Jetson Thor

NVIDIA Jetson Thor (also marketed as Jetson AGX Thor) is a Blackwell architecture edge AI computing module developed by NVIDIA that delivers up to 2,070 FP4 teraflops of AI compute in a 40 to 130 watt power envelope to act as an on-robot "brain" for humanoid systems, autonomous machines, and "physical AI" workloads. The flagship Jetson T5000 module pairs the Blackwell GPU with a 14-core Arm Neoverse-V3AE CPU and 128 GB of LPDDR5X memory, and NVIDIA reports it provides 7.5 times the AI performance and 3.5 times the energy efficiency of the previous Jetson Orin AGX generation. ^[1][2] The platform was previewed by NVIDIA CEO Jensen Huang at GTC 2024 alongside Project GR00T and reached general availability on August 25, 2025, with a developer kit starting at US$3,499. ^[1] At launch Huang said: "With unmatched performance and energy efficiency, and the ability to run multiple generative AI models at the edge, Jetson Thor is the ultimate supercomputer to drive the age of physical AI and general robotics." ^[1] Jetson Thor sits at the top of the Jetson product family and is intended to bring data center class generative AI workloads, including large language models, vision language models, and vision language action policies, onto autonomous machines such as humanoid robots, autonomous mobile robots, drones, surgical robots, and inspection systems. ^[2] NVIDIA expanded the family on July 15, 2026 by announcing the smaller Jetson T3000 and T2000. The T3000 is rated at 865 sparse FP4 TFLOPS with 32 GB of memory, while the T2000 is rated at 400 sparse FP4 TFLOPS with 16 GB. Both remained in development, with availability scheduled for the first quarter of 2027. ^[22][23]

Background and strategic context

The Jetson Thor series belongs to NVIDIA's Jetson family of embedded AI computing platforms, which has been used since 2014 to bring GPU acceleration to autonomous machines, edge servers, and research robotics. By the mid 2020s, the rise of large multimodal models and generalist humanoid robotics created a need for far more on-device compute than the Ampere based Jetson Orin family could provide. NVIDIA framed this transition as the "physical AI" era, in which robots reason about the physical world using foundation models rather than purely deterministic perception and control stacks. ^[2]

When was Jetson Thor announced and released?

Thor was first unveiled on March 18, 2024 at GTC 2024 in San Jose. During his keynote, Jensen Huang introduced Project GR00T, a general purpose foundation model for humanoid robots, and announced that Thor would serve as the on-robot compute platform for next generation humanoids. ^[5] NVIDIA initially described Thor as a system on a chip incorporating a next generation Blackwell GPU with a transformer engine, an integrated functional safety processor, and 100 Gb networking, delivering 800 teraflops of 8-bit floating point performance to run multimodal generative AI models like GR00T. ^[5] Over the following 17 months NVIDIA refined the product into an initial two-module lineup and a developer kit, with revised peak figures based on the FP4 sparse format introduced by Blackwell, raising the headline number from 800 FP8 TFLOPS to 2,070 FP4 sparse TFLOPS. The shipping lineup began with T5000 and T4000; NVIDIA added the lower-memory T3000 and T2000 to its development roadmap in July 2026. ^[22][23]

General availability arrived on August 25, 2025, when NVIDIA announced that the Jetson AGX Thor Developer Kit and Jetson T5000 production module were shipping through authorized distributors worldwide. ^[1] The launch was positioned as the foundation for what NVIDIA called "the age of general robotics," with early adopters spanning humanoid developers, industrial OEMs, surveying firms, and medical device makers. ^[1] At the time of the launch, NVIDIA reported more than 2 million developers active on its robotics stack, over 7,000 customers already using Jetson Orin hardware, and more than 150 hardware, software, and sensor partners offering production ready solutions. ^[1] A second module, the Jetson T4000, was disclosed for release in 2026 at a lower price and reduced power envelope. ^[1]

Announcement timeline

Scroll sideways for more →

Hardware architecture

The released Jetson T5000 and T4000 are implemented as 87 by 100 mm systems on module (SoMs) with 699-pin board-to-board connectors. ^[3][6] The SoM is not pin compatible with Jetson AGX Orin because Thor uses a different connector, a higher peak power budget (up to 130 W versus 60 W for AGX Orin), and additional high speed interfaces such as PCIe Gen5 and 25 GbE. ^[4] NVIDIA publishes a dedicated Orin to Thor migration application note (DA-11926-001) to help integrators move existing carriers to the new platform. ^[4]

GPU subsystem

The GPU is based on NVIDIA's Blackwell architecture, the same family used in data center products such as the B100 and B200 but configured for an embedded power envelope. ^[2] Key GPU traits on the T5000 module include:

• 2,560 CUDA cores running at up to 1.57 GHz
• 96 fifth generation Tensor Cores
• 10 Texture Processing Clusters (TPCs)
• Multi-Instance GPU (MIG) capability that allows the single GPU to be partitioned into isolated instances for mixed criticality robotics workloads
• A second generation Transformer Engine with dynamic precision switching between FP4 and FP8
• Native FP4 quantization, which is the precision behind the headline 2,070 TFLOPS rating

At the T4000 tier the GPU is cut down to 1,536 CUDA cores, 64 Tensor Cores, and 6 TPCs, with a lower 1.53 GHz peak clock. ^[3]

Peak AI performance

NVIDIA publishes peak AI throughput at multiple precisions. The numbers below are taken from the Jetson Thor Series Modules datasheet (DS-11945-001) and the Jetson Thor launch technical blog. ^[2][3]

Scroll sideways for more →

The 7.5 times improvement over Jetson AGX Orin that NVIDIA quotes is calculated using FP4 sparse on Thor against INT8 sparse on Orin, the highest published precision for each generation respectively. ^[2]

CPU subsystem

Thor switches from the Arm Cortex-A78AE used in Orin to the Arm Neoverse-V3AE, an automotive enhanced variant of the Neoverse V3 server class core. ^[8] The T5000 ships with 14 Neoverse-V3AE cores; the T4000 has 12. ^[3] Each core has 1 MB of L2 cache and the cluster shares 16 MB of L3 cache, with a peak frequency of 2.6 GHz. ^[3] The use of Neoverse means that JetPack 7 software aligns with the Arm Server Base System Architecture (SBSA), the same firmware and interface specification used on Arm based server silicon. ^[8] This alignment lets the same CUDA 13 binaries run on Arm servers and on Jetson Thor without recompilation, which is a deliberate move toward unifying the data center and edge developer experience. ^[8]

Memory and storage

The T5000 carries 128 GB of unified LPDDR5X memory on a 256-bit bus at 4,266 MHz, delivering approximately 273 GB/s of bandwidth between the CPU, GPU, and accelerators. ^[3] The T4000 drops to 64 GB on the same 256-bit bus and the same bandwidth. ^[3] Storage is provided externally through an M.2 Key M slot for NVMe SSDs (the developer kit includes a 1 TB drive) and through USB 3.2 mass storage. The 128 GB pool is large enough to keep models up to roughly 100 billion parameters resident with INT4 or FP4 quantization, which is why NVIDIA markets Thor as bringing "frontier scale" inference to the edge. ^[2]

Accelerators and media engines

In addition to the GPU, Thor integrates the following on-die accelerators:

• Third generation Programmable Vision Accelerator (PVA 3.0) for traditional computer vision tasks such as feature tracking and stereo correspondence
• Optical flow accelerator
• Always on DSP for low power audio and sensor preprocessing
• Two NVENC video encoders and matching NVDEC decoders on the T5000

Unlike Jetson AGX Orin, which paired the GPU with two dedicated NVIDIA Deep Learning Accelerators (NVDLA v2), the Jetson Thor SoC moves the heavy lifting onto the much larger Blackwell GPU plus the PVA. ^[4] NVIDIA's developer documentation treats the PVA and Tensor Cores as the primary inference paths on Thor, and the public datasheet lists deep learning acceleration in terms of GPU TOPS and a separate PVA pipeline rather than a discrete NVDLA block. ^[3] Developers migrating NVDLA workloads from Orin are directed to run them on the GPU or PVA on Thor. ^[4]

Multimedia and display

• Video decode: up to four 8K at 30 fps streams, or up to ten 4K at 60 fps streams; H.265, H.264, AV1, VP9, and VP8 are supported
• Video encode: up to six 4K at 60 fps streams; H.265 and H.264
• Camera input: up to 32 simultaneous virtual channels through MIPI CSI-2 or Holoscan Sensor Bridge
• Display: up to four independent displays at resolutions up to 7,680 by 4,320 at 30 Hz via HDMI 2.1 and DisplayPort 1.4a

I/O, networking, and sensor bridge

Networking is a central upgrade. The T5000 exposes four 25 GbE interfaces (aggregated via a QSFP28 connector for 100 Gb total) plus a 5 GbE RJ45 port, while the developer kit adds Wi-Fi 6E and Bluetooth via an M.2 Key E module. ^[11] PCIe Gen5 lanes are configurable up to x8 plus x4 plus x2, supporting NVMe SSDs, accelerator cards, and FPGA peripherals. ^[3]

Industrial interfaces include CAN bus, multiple UARTs, SPI, I2C, I2S, and dozens of GPIO pins. The platform also introduces the Holoscan Sensor Bridge, an Ethernet based time synchronized sensor input path originally developed for the Holoscan medical imaging SDK. ^[13] Sensor Bridge lets cameras and lidar units stream raw pixel and point cloud data over 10 Gb or 25 Gb Ethernet with low and deterministic latency, decoupling the sensor head from the compute module and enabling longer cable runs than MIPI CSI-2 supports. ^[13]

Module variants and developer kit

The Jetson Thor family launched with two production modules and one developer kit. The T5000 is the flagship; the T4000 is a lower cost and lower power option targeted at workloads that do not need full Blackwell scale. ^[3] The announced T3000 and T2000 extend the roadmap below those two shipping variants, with smaller memory capacities and lower peak compute. ^[22][23]

Scroll sideways for more →

The Jetson AGX Thor Developer Kit packages a Jetson T5000 module on a reference carrier with active cooling, a 1 TB NVMe SSD in an M.2 Key M slot, a Wi-Fi 6E plus Bluetooth M.2 Key E module, HDMI 2.0b, DisplayPort 1.4a, 5 GbE, four 25 GbE via QSFP28, multiple USB 3.2 Type-A and Type-C ports, CAN, UART, I2S audio headers, JTAG, and a 140 W power adapter. ^[11] The assembled kit measures roughly 243 by 112 by 57 mm. NVIDIA published the developer kit price at US$3,499, with preorders opening in August 2025 and shipments starting in Q3 2025 and continuing into November 2025 through resellers such as Arrow, RS Components, and Seeed Studio. ^[1][11]

T3000 and T2000 modules

NVIDIA announced Jetson T3000 and T2000 on July 15, 2026 as in-development Thor modules for robotics and edge AI systems that do not need the memory capacity or peak compute of T5000 and T4000. The announcement tied the smaller configurations to both system size and memory cost. Only T3000 was explicitly described by NVIDIA as roughly half the size and power of T5000. ^[22] NVIDIA's official roadmap places both modules in a mainstream tier, below the released T5000 and T4000, and marks them for Q1 2027. ^[23]

The table below separates specifications published directly by NVIDIA from preliminary details supplied by launch partners and technical publications. ^{[26][27][28][29][30]} NVIDIA had not released full module datasheets, design guides, connector pinouts, ordering part numbers, or prices for T3000 and T2000 when it announced them. NVIDIA forum staff said the engineering documents would be published closer to release. ^[28][35]

Scroll sideways for more →

The T3000 retains the 273 GB/s memory bandwidth NVIDIA publishes for T4000 and T5000, but its 32 GB capacity is one quarter of T5000's and its 865 sparse FP4 TFLOPS rating is below the 1,200 and 2,070 TFLOPS ratings of T4000 and T5000. ^[22][27] NVIDIA said T3000 could nevertheless approach T5000 inference performance on selected multimodal workloads. Its launch graphic reported the following normalized results, all produced through emulation and explicitly marked subject to change. ^[22]

Scroll sideways for more →

These figures describe seven selected model workloads under NVIDIA's emulation setup. The published chart gives normalized ratios rather than absolute throughput and does not turn the result into a general equivalence between T3000 and T5000. The lower memory capacity also limits which model sizes and concurrent workloads can fit without additional quantization, offloading, or memory optimization. Independent measurements on production T3000 or T2000 hardware were not possible because the modules were not yet available. ^[22][29]

NVIDIA directed developers to use the T5000-based Jetson AGX Thor Developer Kit to start software work. The company scheduled T3000 emulation for later July 2026 with JetPack 7.2.1, while saying T2000 emulation would follow in a future release. This is an emulation and migration path, not a statement that either lower-tier module can be installed in the existing developer-kit carrier. Hardware compatibility cannot be established until NVIDIA publishes the production mechanical and interface documentation. ^[22][28][35]

The announcement also identified Cosmos 3 Edge, a 4-billion-parameter model built on Nemotron, as an on-device model for T2000 and T3000. NVIDIA said developers could adapt it for specific robots and sensors and deploy it for vision reasoning and robot-policy workloads. ^[24] NVIDIA paired the launch with open-source Jetson agent skills for device diagnostics, memory auditing, inference tuning, model serving, and board-support-package customization across Thor and Orin systems. ^[22][25]

How does Jetson Thor compare with Jetson AGX Orin?

Thor is the direct successor to the Jetson AGX Orin family, which itself replaced Jetson AGX Xavier in 2022. The table below summarizes the principal differences between the highest end Orin and Thor SKUs. ^[16]

Scroll sideways for more →

NVIDIA reports public benchmark speedups for generative AI workloads on the Jetson Thor launch blog. The figures below compare Thor against AGX Orin 64 GB using the latest available kernels at launch. ^[2]

Scroll sideways for more →

In multimodal serving tests with Qwen2.5-VL-3B and Llama 3.2 3B handling 16 simultaneous requests, NVIDIA reported time to first token below 200 milliseconds and time per output token below 50 milliseconds, which it positions as the threshold for natural human and robot interaction. ^[2]

Software stack

Thor uses JetPack 7.0 (and later releases such as 7.2 with CUDA 13.2) as its base software stack. ^[7] JetPack 7 is built on the Linux kernel 6.8 and Ubuntu 24.04 LTS, includes a preemptable real time patch set for deterministic execution, and aligns Jetson firmware and bootloaders with the Arm SBSA specification used on Arm based servers. ^[7][12] The result is that the same Arm CUDA 13.0 binaries used on data center Arm Grace systems will run on Jetson Thor, which simplifies sim to real workflows. ^[8]

Key software components for Jetson Thor include:

• JetPack 7.0: Ubuntu 24.04 LTS, Linux kernel 6.8, real time scheduling, U-Boot, secure boot, and UEFI consistent with SBSA
• CUDA 13.0 unified across Arm targets and Jetson
• TensorRT 10.13 for high performance inference, including FP4 calibration
• cuDNN 9.12 and cuBLAS updated for Blackwell
• VPI 4.0 Vision Programming Interface for accelerated computer vision
• Holoscan SDK for real time sensor pipelines with the Holoscan Sensor Bridge
• CUDA-accelerated ROS 2 packages and the Isaac ROS collection
• DeepStream for multi-stream video analytics
• Native containers via NVIDIA Container Toolkit and prebuilt Triton inference server images

On the model and robotics side, NVIDIA pairs Jetson Thor with several platforms in its Physical AI stack:

• NVIDIA Isaac: A robotics platform that includes Isaac Sim for physics accurate simulation built on NVIDIA Omniverse, Isaac Lab for reinforcement learning and imitation learning, and Isaac ROS for hardware accelerated ROS 2 packages
• Isaac GR00T: A line of open foundation models for humanoid robots; GR00T N1, N1.5, N1.6, and N1.7 were released across 2024 to 2026, each targeting whole body control, manipulation, and vision language action behaviors. GR00T N1.6 integrates Cosmos Reason as its long horizon reasoning brain
• NVIDIA Cosmos: A family of world foundation models for physical AI, including Cosmos Predict for synthetic data, Cosmos Transfer for domain adaptation, and Cosmos Reason, a 7B reasoning vision language model that NVIDIA supports on Jetson Thor for converting vague instructions into step by step plans ^[15]
• NVIDIA Newton: An open source physics engine built with DeepMind and Disney Research, used inside Isaac Sim and announced at CoRL 2025 ^[14]
• NVIDIA Metropolis: Visual AI for smart spaces and industrial sites, with multimodal agents that can run on Thor for on premise inference
• NVIDIA Holoscan: Real time sensor and medical imaging pipelines used in fields such as surgical robotics and medical devices

Jetson Thor supports the major open ecosystem frameworks, including PyTorch, TensorFlow, ONNX Runtime, vLLM, Hugging Face Transformers, llama.cpp, and the Triton inference server, plus model families from Meta (Llama 3 and 4), Alibaba (Qwen 2.5 and Qwen 3), DeepSeek, Google Gemma, and Mistral. ^[2]

Generative AI metrics targeted by NVIDIA

Scroll sideways for more →

Power profiles

Thor exposes a wide range of power and thermal modes that integrators can pick using the nvpmodel tool. NVIDIA describes the envelope from low power deployment to peak compute, with corresponding clock and core configurations. ^[3]

Scroll sideways for more →

The T4000 caps at roughly 70 W, suiting passively cooled and battery powered systems where the smaller GPU and memory pool are sufficient. ^[3] Launch partners list preliminary 70 W and 40 W figures for T3000 and T2000, respectively. NVIDIA itself describes T3000 as operating at roughly half T5000 power but has not yet published the complete power-mode tables for either new module. ^[22][27][30]

What is Jetson Thor used for?

NVIDIA positions Jetson Thor as a single platform spanning several robotics and edge AI segments. Highlighted applications include: ^[1]

• Humanoid robots: Whole body control, manipulation, and language conditioned task execution using GR00T and Cosmos Reason
• Autonomous mobile robots (AMRs): Warehouse logistics, last mile delivery, and inspection robots that run perception, planning, and language models locally
• Industrial automation: Collaborative robots (cobots), pick and place cells, and quality inspection systems with on-device large model inference
• Drones and aerial robotics: Higher autonomy levels in beyond visual line of sight inspection, search and rescue, and agriculture
• Surgical and medical robotics: Real time imaging, augmentation, and assistance using the Holoscan SDK and Holoscan Sensor Bridge
• Autonomous vehicles and off road platforms: Construction, mining, and agriculture equipment using Thor in conjunction with the automotive grade DRIVE AGX Thor variant
• Edge AI servers and digital twins: On-premise inference for vision language models, large language models, and agentic systems where data residency or latency rules out the cloud
• Space and harsh environment compute: Several aerospace integrators are evaluating Thor for satellite and rover payloads due to its single SoM packaging and unified memory architecture

A recurring theme in NVIDIA's launch material is the ability to run multiple AI models concurrently on one device. Thor can simultaneously execute a vision language action policy, a higher level reasoning model, a perception network, and a safety classifier through MIG partitions, an architecture NVIDIA refers to as a "multi-model robot brain." ^[18]

Research deployments

Academic evidence for Jetson Thor currently concerns the shipping generation, not T2000 or T3000. Neither of the new modules was available for independent experiments when announced. Several 2026 preprints nevertheless document the existing platform in on-robot perception and action pipelines:

Scroll sideways for more →

These papers demonstrate workloads that researchers have run on released Jetson Thor hardware. They do not validate NVIDIA's emulated T3000 ratios or establish T2000 performance.

Industry adoption

NVIDIA disclosed a broad list of early Jetson Thor adopters at GA. The roster spans humanoid developers, industrial OEMs, surveying and inspection firms, and medical device makers. ^[1]

Scroll sideways for more →

In aggregate, NVIDIA reported more than 7,000 customers deploying Jetson Orin hardware by 2025, over 150 hardware partners offering production ready solutions, and more than 2 million developers active on its robotics platforms. ^[1] Agility Robotics integrated Jetson into the fifth generation of its Digit humanoid and plans to adopt Jetson Thor as the onboard compute platform for the sixth generation of Digit, while Boston Dynamics is integrating Jetson Thor into its Atlas humanoid. ^[1] Explaining the appeal of the module, Figure CEO Brett Adcock said: "NVIDIA Jetson Thor's server-class performance, delivered within a compact and power-efficient design, allows us to deploy the large-scale generative AI models necessary for our humanoids to perceive, reason and act in complex, unstructured environments." ^[1]

How does Jetson Thor differ from DRIVE AGX Thor?

NVIDIA uses the Thor brand for two distinct but related products. Jetson Thor is the robotics and physical AI variant, while DRIVE AGX Thor is the automotive variant introduced for advanced driver assistance and autonomous driving. ^[8] Both products share the underlying Thor SoC: a Blackwell GPU with the Transformer Engine, Arm Neoverse-V3AE CPU cores, the same memory subsystem, and similar peak compute around 2,000 FP4 TFLOPS. ^[8] They diverge on packaging and software:

• Jetson Thor runs JetPack 7 and the Isaac, Metropolis, Holoscan, and Cosmos stacks. It is optimized for robotics carriers and embedded deployments.
• DRIVE AGX Thor runs NVIDIA DriveOS 7 and adds automotive grade software, deterministic real time scheduling, and certification readiness for ISO 26262 functional safety. Developer kits for DRIVE AGX Thor began shipping in September 2025. ^[9]

NVIDIA's product positioning is that vehicles and robots increasingly share the same software stack, so unifying compute across the two domains lowers cost for OEMs that build both autonomous vehicles and robots. ^[8]

NVIDIA IGX Thor and Halos for Robotics

While Jetson Thor is NVIDIA's robotics edge AI module, NVIDIA also ships a separate industrial grade platform on the same Thor generation called NVIDIA IGX Thor. The two products share the Blackwell based Thor compute foundation but are positioned differently. The Jetson family spans entry to mid range and high end embedded modules for general robotics and autonomous machines, whereas IGX is an industrial grade, enterprise edge AI platform purpose built for industrial, robotics, and medical environments where functional safety and long term enterprise support are required. ^[19] IGX Thor is not the same product as the Jetson Thor module; instead it builds on the Thor SoC and adds dedicated hardware safety features so it can be deployed in safety critical applications. ^[19] The July 2026 roadmap expansion followed the same distinction: NVIDIA announced both Jetson T3000 and a separate IGX T3000. The IGX variant carries the integrated functional-safety features and runs the Halos for Robotics safety stack, while the Jetson variant is the general robotics and edge AI module. ^[22]

The distinguishing feature of IGX Thor is its functional safety architecture. NVIDIA describes IGX Thor as delivering up to 2,070 FP4 TFLOPS of AI compute, with 14 Arm Neoverse cores and 128 GB of memory at 273 GB/s of bandwidth, matching the Thor generation compute envelope, while adding an IEC 61508 SIL 3 capable functional safety island with up to 12,000 DMIPS plus dedicated I/O, power, and clocks. ^[20] NVIDIA states that the IGX Thor SoC includes more than 22,000 safety mechanisms providing diagnostic coverage across the chip, along with logic and memory in system test (IST) for latent fault coverage. ^[20] A separate safety microcontroller (MCU) on the carrier board complements the on SoC safety island, and NVIDIA positions the platform to meet ISO 26262 and IEC 61508 standards. ^[19] This functional safety capability is the central difference between IGX Thor and the standard Jetson Thor robotics module.

Role in NVIDIA Halos for Robotics

On June 22, 2026, NVIDIA announced NVIDIA Halos for Robotics, which it describes as the industry's first full stack, comprehensive functional safety system for robotics and physical AI. ^[20][21] Halos extends safety research and engineering originally developed for autonomous vehicles to industrial robots, humanoids, and autonomous mobile robots. NVIDIA states that the effort draws on more than 18,000 engineering years of autonomous vehicle safety work, the assessment of billions of safety transistors, and millions of lines of safety assessed code. ^[20][21]

IGX Thor serves as the compute layer of the Halos for Robotics stack. NVIDIA organizes Halos into three layers: ^[20][21]

• Platform safety (hardware): NVIDIA IGX Thor paired with the NVIDIA Holoscan Sensor Bridge, providing industrial grade AI compute with built in functional safety and real time, time synchronized sensor connectivity.
• Halos OS (software): Halos Core, a base safety operating layer for safety related functions, together with safety blueprints including the Halos Outside In Safety Blueprint for external camera perception and dynamic robot behavior control.
• Ecosystem safety (certification): the NVIDIA Halos AI Systems Inspection Lab, which NVIDIA describes as the first ANSI/ANAB accredited (ISO/IEC 17020) program for functional and AI safety assessment.

Halos references functional safety standards including IEC 61508, ISO 13849, and ISO/IEC TR 5469, and NVIDIA lists safety certification and inspection bodies such as TUV Rheinland, UL Solutions, TUV SUD, exida, SGS, and CertX as participants. ^[21] Deepu Talla, NVIDIA's vice president of robotics and edge AI, framed the launch around the shift to physical AI in factories, warehouses, and logistics. ^[21] Agility Robotics was named as an early adopter, incorporating Halos elements into its Digit humanoid for industrial logistics and manufacturing customers including Amazon, GXO, Schaeffler, and Toyota Motor Manufacturing Canada. ^[21] Agility chief executive Peggy Johnson said that for humanoids to deliver value at scale, safety has to be built into the robot. ^[21]

In summary, the Jetson Thor module and IGX Thor are distinct products on the shared Thor compute generation. Jetson Thor targets general robotics and physical AI deployments, while IGX Thor adds a certifiable hardware safety island and serves as the functional safety compute foundation underneath NVIDIA Halos for Robotics. ^[19][20]

Comparison with other edge AI silicon

Jetson Thor competes with a range of edge and embedded accelerators, though no other vendor matches its combination of Blackwell class GPU, 128 GB memory, and integrated Arm server class CPU. The table below sketches the broader market context as of late 2025 and early 2026. ^[17]

Scroll sideways for more →

In practice, Thor competes most directly with high end Jetson Orin and DRIVE platforms rather than with low power accelerators like Hailo or Coral. The closer comparisons are NVIDIA's own AGX Orin Industrial and the broader Arm server market, where Thor's SBSA alignment means workloads can move between Grace based servers and the edge module without code changes. ^[8]

Development, security, and lifecycle

NVIDIA provides a 10-year production lifecycle for Jetson Thor modules, matching the long support windows expected by industrial and automotive integrators. ^[3] Secure boot, A or B firmware updates, and a hardware root of trust are standard. JetPack 7 introduces a stronger SBSA-aligned firmware model, with UEFI and ACPI tables on the SoC, which simplifies long term operating system support and field updates. ^[7]

For developers, NVIDIA offers a remote provisioning workflow using NVIDIA Mission Control style tools, a containerized application layer based on the NVIDIA Container Toolkit, and prebuilt model containers in the NGC registry. The Jetson Thor developer kit boots into a desktop Ubuntu environment with CUDA, TensorRT, and Isaac libraries preinstalled, and is intended to act as a "robot brain" out of the box. ^[10]

Reception

Industry coverage at launch focused on three themes. First, Thor's headline 7.5 times AI compute improvement over Jetson AGX Orin was widely cited as a generational leap rather than an incremental upgrade, although several reviewers noted that the comparison mixes FP4 sparse on Thor against INT8 sparse on Orin. ^[10] Second, the introduction of MIG to an embedded platform was seen as a significant step toward multi model robotics, where perception, planning, language reasoning, and safety can coexist on a single SoC. ^[18] Third, the price of US$3,499 for the developer kit and US$2,999 for the bulk module was noted as high by hobbyist standards but competitive against discrete GPU plus host CPU solutions targeted at humanoid robots. ^[11]

Reviews from HotHardware and ServeTheHome highlighted hands-on demonstrations of GR00T N1 running inference on Thor while controlling a simulated robot performing precision tasks, with NVIDIA generating over 750,000 simulated trajectories in roughly 11 hours and reporting 40 percent performance improvements from synthetic data augmentation. ^[10] CNX Software and the Edge AI and Vision Alliance focused on the new networking stack (4 by 25 GbE plus 5 GbE plus Wi-Fi 6E) and the Holoscan Sensor Bridge as a clean fit for camera and lidar heavy robots. ^[11][13]

Coverage of the July 2026 T3000 and T2000 announcement emphasized the lower memory capacities, smaller form factor, and delayed Q1 2027 availability. CNX Software labeled its specification comparison preliminary because NVIDIA had not published full module documents. ServeTheHome noted that NVIDIA's T3000 comparison depends on selected, often memory-bandwidth-sensitive workloads and that pricing was not announced. ^[28][29] These constraints make the official emulation graph useful for development planning but insufficient for judging sustained power, thermals, sensor I/O, or application-level performance on production modules.

See also

• NVIDIA Jetson
• NVIDIA Blackwell
• NVIDIA Isaac
• GR00T
• Cosmos world models
• NVIDIA Omniverse
• Vision language model
• Large language model
• Reinforcement learning
• GPU
• Humanoid robot
• Physical AI
• Edge computing
• Multi-Instance GPU

External links

• Official NVIDIA Jetson Thor page
• Jetson Developer Kits
• Introducing NVIDIA Jetson Thor (NVIDIA Technical Blog)
• NVIDIA Isaac Platform
• NVIDIA Cosmos
• NVIDIA Developer Forums - Jetson

References

1. ^NVIDIA, "NVIDIA Blackwell-Powered Jetson Thor Now Available, Accelerating the Age of General Robotics," press release, August 25, 2025. nvidianews.nvidia.com/...e-age-of-general-robotics
2. ^NVIDIA Developer Blog, "Introducing NVIDIA Jetson Thor, the Ultimate Platform for Physical AI," August 25, 2025. developer.nvidia.com/...e-platform-for-physical-ai
3. ^NVIDIA, "Jetson Thor Series Modules Datasheet," DS-11945-001, v1.4, February 2026. developer.nvidia.com/...downloads
4. ^NVIDIA, "Jetson AGX Orin to Jetson Thor Migration Application Note," DA-11926-001, v1.2, September 2025. developer.nvidia.com/...on-application-note_da.pdf
5. ^NVIDIA, "NVIDIA Announces Project GR00T Foundation Model for Humanoid Robots and Major Isaac Robotics Platform Update," press release, March 18, 2024. nvidianews.nvidia.com/...l-isaac-robotics-platform
6. ^NVIDIA, "Jetson Thor product page," nvidia.com/...jetson-thor
7. ^NVIDIA, "JetPack 7.0 Release Notes," 2025. developer.nvidia.com/...jetpack
8. ^Arm Newsroom, "Arm and NVIDIA Enabling Intelligent Solutions For Roads and Robots," 2025. newsroom.arm.com/...e-agx-jetson-thor-arm-neoverse
9. ^NVIDIA Blogs, "Take It for a Spin: NVIDIA Rolls Out DRIVE AGX Thor Developer Kit," September 2025. blogs.nvidia.com/...loper-kit-general-availability
10. ^HotHardware, "NVIDIA Jetson AGX Thor Tested: Blackwell Brings Physical AI to Life," 2025. hothardware.com/...agx-thor-developer-kit-hands-on
11. ^CNX Software, "$3499 NVIDIA Jetson AGX Thor Developer Kit features 2070 TOPS Jetson T5000 SoM for robotics and edge AI," August 19, 2025. cnx-software.com/...0-som-for-robotics-and-edge-ai
12. ^Canonical, "Canonical's Ubuntu to be supported on NVIDIA Jetson Thor," 2025. canonical.com/...nvidia-jetson-thor-ubuntu-support
13. ^Edge AI and Vision Alliance, "NVIDIA Blackwell-powered Jetson Thor Now Available," August 2025. edge-ai-vision.com/...-the-age-of-general-robotics
14. ^The Robot Report, "NVIDIA launches Newton physics engine and GR00T AI at CoRL 2025," October 2025. therobotreport.com/...cs-engine-gr00t-ai-corl-2025
15. ^NVIDIA, "NVIDIA Cosmos: World Foundation Models Powering Physical AI," nvidia.com/...cosmos
16. ^Forecr, "NVIDIA Jetson AGX Thor vs Orin: Full Series Comparison," 2025. forecr.io/...-family-vs-thor-what-you-need-to-know
17. ^Acrosser, "NVIDIA Jetson AGX Thor vs NVIDIA Jetson AGX Orin," 2025. acrosser.com/nvidia-jetson-thor-vs-orin
18. ^ARC Advisory Group, "NVIDIA Jetson Thor Ushers in the Age of Agentic AI-Powered Robotics," 2025. arcweb.com/...hers-age-agentic-ai-powered-robotics
19. ^NVIDIA, "NVIDIA IGX Platform," product page. nvidia.com/...igx
20. ^NVIDIA Developer Blog, "Inside NVIDIA Halos for Robotics: A Full-Stack Functional Safety System for Physical AI," June 22, 2026. developer.nvidia.com/...ety-system-for-physical-ai
21. ^NVIDIA, "NVIDIA Announces Halos for Robotics, the Industry's First Full-Stack Safety System for Physical AI," press release, June 22, 2026. nvidianews.nvidia.com/...ty-system-for-physical-ai
22. ^NVIDIA Blog, "NVIDIA Introduces New Jetson Thor Computers to Advance Mainstream Robotics and Edge AI," July 15, 2026. blogs.nvidia.com/...on-thor-robotics-edge-ai-agent
23. ^NVIDIA Developer, "NVIDIA Jetson Roadmap," accessed July 25, 2026. developer.nvidia.com/...roadmap
24. ^NVIDIA, "Japan's Robotics and Manufacturing Leaders Build on NVIDIA Cosmos to Advance Physical AI Frontier," press release, July 15, 2026. nvidianews.nvidia.com/...ance-physical-ai-frontier
25. ^NVIDIA Developer Forums, "Jetson Agent Skills: AI-Assisted Workflows for Device & BSP Customization," June 22, 2026. forums.developer.nvidia.com/...374150
26. ^AAEON, "AAEON Outlines Plans for Embedded AI Box PCs Featuring the New NVIDIA Jetson T2000 and T3000 Modules," July 16, 2026. aaeon.com/..._t2000_and_t3000_modules_announcement
27. ^Aetina, "Aetina Extends NVIDIA Jetson Thor Momentum with Support for New NVIDIA Jetson T3000 and T2000 Modules," July 2026. aetina.com/about-news-detail
28. ^CNX Software, "NVIDIA launches smaller, mainstream Jetson T2000 and T3000 modules for Edge AI and robotics applications," July 16, 2026. cnx-software.com/...e-ai-and-robotics-applications
29. ^ServeTheHome, "NVIDIA Announces Expanded Jetson Thor Lineup with Mid-Range T3000 and T2000 Modules," July 2026. servethehome.com/...-range-t3000-and-t2000-modules
30. ^RidgeRun, "RidgeRun Announces Support for NVIDIA Jetson T2000 and T3000: Scalable AI for Robotics and Edge Applications," July 2026. ridgerun.com/...orts-nvidia-jetson-t2000-and-t3000
31. ^Dominic Maggio and Luca Carlone, "VGGT-SLAM 2.0: Real time Dense Feed-forward Scene Reconstruction," arXiv:2601.19887, January 27, 2026. arxiv.org/...2601.19887
32. ^Boseong Jeon, Yunho Choi, and Taehan Kim, "Shallow-pi: Knowledge Distillation for Flow-based VLAs," arXiv:2601.20262, January 28, 2026. arxiv.org/...2601.20262
33. ^Dominic Maggio, Nicolas Gorlo, and Luca Carlone, "FOUND-IT: Foundation-model-first Task-driven 3D Scene Graphs with Granularity on Demand," arXiv:2605.25371, May 25, 2026. arxiv.org/...2605.25371
34. ^Matteo Bortolon, Andrea Caraffa, Alice Fasoli, and Fabio Poiesi, "xperception: Making Robotic Grasping Easier," arXiv:2607.16312, July 14, 2026. arxiv.org/...2607.16312
35. ^NVIDIA Developer Forums, "New Thor T3000/T2000 SoM roll out," July 20, 2026. forums.developer.nvidia.com/...377467