Standing reference. Specifications checked against Nvidia’s own datasheet (DS-10712-001) and developer documentation on 19 August 2026. The weapons-recovery section is checked against Ukraine’s Defence Intelligence publications and Nvidia’s statements to the press. Prices are Nvidia list prices at 1,000-unit volume and are not what a single module costs.
The Jetson Orin NX is a system-on-module the size of a stick of laptop memory — 69.6 by 45 millimetres, on a 260-pin SO-DIMM connector. It is sold for robots, drones, factory vision systems and handheld devices. It is consumer-grade hardware, freely resold, and it is not export-controlled in the United States, the United Kingdom or the European Union.
In August 2026 one was recovered from the wreckage of a Russian cruise missile. Both halves of that sentence are on this page, because they are the same fact: the part is easy to buy precisely because nobody classified it as a weapon component.
Specifications
| Orin NX 16GB | Orin NX 8GB | |
|---|---|---|
| AI performance (INT8, sparse) | 100 TOPS — 157 in MAXN_SUPER | 70 TOPS — 117 in MAXN_SUPER |
| AI performance (INT8, dense) | 50 TOPS — 78 in MAXN_SUPER | 35 TOPS — 58 in MAXN_SUPER |
| GPU | Ampere, 1024 CUDA cores, 32 Tensor cores, up to 918 MHz (1,173 MHz SUPER) | Ampere, 1024 CUDA cores, 32 Tensor cores, up to 765 MHz (1,173 MHz SUPER) |
| Deep learning accelerator | 2x NVDLA, 614 MHz (1,229 MHz SUPER) | 1x NVDLA, 610 MHz (1,229 MHz SUPER) |
| CPU | 8x Arm Cortex-A78AE v8.2, up to 2 GHz | 6x Arm Cortex-A78AE v8.2, up to 2 GHz |
| Memory | 16 GB 128-bit LPDDR5, 3,200 MHz | 8 GB 128-bit LPDDR5, 3,200 MHz |
| Peak memory bandwidth | 102 GB/s | 102 GB/s |
| Power modes | 10 W, 15 W, 25 W, 40 W (MAXN_SUPER) | 10 W, 15 W, 20 W, 40 W (MAXN_SUPER) |
| Power input | 5–20 V, or 8–20 V when MAXN_SUPER is enabled | |
| Module size | 69.6 mm x 45 mm, 260-pin SO-DIMM | |
| Junction temperature | −25 °C to 105 °C; SoC slowdown at 99 °C | |
| Stated operating lifetime | 5 years at 24×7 | |
| Nvidia list price (1KU+) | 999 dollars | 649 dollars |
What the TOPS number actually counts
This is where most specification pages go wrong, including several that outrank nothing. The headline 100 TOPS is not the GPU. It is the GPU plus the deep learning accelerators, added together, and the arithmetic is checkable in Nvidia’s own JetPack 6.2 tables.
| Module | GPU (sparse INT8) | DLA (sparse INT8) | Published total |
|---|---|---|---|
| Orin NX 16GB | 60 TOPS | 40 TOPS | 100 TOPS |
| Orin NX 16GB, Super Mode | 77 TOPS | 80 TOPS | 157 TOPS |
| Orin NX 8GB | 50 TOPS | 20 TOPS | 70 TOPS |
| Orin NX 8GB, Super Mode | 77 TOPS | 40 TOPS | 117 TOPS |
Two consequences worth knowing before you size a workload against the headline figure. A model that does not run on the DLA cannot reach it — on the 16GB module that is 40 of the 100 TOPS sitting idle. And the sparse figures assume structured sparsity in the network; the dense figures, exactly half, are what an ordinary model gets.
Note also that in Super Mode the 8GB and 16GB modules have the same GPU throughput, 77 TOPS. The entire remaining gap between 117 and 157 is the second NVDLA on the 16GB part.
Super Mode is not free
Super Mode is a software unlock — it arrived with JetPack 6.1 and is applied through a new flash configuration file, not new silicon. But on the Orin NX series it needs a 40 W power budget and an HV rail on the carrier board, and Nvidia says plainly that customers may need to redesign carrier boards and thermal solutions to use it. On an existing fielded device it is frequently not available at all.
Measured gains on the 16GB module, from Nvidia’s own benchmarks, are more modest than the 100-to-157 headline implies: Llama 3.1 8B improved 1.11x, Gemma 2 9B 1.26x, and vision transformers between 1.10x and 1.25x. The larger multiples appear in vision-language models — LLaVA 1.6 7B at 2.54x.
A price that went up
Nvidia’s January 2023 announcement post put the Jetson Orin NX 16GB at 599 dollars at 1KU+. Nvidia’s current published price list shows 999 dollars for the same module. We have not found an explanation from Nvidia for the change, and we are not going to invent one; both figures are Nvidia’s, three years apart, and anyone budgeting from the launch article is working from a stale number.
Where it sits in the Jetson line
| Module or kit | Price |
|---|---|
| Jetson AGX Thor Developer Kit | 5,499 dollars |
| Jetson T5000 module | 4,999 dollars |
| Jetson AGX Orin Developer Kit | 3,499 dollars |
| Jetson T4000 module | 2,999 dollars |
| Jetson AGX Orin 64GB module | 2,999 dollars |
| Jetson AGX Orin 32GB module | 1,799 dollars |
| Jetson Orin NX 16GB module | 999 dollars |
| Jetson Xavier NX 16GB module | 899 dollars |
| Jetson Orin NX 8GB module | 649 dollars |
| Jetson Orin Nano Super Developer Kit | 399 dollars |
| Jetson Orin Nano 8GB | 399 dollars |
The Orin NX sits in the middle: far above the Nano, an order of magnitude below the AGX Thor parts that carry the datacentre-adjacent price tags. It is the cheapest Jetson that pairs eight CPU cores with two DLAs, which is what makes it the default choice for autonomous machines that have to run several networks at once inside a small power budget.
The weapons finding
On 12 August 2026, Ukraine’s Defence Intelligence directorate published photographs of a heat-blistered computing module recovered from the wreckage of an S-71 “Monochrome” cruise missile, in the foreign-components section of its War and Sanctions portal. The board marking visible in the photographs is TE980M-A1, which identifies it as a Jetson Orin NX 16GB. According to a later analysis by heise, the module label indicates production in March 2025 — three years after Nvidia withdrew from Russia.
The S-71M is air-launched, reportedly from Su-57 aircraft, and is described as able to search for and engage targets using optical sensors and onboard computing, either under operator control or autonomously. Ukrainian intelligence said the presence of the module “may indicate” the use of AI in the weapon, and was careful not to claim more; it did not say what function the module performed. Nobody outside the programme knows whether it was doing terminal guidance, image recognition, or something mundane.
What Nvidia said
Nvidia’s response to The Register was narrow and, on its own terms, accurate: the company closed its Russian offices and stopped direct sales in 2022, the Orin NX did not ship until 2023, and the module reached Russia through channels it does not control. “Pre-owned Jetsons are available through many reseller channels,” the company said. “Although we cannot track products after they are sold, if we determine that any customer is violating U.S. export controls, we will take appropriate action.”
The part that makes the sanctions argument harder than it looks
Ukrainian intelligence called the find evidence that export controls need strengthening. That framing implies the part slipped through a control regime. It did not, because there is no regime covering it: the Jetson Orin NX is consumer-grade edge inference hardware and is not export-controlled in the EU, the UK or the United States.
So the gap is not enforcement. It is classification. A module sold for robot vacuums, delivery drones and factory vision systems is the same module that fits inside a cruise missile, and no list distinguishes them because the hardware does not. Tightening enforcement against controlled goods and named counterparties — which is what recent EU packages have done — does not reach an uncontrolled part bought second-hand.
That is a genuinely hard policy problem and we are not going to pretend it has an obvious answer. Controlling every edge AI module capable of image recognition would sweep in a very large share of ordinary industrial electronics. Not controlling them leaves the route open. Both of those are true at once.
Where sources disagree
Reports differ on one detail and it is worth being precise. Kyiv Post and Ukrainska Pravda both report that Ukrainian intelligence published data on 35 foreign electronic components identified across samples of Russian weaponry in that release. At least one subsequent write-up compressed this into 35 components found in the missile itself. The first reading is the one supported by the primary reporting, and the difference matters: one is a research batch, the other would be a claim about a single weapon.
This is also not the first Jetson recovery. Ukraine has previously reported Jetson Orin modules in the V2U AI-navigated strike drone.
Questions we get asked
What is the Nvidia Jetson Orin NX?
A compact system-on-module for edge AI — an Ampere GPU, an eight- or six-core Arm CPU, deep learning accelerators and LPDDR5 memory on a 69.6 by 45 millimetre board. It is designed for robots, drones, autonomous machines and camera systems that need to run several neural networks at once inside a 10 to 40 watt budget.
How many TOPS is the Jetson Orin NX?
100 sparse INT8 TOPS on the 16GB module and 70 on the 8GB, rising to 157 and 117 respectively in Super Mode. Those totals combine the GPU and the deep learning accelerators; the dense figures, which is what a model without structured sparsity gets, are exactly half.
How much does a Jetson Orin NX cost?
Nvidia lists 999 dollars for the 16GB module and 649 dollars for the 8GB, both at 1,000-unit volume. Single units through distributors cost more. Nvidia’s own 2023 launch post listed the 16GB at 599 dollars, so any figure you find in older coverage is likely to be wrong now.
What is the difference between the 8GB and 16GB modules?
Memory, CPU cores and accelerators. The 16GB has eight A78AE cores against six, two NVDLA units against one, and a higher base GPU clock. In Super Mode both hit the same 77 TOPS of GPU throughput, so the entire remaining performance gap is the second accelerator.
Is the Jetson Orin NX export-controlled?
No. It is consumer-grade hardware and is not on export-control lists in the United States, the United Kingdom or the European Union. That is the central fact in the debate about it turning up in Russian weapons, and it is the one most coverage leaves out.
What was a Jetson Orin NX doing in a Russian missile?
Unknown, and the people who found it said so. Ukrainian intelligence recovered a Jetson Orin NX 16GB from an S-71M cruise missile in August 2026 and said its presence “may indicate” the use of AI. They did not state what it was doing, and neither will we. The module is capable of onboard image recognition; that is a capability, not a finding.
Can I still get Super Mode on an existing device?
Only if the hardware around the module supports it. Super Mode arrives through a new flash configuration in JetPack 6.1 and later, but on the Orin NX it requires a 40 W power budget and an HV rail on the carrier board. Nvidia says carrier boards and thermal designs may need redesigning, which for a fielded product usually means no.
What we have reported
Continuing coverage sits in Nvidia, Semiconductors and AI.
Sources
- Nvidia Jetson modules and the Jetson Orin NX Series Modules datasheet, DS-10712-001 — every specification on this page.
- Nvidia Jetson FAQ — the price list and the Super Mode carrier-board requirements.
- Nvidia, JetPack 6.2 Super Mode — the GPU and DLA split, and the measured model benchmarks.
- Ukraine Defence Intelligence, War and Sanctions portal, 12 August 2026 — the recovery itself.
- Kyiv Post and Ukrainska Pravda, 12 August 2026 — the 35-component figure.
- The Register, 14 August 2026 — Nvidia’s statement and the TE980M-A1 marking.
- heise online, 17 August 2026 — the March 2025 production date on the module label.
How this page is maintained
Specifications change only when Nvidia revises the datasheet or ships a JetPack release that alters the power modes, so that half is checked monthly. The weapons and export-control section moves with the reporting; we re-check it weekly while the story is live. Prices are Nvidia list prices and are checked with the specifications — if you are buying, check the distributor, not this page.