Google Coral was the quiet workhorse of home lab AI. A cheap USB or M.2 accelerator, a Frigate container, and a Raspberry Pi could watch every camera in the house without frying the CPU. Then Google pulled Coral. Retailers ran out through 2025, and the software repos have not had a meaningful commit in months. Home lab forums have been trading Google Coral alternatives ever since, and the answer is not always “buy a used one on eBay.”
The seven picks below cover the real replacements: dedicated NPUs that ship in 2026, software runtimes that let a plain Intel CPU do Coral’s job, and the NVR stack most people already run on top. Each one is judged on cost per stream, integration with Frigate or CodeProject.AI, and whether it survives a Docker-first workflow.
Quick comparison
| Solution | Best for | Free tier | Starting cost | Standout feature |
|---|---|---|---|---|
| Frigate + Intel iGPU (OpenVINO) | Cheap object detection on Intel N-series | Software free | ~$150 mini PC | Zero extra hardware |
| CodeProject.AI | Multi-backend home security AI | Yes | Free, hardware optional | Blue Iris and Frigate integration |
| Intel OpenVINO Runtime | Any Intel CPU or Arc GPU | Free SDK | Free | Runs on hardware users already own |
| Hailo-8/8L | Dedicated edge NPU, Coral-class TOPS | Free SDK | ~$150 for the module | 26 TOPS on M.2 |
| Nvidia Jetson Orin Nano | Full SBC with GPU inference | Free JetPack | ~$249 | End-to-end dev kit |
| ONNX Runtime | Cross-platform inference on any hardware | Free | Free | One runtime for every backend |
| Ultralytics YOLO | Modern models to swap for Coral’s older ones | Free open source | Free | Direct Frigate integration |
Why people leave Google Coral
The reasons pile up fast once Coral disappears from Amazon and Mouser.
- Supply is dry. Google’s Coral store has been out of stock for the USB Accelerator and the M.2 Dual TPU since late 2024. Third-party marketplaces list used units at multiples of the original price. The XDA piece that flagged Intel CPUs as the alternative also notes the same shortage.
- Software is frozen. The
libedgetpuandgoogle-coral/edgetpurepos on GitHub have not seen a significant release in over a year. New Python and TensorFlow versions do not always work cleanly with the older runtime. - Ecosystem drift. Frigate 0.14 and later added first-class support for OpenVINO, TensorRT, ROCm, and Hailo. Coral is now one option among many, not the default.
- Model ceiling. Coral runs Edge TPU-compiled TensorFlow Lite models with fixed quantization. Modern YOLOv8, YOLOv10, and RT-DETR variants either do not compile for Edge TPU or lose enough accuracy to matter for false-alarm rates.
- Limited compute headroom. A single Coral tops out around 4 TOPS. A modern Hailo-8 hits 26 TOPS in the same M.2 slot for roughly the same money.
The alternatives
Frigate on an Intel iGPU (via OpenVINO) -- Best for cheap object detection
Every recent Intel N-series or Core CPU includes an iGPU that runs OpenVINO models fast enough for a handful of camera streams. Frigate picks up the iGPU automatically once the OpenVINO detector is configured, and a $150 mini PC handles four to six 1080p streams at 5 FPS detection without breaking a sweat. This is the setup the XDA piece points at as the direct Coral replacement.
Where it falls short: OpenVINO’s fastest small models (SSD MobileNet, YOLO-NAS quantized) are not always as accurate as a fresh YOLOv10 running on a bigger card. Older Intel iGPUs (pre-11th-gen) are noticeably slower.
Pricing:
- Free: Frigate is open source, OpenVINO is a free SDK.
- Paid: A mini PC (Beelink S12, Minisforum UN100C, or an old Intel NUC) at around $150 used.
- vs Coral: Uses hardware most home labs already own, no separate accelerator.
Migrating from Coral: Edit Frigate’s config.yml, swap the edgetpu detector for openvino, pull the OpenVINO-format model from the Frigate docs, restart. Existing camera zones, MQTT topics, and Home Assistant automations do not change.
Download: Frigate is open source and ships as a Docker image. OpenVINO Runtime installs via Frigate’s container automatically.
Bottom line: The default answer for anyone with a spare Intel mini PC and up to six cameras.
CodeProject.AI Server -- Best for Blue Iris and mixed backends
CodeProject.AI Server is the AI backend most Blue Iris users pair with their NVR, and it also plugs into Frigate. It supports YOLOv5 and YOLOv8 models, ALPR license plate reading, face recognition, and multiple hardware accelerators through a single web dashboard. Runs on Windows and Linux, in Docker or natively.
Where it falls short: The Coral module for CodeProject.AI still works, but the project’s focus has moved to CPU, DirectML, and CUDA backends. The Windows installer occasionally lags the Linux Docker path.
Pricing:
- Free: The full server and default models.
- Paid: Optional paid add-ons for advanced modules.
- vs Coral: Free like Coral once the hardware was bought, and covers face recognition and plate reading out of the box.
Migrating from Coral: Install CodeProject.AI, install the OpenVINO or CUDA module that matches the host, point Blue Iris or Frigate at the CodeProject.AI URL. No model recompilation.
Download: CodeProject.AI Server provides Docker images for Linux and a native installer for Windows.
Bottom line: Right pick when the NVR is Blue Iris on Windows, or when license plate recognition matters.
Intel OpenVINO Runtime -- Best for hardware users already own
Intel OpenVINO turns any 8th-gen or later Intel CPU, any Intel Arc GPU, and any Neural Compute Stick 2 into an inference target. The runtime is a single Python or C++ library, and it powers the Frigate detector, CodeProject.AI, and countless custom scripts. On an Intel Core i5-1240P laptop CPU it hits Coral-class throughput on the same MobileNet SSD model.
Where it falls short: OpenVINO is Intel-only. AMD Ryzen users get nothing from it. Arm boards (Raspberry Pi, Orange Pi) need a different runtime entirely.
Pricing:
- Free: Full SDK and runtime.
- Paid: N/A.
- vs Coral: Zero hardware cost if the host is already Intel-based.
Migrating from Coral: Convert TensorFlow Lite models to OpenVINO IR format with the mo tool, or grab a pre-converted model from Frigate’s model zoo. Rewire the detector config, restart.
Download: OpenVINO Runtime is available as a Docker image, a pip package, and a native installer for Windows and Linux.
Bottom line: The zero-hardware option. If the NVR box already runs on Intel, install OpenVINO first before shopping for anything else.
Hailo-8 and Hailo-8L -- Best for dedicated edge NPU muscle
Hailo modules are the closest hardware analogue to Coral. The Hailo-8L runs 13 TOPS on an M.2 A+E slot for roughly the same price Coral used to command. The Hailo-8 doubles that at 26 TOPS. Both drop into Frigate through the community Hailo detector plugin and into the official Hailo SDK for custom work.
Where it falls short: The Hailo SDK setup is heavier than Coral’s plug-and-play story. The Hailo Model Zoo covers most common architectures but the conversion pipeline needs practice.
Pricing:
- Free: SDK and runtime.
- Paid: Around $70 for a Hailo-8L, around $150 for a Hailo-8, plus an M.2 slot on the host.
- vs Coral: Similar cost, several times the throughput, actively maintained.
Migrating from Coral: Buy the module, install the Hailo firmware and SDK, replace the Frigate detector block with the Hailo variant, download the compiled model file. Camera configs do not change.
Download: Hailo Developer Zone ships SDKs for Ubuntu and container images.
Bottom line: Best pick when a Coral was already the plan but the throughput or model choice needs to grow.
Nvidia Jetson Orin Nano -- Best for full SBC with GPU inference
The Nvidia Jetson Orin Nano developer kit puts an Arm CPU, an Nvidia GPU with tensor cores, and Jetpack (a CUDA and TensorRT stack) on one small board. It runs Frigate with the TensorRT detector at very high frame rates, and it can also host the NVR itself, so a Coral plus a Raspberry Pi collapses into a single unit.
Where it falls short: Higher upfront cost than a Coral USB. The Jetpack image is Ubuntu-based and requires flashing rather than a plain OS install. Older Jetson Nano boards are much slower and should be avoided in 2026.
Pricing:
- Free: JetPack SDK.
- Paid: Around $249 for the Orin Nano 8 GB developer kit, higher for Orin NX and AGX.
- vs Coral: Replaces both the host and the accelerator, so total cost lines up with a mini PC plus a Hailo.
Migrating from Coral: Reinstall Frigate on Jetpack, use the TensorRT detector, convert models to TensorRT engine format with the trtexec tool. Camera streams and Home Assistant integrations stay untouched.
Download: Nvidia Jetson downloads provide JetPack SDK and pre-flashed images.
Bottom line: The all-in-one path. Right when the NVR host is also up for replacement.
ONNX Runtime -- Best for one runtime across every backend
ONNX Runtime is Microsoft’s cross-platform inference engine. It targets CUDA, TensorRT, DirectML, ROCm, CoreML, and CPU with a single API. Frigate 0.14 shipped ONNX Runtime support that lets the same model file run on an Nvidia GPU today, an AMD GPU next week, and a laptop CPU on the road.
Where it falls short: ONNX Runtime is a runtime, not a model catalog. Users still need a model file (YOLOv8, RT-DETR, MobileNet). Some experimental optimizations lag behind vendor-specific SDKs like TensorRT.
Pricing:
- Free: Open source.
- Paid: N/A.
- vs Coral: Free, hardware-agnostic, and the model choice is much larger.
Migrating from Coral: Export the current model to ONNX format, swap Frigate’s detector to onnx, restart. If the host has a GPU, install the CUDA or ROCm execution provider.
Download: ONNX Runtime is available on PyPI, npm, NuGet, and as pre-built native libraries.
Bottom line: The safest bet for a mixed home lab where the compute might change every year.
Ultralytics YOLO -- Best for modern models to feed the runtime
Ultralytics YOLO ships YOLOv8, YOLOv10, and YOLO11 with a Python API that exports to every runtime on this list. Detection accuracy is well ahead of the older MobileNet SSD models Coral shipped with, and false alerts on a Frigate setup drop noticeably when the model swaps.
Where it falls short: Ultralytics changed its license to AGPL-3.0 for the open source builds. Commercial deployments need the paid Ultralytics license. Some quantized exports still need tuning to match the accuracy of the FP16 originals.
Pricing:
- Free: AGPL-3.0 for personal and open-source use.
- Paid: Ultralytics Enterprise license for commercial use.
- vs Coral: Free for home use, higher-accuracy models than the original Coral zoo.
Migrating from Coral: Pick a YOLO model matching the host (YOLOv8n for CPU, YOLOv8s or larger for GPUs and NPUs), export to the format Frigate expects, restart the container.
Download: Ultralytics is on PyPI and GitHub with weights and export tooling.
Bottom line: The model source once the runtime is settled. Frigate detection quality often improves more from a model swap than a hardware swap.
How to choose
Pick Frigate on an Intel iGPU with OpenVINO if the NVR is a mini PC and the camera count is under six. Cheapest path, zero extra hardware.
Pick CodeProject.AI Server if the NVR is Blue Iris on Windows, or the setup needs face recognition or license plates without extra glue code.
Pick Intel OpenVINO Runtime on any Intel-based host as a starting point before spending on new hardware.
Pick Hailo-8 or Hailo-8L when Coral was the plan and the M.2 slot is free. Same form factor, much more compute, actively maintained.
Pick Nvidia Jetson Orin Nano when the NVR box itself is up for replacement and a full SBC makes sense.
Use ONNX Runtime as the abstraction layer if the hardware might change again in a year or two.
Feed any of the above with Ultralytics YOLO models when the false-alarm rate on the older Coral zoo starts to hurt.
Stay on Coral only if the modules are already installed and working. There is no upgrade to buy, and pushing the setup to a newer Frigate release will keep working for years. Do not start a new home lab on Coral in 2026.
FAQ
Is Google Coral discontinued? Yes. The Coral USB Accelerator, Coral M.2 Dual TPU, and Coral Dev Board have been out of stock through Google’s own retailer channels since late 2024, and the software repos have gone dormant.
What is the cheapest Google Coral alternative? Frigate running on an Intel iGPU with OpenVINO. The software is free, and a used Intel mini PC starts around $100. No dedicated accelerator required.
Can I still buy a Coral USB Accelerator? Only through third-party marketplaces at inflated prices. Buying used in 2026 is not recommended when Hailo-8L, Jetson Orin Nano, and Intel iGPU paths cost the same or less.
Does Frigate support alternatives to Coral? Yes. Frigate 0.14 and later ships detectors for OpenVINO, TensorRT, ONNX Runtime, Hailo, and ROCm. Swapping is a config file change.
What is the best NPU for security cameras in 2026? Hailo-8 for a drop-in accelerator (M.2 slot, 26 TOPS) or Nvidia Jetson Orin Nano when the entire NVR host is up for replacement. Both are actively maintained and cover modern model architectures Coral does not.
Do I need to retrain my models? No. Existing TensorFlow Lite or ONNX models can be exported to the target runtime. Some backends prefer INT8 quantization, which needs a calibration pass but does not require retraining from scratch.
