NVIDIA Launches RTX Spark Windows AI PCs With up to 128GB Unified Memory

RTX Spark laptops went on sale October 16

NVIDIA announced on October 16, 2026 that RTX Spark laptops were available globally through ASUS, Dell, HP, Lenovo, Microsoft and MSI websites and major retailers. The named launch systems are ASUS ProArt P14 and P16, Dell XPS 16, HP OmniBook Ultra and OmniBook X 14, Lenovo Yoga Pro 9n, Microsoft Surface Laptop Ultra, and MSI Prestige N16 Flip AI+ and Prestige N16. NVIDIA’s October 17 follow-up covered retail launch events rather than a separate generation of hardware.

This is a shipping AI-computing product announcement, not merely a future concept. Availability still has a regional and configuration boundary: a family may be listed while one memory capacity, chassis or delivery date is absent in a particular market. Verify the manufacturer part number, memory, operating system, retailer stock and return terms rather than treating the RTX Spark badge as one fixed machine.

The top specification is 128GB unified memory and one FP4 petaflop

RTX Spark joins an NVIDIA Blackwell RTX GPU and NVIDIA Grace CPU in one superchip design. NVIDIA’s product page lists laptop ceilings of a 6,144-core GPU and 20-core CPU, with up to 128GB of unified memory and up to one petaflop of FP4 AI performance. Its visible specification table also presents a lower desktop profile with a 5,120-core GPU and 18-core CPU, reinforcing that the family name does not guarantee the maximum configuration.

Unified memory is the strategically useful part for local AI because the CPU and GPU can address a larger shared pool than a conventional laptop discrete GPU usually exposes as VRAM. It is not all available to model weights: Windows, applications, runtime state, activations, key-value cache, tool processes and display workloads consume capacity too. FP4 peak performance is also a precision-specific theoretical headline, not tokens per second or completed-agent-task latency.

The 120B-model claim is a fit claim, not a speed result

NVIDIA’s product page says RTX Spark can run models up to 120B parameters with a 1M-token context. The page does not tie that statement to a named model, weight format, quantization scheme, runtime version, batch size, usable context, output length, tokens per second or quality measurement. A model fitting into memory does not establish acceptable prefill time, generation speed, thermal behavior or task quality.

The launch release names Nemotron 3.5 Bolt 74B-A9B 24GB for supported RTX Spark configurations and describes it as optimized for tool calling and multistep tasks. That is a more concrete deployment candidate, but its 74B-total, 9B-active label and 24GB artifact size must not be conflated with the generic 120B ceiling. Test the exact checkpoint and runtime you intend to operate.

Windows on Arm gains a more complete CUDA development path

NVIDIA positions RTX Spark as a Windows 11 system with native CUDA, TensorRT, NVFP4 and Windows Subsystem for Linux. The release says NVIDIA and Microsoft prepared a PyTorch-CUDA build for Windows on Arm, a native-default Python build, a native Conda installer and support for 97% of the top 15,000 Python packages. These are vendor-reported ecosystem statements, not a guarantee for a particular wheel, compiler extension, driver or enterprise image.

Before replacing an x86 or Linux workstation, create a dependency manifest for the actual project. Install from a clean image, build native extensions, run the test suite, exercise WSL and Windows paths, validate container and driver behavior, and check every hardware-attached tool. “97%” gives no information about whether the missing 3% contains the one package your stack requires.

The agent layer includes local installers, permissions and network serving

NVIDIA says guided installers support agents including GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy. It also describes NVIDIA OpenShell and Microsoft Execution Containers as permission-control layers, NVIDIA Skills as tool integrations, and NVIDIA PAIR as a way to serve inference to compatible Windows, Linux and macOS computers on a home network. The Build site exposes agent setup, local-LLM and development-workspace playbooks.

Those components form a software stack, not a single security or privacy guarantee. An agent can still call cloud models, external search, package registries or third-party services; locally stored prompts can still enter logs or backups; and tool permissions can expose files or accounts. Record which inference and tool calls remain local, what leaves the device, which identity authorizes each action, and how the runtime contains untrusted code.

NVIDIA publishes two different starting prices

The October 16 newsroom release says RTX Spark laptops start at $2,599 and explicitly notes that pricing and configurations vary by region. The October 17 NVIDIA Blog says they start at $2,399. Both pages were live when reviewed on October 19, and the $200 disagreement cannot be reconciled from the text alone. It may reflect a configuration, promotion, market or correction not identified on the pages.

Use neither number as a universal acquisition cost. Capture a dated quote for the exact manufacturer part number, memory, storage, display, warranty and region. Add tax, docks, storage, support, power, replacement capacity and staff time. “No per-prompt fee” for local generation does not mean no cost per accepted output.

Local execution changes the boundary but does not prove privacy

A local model can keep inference data on the device and avoid a metered model API when the entire configured workflow is genuinely local. That can benefit proprietary code, offline work and fast iteration. It can also move responsibility for encryption, endpoint security, backups, patches, physical loss, model provenance, tool credentials, logs and incident response onto the owner.

Map every boundary in a representative trace. Include model downloads, telemetry, authentication, search, tool APIs, synchronization, crash reports, software updates and remote inference fallback. Make the product claim precise: “this tested workflow kept these data classes on this device under this configuration” is evidence; “RTX Spark is private” is not.

What AI builders and buyers should do next

Shortlist an exact 64GB or 128GB configuration only after checking the model artifact, runtime and context memory budget. Run a bounded evaluation with cold load, realistic prompts, long-context prefill, sustained generation, concurrent tools and unplugged use. Measure accepted task outcomes, time to first token, completion latency, peak memory, thermals, fan noise, battery draw, crashes and recovery—not only whether the model starts.

Buy when the tested configuration clears workload, compatibility, privacy and support gates at a defensible total cost. Constrain it to local prototypes when sustained performance or software support is marginal. Wait when a required SKU or dependency is unavailable. Reject it for the workload when the model, context or tool stack cannot meet the contract. A new high-memory laptop creates a useful local option; it does not remove the deployment decision.

Disclosures

AccessAllGPT received no RTX Spark hardware, software access, model access, credits, benchmark data, private briefing, review, payment or compensation for this article. NVIDIA, Microsoft, the named PC makers, software vendors and agent developers did not sponsor, review or endorse it. AccessAllGPT did not test or rank the platform. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with NVIDIA, Microsoft or the named vendors. Publication-wide relationships are listed on the disclosures page.