NVIDIA just gave local-AI builders a cheaper door into the DGX Spark family. On October 2, 2026, the company said a 64GB unified-memory configuration will ship from Acer, ASUS, Dell, Gigabyte, HP, and MSI starting Friday, October 23, at $4,999. Same GB10 Grace Blackwell Superchip, same DGX OS and NVIDIA AI stack as the 128GB box. Half the memory, and a sticker meant for people who do not need (or cannot find) the fatter SKU while RAM prices stay ugly.
That lands the same week The Register reported NVIDIA lifting the 128GB DGX Spark to about $6,950, nearly 75% above last year's price. So the useful question is not "is $5K cheap." It is whether 64GB plus optional ConnectX-7 clustering beats paying ~$7K for 128GB, and who should wait for stock to settle.
64GB vs 128GB, side by side
| Config | Unified memory | Starting / reported price | NVIDIA-stated model ceiling | Where you buy it |
|---|---|---|---|---|
| DGX Spark 64GB (new) | 64GB | $4,999 (partner MSRP) | Up to ~100B parameters on one unit | Acer, ASUS, Dell, Gigabyte, HP, MSI from Oct 23 |
| DGX Spark 128GB | 128GB | ~$6,950 per The Register; Tom's Hardware cites ~$7K–$9K street | Up to ~200B parameters on one unit | Existing NVIDIA / partner channel |
| Two 64GB units clustered | 128GB pooled | $9,998 before cable/tax | Up to ~200B parameters; NVIDIA cites up to 1.7× vs one unit on Qwen 3.8 27B | QSFP between ConnectX-7 ports + NVIDIA Sync Cluster Assistant |
Our price math on the disclosed numbers: $6,950 − $4,999 = $1,951. That is about 28% less for the 64GB SKU versus The Register's 128GB figure ($1,951 ÷ $6,950 ≈ 0.281). Two 64GB boxes at list are $9,998, which is $9,998 − $6,950 = $3,048 more than one 128GB unit at $6,950, or roughly 44% more cash for pooled memory plus a second Grace Blackwell node and twice the memory bandwidth. You are not buying a clever coupon. You are buying a second computer.
NVIDIA says memory bandwidth on the 64GB config stays at 273 GB/s, which The Register reads as lower-capacity LPDDR5x modules rather than fewer channels. The 20-core MediaTek Arm CPU and ConnectX-7 networking carry over. Storage on the new partner SKU was not spelled out in the launch blog the way the 4TB SSD is on the fuller system, so check the OEM listing before you assume identical disks.
What "up to 100B parameters" actually buys you
Vendor-stated parameter ceilings are marketing boundaries, not a promise that every 100-billion-parameter checkpoint will feel snappy at a long context. NVIDIA's own pitch for the smaller box leans on models in the mid-tens of billions for day-one agent and inference work, with Qwen 3.8 27B called out in the cluster demo. Out of the box you get Agent Toolkit, CUDA-X libraries, Nemotron open models, and runtimes like Ollama, vLLM, and PyTorch with CUDA. Blender support is listed as coming via a prebuilt installer.
If your workload is private coding agents, document analysis, or fine-tuning that fits in 64GB, the new SKU is the practical entry. If you routinely push larger fine-tunes or long-context 100B-class models on one chassis, the 128GB box (or a cluster) still matches the job better, memory crunch or not.
Geeknewz verdict
Geeknewz's view: buy the 64GB Spark if you want a local agent box this month and your models already fit under that ceiling. Treat $4,999 as "accessible for a personal AI workstation," not as a bargain PC. Skip the two-box cluster math unless you already know you need 128GB pooled and prefer two nodes over hunting a single 128GB unit at $7K-plus street. If you only need occasional inference, cloud GPUs or a smaller desktop still win on total cost. Watch partner listings on October 23 for real SSD sizes and availability before you pre-order on the headline alone.
Source: NVIDIA Blog; additional reporting from The Register and Tom's Hardware.