Jetson Orin Nano 2 is the module NVIDIA announced on 25 August 2026 to put physical AI inside delivery drones, inspection drones, home robots and factory vision systems. The headline everyone ran was the same one NVIDIA led with: twice the inference performance of the outgoing Jetson Orin Nano Super, in the same form factor, drawing 40% less power. It is a genuinely interesting part, and the partner list behind it is not decoration.
It is also a product you cannot buy. Availability is the first half of 2027, no price has been announced, and the drone maker quoted in NVIDIA’s own release describes itself as exploring the board rather than shipping it. Between the announcement and the coverage, a carefully hedged claim on NVIDIA’s product page hardened into a flat “doubles” in trade headlines, and in at least one case got attached to a TOPS figure that is 16% higher than its predecessor, not 100%.
This article does the arithmetic that most of the coverage skipped. Every figure below is traced to the page that published it, the 2x claim is checked against each individual specification NVIDIA and its board partners have disclosed, and the power claim is unpacked into the two separate numbers it actually contains. If you are budgeting an edge AI strategy around this part, the gap between the claim and the published specification is the thing worth understanding.
Table of contents
- What NVIDIA Announced With the Jetson Orin Nano 2
- The Jetson Orin Nano 2 Arithmetic: Where Does “2x” Come From?
- How the Jetson Orin Nano 2 Claim Changed as It Travelled
- The Jetson Orin Nano 2 Power Claim Contains Two Different Numbers
- What the Jetson Orin Nano 2 Silicon Tells You
- Jetson Orin Nano 2 Drone Claims Against What Wing Actually Said
- The Jetson Orin Nano 2 Partner List Is the Strongest Evidence
- Aptiv and the Jetson Orin Nano 2 Production Gap
- The Price NVIDIA Did Not Announce
- What the Jetson Orin Nano 2 Software Stack Adds
- How to Read a Jetson Orin Nano 2 Spec Claim
- Frequently Asked Questions About the Jetson Orin Nano 2
- What the Jetson Orin Nano 2 Means for Your Edge AI Plans
- References and Further Reading
What NVIDIA Announced With the Jetson Orin Nano 2
The announcement came as a press release titled “NVIDIA Announces Jetson Orin Nano 2 Robotics Computer to Redefine Entry-Level Edge AI”, published 25 August 2026. It is a module announcement, not a launch: NVIDIA says the part will be available as both a production module and a developer kit in the first half of 2027.
Deepu Talla, NVIDIA’s vice president of robotics and edge AI, framed it around access rather than raw capability: “The Jetson Orin Nano 2 computer puts that breakthrough within reach of millions of developers, delivering the performance and energy efficiency needed for real-time reasoning at the edge.” NVIDIA puts the size of that developer base at more than three million people already building on its robotics stack. The pitch is a module that runs current AI models — language and vision alike — fast enough to make decisions on the device rather than in a data centre.
The specification as published
NVIDIA disclosed four headline numbers for the Jetson Orin Nano 2: 78 TOPS of AI compute, 8GB of memory, an eight-core Arm CPU, and the same compact form factor as the outgoing module. Board-level reporting from ServeTheHome adds the detail NVIDIA left out of the release — 1,536 Ampere CUDA cores, 8GB of LPDDR5X-7500 running at 120GB/s, a 15W to 40W power range, and Samsung’s 8nm process.
Those board-level figures are reported specifications rather than NVIDIA-published ones, and the CUDA core count in particular is carried with a question mark at source. Everything attributed to NVIDIA below is separated from everything attributed to third parties, because the distinction matters for the arithmetic that follows.
What the predecessor established
The comparison point is the Jetson Orin Nano Super, which NVIDIA announced in December 2024. Its published specification is 67 TOPS, a 1,024-core Ampere GPU with 32 tensor cores, a six-core Arm Cortex-A78AE at 1.7GHz, 8GB of 128-bit LPDDR5 at 102GB/s, and power modes of 7W, 15W and 25W. The developer kit sold for $249, down from $499 for the 40 TOPS board it replaced.
That price cut is the reason the Super mattered. It is also the number NVIDIA has not repeated.
The Jetson Orin Nano 2 Arithmetic: Where Does "2x" Come From?
Put every published figure for both modules side by side and the doubling claim has to be located somewhere specific, because no individual specification moves by anything close to 100%.
| Specification | Orin Nano Super | Jetson Orin Nano 2 | Change |
|---|---|---|---|
| AI compute | 67 TOPS | 78 TOPS | +16.4% |
| Memory bandwidth | 102GB/s | 120GB/s | +17.6% |
| GPU cores (Ampere) | 1,024 | 1,536 | +50.0% |
| CPU cores (Arm) | 6 | 8 | +33.3% |
| Memory capacity | 8GB | 8GB | 0% |
| Power range | 7W-25W | 15W-40W | floor +114% |
| Developer kit price | $249 | undisclosed | unknown |
| Available | December 2024 | H1 2027 | – |
No single published figure reaches 2x
The chart below plots the percentage increase in each disclosed specification against the increase the headline claims. The largest real movement is the GPU core count at +50%, and that is the figure NVIDIA did not publish.
ServeTheHome’s Ryan Smith put the same observation directly: “The peak theoretical performance may have only improved by 16%, but NVIDIA says that the amount of real-world work that can be extracted from the hardware has doubled.” That is a coherent position. It is not a claim any published number verifies.
What “inference performance” leaves unspecified
A 2x inference claim is only checkable with three things attached: which model, which precision, and which batch size. NVIDIA has published none of them for the Jetson Orin Nano 2. TOPS figures are normally quoted at INT8 with sparsity, which is why 78 TOPS and “2x inference” can both be true of the same silicon without either explaining the other.
There is a plausible mechanism, and it is worth stating fairly. Small language model decoding is bound by memory bandwidth rather than arithmetic, while vision pipelines are bound by arithmetic. A part that gains 50% more GPU cores and 17.6% more bandwidth simultaneously can compound those gains on a workload that was previously stalling on both. Whether that reaches 2x depends entirely on the workload NVIDIA measured, and that workload has not been named.
How the Jetson Orin Nano 2 Claim Changed as It Travelled
The most instructive thing about this announcement is not the silicon. It is watching one hedged sentence lose its hedge across four publications in a week.
| Source | Wording used | What it measures |
|---|---|---|
| NVIDIA product page | “nearly twice the inference performance” | inference, hedged |
| NVIDIA press release | “2x the inference performance” | inference, unhedged |
| SiliconANGLE headline | “doubles compute” | compute, not inference |
| StorageReview | “78 TOPS, effectively doubling the processing throughput” | attached to the TOPS figure |
From “nearly twice” to “effectively doubling”
NVIDIA’s own Jetson Orin product page is the most conservative version in circulation. It says the module “delivers nearly twice the inference performance of its predecessor in the same form factor and 40W power envelope, consuming 40% less power”. The press release drops “nearly”. The trade coverage drops “inference”. By the fourth step the doubling has been bolted onto the 78 TOPS figure, which is the one number in the announcement that demonstrably did not double.
None of those steps is dishonest on its own. Each is a reasonable compression of the sentence before it. The cumulative effect is a widely repeated claim that the primary source does not make.
Why TOPS and inference are not the same measurement
TOPS is a peak theoretical rate: cores multiplied by operations per clock multiplied by clock speed. Inference performance is throughput on a real model, which is gated by memory, by how well the model fits the tensor units, and by whether the pipeline stalls. The two numbers can move independently in either direction, and conflating them is the single most common error in edge silicon coverage.
For anyone specifying hardware, the practical rule is that a TOPS number tells you the ceiling and nothing about the floor. Benchmarks on your own machine learning workload are the only figure that transfers.
The Jetson Orin Nano 2 Power Claim Contains Two Different Numbers
The efficiency story was reported everywhere as a single achievement. It is actually one measurement stated twice, plus a second fact almost nobody printed.
The 40% figure and the 15W figure are the same measurement
NVIDIA says the Jetson Orin Nano 2 matches its predecessor’s performance in 15W mode while consuming 40% less power. ServeTheHome states the same thing from the other direction: the 15W mode matches the original module’s 25W performance. Divide 15 by 25 and you get 0.6 — exactly the 40% reduction. These are not two independent results; they are one benchmark expressed in two units.
That is a real and useful efficiency gain. It is one data point at one operating mode, not a general 40% saving across the part’s range.
The power envelope moved up, not down
Here is the fact that went almost entirely unreported. The Orin Nano Super’s published power modes are 7W, 15W and 25W. The Jetson Orin Nano 2 is reported at 15W to 40W. The floor more than doubled and the ceiling rose by 60%.
Why the floor matters more than the ceiling on a drone
On a delivery drone, every watt is flight time. A designer who built around the Orin Nano Super’s 7W mode to stretch endurance does not have that option on the new module, whose reported minimum is 15W. The efficiency headline is measured at a mode that used to be the middle of the range and is now the bottom of it.
This is not a criticism of the part. A higher ceiling is exactly what you want in a factory enclosure with mains power and a fan, which is where several of the named partners are building. It does mean the same module reads as an upgrade in one deployment class and a constraint in another, and the drone framing in the headlines is the class where the constraint bites.
What the Jetson Orin Nano 2 Silicon Tells You
The architecture is the quiet answer to the arithmetic question, and it rules out the most obvious explanation for a doubling.
Ampere on Samsung 8nm, not a new architecture
ServeTheHome reports the Jetson Orin Nano 2 as new Ampere silicon on Samsung’s 8nm process — the same architecture generation and the same node family as the outgoing part. Blackwell-class tensor units are not in play here. That is consistent with a 16% TOPS gain and inconsistent with a wholesale generational jump.
The FP4 question answers itself
The cleanest way to double tensor throughput on modern NVIDIA hardware is a narrower numeric format. As ServeTheHome notes, “FP4 support alone brings the potential for twice the tensor throughput as FP8/INT8” — and Ampere does not have it. So the one mechanism that would straightforwardly deliver 2x is the mechanism this part does not appear to include.
That leaves core count, bandwidth and software as the sources of the gain. All three are real. None of them is documented at 2x.
The speculation worth ignoring
Comment threads raised whether this is a repurposed T239, the chip in Nintendo’s Switch 2. Smith’s own answer is the sensible one: Jetson modules carry a 1GbE MAC and GPIO that a games console has no reason to include, so an outright reuse would be more surprising than a new design. It is an entertaining theory with nothing behind it, and it has no bearing on the specification.
Jetson Orin Nano 2 Drone Claims Against What Wing Actually Said
The headline case for this module is drones, and Wing, Alphabet’s delivery drone subsidiary, is the marquee name. The quote in NVIDIA’s release is more careful than the coverage built on it.
“Exploring” is not “deploying”
Dinuka Abeywardena, Wing’s head of perception, is quoted saying: “Wing is exploring Jetson Orin Nano 2 to give us a path to more responsive, energy-efficient drones that can help make deliveries quicker and more dependable for customers.” Every load-bearing word in that sentence is conditional. Exploring, a path to, can help.
Some coverage carried a shorter Abeywardena line — “Drone delivery depends on AI that can enable fast, reliable understanding of the real world” — which is a statement about drone delivery in general and commits Wing to nothing about this module at all.
The drones in the headline fly the predecessor
Wing’s current delivery fleet runs the Jetson Orin Nano Super. No production timeline has been published for moving to the new module, and none could reasonably exist for a part that ships no earlier than 2027. The physical AI genuinely flying on Wing aircraft today is running on the previous generation.
Matic’s chief executive is similarly framed in the future conditional. Navneet Dalal says that “with Jetson Orin Nano 2, Matic can run state-of-the-art AI models at the edge in a compact home robotics platform built for real-time perception, interaction and navigation” — a description of intended capability, not of a shipping product.
The Jetson Orin Nano 2 Partner List Is the Strongest Evidence
Set the performance claim aside and the ecosystem is the part of this announcement that holds up best. NVIDIA named four early adopters and twenty-two hardware partners, which is a meaningful commitment of other people’s engineering budgets.
| Partner | Sector | Stated position |
|---|---|---|
| Wing (Alphabet) | Delivery drones | “Exploring”; fleet runs the predecessor |
| Matic | Home robots | Future capability, no ship date |
| Cognex | Machine vision | Ships In-Sight 6900 on a 157 TOPS Jetson |
| Doosan Bobcat | Heavy equipment | Named adopter, no detail published |
| JWIPC | Industrial computing | JEA fanless 15W vision system in development |
| Aptiv | Sensing and integration | Sensors, compute and lifecycle support |
Cognex ships on a bigger Jetson than this one
Cognex is named as a Jetson Orin Nano 2 adopter, and its shipping product tells you something about where the module actually lands. The In-Sight 6900 vision controller, launched in April 2026, runs 157 TOPS of Jetson silicon — Orin NX class, roughly double the new module’s rating. Cognex uses TensorRT to hold microsecond-level timing for classification, pixel-level segmentation and defect detection on production lines.
That is the useful context for “entry-level”. The Jetson Orin Nano 2 is not the part that runs a demanding computer vision controller; it is the part that makes a cheaper one possible. Cognex’s own survey found 57% of respondents already using AI in machine vision operations with a further 30% planning near-term deployment, which is the demand this tier is aimed at.
The factory-floor partners are the concrete ones
JWIPC’s planned JEA series is the most specific product description attached to the module anywhere: a fanless industrial vision edge computer running at 15 watts and supporting two to four industrial cameras, plus a separate autonomous mobile robot and AGV controller built on the Isaac and ROS ecosystems. Fanless at 15W is exactly the design the new power floor permits and the old 7W mode did not require.
Aptiv and the Jetson Orin Nano 2 Production Gap
Aptiv published its own release the same day, and it is the most revealing document in the set because it describes what is missing rather than what is new.
What Aptiv is actually supplying
Aptiv is offering four things around the module: PULSE surround-view cameras with ultrashort-range radar for 360-degree sensing, its Gen 8 radar for 4D perception, robotics compute solutions spanning Jetson Orin Nano 2, Orin NX and AGX Orin, and Wind River software with lifecycle support for maintenance, security and compliance.
Jay Bellissimo, Aptiv’s senior vice president for intelligent systems, software and services, framed the reason plainly: “Physical AI is moving quickly from experimentation to deployment, and device manufacturers need partners who can help them scale successfully.” Sam Palmisano, vice president of global product development for adjacent markets, was more explicit about the gap: “Developers are building increasingly capable systems on Jetson, but scaling those systems requires a broader production ecosystem.”
Read that as a statement about what a module is not
Both quotes describe the same problem. A compute module is not a product. Sensors, thermal design, functional safety, security patching and a decade of lifecycle support are what separate a working prototype from something a manufacturer can ship, and a tier-one automotive supplier has stepped in specifically to sell that layer. The fact that this layer needed announcing alongside the silicon is a fair measure of how much work sits between the Jetson Orin Nano 2 and a deployed drone.
The Price NVIDIA Did Not Announce
“Entry-level” is a price category, and the price is the one specification missing from an announcement built on that phrase.
What the predecessor’s $249 established
The Orin Nano Super’s significance was never its 67 TOPS. It was 67 TOPS at $249, replacing 40 TOPS at $499 — the cost per TOPS fell from about $12.48 to about $3.72 in a single generation, a 70% reduction. That is what made the previous module a default choice for developers and small manufacturers.
The arithmetic a buyer can do in advance
Holding the predecessor’s $3.72 per TOPS at 78 TOPS implies a developer kit near $290. Holding the $249 headline price instead implies $3.19 per TOPS, a further 14% improvement. Those two figures bracket the range in which the Jetson Orin Nano 2 continues the trend that made the line matter, and anything materially above them makes the entry-level framing hard to sustain.
That is a calculation from published prices and published ratings, not a forecast. NVIDIA may well land inside it. The point is that with no price and an H1 2027 date, nobody outside NVIDIA can currently evaluate the claim the announcement is built around.
What the Jetson Orin Nano 2 Software Stack Adds
The software half of the announcement is less contested than the hardware half, and for most teams it will matter more.
Models named, and what they imply
NVIDIA lists Jetson agent skills alongside its own Cosmos and Nemotron models, and points at Gemma 4 and Qwen 3 as third-party models optimised for edge inference. Two of the four are open-weight models from other vendors, which is a signal about where NVIDIA expects the value to sit: the module is a runtime for whatever wins, not a vehicle for NVIDIA’s own AI models.
That matters for planning. A team building on the Jetson Orin Nano 2 is not betting on one model family, and the 8GB memory ceiling — unchanged from the predecessor — is the real constraint on which models fit, regardless of how fast the tensor units run.
The unchanged 8GB is the specification to watch
Memory capacity did not move. On a part positioned for running small language models and vision language models at the edge, 8GB determines what you can load before quantisation, and no amount of extra bandwidth changes that ceiling. If your workload did not fit on the previous module, it will not fit on this one.
How to Read a Jetson Orin Nano 2 Spec Claim
The method that produced this article transfers to any edge silicon announcement, and it takes about ten minutes.
Separate the vendor page from the press release
Vendor product pages are written by people who expect to be held to them and are routinely more hedged than the accompanying press release. Where the two differ, as they do here on “nearly twice” against “2x”, the product page is the version to quote.
Ask which number the multiplier is attached to
A performance multiple must attach to a named workload at a named precision. If the coverage attaches it to a TOPS figure, check that figure against the previous generation before repeating it. Here that check takes one division: 78 over 67.
Check whether the efficiency claim is one measurement or two
Percentage power savings and matched-performance mode claims are frequently the same benchmark stated twice. Divide the two power figures; if the ratio equals the claimed saving, it is one result.
Read the partner verbs
“Deploying”, “shipping” and “in production” mean something. “Exploring”, “evaluating” and “a path to” mean something different, and the difference is usually a year or more of engineering. It is the cheapest signal in any launch release.
Frequently Asked Questions About the Jetson Orin Nano 2
When can I actually buy one?
NVIDIA says the first half of 2027, as both a production module and a developer kit. No narrower date and no price have been published.
Is the 2x inference claim false?
No. It is unverified. NVIDIA’s own product page states it as “nearly twice”, and no published specification for the Jetson Orin Nano 2 individually doubles, so the claim depends on a benchmark workload NVIDIA has not named.
Should I design a drone around it now?
Design around the 15W power floor rather than the headline efficiency figure, and note the predecessor’s 7W mode has no successor. For anything flying before 2027, the Orin Nano Super remains the shipping option — it is what Wing flies today.
Does it run larger models than the previous module?
Not by capacity. Memory stays at 8GB, so the set of models that fit is unchanged. What improves is how fast the models that already fit are likely to run.
Is this a good part despite all this?
On the evidence, yes. More cores, more bandwidth, a stronger CPU, a broad partner ecosystem and a credible efficiency gain at 15W is a solid generational step. The gap is between that and the way it has been described.
What the Jetson Orin Nano 2 Means for Your Edge AI Plans
Three things carry across from this announcement to anyone specifying edge hardware over the next eighteen months.
First, the module is real and the ecosystem behind it is real, but the timeline is 2027 and the price is unknown. Any 2026 project needs a shipping part, and the shipping part is the predecessor that Wing and the current partner base are already using. Treat the Jetson Orin Nano 2 as a roadmap item, not a procurement option.
Second, the specification that constrains you is the one that did not change. 8GB of memory sets the model ceiling on this tier, and the extra bandwidth makes what fits run faster without making more fit. Size your models against that number, not against the TOPS rating.
Third, the arithmetic in this article is the arithmetic to run on the next announcement too. Every claim checked here was checkable from published pages in a few minutes, and the discrepancies were not hidden — they emerged from ordinary compression as a hedged sentence moved between publications. Whether you are evaluating a robotics module or any other AI tools purchase, the discipline of tracing each number to the page that first printed it is what separates a specification from a headline.
The Jetson Orin Nano 2 looks like a good module. It deserves to be evaluated on what NVIDIA published about it, which is a 16% TOPS increase, 50% more GPU cores, 18% more bandwidth, a measured efficiency gain at one operating mode, and a shipping date eight months out. That is a defensible product. It is simply not the same statement as “doubles”.
References and Further Reading
NVIDIA Jetson Orin Nano 2 brings physical AI to drones and robots – AI News
NVIDIA Jetson Orin product page and module specifications
NVIDIA Announces Jetson Orin Nano 2: Entry-Level Edge Board Gets New Ampere Silicon – ServeTheHome
Aptiv: Accelerating Production-Ready Physical AI with Support for NVIDIA Jetson Orin Nano 2
Nvidia Jetson Orin Nano 2 targets factory-floor edge AI – IoT Tech News
NVIDIA Jetson Orin Nano 2: 78 TOPS and 2x the Inference of Its Predecessor – StorageReview
Nvidia’s $249 AI development board promises 67 TOPS – Tom’s Hardware
$249 NVIDIA Jetson Orin Nano Super Developer Kit – CNX Software
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.