Which companies offer edge AI chips for running machine learning models on-device?

The market for edge AI chips —semiconductors featuring dedicated Neural Processing Units (NPUs) or specialized accelerators designed to run machine learning models locally on-device—is led by several major players spanning mobile, PC, automotive, and industrial Internet of Things (IoT) sectors.

Major Mobile & Consumer Silicon Giants

• Qualcomm: A volume and performance leader in mobile and PC edge AI. Its Snapdragon platforms (including mobile chipsets and Snapdragon X-series PC processors) feature powerful, highly efficient Hexagon NPUs designed to run on-device generative AI and large language models (LLMs) locally.
• Apple: Takes a vertically integrated approach, embedding custom "Neural Engine" cores directly into its A-series (iPhone), M-series (iPad/Mac), and S-series (Apple Watch) chips. This architecture powers on-device machine learning features (under the Apple Intelligence ecosystem) with high performance-per-watt efficiency.
• MediaTek: A dominant force in mobile application processors globally. MediaTek’s Dimensity chipsets incorporate high-performance APUs (AI Processing Units) tailored for localized machine learning, computer vision, and smart-home/wearable implementations.
• Samsung Electronics: Designs Exynos mobile system-on-chip (SoC) platforms featuring integrated NPUs built to handle real-time on-device inferences like voice recognition and computational photography.

PC & Enterprise Edge Providers

• Intel: A major player in the "AI PC" and industrial edge space. Intel's Core Ultra processors feature integrated NPUs capable of high TOPS (Tera Operations Per Second) for local productivity models, while their Movidius and OpenVINO-supported hardware target vision-based edge deployments like smart cameras and retail sensors.
• AMD: Offers Ryzen AI processors with dedicated on-chip XDNA NPUs for local Windows client AI inference, alongside its adaptive computing portfolio (inheriting Xilinx FPGAs and adaptive SoCs) widely deployed in industrial edge robotics and embedded systems.

Robotics, Automotive & Industrial Edge Specialists

• NVIDIA: While famous for data center GPUs, NVIDIA dominates high-end autonomous edge compute via its Jetson platform. Jetson modules bring full CUDA-accelerated machine learning and computer vision to robotics, drones, and smart city infrastructure.
• NXP Semiconductors: Focuses heavily on automotive and industrial IoT edge AI. Following its acquisition of edge AI pioneer Kinara, NXP integrates dedicated NPU tech into its i.MX application processors for localized automotive and industrial decision-making.
• Broadcom: Delivers broadband and networking edge AI solutions—such as home gateway and 50G PON chipsets with integrated NPUs—enabling local inference directly inside telecom and consumer gateway equipment.
• SiMa.ai: A prominent purpose-built startup offering a software-centric Machine Learning System-on-Chip (MLSoC) platform specifically designed to tackle embedded computer vision and edge ML workloads at minimal power consumption.

If you are looking for a chip for a specific use case (such as automotive, robotics, a mobile app, or an industrial IoT sensor), let me know and I can help narrow down which architecture or performance tier (TOPS/watt) fits your project best.

Yes. The edge-AI chip market has two broad groups: large semiconductor companies embedding NPUs/AI accelerators into SoCs, and specialists building dedicated inference chips.

Major companies to know

The specialist field is particularly interesting: current industry maps include Hailo, Kneron, Blaize, Axelera AI, SiMa.ai, MemryX, Rebellions, Mythic, DEEPX and others as dedicated edge-AI silicon vendors.

A few standouts

• NVIDIA is the obvious choice when you need substantial compute and a mature CUDA ecosystem. Its edge strategy now extends beyond Jetson: NVIDIA has announced RTX Spark PCs aimed specifically at local/on-device AI, with the first systems expected in October 2026.
• Qualcomm is particularly strong when the AI accelerator needs to be integrated into a complete, power-efficient SoC for phones, PCs, automotive or embedded devices.
• Ambarella is unusually focused on vision + generative AI at the edge. Its N1 family can run models such as Llama, Gemma and LLaVA without an internet connection, while targeting sub-20-W systems.
• Hailo focuses heavily on dedicated, efficient inference accelerators for embedded devices.
• Axelera AI has the Metis AIPU, including modules consuming roughly 4–8 W and capable of running models locally; its newer Europa architecture scales considerably higher.
• Kneron specializes in NPUs designed from the ground up for on-device inference; its KL1140, for example, was introduced specifically to bring larger AI models onto edge devices.
• MemryX takes a dataflow/at-memory-computing approach and sells MX3 accelerators in M.2 form factors, making them relatively easy to add to an existing ARM/x86 system.

One important distinction: “edge AI chip” can mean very different things. A smartphone NPU doing a 10-million-parameter vision model, a Jetson running a 7B LLM, and a $10 industrial camera doing object detection are all edge AI, but they require very different silicon.

If you're evaluating vendors for a specific product, I can also give you a more useful shortlist by comparing TOPS, power consumption, SRAM/DRAM, supported models (CNN/Transformer/LLM), software stack, price, and availability.

Companies offering edge AI chips (processors/accelerators that run machine learning inference locally on devices rather than sending data to the cloud) span large semiconductor companies and specialized AI-chip startups.

AIMultiple Parse## How the market breaks down

High-performance edge AI

For robots, autonomous systems, and industrial machines:

• NVIDIA Jetson
• AMD Versal AI Edge
• Qualcomm Robotics platforms
• SiMa.ai MLSoC
• Axelera AI Metis

AIMultiple### Low-power AI inference

For cameras, sensors, and battery-powered devices:

• Hailo
• Kneron
• Google Coral Edge TPU
• STMicroelectronics STM32
• Ambiq and BrainChip-class TinyML solutions

Parse EdgeIR### On-device generative AI (small LLMs/VLMs)

For running compact language or vision-language models locally:

• Qualcomm Snapdragon platforms
• Apple Neural Engine
• NVIDIA Jetson
• Hailo-10H
• Intel/AMD AI PC NPUs

AIMultiple## Choosing a vendor usually depends on:

• Model size: tiny neural networks vs. multi-billion-parameter LLMs
• Power budget: milliwatts (sensors) to tens of watts (robots)
• Framework support: TensorFlow Lite, ONNX, PyTorch export, vendor SDKs
• Latency requirements: real-time control vs. batch inference
• Deployment volume: custom ASICs vs. off-the-shelf modules

For embedded products, the most common short list today is often NVIDIA, Qualcomm, AMD/Xilinx, Intel, NXP, Renesas, Hailo, Ambarella, SiMa.ai, and Axelera AI.

The market for Edge AI chips —processors designed specifically to execute trained machine learning models directly on localized hardware like smartphones, smart cameras, and cars—is fiercely competitive.

The primary companies offering Edge AI hardware are categorized below by their market focus and architecture:

🌟 The Big Tech "Titans"

These companies leverage massive developer ecosystems and provide hardware that scales from simple microcontrollers to high-performance autonomous systems.

• NVIDIA: Dominated by the NVIDIA Jetson Platform , including modules like the high-end Jetson AGX Orin (up to 275 TOPS) and the accessible Orin Nano Super for robotics and embedded vision.
• Qualcomm: Offers specialized Neural Processing Units (NPUs) built into their Snapdragon mobile SoCs, as well as discrete edge chips like the Qualcomm Cloud AI 100 and the Robotics RB5 platform , heavily utilized in automotive systems and premium smartphones.
• Intel: Provides the Movidius VPU (Vision Processing Unit) lineup and FPGAs tailored for low-power computer vision, fully supported by Intel's OpenVINO Toolkit for edge model optimization.
• Google: Produces the Google Edge TPU (Tensor Processing Unit), sold via their Coral ecosystem. These chips are purpose-built to efficiently execute quantized TensorFlow Lite models using minimal power. [1][2]
• Apple: Integrates its proprietary Apple Neural Engine (ANE) into all custom Apple Silicon chips (A-series and M-series) to execute on-device AI natively via their Core ML framework.

🚗 Specialized Silicon & Automotive Leaders

• NXP Semiconductors: Prominent in industrial and automotive edge AI via their i.MX application processors integrated with dedicated NPUs, managed through the unified NXP eIQ AI Software Stack.
• Renesas: Offers the RZ/V series microprocessors featuring their proprietary DRP-AI (Dynamic Reconfigurable Processor) accelerator, specialized for low-power real-time vision control.
• Ambarella: Known for high-efficiency AI vision SoCs (like the CV5 series ), widely embedded in multi-camera automotive safety systems and security cameras.
• MediaTek: Produces Dimensity SoCs embedded with APUs (AI Processing Units) targeting flagship smartphones, wearables, and edge AI IoT equipment.

🚀 pure-Play Edge AI Startups & Innovators

A robust roster of dedicated semiconductor companies focus strictly on delivering maximum TOPS-per-watt metrics for localized intelligence.

• Hailo: A major player producing chips like the Hailo-8 and Hailo-10H , offering remarkable performance (up to 40 TOPS) for running local Generative AI and LLMs inside compact edge devices and USB accelerators.
• SiMa.ai: Developed the MLSoC (Machine Learning System on Chip), delivering over 50 TOPS at less than 5 Watts for embedded vision pipelines.
• Axelera: Creators of the Metis AI platform , leveraging digital in-memory computing to achieve up to 214 TOPS for high-throughput vision applications.
• EdgeCortix: Offers the SAKURA-II co-processor, specialized for handling high-volume computer vision and generative AI tasks at the edge with ultra-low latency.
• Kneron: Known for low-power neural processors like the KL730 , built primarily for smart home appliances, IoT sensors, and lightweight vision tasks. [1][2]

If you are trying to pick a chip for a specific project, let me know:

• What type of model you need to run (e.g., Object detection, LLM, or voice recognition)
• Your power budget (e.g., Battery-powered, USB, or wall power)
• The required target frames per second (FPS) or response latency [1][2]

I can help you map out the best hardware options for your requirements!

Yes. The edge-AI chip market has several distinct groups: mobile/PC SoCs with integrated NPUs, embedded/industrial AI accelerators, and automotive/robotics processors.

Major companies

A few particularly interesting specialists:

• Hailo focuses specifically on dedicated edge inference. Its Hailo-8 provides up to 26 TOPS and is designed as a low-power co-processor for edge platforms, including vision transformers and LLM workloads.
• Axelera AI builds dedicated AIPUs around its Metis architecture. Its current products target embedded and edge inference, with versions ranging from constrained M.2 modules to PCIe accelerators.
• Qualcomm is one of the broadest players: its Hexagon NPU is integrated into the Qualcomm AI Engine alongside CPU, GPU and sensing hardware, covering smartphones through PCs, robotics, IoT and automotive.
• MediaTek integrates dedicated NPUs into its SoCs and targets smartphones, tablets, AI+IoT, TVs and networking equipment.
• NXP is particularly relevant if you're building an embedded product rather than a consumer device; its eIQ platform combines ML software with MCUs and application processors for industrial, automotive and robotics applications.

If you're choosing a chip

The shortlist changes substantially depending on what you're building:

• Tiny battery-powered sensor: Ambiq, NXP, STMicroelectronics
• Smart camera / computer vision: Hailo, Axelera AI, NVIDIA Jetson
• Robot: NVIDIA Jetson, Qualcomm, NXP, Hailo
• AI PC: Qualcomm, AMD, Intel, Apple
• Smartphone: Qualcomm, MediaTek, Apple
• Automotive: NVIDIA, Qualcomm, NXP, AMD
• Run an LLM locally: Qualcomm, Apple, NVIDIA, AMD, Hailo, Axelera AI

For example, Qualcomm explicitly supports on-device generative AI, while Axelera and Hailo are more interesting when you want a dedicated accelerator attached to an embedded CPU rather than buying an entire AI SoC.

If you tell me your model size (e.g. 1B/7B), power budget, device type, and target price, I can narrow this to the 5–10 best chips and compare TOPS, memory, power, SDK support, and approximate cost.

A diverse ecosystem of semiconductor companies, major tech giants, and specialized startups offer edge AI chips designed to execute machine learning inference locally on-device. These processors balance low power consumption with high throughput for use cases like smart cameras, autonomous driving, robotics, and mobile/IoT devices.

The leading companies offering edge AI hardware span several categories:

Major Tech Giants & Established Semiconductor Leaders

• NVIDIA: Dominates high-performance edge AI with the Jetson family (such as Jetson Orin Nano, AGX Orin, and newer modules), which pack heavy GPU-based parallel processing for robotics, smart cities, and autonomous machines.
• Qualcomm: A heavyweight in mobile, automotive, and IoT edge devices. Their Snapdragon platforms and dedicated IoT chips feature powerful on-board Hexagon NPUs (Neural Processing Units) built for smartphones, wearables, and advanced driver-assistance systems (ADAS).
• Intel: Offers edge solutions including Intel Core/Xeon processors with OpenVINO support, alongside legacy and specialized vision processors like the Movidius Myriad X VPU used in drones and smart cameras.
• Google: Known for the Edge TPU (available via Coral development boards and accelerators), optimized for ultra-low-power, highly efficient machine learning inference at the extreme edge.
• AMD / Xilinx: Provides adaptive computing and FPGA-based solutions through the Kria system-on-modules and Versal adaptive SoCs, allowing customizable hardware acceleration for deterministic latency in industrial and video analytics.

Automotive & Industrial Edge Specialists

• Ambarella: Renowned for their CVflow AI vision processors (e.g., Ambarella CV5, CV3) heavily adopted in automotive ADAS, autonomous driving, and professional security cameras.
• Infineon & NXP: Focus heavily on ultra-low-power microcontrollers (MCUs) and crossover processors (like NXP's i.MX series and Infineon's PSoC/AURIX) enabling tinyML (machine learning on microcontrollers) for smart home appliances and industrial sensors.
• STMicroelectronics: Offers STM32 microcontrollers paired with specialized software ecosystems for running optimized, lightweight neural networks on battery-constrained edge nodes.

Dedicated Edge AI Startups & Innovators

• Hailo: Specializes in high-performance, low-power edge AI processors (like the Hailo-8 and Hailo-15 vision processors) built to run deep learning on edge devices like smart cameras and automotive systems with high efficiency.
• SiMa.ai: Focuses on MLSoC (Machine Learning System-on-Chip) platforms explicitly designed to effortlessly scale computer vision and ML workloads at the embedded edge for robotics and smart vision.
• Axelera AI: Develops powerful in-memory computing edge AI accelerators (such as the Metis platform) aimed at computer vision and multi-camera processing with extreme energy efficiency.
• MemryX: Creates patented matrix processor technology designed to deliver high-performance, low-power AI acceleration specifically tailored for edge devices without requiring cloud backup.

If you'd like, let me know:

• What specific application or industry (e.g., robotics, battery-powered IoT, automotive) you are designing for
• Whether you are prioritizing lowest power consumption or maximum TOPS (Tera Operations Per Second)

I can recommend the ideal chip architecture or vendor for your project.

Yes. The edge-AI chip market is fairly broad, ranging from tiny NPUs for sensors and wearables to powerful accelerators for robotics and local LLMs. As of 2026, notable companies include:

This isn't just a theoretical market: for example, NXP explicitly markets processors and MCUs for ML running at the edge, while Axelera's AIPUs are purpose-built inference processors.

The companies I'd pay closest attention to

If your interest is specifically running neural networks locally rather than merely having an NPU in a general-purpose SoC, I'd narrow the field to:

• Hailo — particularly interesting for cameras, robotics and embedded GenAI. Its Hailo-10H, for example, targets LLM/VLM inference at roughly 40 TOPS.
• Axelera AI — unusually focused on dedicated edge inference; its Metis-based Edge 130p is advertised at up to 214 TOPS.
• SiMa.ai — combines CPU, ML acceleration and software in an MLSoC aimed at embedded inference.
• Kneron — particularly focused on highly integrated, low-power on-device AI; its platform spans silicon, software and security.
• Ambarella — especially compelling when AI inference needs to be combined with sophisticated video processing. Its N1 family targets multi-camera VLM/NN workloads at under 20 W.
• NVIDIA — the strongest choice when you need a relatively large software ecosystem and substantial compute, especially for robotics and local LLMs.
• Qualcomm — strong when the AI needs to be integrated into a mobile, automotive or highly power-constrained SoC.

There are also a number of newer specialists—including Rebellions, FuriosaAI, Mobilint, Tenstorrent, MemryX, Quadric, Kinara and Expedera—competing in various parts of the edge-inference market.

If you tell me what device/model you're targeting (e.g. camera, robot, Raspberry Pi-class device, automotive ECU, smartphone, industrial PC, or local LLM box), I can narrow this down to the 5–10 most relevant chips and compare TOPS, power consumption, memory, supported models, SDKs, and approximate price.

A wide range of semiconductor giants, mobile system-on-chip (SoC) providers, and specialized hardware startups offer edge AI chips designed to run machine learning and neural network inference locally on-device (low-power, high-efficiency, and without needing a continuous cloud connection).[1][2][3][4][5]

The primary companies offering edge AI chips are categorized below by their core market focus:

1. High-Performance Edge & Robotics

• NVIDIA: Known for the NVIDIA Jetson lineup (such as the Jetson Orin Nano, Orin NX, and AGX Orin modules), which scales from tiny power envelopes up to hundreds of TOPS for autonomous machines, robotics, and advanced computer vision. [1][2][3][4][5]
• Hailo: Specializes in ultra-low-power, high-performance edge AI processors (like the Hailo-8 and Hailo-8L) designed for smart cameras, automotive, and industrial edge devices.[1][2][3][4][5]
• SiMa.ai: Offers purpose-built Machine Learning System-on-Chips (MLSoC) targeted at computer vision and edge-AI appliances prioritizing extreme energy efficiency. [1][2][3]
• Axelera AI: Provides high-performance edge vision processors (such as the Metis AI platform) aimed at accelerating deep learning inference at the extreme edge. [1][2][3][4][5]

2. Mobile, IoT, and AI PCs (Consumer & Industrial)

• Qualcomm: A dominant force in mobile, automotive, and PC edge silicon. Their Snapdragon platforms feature robust, integrated Hexagon NPUs (Neural Processing Units) built to handle on-device generative AI and machine learning locally for smartphones, Windows AI PCs, and IoT.[1][2][3][4][5]
• MediaTek: Integrates advanced custom APUs/NPUs (such as their NPU series) into mobile processors, smart TVs, and IoT devices to run local machine learning and real-time vision/audio workloads.[1][2][3][4][5]
• Intel: Bridges traditional computing and edge AI with processors (like Core Ultra and Lunar Lake series) [4] featuring built-in NPUs alongside versatile CPUs and GPUs optimized for on-device AI workloads in PCs and retail/industrial edge boxes [4].[1][2][3][4][5]
• Apple: Designs custom Apple Silicon (M-series and A-series chips) featuring high-performance, tightly integrated Neural Engines capable of running local on-device machine learning models efficiently for macOS and iOS devices.[1][2][3][4][5]

3. Embedded Systems, Smart Home & Industrial SBCs

• Rockchip: Popular for embedded system-on-chips (like the RK3588) that feature built-in, cost-effective NPUs widely adopted in single-board computers, edge gateways, and smart displays.
• NXP Semiconductors: Offers specialized applications processors (such as the i.MX 8M and i.MX 9 series) with integrated neural network accelerators tailored for industrial IoT, automotive edge, and secure smart-home appliances.
• STMicroelectronics: Provides cost-effective microcontrollers and MPUs (such as the STM32 series with embedded neural acceleration features) for low-power endpoint and appliance AI. [1][2][3][4][5]

If you have a specific use case in mind—such as robotics, smart cameras, mobile development, or an AI PC—let me know and I can help narrow down which specific chip and SDK matches your performance and power budget.

A wide range of semiconductor giants and specialized startups design edge AI chips (often utilizing dedicated Neural Processing Units or NPUs) to execute machine learning models locally on devices like smartphones, IoT gateways, automotive systems, and edge servers.

The prominent companies offering edge AI chips include:

• NVIDIA : Known for the Jetson lineup (such as the Jetson Orin Nano and Orin AGX series), providing scalable, high-performance edge AI and computer vision for robotics, automation, and advanced local inference.
• Qualcomm : Renowned for Snapdragon platforms featuring powerful Hexagon NPUs built into mobile devices, automotive cockpits, and industrial IoT hardware for on-device generative AI.
• Intel : Offers core processors with integrated graphics/NPUs (like Intel Core Ultra) and dedicated vision accelerators (like the OpenVINO-supported Movidius / Intel Gaudi or edge-focused Xeon/Atom lines) for industrial and enterprise edge.
• Apple : Integrates high-performance Neural Engine blocks directly into its M-series (Mac/iPad) and A-series (iPhone) custom silicon to run local transformer and machine learning models efficiently.
• AMD : Provides embedded adaptive computing solutions through its acquisition of Xilinx, offering Versal and Zynq adaptive SoCs with dedicated AI engines for low-latency edge applications.
• MediaTek : Embeds APUs (AI Processing Units) into its Dimensity and Genio platforms, targeting smart home devices, automotive infotainment, and mainstream mobile edge AI.
• Hailo : A specialized edge AI chip startup offering high-performance, low-power dataflow accelerators (such as the Hailo-8 and Hailo-10 series) designed to run heavy deep learning and generative AI workloads on edge devices.
• Ambarella : Focuses on low-power system-on-chips (SoCs) like the CVflow family, heavily utilized in smart security cameras, automotive ADAS, and robotics.
• NXP Semiconductors : Offers the i.MX application processor family integrated with specialized neural network accelerators for industrial and automotive edge intelligence.
• SiMa.ai : Known for their Machine Learning SoC (MLSoC ) purpose-built to scale computer vision and edge AI inference effortlessly in embedded environments.

If you're looking for a chip for a specific project, let me know:

• What is your target application (e.g., robotics, smart camera, mobile app, industrial automation)?
• What is your power budget (battery-powered vs. wall-powered)?

I can help narrow down which of these options fits your performance (TOPS ) requirements.

Yes. The edge/on-device AI chip market is fairly broad, ranging from dedicated neural-network accelerators to full SoCs that combine CPU, GPU and NPU resources.

Major companies to know

A few distinctions are particularly useful:

• NVIDIA, Qualcomm, AMD and Intel are the broad-platform players. Their chips combine conventional compute with AI acceleration, making them attractive when you need to run an entire application—not just an ML model. AMD's current Ryzen AI Embedded processors, for example, combine CPU, GPU and an XDNA NPU in one embedded processor.
• Hailo, Axelera AI and SiMa.ai are more specialized around efficient inference. Hailo's Hailo-10 is explicitly aimed at running generative AI locally, while Axelera's Metis AIPU targets very high inference performance within roughly a 4–8 W envelope.
• Ambarella and Hailo are particularly interesting for cameras and real-time computer vision. Hailo's portfolio spans dedicated AI accelerators and vision processors.
• NXP/Renesas/ST are strong choices when AI is part of a larger embedded-control system rather than the primary computing workload.
• SiMa.ai is targeting more demanding physical-AI workloads; its MLSoC architecture is designed to handle workloads including LLMs and VLMs at the edge.

If you're specifically looking for chips you can buy and integrate

I'd narrow the field to these six:

1. NVIDIA Jetson — best overall ecosystem/software.
2. Qualcomm — excellent for highly integrated, power-constrained products.
3. Hailo — excellent dedicated low-power inference.
4. Axelera AI — compelling performance-per-watt for vision and edge inference.
5. SiMa.ai — particularly interesting for robotics/physical AI.
6. AMD Ryzen AI Embedded — strong if you need CPU + GPU + NPU in one embedded platform.

For example, Axelera currently advertises its Metis-based Edge 130p at up to 214 TOPS, while Hailo's product family spans low-power vision acceleration through local LLM/GenAI acceleration.

If your goal is choosing a chip for a product, the important comparison isn't just TOPS. Power consumption, supported model formats, SRAM/DRAM capacity, INT8/INT4 performance, camera interfaces, software/SDK maturity, unit cost, and availability can change which vendor is actually best.