George Hotz's Tiny Corp pulled off something unexpected: Apple approved their driver that lets you connect Nvidia eGPUs to Arm Macs. The driver, called TinyGPU, works with AMD RDNA3+ and Nvidia Ampere+ GPUs through USB4 or Thunderbolt ports. What's surprising is that Apple signed the driver, meaning you don't have to disable System Integrity Protection like previous workarounds required. Hotz, if you don't know the name, was the first person to carrier-unlock the iPhone and later founded Comma.ai before starting Tiny Corp to build affordable AI hardware.
But temper your expectations. It's not plug-and-play. You need Docker for Nvidia GPUs, and it only works with tinygrad, Tiny Corp's own deep learning framework. No CUDA, no PyTorch. The Thunderbolt connection also creates bandwidth bottlenecks that limit performance compared to a dedicated PC setup. As one Hacker News commenter put it, you get neither the full performance of a proper workstation nor the tight integration of Apple's native tools.
For the AI agent crowd, it's more of an experimental toy than a production tool. If you're already running LLMs locally on Mac hardware, Apple's native MLX framework will probably serve you better for most tasks. But if you've got a spare Nvidia GPU collecting dust and want to tinker with LLM inference on your Mac, Tiny Corp just made that possible without the security compromises that used to be necessary. The real story here might be Apple's willingness to sign a third-party driver at all, which suggests a slight softening of their notoriously tight grip on hardware expansion.