Rent or buy a DGX Spark
Short answer: if you'll use a Spark under about 40 hours a week, rent it, per-minute billing beats ownership by a wide margin for years. If you'll run one flat out, 24 hours a day, it pays for itself against our $0.79/hour rate in roughly seven and a half months, and buying is the better deal from there on. Everything in between is arithmetic, not opinion, and the table below shows it.
The purchase price this page uses
We sell and manage DGX Sparks on-premise ourselves, but the range below is not our sale price, it's the EU street price range we already publish in our before-you-buy guide: €3,800 to €4,800, spanning ASUS's Ascent GX10 board at the low end and the NVIDIA Founders Edition at roughly €4,800 on the EU marketplace. Rental is billed in USD and this purchase range is in EUR; we compare the raw numbers below rather than guessing an exchange rate, so convert to your own currency before you decide.
The break-even table
All rows use the single-node on-demand rate of $0.79/hour and the €4,300 midpoint of the purchase range above, shown at face value against the dollar figure. The arithmetic: hours per week × 52 = hours per year; hours per year × $0.79 = cost per year; purchase price ÷ (cost per year ÷ 12) = months to break even.
| Usage pattern | Hours/year | Rental cost/year | Months to reach €4,300 |
|---|---|---|---|
| 20 hours/week, part-time project | 1,040 | $821.60 | ~62.8 months (5.2 years) |
| 8 hours/day, weekdays only | 2,080 | $1,643.20 | ~31.4 months (2.6 years) |
| Continuous, one Spark, 24x7 | 8,760 | $6,920.40 | ~7.5 months |
| Continuous, two-node 256 GB cluster, 24x7 ($1.79/hour, purchase €8,600 for two machines) | 8,760 | $13,840.80 | ~7.5 months |
The two continuous rows land on the same break-even period because the cluster rate is exactly double the single-node rate and two machines cost roughly double one. Scaling up doesn't change when renting stops being the cheaper option, it only changes how much money is on the table either way.
Try your own numbers
The table above covers four fixed patterns. If your week doesn't look like any of them, use this instead. It runs the same formula client-side; nothing is sent anywhere.
Enter hours per week and a purchase price above zero.
If JavaScript is off, the table above already covers the four patterns most people ask about.
What the rental price includes and excludes
From our own pricing page, so this stays consistent with what you'd actually be billed:
- Included: the whole compute rate, no separate egress or ingress fees, 1 TB NVMe storage, and a stopped instance is billed at 75% of the running rate rather than the full rate.
- Excluded: applicable tax, which is added on top of the $0.79/hour figure. There is also no backup service: terminating an instance deletes the workspace with no copy kept, and if your credit runs out the machine is released and the workspace is only saved for 7 days before the same permanent deletion.
Costs of owning that renters never see
The purchase price is the start of ownership costs, not the end of them.
- Electricity. NVIDIA's own DGX Spark specification lists a 240 W system power supply and a 140 W GB10 chip TDP; day-to-day draw sits somewhere below the 240 W peak depending on load, and we don't have an instrumented idle-power measurement we're willing to publish as fact, so we're leaving a firm number out rather than repeating an unverified one.
- No warranty stopgap. A rented Spark that fails is replaced by us, usually within the hour, and you're on a different machine while we sort it out. A Spark you own that fails is yours to get repaired, and you have no compute in the meantime unless you also keep a rental account funded as a fallback.
- Idle time. A rented Spark you're not using costs nothing beyond the 75% stopped rate if you leave it reserved, or nothing at all if you terminate it. An owned Spark sitting in a rack between projects has already been paid for either way, so its idle hours are simply capacity you bought and aren't using.
The middle path: rent first, then decide
Most rent-vs-buy pages assume the two options are exclusive. Ours don't have to be, because GPUwerk is on both sides of this trade:
- Rent, then buy. Run your actual workload on a rented Spark for a few days or weeks, and if it fits, we'll deliver and install the same class of machine on-premise, managed, with the transition planned rather than guessed at.
- Buy, then rent it out. If you already own a Spark and it spends most of its time idle, you can send it to us to rent out to other customers instead of leaving it depreciating on a shelf.
Not sure which side of this you're on yet? A first engagement scopes your actual workload before you commit to either a rental budget or a purchase order.
Read the full before-you-buy guide for the memory-bandwidth arithmetic, published benchmark numbers, and the total cost of ownership case against a rig built around an RTX 5090 instead.
FAQ
What's the actual break-even point?
Continuous 24x7 use crosses a mid-range EU purchase price in about seven and a half months. Below roughly 40 hours a week, renting stays cheaper than owning for years, because the machine spends most of its time idle either way.
Does this include tax?
No. GPUwerk's $0.79/hour rate excludes applicable tax, as stated on the pricing page. Purchase prices in the EU normally include VAT. Compare ex-tax to ex-tax or the gap will look smaller or larger than it is.
Can I rent first and then buy the same machine?
Yes. GPUwerk is the only party in this comparison that both rents Sparks by the hour and sells and manages them on-premise, so you can test on a rented node and then have the same class of machine delivered and installed once you know it fits your workload.
What if I already own a Spark and it's sitting idle?
You can send it to GPUwerk to rent out to other customers instead of leaving it idle. See the host-your-Spark page for how that works and what it pays.
Does the calculator account for currency conversion?
No. Rental is billed in USD and the EU purchase price is normally quoted in EUR. The calculator compares the two numbers at face value rather than applying an exchange rate, so convert the purchase price to USD yourself first if you want an exact answer.