NVIDIA pivots to local AI: DGX Spark clusters to 8 units with 1TB unified memory
ryanshrout · x · 2026-10-21
At GTC Berlin, NVIDIA's focus shifted from new silicon to local AI: making its three-tier stack—RTX Spark (Windows PCs), DGX Spark (Linux devs), DGX Station (teams)—clusterable, IT-manageable, and easy to stand up, turning local AI from hobbyist boxes into real infrastructure.
Key updates:
- DGX Spark 64GB ships Friday, Oct 23, running Nemotron 3.5 Bolt 74B, Qwen3.8 27B, and Gemma 4 26B-A4B at a lower entry price
- DGX Spark clusters up to 8 units via built-in ConnectX-7 and NVIDIA Sync, reaching 1TB unified memory to run frontier open models like GLM 5.3 as a shared team backend
- NVIDIA Sync Model Launcher arrives this month, pulling the right model and container for a single Spark or cluster
- UnderdogAI local agent lands on RTX Spark; DGX Station gets Windows systems in December
More from Infra
- Shift-1 coding router cuts agent costs up to 55%, custom router training opens on AC2 — ypatil125 · 2026-10-22
- Lam Research pre-earnings checklist: how much of the memory capex boom reaches tool deliveries by 2027 — tengyanAI · 2026-10-22
- PyTorchCon 2026 opens with vision of training frontier models on any chip, in any cloud — PyTorch · 2026-10-22
- Cloudflare ships one-click mitigation for php2shell vulnerability — cyb3rops · 2026-10-22
- A beginner-written CUDA guide that finally builds the full mental picture — goyal__pramod · 2026-10-22
- Shadeform adds CPU compute today, Apple Silicon instances coming this quarter — socialwithaayan · 2026-10-22