NVIDIA puts a Blackwell brain inside robots with the $3,499 Jetson Thor as Figure, Agility and Boston Dynamics sign on
NVIDIA made its Jetson AGX Thor developer kit and Jetson T5000 production modules generally available on 25 August 2025, pricing the kit from $3,499 and claiming 7.5 times the AI compute of Jetson Orin with 128 GB of memory. Figure, Agility Robotics, Boston Dynamics and Amazon Robotics are among eight named early adopters, while 1X, OpenAI and Physical Intelligence are evaluating the computer.
NVIDIA on 25 August 2025 began selling Jetson Thor, a Blackwell-based computer designed to sit inside humanoid and other autonomous robots, giving developers of robot foundation models a standard onboard platform with enough memory to run several generative AI models at once. The Santa Clara company priced the Jetson AGX Thor developer kit from $3,499 and said Jetson T5000 production modules were available from its worldwide distribution partners, with carrier boards and finished systems sold by embedded partners.
According to NVIDIA's announcement, Thor delivers up to 2,070 teraflops of FP4 AI compute within a 130-watt power envelope and carries 128 GB of memory. The company said that is 7.5 times the AI compute of Jetson Orin, its previous robotics platform, with 3.5 times the energy efficiency. NVIDIA's technical blog specifies that the 2,070 teraflops figure applies to sparse FP4 arithmetic; dense FP4 performance is 1,035 teraflops.
The module pairs a 2,560-core Blackwell GPU with 96 fifth-generation Tensor Cores and a 14-core Arm Neoverse-V3AE CPU running at up to 2.6 GHz, the blog said. Memory is 128 GB of LPDDR5X on a 256-bit bus with 273 GB/s of bandwidth, and power can be set anywhere from 40 to 130 watts. The GPU supports Multi-Instance GPU partitioning across 10 TPCs, which lets a robot fence off parts of the chip for separate jobs such as perception and motion planning. The developer kit adds a QSFP28 port carrying four 25 GbE links for camera and sensor traffic, and the software release, JetPack 7, runs on Ubuntu 24.04 LTS with Linux kernel 6.8 and CUDA 13.0.
NVIDIA named Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic and Meta as early adopters and said 1X, John Deere, OpenAI and Physical Intelligence were evaluating Thor. Four of the twelve, Figure, Agility, Boston Dynamics and 1X, build humanoid robots. Physical Intelligence develops general robot models. The others range from Amazon's warehouse robotics arm and Caterpillar's construction and mining machines to Deere's farm equipment and Medtronic's medical devices. The release said NVIDIA's robotics stack is used by 2 million developers and backed by more than 150 hardware, software and sensor partners, and put the Jetson Orin customer base at more than 7,000.
Founder and chief executive Jensen Huang described Thor as the supercomputer for physical AI and general robotics, pointing to its ability to run multiple generative AI models at the edge. Figure founder and CEO Brett Adcock said in the release that Thor's server-class performance in a compact, power-efficient design lets the company deploy large-scale generative AI models on its robots. Amazon Robotics chief technologist Tye Brady cited the computing power and energy efficiency needed to develop and scale the next generation of AI-powered robots, and Caterpillar CEO Joe Creed linked the module to the company's future construction and mining equipment.
NVIDIA said Thor runs its full robotics software set, including Isaac for simulation and development, Isaac GR00T humanoid foundation models such as GR00T N1.5, Metropolis vision AI and Holoscan real-time sensor processing. The developer blog benchmarked models including Llama 3.1 8B, Qwen3-32B, Qwen2.5-VL-7B, GR00T N1.5 and Cosmos Reason 7B, and said generative reasoning runs up to five times faster than on Orin, with DeepSeek-R1-Distill-Qwen-32B running 4.87 times faster.
The memory figures matter more than the headline teraflops. A vision-language-action model of the type humanoid developers now use reads camera frames and an instruction and turns them into joint commands many times a second. Running such a model alongside a language model and a perception stack on one robot is limited first by how much model fits in memory and how fast weights can be read. Thor doubles the 64 GB of the top Jetson AGX Orin module and raises bandwidth from 204.8 GB/s to 273 GB/s, a 1.33-fold gain by MEKAWIRE calculation, while its maximum power draw rises from 75 watts to 130 watts.
- GPU cores
- 2560
- 1792
- 1.43
- Tensor Cores
- 96
- 56
- 1.71
- CPU cores
- 14
- 8
- 1.75
- CPU max frequency (GHz)
- 2.6
- 2
- 1.3
- Memory (GB)
- 128
- 64
- 2
- Memory bandwidth (GB/s)
- 273
- 204.8
- 1.33
- Max power (W)
- 130
- 75
- 1.73
- Min power (W)
- 40
- 15
- 2.67
Specifications from NVIDIA's Jetson Thor developer blog (25 August 2025) and NVIDIA's Jetson Orin product page. Ratio column is a MEKAWIRE calculation: Thor value / Orin value, rounded to two decimals. AI compute is not compared row by row because NVIDIA quotes Thor in FP4 teraflops and Orin in INT8 TOPS; NVIDIA's own comparison is 7.5x AI compute and 3.5x energy efficiency. Figures disclosed, not estimated.
As of Aug 25, 2025
Two days after the launch, NVIDIA reported second-quarter fiscal 2026 revenue of $46.743 billion, up 56% from a year earlier. Its automotive and robotics segment brought in $586 million, up 69% year on year and 3% from the previous quarter, and the results release cited Thor's availability as a platform meant to power millions of robots. By MEKAWIRE calculation, that segment equalled about 1.25% of group revenue, and NVIDIA does not report robotics separately from automotive.
The list leaves gaps. The release named no Chinese humanoid maker and did not include Tesla's Optimus programme. NVIDIA also gave no volume price for T5000 modules, no shipment figures and no indication of which adopters will put Thor into production robots rather than development units. On a battery-powered humanoid, a computer that can draw up to 130 watts competes with the actuators for energy, a trade-off each maker will have to settle in its own design.
Investors and rivals will look for three things in the coming quarters: whether named humanoid makers publish robot specifications that list Thor as the onboard computer, how NVIDIA's automotive and robotics revenue moves in its next quarterly reports, and whether NVIDIA follows Thor with cheaper modules for lower-cost robots.
MEKAWIRE analysis
Jetson Thor makes NVIDIA the default onboard computer for most of the best-funded Western humanoid makers, but it will move NVIDIA's revenue only once robots ship in the tens of thousands.
The evidence for the first half is in the customer list. Figure, Agility Robotics and Boston Dynamics are adopters and 1X is evaluating, which takes in all three US humanoid developers in the table below that have disclosed nine-figure rounds. Thor's 128 GB of memory is aimed squarely at the multi-model workloads those companies describe, and NVIDIA ties the hardware to its own GR00T models and Isaac simulation tools, so a team that trains in NVIDIA's stack faces the least friction deploying on it.
The strongest counter-argument is that an early-adopter line in a vendor press release is not a design win. None of the named humanoid makers disclosed volumes, and companies with the scale to do so may design their own inference silicon to cut cost and power. No Chinese humanoid maker appears on the list either, so a large group of developers has made no visible commitment.
Bull case. Humanoid production moves from hundreds to tens of thousands of units over the next few years, most Western makers standardise on T5000-class modules, and robotics becomes a reported line of its own inside NVIDIA. The software lock-in from GR00T and Isaac then makes switching costly for customers.
Bear case. Volumes stay small, the biggest makers bring inference in-house, and cheaper or domestic chips win in China. Thor then remains a developer platform that sells kits to labs while the automotive and robotics segment stays near 1% of NVIDIA's revenue.
- Figure AI
- Early adopter
- Series B
- 2024-02-29
- 675,000,000
- 2,600,000,000
- Agility Robotics
- Early adopter
- Series B
- 2022-04-22
- 150,000,000
- No data
- 1X Technologies
- Evaluating
- Series B
- 2024-01-13
- 100,000,000
- No data
- Boston Dynamics
- Early adopter
- Subsidiary, no round disclosed
- No data
- No data
- No data
Status from NVIDIA's 25 August 2025 release. Funding from each company's own announcement; Figure's $2.6 billion valuation was not specified as pre- or post-money. Figure's later Series C (16 September 2025) is excluded because it came after this launch. — = not disclosed. No estimates.
As of Aug 25, 2025
Signals to watch:
- November 2025: NVIDIA's third-quarter fiscal 2026 results, and whether the automotive and robotics segment keeps growing faster than the group.
- During 2026: production specifications from Figure, Agility Robotics, Boston Dynamics and 1X that name the onboard computer.
- During 2026: any lower-priced Thor-family module that would bring Blackwell into cheaper robots.
- 40 to 130 W configurable
- 128 GB LPDDR5X, 273 GB/s
- Jetson AGX Thor developer kit and Jetson T5000 production module
- 2,070 FP4 TFLOPS sparse (1,035 dense)
- 1X, John Deere, OpenAI, Physical Intelligence
- General availability announced 25 August 2025
- Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic, Meta
- 7.5x AI compute, 3.5x energy efficiency (NVIDIA)
- From $3,499
- $586 million, up 69% year on year
Why it matters
Until Thor, most humanoid developers ran their largest models off-board or squeezed them into Orin-class modules with 64 GB of memory. Thor gives them a catalogue part with twice the memory, a Blackwell GPU with native FP4 and a software stack that matches the one many of them already use for training and simulation. That lowers the cost of putting a vision-language-action model, a language model and perception on the robot itself, which matters for latency, for privacy in homes and for operating where connectivity is poor.
For NVIDIA, Thor closes the loop on a strategy that already sells data-centre GPUs for training robot models and simulation software for generating data. Owning the onboard computer as well means each robot that ships with Thor adds hardware revenue and deepens dependence on NVIDIA's tools.
Rival analysis
The release names no competing robot computer, and NVIDIA's customers have options. Large robot makers can design their own inference chips once volumes justify it, and in China domestic edge AI suppliers are positioned to serve local humanoid makers, none of which appeared on NVIDIA's list. Mobile chip makers that already sell edge AI processors are a further source of competition at lower price and power points.
Among humanoid makers, the list splits the field. Figure, Agility Robotics and Boston Dynamics are named adopters and 1X is evaluating; Apptronik, which raised a $350 million Series A in February 2025, and Tesla were not named. That does not mean they reject Thor, but it leaves their onboard compute choices undisclosed.
Valuation context
NVIDIA reported group revenue of $46.743 billion in the second quarter of fiscal 2026 and $586 million from automotive and robotics, about 1.25% of the total by MEKAWIRE calculation. Robotics revenue is not broken out. At a developer kit price of $3,499, the hardware is a rounding error for NVIDIA today, and its value lies in the platform position it buys.
On the customer side, the humanoid makers on the list have raised large rounds on the expectation of volume production. Figure disclosed a $675 million Series B at a $2.6 billion valuation in February 2024, Agility a $150 million Series B in April 2022 and 1X a $100 million Series B in January 2024. Their production plans are the demand that would turn Thor into material revenue.
Supply-chain implications
Thor combines a Blackwell GPU, a 14-core Arm Neoverse-V3AE CPU and 128 GB of LPDDR5X memory on one module, and NVIDIA sells T5000 modules through distribution partners while embedded partners build carrier boards and complete systems. That model lets small robot companies buy a finished computer instead of designing boards, but it also means module pricing, memory supply and partner lead times shape their bill of materials.
The 130-watt ceiling has knock-on effects for battery packs and thermal design in humanoids, where every watt spent on compute reduces walking and working time. Developers can configure Thor down to 40 watts, and the practical trade-off between model size and runtime will show up in the robot specifications makers publish.
Signals to watch
The first signal is NVIDIA's third-quarter fiscal 2026 report, expected in November 2025, and whether automotive and robotics growth stays above the group rate. The second is robot specifications. When Figure, Agility, Boston Dynamics or 1X publish production specs, the onboard computer they name will show whether early adoption became a design win.
The third is the product roadmap. A cheaper Thor-family module would extend Blackwell to lower-cost robots and would indicate that NVIDIA wants unit volume, not only the premium segment. The fourth is China, where any named Chinese humanoid adopter would change the reading of the customer list.
Analyst view
Thesis: Thor gives NVIDIA the onboard slot in most leading Western humanoid programmes, but the revenue impact is a 2026-and-later story that depends on robot volumes the industry has not yet demonstrated. Confidence: medium.
The reasons for medium rather than high confidence are that the adopter list carries no volume or exclusivity terms, that NVIDIA does not report robotics revenue separately and that no Chinese humanoid maker is represented. The reasons it is not low are the breadth of named adopters across humanoids, logistics, construction and medical devices and the coupling of Thor with NVIDIA's training and simulation tools.
Questions you should be asking
Which of the named adopters will ship production robots with T5000 modules, and in what quantity? What will NVIDIA charge for T5000 modules at volume, and how does that compare with the cost of a custom inference chip for a maker building tens of thousands of robots?
How much of a humanoid's battery budget will Thor consume in typical use, and will makers run it near 40 watts or near 130 watts? Will NVIDIA start reporting robotics revenue separately, and will any Chinese humanoid maker be named as a Thor customer?