The cheap one costs five grand

NVIDIA has added a 64GB configuration to the DGX Spark line, shipping from Acer, ASUS, Dell, Gigabyte, HP and MSI, starting at $4,999. Same GB10 Grace Blackwell silicon, same claimed petaflop of AI compute, same chassis, same software stack. The only meaningful change is that the unified memory pool is half what it was.

And this is the budget option. When a vendor cuts a product’s headline capacity in half and the result still lands at five thousand dollars, the savings got eaten somewhere between the factory and the invoice. Memory contracts are where. LPDDR5X pricing has been brutal all year, and a 128GB coherent pool is now a genuinely expensive bill of materials, not a rounding error on a GPU die.

So the honest read on the 64GB Spark is that it’s a concession, not a product decision. NVIDIA didn’t segment the lineup because developers were asking for less memory. It segmented because the price of the thing people actually wanted drifted out of reach.

Two boxes are not one box

The pitch alongside the launch is that you can start with one unit and cluster two of them for 128GB across the cluster. Treat that carefully. 128GB spread over two machines and an interconnect is not 128GB of unified memory on one chip. For data-parallel fine-tuning or running several agents at once, clustering is fine and genuinely useful. For loading a single large model that needs to sit in one address space, it isn’t a substitute, and anyone buying two 64GB units to fake a 128GB machine has spent more money for a worse version of the box they skipped.

Which leaves a narrow but real buyer: someone whose workloads fit comfortably under 64GB. Quantized models in the 20B to 30B range, LoRA runs, agent loops that need to hammer a local endpoint without a metered API in the path. That describes a lot of practical local AI work in 2026, as the interesting small models have gotten considerably better at reasoning and tool use. If that’s your workload, the 64GB Spark is the Spark you should have been looking at anyway.

The comparison NVIDIA would rather you skip

At this price, the Spark competes less with cloud instances than with a desktop holding a consumer card. Memory bandwidth on the GB10 has always been the compromise in this design, far below what a discrete gaming GPU delivers, and that gap shows up directly in tokens per second. You buy a Spark for capacity, coherence, and the fact that it runs the same stack as the big machines. You do not buy it for raw speed per dollar.

Six OEMs are shipping the 64GB configuration. Whether it still costs $4,999 in six months depends on DRAM contract pricing, which is not a thing NVIDIA controls.