The home-local-AI push gained two important pieces in September 2026. NVIDIA presented its consumer direction around DGX Spark and planned RTX Spark systems on September 20, while Perplexity made Portable Computer available on Windows in mid-September for qualifying users with compatible NVIDIA hardware. Together, they point to a broader ecosystem: hardware runs models close to the user, software manages them, and cloud services remain available for tasks that need them.
That is a more practical proposition than simply putting a smaller data center on a desk. NVIDIA is combining large memory pools, CUDA software, local runtimes and agent tools. Perplexity adds an agent that starts locally and can request permission to send an individual task to the cloud. The result is a new set of choices for developers and enthusiasts—not a universal replacement for cloud AI.
The local-AI shift is moving beyond specialist setups
Local AI means running an AI model on hardware you control rather than sending every prompt and file to a remote service. NVIDIA’s local-AI stack supports tools including Ollama, llama.cpp, TensorRT, SGLang, vLLM, WindowsML and PyTorch with CUDA.
That software layer matters because the hardware is only half the problem. A large model must fit in memory, the runtime must support it, and the application must be able to use the available accelerator. A local system can keep prompts and files within a home network, while also giving the user more control over the model and its runtime.
NVIDIA’s vision is now reaching beyond a Linux workstation assembled by an AI hobbyist. DGX Spark targets compact local development and inference. RTX Spark brings the same broad idea to Windows 11 laptops and desktops. Perplexity Portable Computer adds an agent interface for qualifying Windows systems.
What NVIDIA DGX Spark actually brings to a home setup
NVIDIA DGX Spark uses the GB10 Grace Blackwell Superchip and 128 GB of coherent unified LPDDR5x memory. Its listed configuration includes a 20-core Arm CPU, 273 GB/s of memory bandwidth, 4 TB of NVMe storage, Wi-Fi 7, 10GbE and a 200Gbps ConnectX-7 network interface. The system measures 150 × 150 × 50.5 mm and weighs 1.2 kg.
NVIDIA claims up to 1 PFLOP of theoretical FP4 AI performance using sparsity. It also positions DGX Spark for inference models of up to 200 billion parameters and fine-tuning models of up to 70 billion parameters. Those figures describe NVIDIA’s supported or targeted model sizes; they do not describe a guaranteed token rate, context window or response time.
The important distinction is between capacity and speed. Unified memory can let a model fit when a conventional graphics card’s dedicated VRAM would not be enough. Memory bandwidth and the workload then influence how quickly that model runs. A model that fits is not automatically a fast model.
A practical coding workflow demonstrated on DGX Spark used local model management, coding tools and remote access from another computer. That makes the system easier to imagine as a private AI server: the demanding work stays on the Spark while another device provides the keyboard, display and browser.
The system uses NVIDIA DGX OS, so DGX Spark is not a conventional Windows desktop. That distinction matters if your preferred tools, games or creative applications depend on Windows compatibility.
DGX Spark and RTX Spark: two routes to local AI
NVIDIA RTX Spark extends the local-AI concept to Windows 11 laptops and desktops. The laptop platform combines a Blackwell RTX GPU with a Grace CPU and offers configurations with up to 128 GB or up to 64 GB of LPDDR5X unified memory. NVIDIA lists laptop TDP configurations from 45–80W and a 140W desktop configuration.
The higher-memory laptop configuration can suit users who want local model capacity in a Windows machine. A conventional GeForce RTX desktop or laptop remains a different proposition: dedicated GPU VRAM can deliver strong throughput for models that fit, while the available capacity depends on the exact graphics card.
| System | Unified memory | Operating system | Documented role | Practical consideration |
| NVIDIA DGX Spark | 128 GB coherent unified LPDDR5x | NVIDIA DGX OS | Compact local AI development, inference and fine-tuning | High memory capacity in a small system; model speed still depends on bandwidth and workload |
| NVIDIA RTX Spark | Up to 128 GB on the higher listed laptop configuration | Windows 11 | Local AI, agents, gaming and conventional PC work | Windows compatibility and portability; laptop configurations differ in memory and power |
NVIDIA has described RTX Spark systems as a platform for AI development, agents, gaming and ordinary PC work. Their distribution, final prices and country availability are separate questions from the specifications themselves, so the platform description should not be treated as a current retail offer.
The software makes local AI usable
Perplexity Portable Computer for Windows requires a compatible NVIDIA GeForce RTX or RTX PRO GPU with at least 24 GB of VRAM. It uses the PPLX 27B local model in the documented Windows configuration and runs locally by default.
The important word is “default.” When a task needs advanced reasoning or web research, Portable Computer can ask the user for permission to escalate that individual step to cloud models. The user remains in the loop, and the escalation consumes Computer credits. This local-first design preserves a route to cloud capability without sending every task away automatically.
NVIDIA’s own software aims at a similar practical problem from another angle. NVIDIA PAIR, or Personal AI Router, connects compatible Windows, Linux, macOS and DGX Spark systems through a local endpoint. It routes requests between machines, but it does not pool their VRAM or create one virtual GPU.
In plain English: PAIR can send separate jobs to the computers that are available, but it cannot turn three machines into one giant graphics card. That makes it useful for a home network with several systems, not a shortcut around the memory limit of one model.
PAIR’s documented minimum is 8 GB of system RAM, and its supported software environment includes Windows 11, DGX OS, Ubuntu, macOS Tahoe and compatible NVIDIA systems. Its role is coordination. The model still runs on the machine that receives the request.
What local AI gains—and what it gives up
The strongest case for local AI is control. Models and files can remain inside a home network, and the owner can choose the runtime, model and software stack. That matters for private documents, development work and recurring workloads where sending every request to a cloud service is undesirable.
Capacity is another clear advantage for systems such as DGX Spark. Its 128 GB unified memory gives it a different model-size profile from many consumer GPUs with smaller dedicated VRAM pools. That does not make every workload faster. Dedicated RTX hardware can provide higher throughput when a smaller model fits comfortably in its VRAM, while a large unified-memory system can prioritize fitting a larger model at all.
The costs are equally concrete. DGX Spark is expensive hardware, and NVIDIA’s US marketplace snapshot marks the DGX Spark listing out of stock. The same marketplace lists ASUS Ascent GX10 systems at $5,999 for 1 TB and $7,999 for 4 TB. Those are prices for the named ASUS configurations, not a universal price for every DGX Spark system.
Software compatibility also matters. A local setup may involve selecting a model, configuring a runtime and matching the application to the available hardware. NVIDIA’s push toward easier runtimes and graphical tools is designed to reduce that friction, but the underlying constraints remain: memory, bandwidth, power, thermals and application support.
Cloud AI retains a practical advantage when a user needs current web research, a specialized hosted model or a workload that exceeds the local system. Perplexity’s permission-based escalation acknowledges that boundary directly. Local AI is best understood as an additional execution path—one that can handle suitable work privately and locally—rather than as a promise that every cloud capability now fits under a desk.
What this means for people who want AI at home
For a developer, DGX Spark is the clearest high-memory option in NVIDIA’s current local-AI lineup. It can host larger models than many conventional consumer GPU configurations and serve as a remote coding or inference machine. The trade-off is the price, DGX OS and the difference between fitting a model and running it quickly.
For someone who wants a familiar Windows computer, RTX Spark is the more direct concept. Its documented configurations target local AI alongside gaming, agents and regular PC work, with memory options that vary by model. A conventional GeForce RTX system remains relevant when the preferred models fit in dedicated VRAM and throughput matters more than maximum capacity.
Perplexity Portable Computer makes the decision more concrete for Windows users: compatible NVIDIA hardware with at least 24 GB of VRAM can run the local workflow by default, while selected tasks can move to the cloud after permission is granted. PAIR addresses a different need by coordinating several machines without pretending that their memory becomes one pool.
NVIDIA’s September direction therefore looks less like one home supercomputer and more like a stack: hardware for capacity, runtimes for execution, routers for coordination and agents for everyday use. The home becomes another place to run AI—but the right balance between privacy, model size, speed, Windows compatibility and cloud access still depends on the workload.
