NVIDIA DGX Spark (GB10 Grace Blackwell)

The DGX Spark is one of the more interesting product ideas NVIDIA has shipped in years: a 150 mm cube weighing barely over a kilogram that puts a Grace Blackwell superchip, a 20-core Arm CPU and 128 GB of coherent unified memory on your desk. The pitch is not "fastest", it is "biggest model you can fit in something this small", and understood on those terms it makes a lot of sense. The champagne-gold metal chassis with its foam end panels is a deliberate nod to the rack-scale DGX machines, and the I/O tells you who this is really aimed at — a 10 GbE port is the boring option, while dual QSFP cages fed by a ConnectX-7 NIC at up to 200 Gbps exist so you can pair units or hang the thing off serious infrastructure.

This is a machine for model developers, researchers and ML engineers who need to prototype against large models locally — people who care far more about whether a model loads at all than about how many tokens per second it emits. If you want maximum throughput per dollar for serving, the numbers below will not flatter it. If you want a quiet, portable, CUDA-native sandbox that behaves like a small slice of a datacenter, there is very little else shaped like this.

The central trade-off is impossible to miss. That 128 GB unified pool is the whole reason to buy a Spark, but it is fed by only 273 GB/s of bandwidth — an order of magnitude below an H100's ~3.35 TB/s and well under half of what an M3 Ultra Mac Studio manages. Memory-bound large-model decode is therefore where the Spark struggles, and the benchmark picture backs that up: LMSYS measured GPT-OSS 20B in Ollama at around 2,053 tokens/s prefill and 49.7 tokens/s decode, roughly a quarter of an RTX Pro 6000 Blackwell Workstation card and comfortably behind a single RTX 5090. For pure inference throughput per dollar, this is not the machine.

It looks much better when you play to its strengths. Smaller models and higher batch counts scale well — Llama 3.1 8B at batch 32 hitting 368 tokens/s decode is a respectable result for a 1.2 kg box — and single-sample PyTorch inference has been found roughly on par with vastly more expensive hardware, which is exactly the workload a developer iterating on code actually runs. Software support is also unusually mature for a first-generation device: SGLang came to consumer-class hardware for the first time here, speculative decoding via EAGLE3 is supported, and Ollama-driven local coding workflows work out of the box. Tom's Hardware rates it above AMD's Ryzen AI Max+ 395 as an all-round local AI toolkit, which matches our read of it.

The sting is the price. At the original $3,999 with a 4 TB SSD, this was a fairly easy recommendation for anyone doing serious local model work. The move to $4,699 in February 2026 — an 18 percent jump NVIDIA blames on memory supply constraints — pushes it into territory where a workstation GPU starts to look tempting for anyone whose models fit in 24–48 GB. Note also that we have not seen verified figures for power draw, thermals or acoustics, so we cannot speak to how it behaves under sustained load.

Our rating reflects a genuinely well-executed and, in its niche, near-unique product that is held back by a hard bandwidth ceiling and a price increase that erodes its value case. Buy it for the 128 GB and the software stack; do not buy it expecting datacenter throughput.