Your idle GPU box is a tax move and a private inference cloud at the same time

For US indie AI/ML devs, small studios, SMB IT buyers

If you're a small AI/ML shop in the US and you were already planning to buy a GPU box or a few Mac minis this year, there are two separate reasons to do it before your tax year closes instead of "whenever" — and they compound.

Reason 1: depreciation timing is not a rounding error

Equipment you place in service — GPUs, servers, workstations — doesn't have to sit on a multi-year depreciation schedule. Between Section 179 expensing and bonus depreciation (100% for qualifying property placed in service after January 19, 2025, under current law), a lot of small businesses can write off the entire purchase in the year they buy it, instead of spreading it over 3-5 years. The exact limits and phase-outs change and depend on your situation — this isn't tax advice, talk to your accountant — but the shape of the incentive is real and it's timing-sensitive: it only helps if the asset is in service before your tax year ends.

Reason 2: the box doesn't have to sit idle

Here's the part most of this conversation skips: buying the hardware for the depreciation timing doesn't mean it has to become a space heater under your desk. We built cloud-murakumo on a simple observation — most GPU/Apple Silicon boxes bought by small teams run hot for a few hours a day and idle the rest. Ours does too (11 Mac minis, real fleet, not a slide). Point the same box at cloud-murakumo's fleet and it serves OpenAI-compatible inference the rest of the time, earning a compute-revenue share instead of just depreciating on a spreadsheet.

We're not going to pretend this beats a rack of H100s on raw $/token — it doesn't, and we said so from day one (see the inference economy positioning). What we're selling isn't FLOPS density. It's:

• the asset you were already buying, doing something instead of nothing between your own workloads
• an audit-able run ledger — every inference is signed and appended, so you (or your accountant, or a future auditor) can see exactly what ran and what it earned
• try before you buy — the live playground runs the real scheduler for free, no signup, no purchase

What this actually looks like

murakumo.cloud/#store/us sells two things right now:

1. Credits — top up with a card if you want to use the fleet without contributing hardware.
2. GPU/PC node ownership shares — buy the physical node, it gets deployed into the fleet, you get a compute-revenue lease arrangement on it. This is the one that lines up with the depreciation timing above.

Card payments go through Stripe today (no installment financing at this time). Crypto payment is not live yet — there's a placeholder button so you know it's coming, not a promise of when.

The live playground runs the actual scheduler that places functions on the fleet — it's not a mockup, it's the same .cljc code the CLI uses. Slide the load up and watch real placement/autoscale/cost math happen in your browser.