical AI Bet

In discussions about AI chips, NVIDIA and AMD capture most of the attention. But one company is pursuing a fundamentally different AI strategy — and it’s increasingly worth engineers paying attention to. That’s Qualcomm.

Qualcomm isn’t trying to compete with NVIDIA head-on in data center model training. Its bet is in a different direction: running AI on the phone in your pocket, the PC on your desk, the car on the road, and the robots soon to be mass-produced.

TL;DR

Qualcomm’s AI strategy is centered on edge inference, not cloud training. The Snapdragon X Elite’s NPU lets 7B-parameter models run locally on a laptop. 6G’s core design goals include low-latency AI inference loops. Physical AI is the vision of integrating perception, decision-making, and actuation onto a single chip. All three point in the same direction: AI that runs in real time on-device without requiring a network connection.

The Business Logic of Edge AI

Why run AI on the device rather than in the cloud? There are clear reasons:

Latency: Having a self-driving car’s decision system query cloud AI for every decision is infeasible. On-device inference can bring latency from hundreds of milliseconds down to single digits.

Privacy: More users and enterprises don’t want to send data to the cloud. On-device AI makes “data never leaves the device” possible.

Cost: API call costs at scale are substantial. The marginal cost of running a local model approaches zero.

Offline capability: Remote areas, underground, on aircraft — many scenarios lack stable network connectivity.

Qualcomm’s Snapdragon X Elite integrates CPU, GPU, and NPU (Neural Processing Unit), optimized for on-device LLM inference. According to Qualcomm, it can run a quantized version of Llama 3 70B at approximately 30 tokens/second on a Windows laptop.

Snapdragon’s AI Architecture

Snapdragon X Elite’s NPU is called the Hexagon NPU, delivering 45 TOPS (Tera Operations Per Second) of AI compute. This figure became the marketing benchmark for “AI PCs” in 2024 — Microsoft requires Copilot+ PCs to have at least 40 TOPS NPU capability.

graph LR
    A[Input Data] --> B[Hexagon NPU]
    A --> C[Adreno GPU]
    B --> D[AI Inference Results]
    C --> D
    D --> E[Application]
    F[Kryo CPU] --> E

Qualcomm is also advancing Qualcomm AI Hub on Snapdragon, letting developers deploy models from Hugging Face directly to Snapdragon devices without manual model optimization.

6G: AI-Native Communications

6G isn’t just “faster 5G.” Qualcomm emphasizes several AI-native design goals in 6G standards work:

AI-assisted radio resource management: Base stations can use AI to predict user movement trajectories and proactively allocate bandwidth and handoff — rather than reacting after signal degrades.

Sensing integration: 6G base stations can simultaneously handle communication and environmental sensing, using radio signals to build a 3D map of surroundings. This matters for autonomous vehicles and robot localization.

Ultra-low latency: Target end-to-end latency below 1ms, making remote AI inference assistance for local decision-making viable.

6G is expected to see commercial deployment around 2030. Qualcomm’s patent portfolio from 5G gives it significant leverage in 6G standard-setting.

Physical AI: From Perception to Action

Physical AI means AI directly controlling physical-world machines — robots, drones, industrial equipment. Qualcomm’s entry point is providing the edge compute platform these devices need.

Traditional robot control systems and perception systems are often separate: camera images go somewhere for processing, results come back to control motors. This loop’s latency and power consumption are serious problems for mass-market hardware.

Qualcomm’s Robotics RB platform series (based on Snapdragon) integrates:

• Multi-camera inputs
• Computer vision NPU
• Low-latency CPU cores for real-time control
• 5G connectivity

Boston Dynamics and multiple Chinese humanoid robot manufacturers are using or evaluating Snapdragon-series chips as the compute core for their robots.

Qualcomm vs. NVIDIA

Dimension	NVIDIA	Qualcomm
Main battleground	Data center training and inference	Edge device inference
Core product	H100/H200 GPU + CUDA	Snapdragon SoC
Power consumption	300–700W	5–20W
Price	$10,000–$40,000/chip	$50–$200 (SoC at volume)
AI software ecosystem	CUDA near-monopoly	Weaker, actively building

These two aren’t direct competitors — they serve different markets. In edge AI, Qualcomm’s closer competitors are Apple (A18 Pro, M-series) and MediaTek.

Summary

Qualcomm’s AI strategy has a clear logic: use AI to strengthen its existing core capabilities (low-power SoC design and wireless communications), rather than attacking NVIDIA’s data center market. 6G and Physical AI are extensions of this logic, not buzzword-chasing.

For engineers, Snapdragon’s edge AI capabilities are now at a level worth serious consideration. If you’re building applications that need on-device AI inference — privacy protection, low latency, offline capability — Qualcomm AI Hub and Snapdragon’s NPU are options you can actually test today.

References

• Qualcomm Snapdragon X Elite specs
• Qualcomm AI Hub
• 6G vision whitepaper - Qualcomm
• Qualcomm Robotics RB platform