DGX Spark ships — $3,999
NVIDIA’s DGX Spark personal AI supercomputer is listed at $3,999 and is shipping now — pitched as a countertop AI box for edge and developer setups (x.com). The announcement was contrasted directly with Apple’s upcoming Mac Studio M5 Ultra, framing a broader edge‑AI arms race between single‑vendor superboxes and traditional workstation makers (x.com).
NVIDIA began customer fulfillment on Oct. 15, 2025 and lists availability through Nvidia.com and partner OEMs including Acer, ASUS, Dell Technologies, GIGABYTE, HPI, Lenovo and MSI. (pcmag.com)) The Spark is built around NVIDIA’s GB10 “Grace Blackwell” superchip and is advertised as delivering up to one petaFLOP of FP4 AI performance with 128 GB of coherent, unified system memory. (nvidia.com)) NVIDIA ships the machine with DGX OS on Ubuntu and a preinstalled NVIDIA AI software stack that includes the Agent Toolkit’s NemoClaw, developer access to NVIDIA NIM and Blueprints, and compatibility with tools such as PyTorch, Jupyter and Ollama. (nvidia.com)) Hardware documentation and vendor listings show the Spark supports up to 4 TB of NVMe storage, uses a ~120W TDP GB10 SoC design, and can be linked via NVIDIA ConnectX networking so two units together can target roughly 405‑billion‑parameter model workloads. (resource.naddod.com)) Independent Ollama inference tests and early lab writeups report the DGX Spark outperforming a tested Mac Studio and RTX‑4080 rigs in real‑world token throughput for a 120B GPT‑style model (example measurements: ~1,159 tokens/sec prompt eval and ~41 tokens/sec generation on the DGX Spark in one benchmark). (glukhov.org)) Apple’s Mac Studio M5 Ultra is still reported from leaks and analyst coverage as a mid‑2026 arrival rather than an immediate rival product, which frames the current contest as one between Nvidia’s single‑vendor “superbox” launches and forthcoming workstation updates from traditional OEMs. (macworld.com))
