In April 2026 NVIDIA moved up the end-of-life dates for several LPDDR4-based Jetson modules, citing memory pricing and supply. For anyone building a product with a five- to ten-year field life on an embedded vision platform, that announcement changed the decision from "which board is fastest" to "which board will still be purchasable when we are on revision three". Here is how we think about platform selection after the EOL wave.
Lifecycle first, benchmark second
A module that benchmarks 20% faster but has a last-time-buy date inside your product's ramp is not the faster option. Before comparing frames per second, establish for each candidate:
- Published lifecycle commitment and any announced last-time-buy date.
- The memory technology it depends on — the 2026 phase-outs were specifically LPDDR4 parts, and current modules on newer memory were not affected.
- Whether a pin-compatible successor exists on the same carrier board.
- The software support window (JetPack/BSP releases) for the module, which often ends before the hardware does.
Only then rank the survivors on performance and price.
Orin versus Thor
Jetson Orin (Nano, NX, AGX) is the mainstream 2026 edge-vision platform: a mature CUDA and TensorRT stack, a broad carrier-board ecosystem, and a range that scales from a single-camera Orin Nano to multi-model AGX Orin deployments. For conventional vision — detection, segmentation, tracking, measurement — Orin is the default answer and will be for a while.
Jetson Thor is the Blackwell-based successor, now shipping. It is positioned for humanoid and general robotics and for workloads that combine vision with generative models on the device. JetPack 7 supports both Orin and Thor, and Thor aligns Jetson software with the Arm server system architecture. Choose Thor when your roadmap includes on-device vision-language models or large multi-sensor fusion; choose Orin when it does not, and pocket the difference.
Neither family was the target of the 2026 LPDDR4 phase-outs, but the episode is the reminder: check the lifecycle page, not the marketing page, and re-check it annually.
The alternatives
- Raspberry Pi 5 + AI Kit / AI HAT+ (Hailo NPU): excellent price/performance for single-camera detection and classification, a very large community, and Raspberry Pi's own long-availability commitments. No CUDA, so CUDA-dependent code does not carry over.
- NPU-based SBCs (Rockchip, Qualcomm, Intel): often the lowest power per inference, but vendor SDK quality and lifecycle commitments vary widely. OpenCV 5's new hardware acceleration layer and OpenVINO/LiteRT runtimes make the application layer more portable than it used to be.
- x86 with a discrete GPU or Intel iGPU/NPU: still the right answer for high-throughput fixed installations where power and size are not constraints.
Lifecycle-aware engineering
The platform choice matters less if the software is built to move:
- Abstract the accelerator. Keep the model in ONNX as the source of truth and generate TensorRT, OpenVINO or LiteRT artifacts per target in CI. A module swap becomes a rebuild, not a rewrite.
- Keep capture behind a config. GStreamer pipeline strings and camera parameters belong in configuration so a sensor or carrier change does not touch application code.
- Containerize the JetPack userspace and pin every version; rebuild engines when any of them changes.
- Qualify a second source early. The time to prove the application runs on a Raspberry Pi 5 or a second Jetson module is while the first one is still in stock.
- Buy ahead on announced EOLs only after the second source is qualified; a last-time-buy is insurance, not a strategy.
A procurement checklist
- Lifecycle and last-time-buy dates for every candidate module, in writing.
- Memory technology and a note on whether the module was touched by the 2026 phase-outs.
- JetPack/BSP support window versus your maintenance commitment.
- Pin-compatible successor on the same carrier: yes or no.
- Benchmarks of your model, on your camera configuration, at the power mode you will actually ship.
- A qualified second platform before volume production.
Not sure which way to go for a new product, or holding a design on a module with a newly announced EOL? Our edge AI and embedded vision practice offers a fixed-scope feasibility study that answers exactly these questions.