Executive summary
1,000+ TOPS
AI compute on NVIDIA's Jetson Thor module, generally available since August 2025
7.5x
the compute of the previous Jetson Orin generation, per NVIDIA
Key conclusions
- What a machine can decide without a live connection depends on the compute physically inside it.
- The named early adopters include humanoid platforms, mobile and heavy machinery, fleet robotics and regulated precision applications.
- Ask what compute is physically present to keep a system working through a dropped connection. The answer should be a stated design choice, made before the first outage.
An autonomous machine that needs a live cloud connection for its next decision stops being autonomous where the connection is weakest. Examples include the inside of a steel-hulled vessel, the interior of a large plant, or a port berth where the network was specified for email. What the machine can decide on its own depends on the compute physically inside it, and that compute has changed significantly in the last year.
NVIDIA's Jetson Thor is built on the company's Blackwell architecture. It became generally available in August 2025, with more than 1,000 TOPS of AI compute and 128GB of memory, on a module designed for this kind of embedded deployment. The flagship T5000 configuration has 2,560 CUDA cores and 96 fifth-generation Tensor Cores. NVIDIA states that it delivers 7.5 times the compute and 3.5 times the energy efficiency of the previous Jetson Orin generation.
The early adopters show where onboard compute matters most
- Agility Robotics and Figure: humanoid platforms, where onboard compute must run perception and manipulation together in real time.
- Boston Dynamics and Caterpillar: mobile and heavy machinery, where the connection is intermittent because of the nature of the work site.
- Amazon Robotics: fleet-scale deployment, where the case for edge compute is latency and fleet-wide reliability.
- Medtronic and Hexagon: regulated and precision-measurement applications, where a round trip to the cloud creates both a latency problem and an audit-trail problem.
Edge compute is now a specific line item in any autonomy specification
This does not mean every plant needs a Jetson Thor. It means the edge-compute budget is now a specific line item that vendors can quote. If a system claims autonomous decision-making without this compute or its equivalent, the hardware on site cannot support the claim.
Before commissioning any system described as autonomous, ask what happens to its decision-making when the network connection is cut. Also ask what compute is physically present to keep it working during that gap. If the honest answer is 'it stops safely and waits', that may be the right design. It should still be a stated design choice, made before the first outage happens.
