What is GPU rental?
GPU rental is paying a provider for temporary access to GPU hardware you don't own, typically billed by the hour or minute, instead of buying the hardware yourself. The category spans a wide range: a fraction of a shared card split across many renters, a whole card in a shared server, or an entire dedicated machine allocated to one renter at a time.
The core trade: access without ownership
Buying a GPU means a large upfront cost, plus the ongoing burden of powering, cooling, housing, and maintaining it, and it sits idle (and still depreciating) whenever you're not using it. Renting shifts that to a running hourly cost with no upfront hardware spend, no physical maintenance, and the ability to stop paying the moment you stop using it. The trade-off is that you don't own anything at the end, and your access depends on the provider staying available and priced reasonably. Which side wins depends on how continuously the hardware gets used and over what time horizon.
Shared vs dedicated allocation
Within GPU rental, the meaningful split is how much of the hardware is actually yours at any moment. Shared or fractional rental splits a GPU's compute or memory across multiple renters, using scheduling or partitioning to keep them from directly colliding, usually at a lower price reflecting that the hardware isn't exclusively yours. Dedicated rental allocates a whole GPU, or a whole machine, to one renter, with no other tenant's workload running on it concurrently. The two aren't better or worse in the abstract; a batch job tolerant of variable latency might be fine on shared capacity, while a production service with predictable latency requirements usually wants dedicated. See dedicated vs shared GPUs for the isolation mechanics involved.
Billing models vary too
Most rental providers bill by the hour or by the minute, with usage-based rates that scale with time actually running. Some also offer reserved or committed pricing at a discount in exchange for a longer minimum commitment, which trades flexibility for a lower rate. The right choice depends on whether the workload is intermittent, spiky, or genuinely continuous.
What to check before renting
Hourly rate alone doesn't tell the whole story. It's worth checking whether billing is per-minute or rounds up to the hour, whether there's a minimum commitment or top-up requirement, whether egress bandwidth is billed separately, and what happens to your data and running workload if your balance runs out mid-session. Those terms differ meaningfully between providers even when the advertised hourly rate looks similar, and they're usually stated plainly on a provider's own pricing or terms page rather than hidden. It's also worth checking where the hardware physically sits, since data residency requirements can rule out a provider whose region doesn't fit, regardless of price.
Where GPUwerk fits in this category
GPUwerk rents dedicated DGX Spark machines, not fractional or shared GPU slices: one Spark, with 128GB of unified memory, allocated entirely to one customer at $0.79/hour from EU-Central, or a two-node cluster at $1.79/hour. That places it at the dedicated, hourly-billed end of the GPU rental spectrum, alongside other providers offering dedicated instances, rather than in the fractional-sharing segment of the market. It's one option among several in a broader category; see the hardware page for the full specification and pricing for current rates.