Jetson Price per TOPS 2026 — Cost-per-TOPS Across the Orin & Thor Lineup

"What does a Jetson TOPS actually cost?" is the first question most edge-AI procurement teams ask — and the answer is a U-shaped curve that surprises people. Divide NVIDIA's H2 2026 module prices by their published AI TOPS and the Orin Nano Super comes out at roughly $3.72 per TOPS, the AGX Orin 32 GB at $8.00, and the ECC-equipped AGX Orin Industrial at $11.29. The entry-level Orin Nano 4 GB is a worse deal per TOPS than the module two tiers above it.

But price-per-TOPS is also the most misleading spec sheet in edge AI. This guide ranks every current Jetson SKU by dollars per published TOPS at one-unit and volume pricing, then explains why the ranking alone should not choose your platform.

Cost-per-TOPS Table — H2 2026 List Prices

Prices below are NVIDIA's H2 2026 one-unit list for modules, plus the 1,000-unit tray tier, consistent with our Q3 street-price tracking. The Thor dev kit is priced as a complete kit; the T5000 module price is the volume quote.

Sources: NVIDIA H2 2026 module price list and Q3 2026 distributor street pricing, as published in our Jetson price tracking posts.

Three patterns stand out. First, the value center of the lineup is the Nano Super: NVIDIA effectively repositioned the $249 slot with 67% more compute than the Nano 8 GB it replaced. Second, the true entry tier (Nano 4 GB) is poor value — you pay a $7.45/TOPS penalty for a 4 GB module that fits almost nothing modern. Third, the Industrial module carries a ~54% per-TOPS premium over the standard AGX Orin 64 GB — you are paying for ECC memory, wide-temperature operation and a 10-year lifecycle, not raw TOPS.

Why the $/TOPS Ranking Misleads

1. The TOPS numbers are not all the same currency. The Nano Super's 67 TOPS is dense INT8. AGX-class published figures are typically quoted with 2:1 structured sparsity, which roughly doubles the headline number over dense throughput. Thor's headline is FP8, a different precision again. Comparing them with a single division sign quietly inflates the AGX and Thor rows relative to the Nano Super.

2. Memory bandwidth caps what TOPS can deliver. The Nano Super has ~102 GB/s of memory bandwidth; the AGX Orin 64 GB ~205 GB/s and Thor ~276 GB/s. Most real inference is memory-bound — a detector or token stream scales with bandwidth, not peak matrix TOPS, so a bandwidth-limited 275-TOPS module can deliver less useful throughput per dollar than a 67-TOPS one.

3. Capacity decides which models fit at all. An 8 GB module cannot hold a 32 GB model no matter how cheap its TOPS are. If your workload is a quantized 7–8B LLM or a 16-camera vision stack, the NX 16 GB or AGX tiers are the only rows that qualify — per-TOPS ranking is irrelevant among disqualifiers.

4. Real workloads rarely hit peak TOPS. Published TOPS are dense-math, best-case matrix numbers. A YOLO detector or Whisper stream uses a fraction of peak; the honest denominator is measured FPS or tokens per second on your actual model.

Which SKU Actually Wins, by Deployment

The Volume Reality Check

At 1,000 units every row drops 20–30% and the ordering barely changes — the Nano Super still leads at ~$2.97/TOPS and the Industrial module stays the premium tier at ~$9.03/TOPS. What volume changes is the conversation: at tray quantities you negotiate availability as much as price, and Thor T5000 modules (~$2,999 at 1,000 units) are quote-based allocations. If your Q4 plan touches AGX or Thor, lock quotes early — memory-supply pressure is still stretching lead times.