Around 2012, there was a time when the CuBox seemed like a tiny miracle. Barely heavier than a deck of cards, this two-inch cube can stream 1080p video to a living room flat screen. It ran XBMC, cost about $100, and looked ridiculous compared to the heavy home theater PCs of the time.
It was a hit with hobbyists. Build guides lit up forums. It then disappeared from the cultural discourse, much like a lot of specialized hardware. The majority of the oxygen went to the Raspberry Pi. Nevertheless, the CuBox continued to ship.
| Product Name | CuBox |
| Manufacturer | SolidRun Ltd. |
| Headquarters | Yokneam Illit, Israel |
| First Released | 2011 |
| Original Use Case | Home media center / XBMC streaming |
| Current Generation | CuBox-M and CuBox Pulse |
| Processor Family | NXP i.MX (ARM-based SoC) |
| Form Factor | Roughly 2 inches per side |
| Operating Systems | Linux, Android, OpenWrt |
| Key Modern Application | Edge AI, IoT gateways, embedded inference |
| Power Draw | Under 10 watts typical |
| Founders | Dr. Rabeeh Khoury, Atai Ziv |
What’s intriguing is what transpired next, which was largely overlooked at the time. The device’s Israeli manufacturer, SolidRun, did not make a dramatic Silicon Valley turn. They simply continued to refine. NXP offers better SoCs. More memory. quicker networking. The silicon inside the cube continued to advance generation after generation, but the cube itself remained about the same size. It’s the kind of slow, unglamorous engineering that ages well but doesn’t garner much attention.
There’s a good chance that something CuBox-shaped is bolted to a wall and blinking silently if you walk into any small industrial deployment today, whether it’s a factory floor in Stuttgart, a logistics hub outside of Memphis, or a wind farm in coastal Denmark. In recent years, edge computing has evolved from a catchphrase to an operational reality, and gadgets like this one found their way to the ideal location at the ideal moment. ARM-based, compact, fanless, and low power. precisely the profile that edge AI inference seeks.
The computation for such hardware has changed due to the trend toward running models locally, on the device itself rather than in some hyperscale data center. It is not necessary for a loading dock camera to call home to AWS in order to detect a damaged pallet in real time. All it needs is enough processing power to run a vision model on a limited NPU and enough input/output to use the output. The CuBox stealthily entered that area.
Recently, Charles Giancarlo of Pure Storage said something that really resonated with me. He claimed that the excitement surrounding AI and the realization that data centers are growing faster than anyone could have predicted have made hardware and systems cool once more. At the edge, the same reasoning holds true in reverse. Watching this unfold gives the impression that the industry neglected the importance of physical devices for ten years in favor of cloud computing. A tiny reminder of that is the CuBox.
Whether SolidRun can compete with more recent players is still up in the air. The CuBox lacked the marketing power of Nvidia’s Jetson line. Cheaper alternatives are being flooded into the edge AI market by Chinese manufacturers. To be honest, the Raspberry Pi 5 is now a powerful device. However, a product that has been quietly refining for fifteen years seems to be here to stay. The edge AI market appears to be real, according to investors. The businesses that are actually implementing these strategies on a large scale are already aware of this.
As this develops, it’s difficult not to consider how many “obsolete” technologies were actually just early. Not many people are familiar with the CuBox. Most likely, it will never be. However, the next time someone inquires about the actual location of the AI revolution—not the training, but the running—the response might be that it’s humming along on a shelf in a server closet, performing precisely the function for which it was redesigned.
