HWBRIEF
The smaller AI 'models' handling tasks like offline translation, local motion filtering, and real time sign reading increasingly run on devices themselves, cutting latency and keeping data on device.
The AI you already use (the phone translating a menu offline, the doorbell filtering motion events locally, the dashcam reading a sign in real time) is increasingly not in the cloud. Heavy training still happens in remote data centers, but the model you actually talk to is moving onto the device, and one industry-cited estimate puts that shift at 100 billion models running locally today.
That architectural split is the subject of a Semiconductor Engineering analysis published Tuesday by Ed Sperling. The piece frames edge AI not as a replacement for cloud computing but as a relocation: data centers still train the largest language models on fleets of GPUs, while phones, cars, sensors, and appliances run smaller, more targeted models on their own silicon.
Nigel Drego, CTO and co-founder of Quadric, told the publication that the two sides are being co-optimized rather than swapped. Running inference on-device, he said, cuts response time and keeps user data from leaving the device, which shows up as faster replies and quieter privacy trade-offs.
The 100-billion-models figure is the load-bearing claim in the piece, attributed inside the interview to unspecified "recent estimates." It has not been independently verified in the visible source, and the analysis rests on a single vendor CTO voice. What is solid is the architectural shape: training in the cloud, inference on the device, and a middle layer of co-optimized systems the rest of the industry is now building toward.