Apple Macs Can Now Train Private AI for Less Than Nvidia Chips

WHAT HAPPENED: A team of researchers took four Mac Studio desktop computers, linked them together with Thunderbolt cables (the same kind of cable you might use for a monitor), and used them as one combined machine to fine-tune an AI model called Qwen3-9B. Fine-tuning means teaching an already-trained AI your own specialized material — company documents, legal files, medical records — rather than building an AI from scratch.

The problem they were solving is money and memory. Big AI models are usually customized on Nvidia's H100 chips, the gold-standard hardware inside data centers. Those chips are expensive and have limited memory, so a model that needs to absorb lots of private data may simply not fit. Apple's chips take a different approach: one large shared pool of memory (called unified memory) that can hold more at once, at a much lower total system cost.

The team's engineering work mattered. They found that real-world Thunderbolt speeds were far below what the spec sheet promises, so they wrote software tricks — using multiple cable connections at once and overlapping work between chips — to squeeze out more performance. The result was up to 1.6 times faster than their starting setup, reaching 936 tokens per second on long documents. That's still slower than an H100 in raw speed, but the researchers argue the trade-off favors Apple when you weigh memory capacity, speed, and purchase price together.

WHY IT MATTERS: This is a laboratory result, not a product you can buy today, and setting it up takes real technical skill. But the direction is clear. If private AI customization gets cheap enough, more hospitals, law firms, banks, and small companies can build AI that knows their business without ever handing their secrets to an outside vendor.