Nvidia is bringing out a cheaper 64GB version of its DGX Spark desktop AI computer, with a starting price of $4,999 and an unusual promise: if one box is not enough, two can be linked together to pool memory and run larger models. The new configuration is due to go on sale through hardware partners on 23 October. Nvidia says a single 64GB system can support models of up to 100 billion parameters, while two connected machines can provide 128GB of pooled memory for models up to 200 billion parameters.
The appeal is straightforward. Instead of sending every AI request to a cloud service, developers can keep models and data on a machine sitting beside them. That can matter for privacy, predictable costs, development speed and work that cannot easily leave a local network. It does not make cloud AI obsolete, but it gives serious local AI development a more practical route than building a conventional multi-GPU workstation from parts.
Two small boxes can act like one larger machine
The 64GB DGX Spark uses Nvidia’s GB10 Grace Blackwell Superchip, unified memory, DGX OS and the company’s AI software stack. The DGX Spark product page lists the 64GB machine as supporting models up to 100 billion parameters. Nvidia says two 64GB units can be joined through their ConnectX-7 networking and configured by Nvidia Sync Cluster Assistant, creating a 128GB pool with support for models up to 200 billion parameters.
Pooling memory is useful because the size of a model is often limited less by raw arithmetic speed than by whether its weights and working data fit into available memory. A developer who can run a model locally avoids sending prompts, documents or code to a remote inference service, although local operation does not automatically make an application private or secure. The software around the model still has to be configured sensibly.
Nvidia is also making a performance claim for the paired setup. In its own Qwen 3.8 27B test, the company says two clustered 64GB systems delivered up to 1.7 times the performance of one. That is a vendor benchmark for a particular workload, not a guarantee that every model or application will scale by the same amount.
Local AI is getting less experimental
For many people, local AI still means running a small model on a laptop. DGX Spark sits in a different category: it is aimed at developers, researchers and enthusiasts who want to run agents, inference, fine-tuning and data work without renting a large cloud instance for every experiment. That direction connects with the wider shift towards agents that continue working after a user leaves, a theme LiveAIWire recently covered with OpenAI’s always-on Dots and with security problems created by phone-based agents.
Nvidia says the machine ships with support for common local AI tools including Ollama, vLLM, PyTorch and its own agent software. It is also promoting a model launcher that is meant to simplify spreading a model across two machines. That matters because cluster software can be a barrier of its own. A hardware product that technically supports a large model is much less useful if configuring it requires specialist networking and distributed-computing knowledge.
The price still puts it far above a normal PC
At $4,999, the 64GB DGX Spark is not a mass-market home computer. Two units would mean roughly $10,000 before taxes, accessories or any other equipment. The value proposition therefore depends on what it replaces. For a developer who would otherwise pay recurring cloud GPU charges, local hardware can have a clear attraction. For somebody who only occasionally asks a chatbot questions, it would make little economic sense.
The 64GB model is also not the top of the range. Nvidia continues to sell a 128GB configuration, and its product page lists larger cluster combinations as well. The new model is better understood as a lower entry point into the same desktop AI platform rather than a new generation of hardware.
A useful test of where personal AI computing is going
The interesting question is whether local AI becomes a normal part of professional computing rather than a niche for enthusiasts. LiveAIWire has covered how model capability and price can change quickly, including frontier models moving closer together on capability and cost. Hardware has to keep up with that movement. If more useful open models can fit inside 64GB or 128GB of memory, a desktop box becomes more attractive. If the frontier keeps demanding much larger systems, cloud services retain the advantage.
Nvidia is betting on the first path strongly enough to build clustering into the product story. The 64GB DGX Spark is therefore more than a cheaper SKU. It is a test of whether developers want AI compute to behave like a local appliance: always available, private by default, and expandable by adding another box when the workload grows.
What 64GB does not mean
The headline memory figure is useful, but it should not be read as a promise that every model advertised at a particular parameter count will run equally well. Local model size depends on the numerical format used to store the weights, the length of the context, the software runtime and how much memory is needed for temporary working data. Quantisation can make a large model fit into less memory, but that can also change speed or quality.
That distinction matters because Nvidia’s published model-size claims describe what the DGX Spark platform can support, not a universal performance guarantee for every workload. A developer running a compact coding model may care about response speed, while somebody experimenting with a very large model may accept slower generation in exchange for keeping data on the desk. Pairing two machines expands the memory pool further, but it also adds networking and software considerations. The useful question is therefore not simply how many parameters fit, but whether a particular model, precision and workload perform well enough for the job.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.