Mac eGPU and Apple’s local‑AI rise

A Mac used to have a simple rule for external graphics processors: if you wanted one, you needed an Intel Mac. Apple’s current support page still says external graphics processors require “a Mac with an Intel processor,” which is why the new Apple Silicon driver story is getting so much attention. (support.apple.com) An external graphics processor is a laptop-sized computer hitching a ride on a bigger graphics card in a box. The box connects over Thunderbolt or Universal Serial Bus 4, so the Mac keeps its keyboard, battery, and screen while the outside card does the heavy math. (support.apple.com) That heavy math matters more for artificial intelligence than for games right now. Apple’s own machine learning team says people are increasingly using Macs to run large language models locally with MLX, its open-source framework built for Apple silicon. (machinelearning.apple.com) MLX works because Apple silicon uses unified memory, which means the central processor and graphics processor share one pool instead of shuttling data back and forth like two cooks using separate kitchens. Apple says that cuts memory movement and makes local inference and fine-tuning practical on a Mac. (machinelearning.apple.com) The new twist is that Tiny Corp said on March 31, 2026 that Apple approved its DriverKit extension for external Advanced Micro Devices and NVIDIA cards on Apple Silicon Macs. NYU Shanghai’s research technology group summarized the change as the first officially sanctioned path for third-party discrete graphics processors on Apple silicon macOS. (rits.shanghai.nyu.edu) The approval is narrower than “external graphics processors are back.” Multiple reports say the signed driver is for compute, not for accelerating monitors or bringing back the old one-cable gaming setup, and AppleInsider says external-display video is not accelerated at all. (appleinsider.com) (siliconreport.com) The security detail is a big part of why developers care. Tiny Corp’s driver reportedly loads with Apple’s approval through DriverKit, so users do not have to disable System Integrity Protection, which is one of macOS’s main guardrails against low-level tampering. (rits.shanghai.nyu.edu) (siliconreport.com) The hardware support is also modern and selective. Tiny Corp’s published requirements, as relayed by AppleInsider and Silicon Report, start at Advanced Micro Devices RDNA 3 cards and NVIDIA Ampere cards, require macOS 12.1 or later, and use Thunderbolt 3 or Universal Serial Bus 4 enclosures. (appleinsider.com) (siliconreport.com) Tiny Corp’s own early benchmark gives a sense of the target user. NYU Shanghai’s summary says a Mac mini with an M4 chip connected to a Radeon RX 7900 XTX reached 18.5 tokens per second on Qwen 3.5 27B, which is the kind of number local-model users watch because it changes whether a tool feels instant or sluggish. (rits.shanghai.nyu.edu) That speed story fits a software stack that is already moving toward local use on Macs. Apple used its 2025 Worldwide Developers Conference session to teach developers how to run and fine-tune large language models with MLX on a Mac, and LM Studio now supports Apple’s MLX runtime on Apple Silicon Macs for local chat, document work, and OpenAI-like local application programming interfaces. (developer.apple.com) (lmstudio.ai) So the shift is not “Mac beats a desktop graphics card tower.” The shift is that a MacBook Pro or Mac mini can now sit in the middle of a local artificial intelligence workflow, use Apple’s own silicon when that is enough, and reach for an external NVIDIA or Advanced Micro Devices box when the model gets too large or the response needs to get faster. (machinelearning.apple.com) (rits.shanghai.nyu.edu) For creative work, that changes the tradeoff from “cloud or nothing” to “local first, cloud when needed.” A video editor, designer, or app developer can now plausibly keep drafts, prompts, and private files on the Mac, get low-latency replies from local models, and add an external compute box only when the project outgrows the machine on the desk. (machinelearning.apple.com) (lmstudio.ai)