M5 Ultra Mac Studio hits 50 tokens/sec as prices jump

Apple discontinued the Mac Pro this spring, marking the end of its expandable tower era. Apple Silicon's unified memory architecture has made high-end Macs a good fit for running language models locally, and Apple was caught off guard by demand from users running software like OpenClaw and open-weight models. Combined with an AI-fueled memory shortage, this made most Mac Studio systems nearly impossible to buy for months.

The M5 Ultra is a two-generation jump because there was no M4 Ultra, and it doubles the major specs of the M5 Max by pairing two M5 Max chips. In practice, local AI coding performance reaches just over 50 tokens per second, compared with around 31 for the M5 Max. The M5 generation's neural accelerators also improve time to first token, making models feel more responsive. However, the M5 Ultra consumes about 75 W during video transcoding, similar to the M5 Max, suggesting it may be power-limited to fit the existing Studio design.

For those willing to pay, the M5 Ultra Mac Studio offers a way to run capable AI models locally without data leaving the home. But the 25 percent and up price increases, driven by the memory shortage, make the high-end configurations a significant investment. Whether these prices hold depends on how the memory supply situation evolves, and for many users, cloud-based AI services may remain the more economical choice.