Why Spend This Kind of Money on Local AI Hardware?

If you looked at my desk from the outside, or if my wife ever saw the full running total for the Mac Studio M3 Ultra 256GB together with the two DGX Sparks (and the Miniforums Ubuntu server running the apps), it would look idiotic. A grown man putting serious money into computers that just sit there while he photographs wildlife or runs companies. For pure personal fun projects that judgment would probably be fair. Unless the work itself is your absolute love and hobby, the numbers rarely make sense.

That is not what this is.

I run several companies and the family’s investments. This hardware is infrastructure for that work, not a collection of toys. Most of what the agents do is coding. Systems that have already replaced SaaS tools we used to pay large yearly fees for. The recurring costs disappear. The capability stays.

One concrete system is the one we call Money Maker. Every day at market close it imports the closing prices. It looks at the portfolio and surfaces what we should consider selling or buying more of. During the day it watches for large moves and sends a Slack message when something needs attention. It produces monthly and quarterly reviews. At the end of the year it recommends tax optimizations — harvesting losses against gains in a clean way. All of it runs locally, around the clock. Nothing leaves the machines. No token counter is ticking while the market is open or while we sleep.

The cloud side is the other half of the picture. I already sit on the maximum plans for Claude Code, ChatGPT Pro, Kimi K3 and Grok Heavy. I still run out of tokens. Constantly. I could buy more subscriptions. I could put my wife on the same plans and double the capacity. Both work in theory. In practice they stay clumsy, still rate-limited, and still send investment logic and company code out to other people’s servers. The cleaner solution is capacity that lives on the desk and does not reset every month (luckily ChatGPT Pro did one of their resets late last night, so I have tokens to use, until the other resets)

Local coding models have reached a point where they are good enough for the heavy lifting that only a few months ago still needed the absolute frontier. Pair them with the cloud models for planning and the genuinely hard problems, and you get a working team: local for volume and privacy, cloud for the sharp edge. The 512GB Mac Studio and the Sparks make that combination practical instead of theoretical.

There is a small personal side as well. My kid and I still build things just for the fun of it. He runs funwildfacts.com, decent traffic, built with AI, and he learns while he enjoys it. That is the minority of the compute, and it is only possible because the majority is already justified by the real work.

The earlier post was about the hardware decision itself. This one is about the quieter question underneath it: why any of this is rational. Fixed cost instead of endless SaaS and token bills. Continuous local analysis of real money instead of waiting for the next reset. Privacy for investment logic and company systems instead of shipping everything out. Once you look at it that way, the spend stops looking like excess and starts looking like the simplest way to keep the actual work moving.