Apple just approved a driver that lets Nvidia eGPUs work with Arm Macs, and if you think this means Apple and Nvidia kissed and made up, you’re reading this story completely wrong.
The driver came from Tiny Corp, not Nvidia. Apple didn’t suddenly develop a warm relationship with its longtime GPU rival. What actually happened is that a third-party developer did the work both companies refused to do, and Apple’s approval process—for once—didn’t stand in the way.
Why This Matters for AI Work
If you’re running AI models locally, you know the pain of Apple Silicon’s memory limitations. The M-series chips are fast, sure, but when you’re trying to run larger language models or do serious machine learning work, you hit a wall. An eGPU should theoretically solve this problem by giving you access to Nvidia’s CUDA ecosystem and significantly more VRAM.
Should. Theoretically.
The reality is messier. You’re still bottlenecked by Thunderbolt bandwidth, which means you’re not getting anywhere near the performance you’d see from a native PCIe connection. This isn’t a desktop replacement—it’s a compromise that might work for specific workflows but won’t suddenly turn your MacBook into a proper AI workstation.
The Tiny Corp Factor
Tiny Corp deserves credit for actually shipping this. They claim the installation is simple enough that “a Qwen could do it”—a cheeky reference to the open-source language model. The driver supports both AMD and Nvidia GPUs, which is more than Apple has bothered to do since they transitioned to Arm.
But let’s be clear about what this represents: a band-aid solution for a problem Apple created by abandoning x86 without providing adequate GPU options for professionals who need more than integrated graphics. The fact that a small company had to step in and build this bridge tells you everything about Apple’s priorities.
What You’re Actually Getting
An eGPU setup on an Arm Mac is half-baked by design. The Thunderbolt connection limits bandwidth to roughly 40 Gbps, compared to the 256+ Gbps you’d get from a PCIe 4.0 x16 slot. For AI inference, this means slower model loading times and reduced throughput. For training, forget about it—you’re better off renting cloud compute.
You also don’t get access to all the GPU’s capabilities. Some CUDA features won’t work properly over Thunderbolt, and you’ll run into compatibility issues with certain frameworks and libraries that expect a native connection. This isn’t Tiny Corp’s fault—it’s a fundamental limitation of the architecture.
The Real Question Nobody’s Asking
Why are we celebrating the ability to duct-tape an external GPU to a Mac in 2026? Apple has had years to address the needs of users who require serious GPU compute. Instead, they’ve pushed the narrative that their unified memory architecture is sufficient for everyone, which is true only if “everyone” means video editors and casual developers.
For AI researchers, data scientists, and anyone doing GPU-intensive work, Apple Silicon remains a beautiful prison. You get incredible efficiency and battery life, but you’re locked into Apple’s ecosystem with limited expansion options. This driver doesn’t change that fundamental reality—it just gives you a slightly longer leash.
Should You Bother?
If you already own an Arm Mac and need occasional access to Nvidia’s CUDA ecosystem, this driver might be useful. The installation process is apparently straightforward, and having the option is better than not having it.
But if you’re considering buying a Mac specifically because this driver exists, stop. You’re better off building a dedicated Linux workstation or using a cloud provider. The performance compromises and compatibility headaches aren’t worth it unless you’re already committed to the Apple ecosystem for other reasons.
Tiny Corp built something genuinely useful here, and Apple deserves minimal credit for approving it. But this is a workaround, not a solution. The fact that we needed a third-party developer to make this happen in 2026 says more about Apple’s blind spots than it does about their openness to external GPUs.
đź•’ Published: