From 23 October at Acer, ASUS, Dell, Gigabyte, HP and MSI. The GB10 Grace Blackwell chip and 64 GB of unified memory. Sync Cluster Assistant links two boxes with a QSFP cable and pools the memory to 128 GB: per NVIDIA, models of up to 200 billion parameters in the office, no cloud.
- DGX Spark 64GB: GB10 Grace Blackwell Superchip, from $4,999, on sale from Friday, 23 October 2026.
- Sync Cluster Assistant: two machines over a QSFP cable and the built-in ConnectX-7 at 200 GbE, 128 GB of pooled memory. Per NVIDIA, up to 1.7x the performance of one, measured with Qwen 3.8 27B.
- One machine: models up to 100 billion parameters. Two: up to 200 billion. The numbers are NVIDIA's.
Four thousand nine hundred and ninety-nine dollars. That is the price of a high-end laptop, and from 23 October it is the price of a machine that, per NVIDIA, runs a 100-billion-parameter model on your desk without sending anything out.
The numbers are NVIDIA's and were measured on one model, Qwen 3.8 27B. With another model and another workload they will differ. 1.7x from two machines means the cable eats about 15 percent. Normal for a link between two boxes, and better than I expected from a cable on a desk.
The important part is further down: who it is for. A studio of ten, a law office, a medical practice. Anyone who wants a model over their own documents and does not want the documents to leave the room. Until now that meant either the cloud or a server with a rack and air conditioning. Now it is a box on the desk, and a second box if needed.
What NVIDIA does not say: how fast a 200-billion model answers on two of these. "Runs" and "runs nicely" are different words. We find out on 23 October, when someone puts it on a desk and loads their own model, not the benchmark.