Systems integration
The chip decides what your machine can still do when the network drops
NVIDIA's Jetson Thor, generally available since August 2025, puts more than 1,000 TOPS of inference compute onto a module built to sit inside a machine rather than a server rack. That number is the actual constraint on how autonomous a piece of equipment can be without a live connection.
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
Core conclusions
- What a machine can decide on its own without a live connection comes down to the compute physically inside it — not a network assumption.
- Named early adopters span humanoid platforms, mobile and heavy machinery, fleet robotics, and regulated precision contexts.
- Ask what compute is physically present to keep a system functioning through a dropped connection — that should be a stated design choice, not a discovery made during the first outage.
An autonomous machine that needs a live cloud connection to make its next decision is not autonomous where the connection is weakest — inside a steel-hulled vessel, deep in a plant, or at a port berth where the network was specified for email, not inference. What the machine can decide on its own comes down to the compute physically inside it, and that has changed materially in the last year.
NVIDIA's Jetson Thor, built on its Blackwell architecture, became generally available in August 2025 with more than 1,000 TOPS of AI compute and 128GB of memory on a module designed for exactly this kind of embedded deployment. The flagship T5000 configuration carries 2,560 CUDA cores and 96 fifth-generation Tensor Cores, and NVIDIA states it delivers 7.5 times the compute and 3.5 times the energy efficiency of the previous Jetson Orin generation.
Named early adopters, and what that tells you
- Agility Robotics and Figure — humanoid platforms, where onboard compute has to run perception and manipulation together in real time.
- Boston Dynamics and Caterpillar — mobile and heavy machinery, where the connection is intermittent by the nature of the work site.
- Amazon Robotics — fleet-scale deployment, where the case for edge compute is latency and fleet-wide reliability rather than novelty.
- Medtronic and Hexagon — regulated and precision-measurement contexts, where a round trip to the cloud is a latency and an audit-trail problem at once.
The specification question this settles
None of this means every plant needs a Jetson Thor. It means the edge-compute budget is now a specific, quotable line item rather than a hand-wave in a proposal, and a system that claims autonomous decision-making without one — or without its equivalent — is making a claim the hardware in the room cannot support.
Before commissioning anything described as autonomous, ask what happens to its decision-making the moment the network connection is cut, and what compute is physically present to keep functioning through that gap. If the honest answer is 'it stops safely and waits', that may be the right design — but it should be a stated design choice, not a discovery made during the first outage.