No, you're not going to run Claude locally on your Mac Studio
The math nobody does on local models: RAM, bandwidth, throughput. The core problem is physical, not software.
No, you’re not going to “run Claude locally on your Mac Studio”. Sorry, had to get it out.
I keep seeing very serious folks explaining that they’ve cut their AI bill by X running their agents locally. That APIs are done.
My friend, 2 minutes.
Yes, for async doc processing, it’s not dumb. But that’s a tiny slice of what people actually do at work. For the rest (real-time, multiple people, real agentic work), we need to talk orders of magnitude.
A million seconds is 11 days. A billion? 32 years. The brain doesn’t grasp the difference, and yet it’s huge.
Same with models. People run Llama 70B locally, and it works. 70 billion parameters. So others tell you all you need is a Mac Studio to run GLM-5.2, the best open-source. 750 billion. Ten times more. And that “ten times”, nobody visualizes it.
How does it fit on a Mac Studio? By compressing it. At full precision, GLM weighs 1.5 TB (of RAM). In 4-bit (the trimmed-down version), 375 GB. On a 512 GB Mac Studio (€10,000 if not out of stock), it fits.
Your €3,000 MacBook has 32 GB. Here we’re talking 512.
That leaves ~130 GB for context. But GLM handles 1M tokens: a single big agentic session can eat 100 GB. Translation: your €10,000 machine is saturated by ONE conversation. A single-task workstation, not a server.
And speed? 819 GB/s for the Studio, 3.35 TB/s for an H100. 4 times slower. In async you don’t care. In real-time agentic it’s dead.
“Fine, a lighter model then.”
Test it, no need to wait: grab Haiku, Anthropic’s small model, run it for 10 minutes and tell me what you can actually delegate to it in your job without going back over its work.
The core problem is physical, not software.
And our appetite for tokens is exploding faster than the hardware. When it launched, ChatGPT answered ONE question in 15 seconds, and that was already wild. Today, a single prompt fires 5 “ultrahard” reasoning cycles followed by a 4-phase workflow, 3 to 5 agents each, running for 1h30. For ONE task. The hardware improves, but what we ask of it moves much faster.
And to serve such a model in prod (not load it: serve it, with throughput), you need a whole node, like 8 B200 cards, 1.44 TB of VRAM, ~$500,000. The price of a house. And even that holds a few dozen sessions in parallel. Not thousands.
Multiply by millions of people, each launching several agents. That’s why a datacenter is hundreds of thousands of GPUs (xAI lines up 500,000+ on a single site).
Local doesn’t replace datacenters. Confusing the two is just not knowing how to read orders of magnitude.