Solved NVIDIA DGX Spark (GB10)

On-demand is described as pay per started hour while you use it, while reserved week/month/year capacity is booked as an always-on term.

For a scheduled reservation, the booked period is what matters. If you need exact billing behavior for a specific schedule type, check the current console terms before starting it.

Can I run a huge MoE model on a GB10?

Possibly. MoE models can be much larger on paper than the number of parameters actually used for each token.

That is one of the more interesting GB10 use cases: the 128 GB pool can hold large quantized weights, while only a subset of experts are active during generation.

The exact answer still depends on the model, quantization and serving engine. AxForge lists community GB10 results for MoE models well above 100B total parameters.

Can AxForge generate images?

Yes. The serverless catalogue currently includes ERNIE Image Turbo for image generation.

The current public price is €0.02 per generated image.

For editing an existing image, AxForge separately lists FLUX.2 Klein 4B on the image-edit endpoint.

What GPU is better for really long context?

Usually the one with more usable memory.

Long context can consume a lot of KV-cache memory, so GB10's 128 GB unified pool is useful even when the model weights themselves would fit on a smaller GPU.

A 5090 can still be faster if the model plus the context window fits into its 32 GB, but once memory gets tight the comparison changes.