On October 1, SEMIFIVE signed a deal with Mobilint to build an AI chip for robots that must keep working when the network goes down.

On October 1, SEMIFIVE signed a deal with Mobilint to build an AI chip for robots that must keep working when the network goes down.

The first target is agricultural robots that process vision and sensor data outdoors, with no cloud to call. The chip pairs LPDDR6 memory with PCIe Gen6 and UCIe-S chiplet links. Korea's Ministry of Trade, Industry and Resources backs it through the K-On-Device AI Semiconductor program. SEMIFIVE says the deal takes its business "beyond data center, HPC and edge AI into Physical AI."

A chip company just named Physical AI as its next category, in a press release.

Every production robot does two jobs at once. The Slow Brain reasons: vision, language, planning, foundation models. The Fast Brain controls: motor signals, safety loops, guaranteed timing, every time. I have written about that split since May. In June I walked the floor at Automate looking for anyone building the layer between the two.

I keep a count of who is. In March, Morgan Stanley's "When Factory = Robot" report profiled 18 private startups building the agentic factory. By my read, none addressed deterministic real-time control. Neither did the dozens of startups that pitched at a Plug and Play Physical AI day in April.

The SEMIFIVE chip moves the thinking onto the machine, which is the right call. The announcement describes compute for vision and sensor data, not guaranteed timing. One day later, AMD's Salil Raje described a robot as a brain, a spine and joints, and put deterministic, low-latency control in the joints. AMD's answer is to sell every layer itself. Korea's answer is to fund the silicon. I have not found anyone building the independent layer that lets a robot maker pair one vendor's models with another vendor's chip and still guarantee timing.

I am a co-founder of a venture in stealth, and that is the layer we are building.

I will be at TechCrunch Disrupt in San Francisco on October 13 to 15. If you build robots, chips, or the software between them, I would like to compare notes. Thoughts?

This perspective was originally published on LinkedIn. Join the discussion, add your thoughts, and follow for regular updates.

View on LinkedIn