vs renting
Rent for bursts. Own the baseline.
Rented datacentre cards are faster than this machine and perfect for intermittent heavy lifts — we say that first, so the payback numbers below mean something.
The payback table, at 24/7 duty
Verified hourly rates (community tier), monthly cost around the clock, and months until a $4,699 Spark has cost less. At office-hours duty (176h/mo) every payback stretches 3–4×: 13–39 months — renting wins there.
| Rented GPU | $/hour | 24/7 monthly | Spark pays back in |
|---|---|---|---|
| RTX PRO 6000 96GB | $1.69 | $1,234 | 3.8 mo |
| H100 80GB PCIe | $1.99 | $1,453 | 3.2 mo |
| A100 80GB PCIe | $1.19 | $869 | 5.4 mo |
| RTX 5090 32GB | $0.69 | $504 | 9.3 mo |
What renting costs that isn't on the invoice
The part the hourly rate hides.
- Your prompts, weights and data live on a provider's fabric under a provider's terms — the sovereignty argument does not transfer to a rented card.
- Storage bills separately and idles at double rate; ephemeral instances mean re-downloading 100GB models on every cold start.
- Spot/interruptible pricing looks great until the instance vanishes mid-run — the cheap tiers carry no SLA.
- The meter shapes behaviour: experiments you would run freely on owned hardware quietly stop happening at $2/hour.
Questions, answered straight
- When is renting simply the right answer?
- Intermittent heavy work: a fine-tune weekend, a one-off batch job, evaluating whether local AI fits you at all. A rented H100 is several times faster than this machine and costs a few dollars an hour. Rent first, buy when the meter starts running monthly.
- Why not rent long-term instead of buying?
- At continuous duty the crossover is brutal: 3–5 months against every big-memory card, and the rented option still holds your data on someone else's machine. Past the crossover you are paying monthly for less sovereignty.