Jetson Generational Price History 2017–2026 — TX2 to AGX Thor Value Curve & When to Upgrade

Every NVIDIA Jetson generation has reset the price of a given amount of edge AI. A dollar bought roughly 1/450th of an FP16 TFLOP when the TX2 shipped in 2017; by 2025 the $3,499 AGX Thor developer kit delivered about 3.4× more INT8-class TOPS per dollar than the 2022 AGX Orin — on top of a 5–10× jump in absolute throughput per generation. This guide reconstructs the full Jetson price history generation by generation, shows what the value curve really measures, and maps the software-support horizon that quietly decides when an older Jetson fleet must be upgraded.

The Jetson price curve, generation by generation

Each row uses the developer kit's launch price and NVIDIA's headline AI figure of its era. Note that the headline metric changed over time — FP16 TFLOPS on Pascal/Maxwell, dense INT8 TOPS on Volta, sparse INT8 TOPS on Ampere, and FP8/FP4 on Blackwell — which is itself the first lesson in comparing Jetson generations (see below).

The Orin Nano Super remains the standout price event of the lineup: NVIDIA shipped the 8GB kit at $499 in 2023, then cut it to $249 with a claimed 1.7× performance uplift in late 2024 — turning the entry tier into a 67-TOPS board at $3.72 per TOPS. At the top end, the AGX Thor developer kit's $3,499 buys 2,070 FP4 TFLOPS and 128 GB of LPDDR5x at 276 GB/s, roughly 7.5× the AGX Orin's compute at 3.5× the efficiency.

Why the value curve is not as smooth as it looks

Three distortions separate a real workload from the headline curve. First, the metric moved: TX2's 1.33 "TFLOP" is dense FP16; Ampere and Blackwell headlines count sparse INT8 or FP8/FP4, which inflate apparent throughput 1.5–2× over dense operations — an AGX Orin 64GB "275 TOPS" is ~137 dense INT8 TOPS. Second, TOPS is not FPS: memory bandwidth is what actually bounds many edge workloads, and it has grown far slower than TOPS (59.7 GB/s on TX2, 102 GB/s on Orin Nano Super, 204.8 GB/s on AGX Orin 64GB, 276 GB/s on Thor). A detection model that is bandwidth-bound on Orin will not run 4× faster on a paper-4×-TOPS Thor without also moving to FP8/FP4 weights. Third, software follows hardware — and that is where old Jetsons truly die.

Software support is the hidden EOL

NVIDIA's module lifecycle page gives commercial availability dates, but the practical support horizon is set by JetPack. Each major JetPack generation ships with a fixed CUDA line, and NVIDIA's PyTorch and TensorRT releases follow JetPack — not the other way around. A TX2-era product is not merely old hardware; it is stuck on a toolchain that cannot build or run any current model.

Three consequences follow. A Jetson on JetPack 4 (TX2, Nano) cannot run models that need CUDA 11/12 features, and its stack no longer receives security or driver fixes — a real liability for a connected device. Xavier-family modules are orderable through July 2027 but development on JetPack 5 is effectively frozen, so new model work should start on Orin. And any "but it still works" argument has a date: the last TX2 modules reached EOL in January 2025, so spares and RMAs for those fleets now come only from the aftermarket.

When to upgrade: a decision table for 2026

The upgrade math is rarely about the old board failing — it is about the model you can no longer run and the JetPack you can no longer patch. NVIDIA's own lifecycle dates (Orin modules through January 2032, T5000 through August 2035) make the Orin generation the safe 2026 procurement target for new edge products, with Thor reserved for workloads that genuinely need Blackwell-class throughput at the edge.