Apple Just Made the Best Local AI Machine. Do Not Buy It Yet.
🤝 Join the Community: https://discord.gg/MRESQnf4R4
🔥 Apple just announced the new Mac Studio with M5 Max and M5 Ultra, and the new Mac mini with M6 and M5 Pro. Pre-orders are open now, shipping September 22. I have spent the last 24 hours running the numbers, and on paper, the base M5 Ultra Mac Studio with 96GB of unified memory is the best value machine for local AI that has ever been made. So why am I not recommending you buy it? Because this is paper data. There are no tokens-per-second benchmarks yet. Let me show you the math.
Here is the deal. Memory bandwidth is the whole game for local AI token generation. The M5 Ultra ships 1.2 terabytes per second of bandwidth, that is 50% more than M3 Ultra, with up to 512GB of unified memory, a 36-core CPU, and an 80-core GPU with Neural Accelerators for the first time on Ultra. Apple says it delivers up to 4.3x the peak AI compute of M3 Ultra. The Mac mini M5 Pro ships 307GB/s with up to 64GB of unified memory. Compare that to the RTX 5090’s 1.79TB/s and the DGX Spark’s 273GB/s, and you see the shape of it. The M5 Ultra is slower than a 5090. My estimate is around 30% slower. That is an estimate, not a benchmark. But at $5,499 for 96GB, complete and quiet and on your desk, no GPU in the gap below the 5090 and the Pro 6000 beats it on cost-per-token. The Mac mini M5 Pro with 64GB is your entry point if you want Spark-class capability for under two grand.
The choice inside the M5 Ultra is interesting. There are two CPU configurations. My understanding, and I need to double-check this, is that token generation is bandwidth-bound and not CPU-bound, so both tiers sit on the same 1.2TB/s. If that holds, the cheaper CPU does not cost you speed, and the base Ultra with 96GB is the deal. I will confirm this when real benchmarks land. People love to FOMO. Most buyers will max everything. My take: the base model is the one to watch, and the pre-order will not go as insane as you think.
These are my early reflections. There will be a lot of talk about this on the AI stack call we have every Wednesday on our Discord channel. If you want to argue the math with me, the link is in the description. I am not recommending buying today, I would wait for the numbers. But if you do buy, buy the boring config: base M5 Ultra, 96GB, no maxing.
When the first real benchmarks land, I will do the math video: Mac Studio M5 Ultra versus RTX 5090 versus DGX Spark, tokens per second per dollar, head to head. Subscribe so you do not miss it. The browser-side AI agent I run every day (the augmenter extension I built) is exactly the kind of workload that thrives on a machine like this, so this story will keep evolving.
⸻⸻⸻⸻
🚀 Build Your AI Creative Collaborator, an Augmentor, with ResonantOS Open (Free)
Architect an AI that remembers, aligns with your values, and tunes to your creative DNA.
http://resonantos.com/
⸻⸻⸻⸻
📬 Stay Connected
📝 Newsletter: http://augmentedmind.co/
🌐 Website: https://manoloremiddi.com/
💼 LinkedIn: https://www.linkedin.com/in/manoloremiddi/
🤖 Augmentatism Philosophy: https://augmentatism.com/
🤝 Join the Inner Circle: https://discord.gg/MRESQnf4R4
🪐 Cosmodestiny Philosophy: https://cosmodestiny.com/
⸻⸻⸻⸻
📕 My Free eBook
The Last Human Teacher: A Survival Guide to the Age of AI and the TikTok Brain.
The Last Human Teacher: A Survival Guide to the Age of AI and the TikTok Brain
⸻⸻⸻⸻
🛠 Tools:
🎁 GLM Coding Plan 10% OFF for friends:
https://z.ai/subscribe?ic=4M3MEVMSA4
🎁 MiniMax Coding Plan 10% OFF for friends: https://platform.minimax.io/subscribe/coding-plan?code=d5bS2lwRtz&source=link
🧠 AI Voice: https://try.elevenlabs.io/gkb7rak7o4zi
🔍 AI Research: https://perplexity.ai/pro?referral_code=VNRVYO20
🌐 VPN: https://refer-nordvpn.com/SOaIHsnlppb
✂️AI Video Editing: https://gling.ai/?via=manolo
📡 One free month of Starlink service! https://www.starlink.com/residential?referral=RC-DF-6112588-38079-56&app_source=share