Edge AI coverage across on-device models, wearables, embedded systems, private AI, phones and PCs, and the hardware needed to run AI closer to users.
Stories
567
Latest source update
July 15, 2026
Coverage
Live
Topic brief
What to know about Edge AI
Brief updated Jul 10, 2026
Edge AI refers to running machine learning inference, and increasingly small language models, directly on devices such as phones, wearables, cameras, vehicles, robots, and industrial gateways, instead of sending data to cloud data centers. The appeal is a mix of lower latency, offline capability, predictable cost, and privacy, because raw sensor data like audio, video, and biometrics can be processed locally and never leave the device.
For practitioners, edge AI is a systems discipline that spans model compression and quantization, specialized silicon such as NPUs, custom ASICs, and memory-on-package designs, and the toolchains that map models onto constrained hardware. Techniques like int8 quantization, distillation, and sub-2B-parameter models make it possible to run useful assistants, speech recognition, and vision on CPUs, mobile SoCs, and single-board computers. The trade-offs, including accuracy versus power draw, memory bandwidth, and thermal limits, are different from cloud ML and reward careful benchmarking.
The business stakes are large because edge AI touches consumer hardware cycles across phones, watches, glasses, and PCs, as well as automotive and robotics autonomy, defense, and industrial IoT. Whoever controls the on-device stack, meaning the silicon, the OS-level model, and the developer APIs, captures durable platform advantage, which is why Apple, Samsung, Google, Meta, Amazon, and a wave of chip startups are all investing heavily.
What changed recently
The clearest through-line right now is the shift from on-device inference as a feature to on-device agents as a product. Counterpoint Research is framing the next smartphone generation as Agent Phones that interpret intent and execute multi-step tasks, Samsung is building its July 22 Galaxy Unpacked around agentic AI and new foldables, and Apple is both widening Siri AI's reach into third-party apps in the iOS 27 beta and reportedly eyeing PrismML compression to run larger models locally. Underneath the assistants, the model layer keeps getting smaller and better: OpenBMB's MiniCPM5-1B targets 1B-class on-device reasoning and a Parakeet speech model beat Whisper on a cheap 2-vCPU box, showing that capable reasoning and ASR now fit on modest hardware.
The hardware and money side is moving in parallel. Syntiant filed for a Nasdaq IPO focused on edge-AI chips, Broadcom extended its custom Apple ASIC deal through 2031, AMD shipped a Ryzen AI Halo local workstation and introduced Versal Gen 2 memory-on-package parts, and Ceva licensed its NPU IP to a major US platform company. Wearables and physical AI are the other fast-moving front, with Meta's smart glasses reportedly containing unreleased NameTag facial-recognition code, edge-AI smartwatches reaching a quarter of global shipments with Apple taking roughly 90 percent of that segment, and autonomy arriving in trucks via Hanjin, maritime vessels via Hanwha, and Army drone swarms via Palladyne. Privacy tension runs through all of it, from Apple's consent prompts when requests route to Google Cloud to the facial-recognition debate around glasses.
What to watch
Near-term signals to track from this batch: Samsung's Galaxy Unpacked on July 22, 2026 is expected to detail its foldables and agentic-AI strategy; Apple's Siri AI remains blocked from the EU pending the Tim Cook and EU discussions over Digital Markets Act obligations, and its third-party Siri access and reported PrismML on-device LLM work are still in beta or exploratory rather than shipped; Syntiant's IPO under ticker SYTN is filed but has no share count or price range yet; Snap's Specs AR glasses are slated to ship this fall in the US, UK, and France; and Meta's unreleased NameTag facial-recognition code has not shipped, leaving its rollout and the surrounding privacy questions open.
Frequently asked questions
What is edge AI and how is it different from cloud AI?+
Edge AI runs model inference directly on a device such as a phone, watch, camera, vehicle, robot, or industrial box, instead of sending data to a remote data center. That lowers latency, works offline, and keeps raw data local for privacy, at the cost of tighter compute, memory, and power budgets than cloud GPUs.
Can real language models actually run on-device today?+
Yes, within limits. Small models such as OpenBMB's MiniCPM5-1B target 1B-class on-device reasoning, and Apple is reportedly exploring PrismML compression to run larger models locally. Sub-4B models on single-board computers like the Raspberry Pi 5 and on mobile NPUs are practical for assistants, speech, and vision, while frontier-scale reasoning still favors the cloud.
Which companies lead the on-device AI stack?+
Apple dominates edge-AI smartwatch shipments and is rebuilding Siri and Spotlight around on-device intelligence, while Samsung, Meta, Google, and Amazon push phones, glasses, and custom silicon. Chip suppliers including AMD, Broadcom, Nvidia, Ceva, Ambarella, and Syntiant provide the NPUs and ASICs underneath.
What are Agent Phones?+
Per Counterpoint Research, Agent Phones are the next step beyond AI phones: devices whose on-device assistant can interpret user intent and carry out multi-step tasks across apps automatically, rather than just answering single prompts. Samsung and Apple are both moving toward this agentic, cross-app model.
Why does edge AI matter for privacy?+
Because processing happens locally, sensitive audio, video, and biometric data can stay on the device. But edge hardware also enables sensitive capabilities like facial recognition in glasses, and some features still route prompts to the cloud, so local processing is a tool rather than an automatic guarantee. Apple now shows a consent prompt when AI prompts are sent to Google Cloud.
What hardware do I need to build an edge AI project?+
It ranges widely. Hobbyists run offline assistants on a Raspberry Pi 5 or Rockchip boards with small NPUs, developers can use appliances like AMD's Ryzen AI Halo workstation or Nvidia Jetson Orin boxes, and industrial deployments use ruggedized boxes with dedicated NPUs delivering tens to hundreds of TOPS.
