In one line: NVIDIA announced the DGX Spark 64GB on Oct 2, shipping Oct 23 from Acer, ASUS, Dell, Gigabyte, HP and MSI at $4,999 — while the 128GB Founders Edition moves up to $6,950. Local AI is getting cheaper and bigger, and it is directly competing with cloud GPU workloads.
What this machine is
The DGX Spark 64GB keeps the same GB10 Grace Blackwell Superchip as the 128GB model (20-core Arm CPU + Blackwell GPU with 5th-gen Tensor Cores, up to 1 PFLOP FP4), ships with DGX OS and the full NVIDIA AI software stack — Agent Toolkit, CUDA-X libraries, Nemotron open models, plus runtimes like Ollama, vLLM and CUDA-enabled PyTorch out of the box.
Memory is the story here: 64GB of LPDDR5x unified memory, roughly 56GB usable for model weights and KV cache after system reservation, with NVFP4 quantization support. Per NVIDIA, a single unit can run up to 100-billion-parameter models and agentic applications entirely on device, with no cloud dependency.
Adapted open models already include Gemma4 26B, Qwen3.8-27B, Meta Muse Glimmer, Nemotron3.5 Lightning, and Perplexity's portable computer agent Laguna S2.1 118B.
Two-unit clustering: pooling memory as a feature
Every DGX Spark ships with a built-in ConnectX-7 NIC. Two 64GB units connect directly over a QSFP cable (200 GbE); NVIDIA Sync Cluster Assistant auto-discovers the units, validates configuration and sets up the network — pooling memory to 128GB and extending model support to 200-billion parameters, with doubled memory bandwidth.
In NVIDIA's Qwen 3.8 27B test, two clustered 64GB systems delivered up to 1.7x the performance of a single system, with room to keep scaling (up to 4 units). Because every node runs the same software stack, workloads scale from one unit to two without reconfiguration. The NVIDIA Sync Model Launcher lands later this month to make local model deployment one-click.
Heads-up: two clustered 64GB units pool 128GB, but that is not the same as one physical 128GB machine — whether a model runs well across nodes still depends on software and workload partitioning.
Local AI vs cloud GPU: the real trade-off
| Dimension | Local (DGX Spark 64GB) | Cloud GPU |
|---|---|---|
| Cost model | One-time hardware cost (from $4,999); cheaper at sustained high utilization | Pay-as-you-go or reserved; cheap for spikes and short jobs |
| Privacy & compliance | Data stays on-prem; good for sensitive or offline workloads | Data travels to the cloud; evaluate residency and compliance |
| Elasticity | Scale by buying hardware; 2-4 units has a ceiling | Minute-level scaling, broad regions and GPU SKUs |
| Operations | You own hardware, network, environment and upgrades | Mostly managed; no hardware depreciation or failures |
| Best fit | Always-on agents, privacy-sensitive inference, model experiments, offline dev | Training & fine-tuning, bursty traffic, multi-region low latency, flexible SKUs |
Practical guidance: long-running fixed workloads + sensitive data → local wins; training/fine-tuning + variable traffic + multi-region needs → cloud GPU stays the workhorse. Many teams end up with local for development and always-on agents, cloud for training and elastic capacity.
Getting started with cloud GPU
If you choose cloud, both major Chinese providers offer new-user deals. Direct entries below (with referral tracking):
Cloud GPU on demand — new-user deals live
Tencent Cloud GPU servers · Alibaba Cloud ECS/GPU — start hourly or reserved
Links land on official campaign pages; pricing follows the vendors' sites.
Sources
- NVIDIA Blog: DGX Spark 64GB specs, clustering and the 1.7× measured result
- PCOnline (CN): $4,999 price, Oct 23 ship date, hardware and software ecosystem details
- Wall Street CN (华尔街见闻): 128GB price move to $6,950 and memory-supply context
- CSDN: pooled-memory clustering and model-size scaling analysis
- AI Chat Daily: English coverage of 100B/200B parameter support and availability