NVIDIA Jetson Thor Review: 130W Rated, 330W Transient

NVIDIA rates the Jetson AGX Thor T5000 at 40W to 130W on its product page. On NVIDIA's own developer forum, in thread 380561, NVIDIA staff put the transient figure much higher: "the 1ms system EDP can reach up to 330W for Jetson T5000", with a recommendation to use a supply with enough headroom for that demand. The thread came out of developers watching T5000 modules power themselves off mid-inference. The reporter in that thread measured peak draw around 28V at 5.5A, about 154W, during model loading.

That is the review in one paragraph. A battery or PSU sized to the 130W ceiling on the marketing page can brown out the robot under exactly the workload the part is sold for. The 330W number is 2.5x the rated ceiling, and it does not appear on the product page.

The second theme is the gap between peak throughput and delivered throughput. Thor is currently the Jetson that holds a 120B-parameter model in unified memory. It is also the part whose 7.5x compute claim rests on comparing two different numeric formats, and whose own sibling module undercuts the headline.

Companion reading: edge AI robot compute, foundation models and VLA robotics, humanoid robot hardware, robot cybersecurity, Boston Dynamics Spot review.

Table of contents

1. Key takeaways
2. The power envelope problem
3. The 7.5x claim and the bandwidth ceiling
4. The tier question: T5000 against T4000
5. Security disclosures
6. Thor against its siblings and Orin
7. Deployment reality and the Xavier squeeze
8. Who should buy it
9. Verdict
10. FAQ

The power envelope problem

NVIDIA's Jetson Thor product page gives the T5000 a 40W to 130W range and 2,070 FP4 (sparse) TFLOPS. That range is the sustained configuration envelope. The transient figure is separate, and NVIDIA staff stated it directly in developer forum thread 380561: the 1ms system EDP can reach up to 330W for the T5000, and the guidance is to specify a supply with headroom for that.

The practical consequence lands on carrier board and power system design rather than on benchmark scores. The failure mode reported in that thread is the module shutting itself down mid-inference on a supply that satisfies 130W and nothing more. The measured peak in that report, about 154W at 28V and 5.5A during model load, is already well past the rated ceiling and still far below the 330W transient design point.

Two caveats on the 330W number. It is a 1ms system-level EDP figure, so it describes what the power system must survive, not what the module dissipates continuously. And it is a single forum statement from NVIDIA staff. It is the strictest published number, which makes it the one to design against.

The 7.5x claim and the bandwidth ceiling

NVIDIA's developer site says Thor modules deliver "over 7.5x higher AI compute than NVIDIA AGX Orin, with 3.5x better energy efficiency". The 7.5x compares Thor's 2,070 FP4 sparse TFLOPS against AGX Orin's 275 INT8 TOPS. Different numeric formats and a sparsity assumption on one side only, so it is a peak-versus-peak number rather than a like-for-like one.

The number that governs token generation is memory bandwidth. Thor's 273 GB/s against AGX Orin 64GB's 204.8 GB/s is a 33 per cent lead. Autoregressive decode reads the full weight set per token, so on a model that fits both parts, that 33 per cent is closer to the ceiling on throughput gain than 7.5x is.

NVIDIA's own lineup makes the same point without any third-party benchmark. The T4000 lists 1,200 FP4 sparse TFLOPS, 58 per cent of the T5000's, with 64GB instead of 128GB, in a 40W to 70W envelope. It runs the same 273 GB/s memory bandwidth as the flagship. Two modules with a 1.7x gap in peak compute share the bandwidth that bounds LLM decode.

What only the T5000 does is hold the large model. 128GB of unified LPDDR5X is double the T4000 and double AGX Orin 64GB. If your policy or VLM does not fit in 64GB, capacity is the reason to buy this part, and the throughput argument is secondary.

One I/O caveat worth your own testing. In developer forum thread 373776, users report the T5000's QSFP28 MGBE capping around 12.7 Gbps on a single stream and around 12 Gbps aggregate across four parallel streams, against an expected 25 Gbps per link. The thread is substantially unanswered, with no NVIDIA confirmation and no published root cause. Test it yourself before planning multi-camera aggregation over that link.

The tier question: T5000 against T4000

NVIDIA does not publish module prices on the Jetson Thor product page or the Jetson modules page, so any budget has to come from a distributor quote rather than from a public list price. Treat any figure you have seen quoted as a dated band from one reseller, not a fixed price.

The specification gap is public, and it is the part that should drive the tier decision:

The T5000 buys 1.7x the peak FP4 compute and 2x the memory. It does not buy any memory bandwidth. For a workload that fits in 64GB and spends its time generating tokens, the T4000 is the part that matches the constraint, at roughly half the peak power. For a workload that does not fit in 64GB, the T4000 is not a candidate at any price, and the comparison ends there.

Security disclosures

Two CVEs published 31 March 2026 name Jetson Thor in their affected product lists. Both are recorded against Jetson Linux 38 version 38.2 (and against Jetson Linux 35 before 35.6.4 and Jetson Linux 36 before 36.5 for the older Jetson families).

Two notes for operators.

The CVE records list fixed versions for the Jetson Linux 35 and 36 branches (35.6.4 and 36.5). For the Jetson Linux 38 branch they name 38.2 as the affected version and do not name a fixed release, so confirm your target release against NVIDIA's bulletin rather than assuming a specific 38.x fix level.

NVIDIA and NVD disagree on CVE-2026-24154, in NVIDIA's more conservative direction: NVIDIA scores it 7.6 High with a scope change, NVD scores it 6.8 Medium without one. The vector is identical otherwise and both require physical access.

The physical-access precondition deserves less comfort than usual on this class of product. A humanoid or an AMR is an unattended machine in a semi-public space, which is a materially easier physical-access story than a rack in a locked room. No public exploit or in-the-wild exploitation was found for either CVE.

One CVE from the same March 2026 batch, CVE-2026-24148, is frequently grouped with these two. Its affected product list names only the Jetson Xavier and Jetson Orin series, on Jetson Linux 35 and 36. It does not list Jetson Thor and it does not list Jetson Linux 38 in any form, so it is not treated as a Thor issue here.

Thor against its siblings and Orin

Read that as one decision. If the model fits in 64GB and you have a working stack, AGX Orin remains available through January 2032 and the industrial variant through July 2033. If it fits in 64GB and you want the newer architecture at half the power ceiling, the T4000 gives you the flagship's memory bandwidth. If you need 128GB of unified memory on a robot, the T5000 is the part, and you are paying the tier premium for capacity.

Deployment reality and the Xavier squeeze

The lifecycle side of the decision is documented on NVIDIA's Jetson product lifecycle page, and it pushes in one direction. Jetson AGX Xavier 32GB, Xavier NX (8GB and 16GB), TX2 NX and TX2i are all listed as available through July 2027. Jetson AGX Orin 32GB and 64GB run through January 2032, and Jetson AGX Orin Industrial through July 2033.

For anyone still shipping on Xavier, that is a migration with a fixed deadline: Orin buys the longest runway, Thor buys the newest architecture and the memory capacity. Orin's January 2032 date is far enough out that "stay on Orin" is a legitimate answer for a program starting today, not a stopgap.

Claims about specific named robot programs shipping on Thor, about JetPack version transitions, and about unannounced module pricing appear widely in secondary coverage. None of that could be verified against a primary source for this edition, so it is left out. Treat partner-logo slides as intent until a shipping product names the module.

Who should buy it

Buy the T5000 if your policy or VLM genuinely needs more than 64GB of unified memory on the robot, and you are designing your own carrier with a supply budgeted for a 330W transient rather than a 130W sustained load. Multi-camera perception stacks running large world models fit this.

Buy the T4000 instead if your models fit in 64GB. You keep the same 273 GB/s memory bandwidth, you give up peak FP4 compute and half the memory, and you halve the top of the power envelope.

Stay on AGX Orin if your workload fits 64GB and you already have a working stack. It is listed as available through January 2032, and the industrial variant through July 2033.

Re-quote before you budget either module. NVIDIA does not publish Thor module prices on its product pages, so a number from a reseller is a dated quote and nothing more.

Verdict

Jetson Thor is real, shipping, and the Jetson that holds a 120B-parameter model in unified memory. The 7.5x headline is an FP4-sparse-versus-INT8 peak comparison, and the number that actually bounds token generation is the 33 per cent memory bandwidth lead over AGX Orin. NVIDIA's own T4000 makes that concrete: 58 per cent of the peak compute on identical bandwidth. Buy the T5000 for the 128GB.

Then size the power system for the 330W transient that NVIDIA's own engineers describe rather than the 130W on the product page, and check your Jetson Linux release against the March 2026 CVEs before the robot leaves the building.

Frequently asked questions

How much power does a Jetson T5000 actually need? The product page rates it 40W to 130W. NVIDIA staff on the developer forum put the 1ms system EDP at up to 330W and advise specifying a supply with headroom for it. One developer reported roughly 154W peak at 28V during model loading, with modules shutting down on undersized supplies.

Is Thor really 7.5x faster than Orin? The 7.5x compares Thor's 2,070 FP4 sparse TFLOPS to AGX Orin's 275 INT8 TOPS, which are different numeric formats. For token generation the governing spec is memory bandwidth, where Thor's 273 GB/s leads AGX Orin's 204.8 GB/s by 33 per cent.

What is the difference between the T5000 and the T4000? The T5000 has 2,070 FP4 sparse TFLOPS and 128GB; the T4000 has 1,200 and 64GB, in a 40W to 70W envelope. Both run 273 GB/s. The T5000's real advantage is memory capacity.

Which CVEs affect Jetson Thor? Two published 31 March 2026 list Jetson Thor: CVE-2026-24154 (initrd command line injection, physical access, NVIDIA 7.6 / NVD 6.8) and CVE-2026-24153 (nvluks trusted application not disabled, information disclosure, Medium). Both are recorded against Jetson Linux 38.2.

Does CVE-2026-24148 affect Thor? Its affected product list names only the Jetson Xavier and Jetson Orin series, on Jetson Linux 35 and 36. Jetson Thor and Jetson Linux 38 do not appear in it.

How long is AGX Orin supported? NVIDIA's Jetson product lifecycle page lists AGX Orin 32GB and 64GB as available through January 2032, and AGX Orin Industrial through July 2033. The Xavier family and TX2 NX and TX2i run through July 2027.

Are the QSFP28 links full speed? Developer forum reports put the T5000's QSFP28 MGBE around 12.7 Gbps on a single stream and around 12 Gbps aggregate over four streams, against 25 Gbps per link expected. The thread is unanswered and there is no NVIDIA confirmation, so validate it on your own hardware.