Even if veterans leave, the equipment cannot stop. The era where factory AI has started to 'think for itself'
The other day, I wrote about the basic concept of edge AI and the idea that 'even small factories can start with just one sensor.' This time, I am diving deeper into that. However, instead of digging into the same topic, I will shift the perspective entirely.
Attaching sensors to a factory, collecting data, and issuing alerts when there is an anomaly... Isn't this what many people imagine when they hear 'edge AI'? Until very recently, I also perceived edge AI for small and medium-sized manufacturers as little more than an 'extension of smart sensors'.
However, the situation has changed completely as of 2026. Generative AI is running on palm-sized modules. The 'brains' of industrial robots are starting to be replaced entirely. To top it off, edge AI is even being installed on satellites in outer space.
This time, I will set aside the topic of company size and share, based on recent examples, just how far edge AI hardware and systems have come as of 2026. I hope you will stick with me, as knowing this trend will completely change your judgment on capital investment over the next two to three years.
March 2026: NVIDIA updated the 'factory brain' all at once
At NVIDIA's technology conference 'GTC 2026' held this March, the edge AI community was buzzing. According to information released by NVIDIA, the new Jetson Thor achieves an AI processing performance of 2,070 TFLOPS (FP4) (a processing speed that allows a computer to perform approximately 2.07 quadrillion calculations per second), which is a 7.5x jump in AI performance and a 3.5x increase in efficiency compared to the previous generation Jetson AGX Orin. Moreover, the power consumption is 40-130W. This is not about a rack-mounted machine you would put in a server room, but a module about the size of a GPU graphics board.
According to a teardown report, the Jetson AGX Thor development kit released in late 2025 fits into a chassis the size of a mid-range GPU board and even includes an optical transceiver slot supporting 100Gbps.
What is amazing is that you can run generative AI models like VLA (Vision-Language-Action) models, LLMs (Large Language Models), and VLMs (Vision-Language Models) directly on this module. In other words, without connecting to the cloud, an AI that can 'see, think, and judge' runs inside a small box placed on the factory floor. This is fundamentally on a different level from edge AI that just 'issues an alert when a threshold is exceeded'.
The 'brains' of industrial robots are starting to be replaced
What drew attention at GTC 2026 was the announcement that global industrial robot manufacturers such as FANUC, ABB Robotics, Yaskawa Electric, and KUKA are integrating NVIDIA's Isaac simulation framework and Jetson edge AI inference modules into their own controllers. These four companies alone have over 2 million robots operating worldwide.
This is a significant turning point. Until now, industrial robots were basically 'machines that accurately repeat what they are taught.' You teach them movements through programming, and they perform the same actions on the same line over and over. In high-mix, low-volume production sites, re-teaching was required every time there was a setup change, which became a bottleneck.
When robots can execute AI inference locally via Jetson modules, they can make high-speed judgments without being constantly connected to the cloud, allowing them to adapt to more complex production environments. Robots are starting to change from 'tools that move according to a program' to 'colleagues that think for themselves'.
The era where edge AI flies into space
The scale of the story changes drastically, but this was also announced at GTC 2026 in March 2026. NVIDIA announced that it would provide edge AI inference for orbital data centers and autonomous space operations using the IGX Thor and Jetson Orin platforms. Space startups such as Aetherflux, Axiom Space, and Planet Labs have adopted this platform.
The Jetson Orin is an ultra-compact, power-efficient module that is said to be capable of performing real-time processing of visual data, navigation, and sensor data directly on satellites.
...The reason I brought up the space story is that I wanted to convey that 'the reliability of edge AI has reached that point.' Outer space is an extreme environment where if the line is cut, it's over, and if the equipment breaks, you cannot go to repair it. Edge AI modules at a level that can be used there are coming down to factories with the same architecture. This is extremely persuasive as proof of reliability.
Caterpillar, a construction machinery giant, is operating an 'excavator with an AI assistant'
Here is another example close to the manufacturing floor. Construction machinery giant Caterpillar showcased a demo at this year's CES of a small excavator, the 'Cat 306 CR,' equipped with an AI assistant running on Jetson Thor. It uses NVIDIA's Nemotron model for voice recognition and runs Qwen3 4B locally for language processing, a system that responds to operator instructions without a cloud connection.
Inside the cramped cockpit of a mini-excavator under 8 tons, the AI responds by voice to support the operator. Moreover, it does not need to be connected to the cloud. Since construction sites are prime examples of places where communication environments are unstable, completing everything at the edge is extremely practical.
💡 This is actually the beginning of an era where 'AI lives on the factory floor', isn't it?
Looking at the examples so far, what becomes clear is that the phase of edge AI has completely shifted from being an 'extension of sensors that monitor data' to 'AI agents that make autonomous decisions on-site'.
Taiwan's Advantech showcased an edge AI system equipped with Jetson Orin at GTC 2026, realizing LLM interaction on the device without cloud connectivity. It is designed for use in industrial edges and retail stores.
Meanwhile, AMD also announced the expansion of its 'Ryzen AI Embedded P100' series of processors for edge AI, with mass production shipments scheduled within 2026. It is also worth noting that it is no longer just NVIDIA's world.
In other words, options are increasing, price competition is beginning, and the possibility of access for small and medium-sized enterprises is steadily expanding. This trend is irreversible.
Roadmap to Tomorrow
Let's return to the question, 'So, what should we do?' This does not mean you should go out and buy a Jetson Thor right now.
Step 1:
Watch one GTC 2026 keynote archive. NVIDIA Japan has also released blog posts in Japanese. The starting point is to first know 'where the world's leaders are heading.' Since they specifically discuss manufacturing, it is less boring than you might imagine.
Step 2:
Ask your equipment manufacturer, 'Do you have any maintenance options that support edge AI?' Major manufacturers like FANUC, Yaskawa Electric, and Omron are advancing collaboration with the NVIDIA platform. Cases where you can add it to existing equipment as an add-on are emerging, so it would be a waste not to know about them.
Step 3:
Buy one Jetson Orin Nano Super developer kit (in the tens of thousands of yen range) and let the young staff in your company play with it. Models have been released with up to 142 times the performance of the previous generation Jetson Nano. It can even run generative AI. Even just letting them play with it as a 'toy' is well worth the value of sowing the seeds of AI literacy within the company.
An era where 2,070 TFLOPS (a processing speed where a computer can perform approximately 2.07 quadrillion calculations per second) runs in the palm of your hand. Edge AI that works in space or inside an excavator. I don't think there is any industry left that can claim this trend 'has nothing to do with us'.
…I have talked about some big things, but the important thing is 'knowing.' If you know, you gain one more option when it comes time to update equipment. You can write one line in a subsidy application. You can bring it up in negotiations with vendors by asking, 'Do you know about Jetson Thor?' That alone changes the landscape.
If you could support me with a like or comment, it will be fuel for the next article. 🙌