NVIDIA DGX Spark

NVIDIA shipped the DGX Spark in October 2025 — a Grace Blackwell GB10 superchip, 128 GB of unified LPDDR5X memory, and a petaflop of FP4 compute in a 150 mm desktop box that plugs into any household outlet. It is NVIDIA’s answer to the question the unified-memory Macs and Strix Halo boxes had been posing: what does a purpose-built local-AI appliance look like when it comes from the CUDA side of the fence?

D-Central sells and configures the DGX Spark. It’s a build-to-order, quote-first product like the rest of our AI line — see the NVIDIA DGX Spark product page to request a quote, or read the Canadian buyer’s guide for sovereignty and Law 25 context.

What’s in the box

• GB10 Grace Blackwell superchip: Blackwell GPU with 5th-gen Tensor Cores plus a 20-core Arm CPU (10× Cortex-X925 + 10× Cortex-A725, co-designed with MediaTek)
• 128 GB unified LPDDR5X on a 256-bit bus at 273 GB/s — CPU and GPU share one pool, no PCIe shuffling
• Up to 1 petaflop of FP4 AI compute (with sparsity), with NVFP4 as the native low-precision format
• 4 TB NVMe storage, Wi-Fi 7, 10GbE, and dual ConnectX-7 200 Gb ports — two Sparks link directly to run models up to the ~405B class
• 240 W external power supply (the SoC itself is a 140 W part) — trivial on any 120V/15A circuit
• Runs NVIDIA’s DGX OS (Ubuntu-based) with the full CUDA stack preinstalled

Honest math: capacity king, not speed king

The spec that defines this machine is not the petaflop — it’s 273 GB/s of memory bandwidth. Token generation on large models is bandwidth-bound, and 273 GB/s is roughly a third of an M3 Ultra (~800 GB/s) and a quarter of an RTX 5090 (1,792 GB/s). A dense 70B model at Q4 fits easily in 128 GB but generates noticeably slower here than on the Mac.

What the Spark is genuinely excellent at is the workload 2025–2026 actually produced: sparse MoE models in the 60–120B-total class. gpt-oss-120b is the poster child — 117B total parameters that need ~61 GB in native MXFP4, but only 5.1B active per token, so the bandwidth ceiling barely hurts while the 128 GB pool holds the weights, a long context, and your whole serving stack at once. Prompt processing and fine-tuning lean on compute rather than bandwidth, and that petaflop of FP4 plus the CUDA ecosystem is where the Spark pulls away from the Mac — LoRA fine-tunes, batch inference, and anything that expects nvidia-smi to exist just work.

Pricing, as of July 2026: the Founders Edition launched at $3,999 USD in October 2025; the 2026 DRAM shortage pushed current pricing to ~$4,699 USD at major resellers. Partner variants (ASUS, Dell, Lenovo, HP) with smaller SSDs list somewhat lower. Canadian pricing moves with the exchange rate — request a current quote and we’ll spec it against your actual workload rather than the spec sheet.

Who should buy it — and who shouldn’t

Buy it if: you want the largest open-weight models a single quiet box can hold, you fine-tune or prototype against the CUDA stack that production clusters run, or you need 128 GB of model memory on a 120V circuit without building a multi-GPU rig. It is also the rare AI appliance that draws less than a gaming console under load — heat and noise are non-issues.

Skip it if: your models fit in 24–32 GB (a used RTX 3090 or a 5090 gives you several times the token speed for less money), or your priority is maximum tok/s on dense 70B chat — the M3 Ultra’s bandwidth wins that fight. We wrote up the full trade-off, including total cost of ownership against a self-built workstation, in DGX Spark vs custom AI build.

The sovereignty frame, as always: this is compute you own, in your building, with no per-token meter and no retention policy but your own — the argument in Sovereign AI in Canada and the Sovereign AI Manifesto. Credit where due: the Spark stands on Grace Hopper’s shoulders, MediaTek’s Arm work, and a decade of DGX systems engineering shrunk to desk size. We’re happy to document it — and happier to put one on your desk.