Direct answer: If you cannot get a 128 GB DGX Spark right now, you do not have to stop working. Renting GPU time lets you start this week, find out what your workload really needs, and decide later whether buying makes sense. Vast.ai’s pricing page, read on October 23, describes prices “set by supply and demand across 40+ data centers” and three plan types: on-demand, interruptible, and reserved. We could not read live per-GPU prices from that page, so the numbers in our worked example below are illustrative placeholders, not Vast.ai quotes. Replace them with the current offer you actually pick.
This is a starting guide, not a verdict. Renting is a good first step for many buyers and a bad long-term answer for others; we explain which is which.
Why rent first
The last NVIDIA Marketplace reading we could take (October 6) showed the 128 GB DGX Spark at $6,950 and out of stock. Aradia’s turnkey Spark, read October 23, was listed at $15,125. Spending several thousand dollars before you know your real memory needs, context lengths and usage hours is a risk. A rental pilot answers those questions for the price of a few hours of compute. Our out-of-stock options guide compares all the ways to get started; this page goes deep on the rental route.
How Vast.ai’s three tiers differ
According to Vast.ai’s pricing page on October 23:
| Tier | What the page says | Good for |
|---|---|---|
| On-demand | Guaranteed uptime, per-second billing, no interruptions, spin up and down anytime | Interactive work, demos, anything that must not be cut off |
| Interruptible | “50%+ cheaper”; preemptible, so it may be reclaimed | Fault-tolerant batch jobs where you can checkpoint and resume |
| Reserved | “Up to 50% off” for 1, 3 or 6 month commitments; contact sales | Steady, predictable workloads |
Because prices are set by supply and demand, the same GPU can cost different amounts on different days. Compare offers at the moment you start, and re-check when you scale up. Prices on the page are not a quote.
A break-even worksheet with illustrative numbers
The question “when does buying beat renting?” depends on three inputs: the hourly rate you actually pay, how many hours you use per month, and what a comparable system costs. The table below uses placeholder rates to show the shape of the answer. It ignores electricity, storage, networking, taxes, resale value, your time, and the fact that a rented GPU and a DGX Spark are not the same machine.
| Illustrative hourly rate | Hours to spend $6,950 | At 40 hours/month | At 200 hours/month |
|---|---|---|---|
| $0.50 | 13,900 hours | about 348 months | about 70 months |
| $1.00 | 6,950 hours | about 174 months | about 35 months |
| $2.00 | 3,475 hours | about 87 months | about 17 months |
(The $6,950 input is the last NVIDIA Marketplace price we could read, which was out of stock; a different price changes every row.)
The pattern is what matters. At low usage, renting stays cheaper for years. As usage approaches full-time, buying starts to win within a couple of years. Measure your real hours during a pilot before you trust any break-even number, and use our interactive break-even calculator or the cloud GPU cost calculator article to build the worksheet with your own rate.
A practical pilot plan
- Pick the model you actually want to run and note its memory need; our GPU memory guide explains how to estimate it.
- Rent the smallest offer that fits that memory, using on-demand for the first sessions so interruptions do not confuse your results.
- Log hours, not guesses. Write down start and stop times and what you ran.
- Switch batch work to interruptible once you can checkpoint, to test whether the lower price is worth the interruptions.
- After two to four weeks, compare your measured hours and rate with the worksheet above.
What to check before you rent
- Who runs the machine. Vast.ai describes a marketplace of data centers. Do not upload confidential or regulated data until you have checked who operates the machine and whether the offer meets your requirements. This is our general caution, not a statement from Vast.ai.
- Storage keeps costing money. Stopping an instance does not necessarily stop every charge; read the billing documentation for your offer.
- Interruptions. Interruptible instances can be reclaimed. Only use them for jobs you can resume.
- Performance is not comparable by hourly price. A rented GPU and a DGX Spark have different memory, speed and software behavior. Compare completed work, such as tokens per second or job time, not just price per hour.
When buying still wins
Rent first, but plan to buy if you will run the system most hours of most days, if data must stay on premises, or if you want a fixed monthly cost and an asset you control. Our DGX Spark versus cloud GPUs comparison covers those cases, and our buying checklist lists the questions to settle before spending money. If you want a finished private system instead of running it yourself, see the DGX Spark buyer’s guide.
What we could not verify
- Live Vast.ai GPU prices: the pricing page loads them dynamically, so we did not record any, and the worksheet above uses placeholders.
- Current DGX Spark price and stock: NVIDIA Marketplace did not load on October 23; the $6,950 input is from October 6.
- Your real utilization and data constraints: only a pilot can tell you.
Sources and verification note
Vast.ai’s pricing page and Aradia’s pricing page were read on October 23, 2026. The NVIDIA Marketplace figure is from October 6. Prices and offers change continuously; check the live offer before you rent.
Affiliate disclosure: AI Compute Scout may earn a commission if you sign up through a marked Vast.ai referral link. This does not affect the analysis, and we recommend buying hardware when your usage justifies it.