Sony and Mitsubishi Electric are putting AI inside the sensor itself to rewire factory automation

Sony Semiconductor Solutions and Mitsubishi Electric will launch a joint venture, Advanced Vision Solutions Co., Ltd., in October 2026 to embed AI processing directly into image sensors for factory automation, combining Sony's IMX500 on-chip AI sensor with Mitsubishi Electric's industrial control systems. The venture aims to reduce latency and bandwidth needs by performing inference at the sensor, competing with Nvidia's recent industrial AI partnerships with Fanuc, Yaskawa Electric, Fujitsu, Hitachi, and Kawasaki Heavy Industries.

Advanced Vision Solutions Co., Ltd., the joint venture Sony Semiconductor Solutions and Mitsubishi Electric agreed to form this month, launches in October 2026 with a single bet: that the smartest place to run AI on a factory floor is inside the camera, not somewhere above it in the cloud. The deal pairs two things that have rarely sat in the same company. Sony Semiconductor Solutions built the IMX500, the world's first image sensor with on-chip AI processing, an architecture that stacks a neural network accelerator and dedicated SRAM directly onto the pixel layer so inference happens at the sensor itself, not after the data travels anywhere. Mitsubishi Electric has spent decades wiring up factory control systems. Put them together in a joint venture and you get something that neither company could credibly sell alone: a vision system that sees, thinks, and signals, all before a single byte of image data leaves the camera housing. That architecture matters more than it might sound. Conventional machine vision in manufacturing still works roughly the way it did fifteen years ago: cameras capture frames, those frames go to a server or a cloud endpoint, software runs the inspection or the anomaly-detection model, and a verdict comes back. It works, but it's slow by industrial standards, it depends on network reliability, and it scales poorly when you're trying to instrument hundreds of stations across a plant floor. On-sensor inference cuts that loop entirely. The IMX500 outputs metadata, classification labels, bounding boxes, anomaly flags, rather than full image streams, which shrinks bandwidth by orders of magnitude and brings latency down to the kind of real-time response a moving assembly line actually needs. The timing is not accidental. Nvidia announced on July 16, just days before this deal closed, a sweeping set of partnerships with Fanuc, Yaskawa Electric, Fujitsu, Hitachi, and Kawasaki Heavy Industries, centered on its Cosmos 3 Edge world model for physical AI in factories. That puts direct competitive pressure on every Japanese industrial player that doesn't already have a credible AI-on-hardware story. Sony and Mitsubishi Electric both do, separately. Together, according to reporting by The AI Journal on the joint venture announcement, they're positioned to offer vision sensor solutions that capture product quality, equipment operations, and maintenance-related conditions that have traditionally been difficult to visualize, and connect them directly to decision-making systems on the factory floor. Sony brings a genuine technical moat here. The IMX500 wasn't a concept demo; it's in commercial products and the AITRIOS platform already lets developers deploy custom neural network models directly to the sensor. Mitsubishi Electric brings the harder-to-replicate asset: deep integration with FA control systems and decades of customer relationships with the manufacturers who actually run plants. Neither of those advantages transfers easily to a new entrant. Chinese robotics challengers complicate the picture further. Chinese manufacturers have moved fast on lower-cost vision and automation hardware, and the competitive pressure in industrial AI is no longer purely an American story. A joint venture that combines Sony's sensor IP with Mitsubishi Electric's installed base in Japanese and global manufacturing creates a defensible position that would take years to replicate from scratch. What this means for the factory floor For a plant manager evaluating vision inspection today, the economic case for on-sensor AI is straightforward. You don't need GPU servers at every station. You don't need high-bandwidth network infrastructure to pipe raw video around. You don't take on cloud dependency for a process that can't tolerate latency spikes. A system built on Sony's stacked sensor architecture with Mitsubishi Electric's FA integration layer is, in theory, something a factory can trust the same way it trusts a PLC: it runs at the edge, it runs deterministically, and it doesn't go offline when the WAN does. The harder question for founders building in the manufacturing and robotics stack is where this joint venture's boundaries sit. Advanced Vision Solutions is focused on vision sensor solutions for manufacturing, which means the sensors, the inference layer, and the connection to factory control systems. That's not the same as building the application layer: the scheduling software, the yield optimization models, the digital twin integrations. Those remain open territory, and a sensor platform with Sony's image quality and Mitsubishi Electric's FA reach underneath it is exactly the kind of infrastructure layer that makes adjacent application businesses easier to build on. The joint venture is subject to regulatory clearance before its October 2026 launch. Neither company commented on revenue projections for the new entity. But the bet is visible in the structure itself: two companies that don't need each other for survival decided they needed each other for this particular problem. That's usually a sign the problem is real. Also read: Wistron opens its first US factory in Fort Worth and starts stamping out Nvidia's most advanced AI chips https://startupfortune.com/wistron-opens-its-first-us-factory-in-fort-worth-and-starts-stamping-out-nvidias-most-advanced-ai-chips/ • Samsung is in talks to take an equity stake in Mistral AI as the French startup seeks €3 billion at a €20 billion valuation https://startupfortune.com/samsung-is-in-talks-to-take-an-equity-stake-in-mistral-ai-as-the-french-startup-seeks-3-billion-at-a-20-billion-valuation/ • Fireworks AI closes $1.5 billion Series D at a $17.5 billion valuation as enterprises flee frontier API pricing https://startupfortune.com/fireworks-ai-closes-15-billion-series-d-at-a-175-billion-valuation-as-enterprises-flee-frontier-api-pricing/