Solved NVIDIA DGX Spark (GB10)

For LoRA or QLoRA on a model that fits in 32 GB, a 5090 can be very attractive. For larger models or experiments that need more memory in one pool, GB10 can be easier.

If you tell us the model, precision, sequence length and whether you are doing full fine-tuning or adapters, we can give a much better answer.

How much does a GB10 cost to rent?

AxForge currently lists DGX Spark / GB10 from €0.55 per hour, excluding VAT.

The rate depends on how long you book: on-demand is higher, with lower effective hourly rates for week, month and year terms.

The DGX Spark page currently lists €0.69/hour on demand, €0.66/hour by the week, €0.62/hour by the month and €0.55/hour by the year.

Should I rent a GB10 or an RTX 5090 for AI work?

If you need a lot of memory, start with GB10. If the model fits comfortably in 32 GB and you want maximum speed, RTX 5090 is usually the more interesting option.

The big difference is the shape of the hardware: GB10 gives you one 128 GB unified memory pool, while each RTX 5090 has 32 GB of GDDR7.

For large models, long context, or workloads that are awkward to split across GPUs, I would normally start with GB10. For smaller models, image workloads, or high-throughput jobs that fit on a 5090, the 5090 can be much faster.