NVIDIA's RTX Spark Brings 120B AI Models to Windows Laptops

For years, running a serious large language model locally meant either a beefy Linux workstation or a Mac with Apple Silicon. That changes this fall. NVIDIA unveiled NVIDIA RTX Spark, a new superchip that reinvents Windows PCs for the era of personal AI agents, offering a new class of computer that moves from tool to teammate. The announcement, made jointly with Microsoft at Computex 2026 and followed up at Microsoft Build, is the most significant shift in PC architecture in years.

What is actually inside the chip

NVIDIA is bringing two processors to market: the N1 and the higher-end N1X. The N1X features a 20-core Arm v9.2 CPU split between 10 high-performance and 10 energy-efficient cores, paired with 6,144 CUDA cores on NVIDIA's Blackwell GPU architecture. At full strength, this chip offers up to 20 Arm CPU cores, a Blackwell GPU with 6,144 CUDA cores, 128GB of LPDDR5X RAM, and up to 300 GB/s of memory bandwidth. That powerful CPU and GPU, connected over NVLink C2C, and the large memory pool give AI agents and 120-billion-parameter models plenty of power and space for long-running tasks with context lengths stretching to a million tokens, according to NVIDIA.

RTX Spark powers the world's first Windows PCs purpose-built for personal agents, featuring 1 petaflop of AI performance, industry-leading power efficiency, full-stack NVIDIA AI and graphics technology, and up to 128GB of unified memory. The unified memory architecture -- where the CPU and GPU share a single high-bandwidth pool -- is the same design principle Apple uses in its M-series chips, and it is what makes running large models on a laptop feasible.

Why CUDA changes everything here

Every other Windows-on-Arm chip -- Qualcomm Snapdragon X, AMD Strix, even Apple's M-series -- runs on proprietary AI stacks. Most laptop AI chips have capable NPUs, but they run on proprietary AI stacks like QNN, ROCm, and Core ML. CUDA tooling does not run natively on any of them. RTX Spark changes this. If your pipeline uses PyTorch with CUDA, llama.cpp with CUDA, Flash Attention, TensorRT, or the broader NVIDIA inference stack, that code runs on RTX Spark without recompilation. The same CUDA binary that runs on an H100 runs on RTX Spark.

That is a genuinely big deal. It means the entire ecosystem of CUDA-accelerated libraries -- the tools most ML practitioners already use -- just works on a laptop, without porting or rewriting. NVIDIA's CUDA ecosystem, used by millions of developers for AI, scientific computing, and creative applications, gives these laptops a built-in software advantage that Qualcomm's chips lacked at launch.

The agent security layer

Raw compute is only half the story. The bigger architectural bet is on local AI agents -- autonomous software that can take actions on your behalf across apps and files. The problem has always been security: how do you let an agent touch your filesystem, calendar, and browser without it becoming a liability? At Microsoft Build, Microsoft announced a set of security primitives to allow agents to execute code, operate on files, and orchestrate tasks across systems with built-in identity and policy execution.

Microsoft announced security primitives to allow agents to execute code, operate on files, and orchestrate tasks across systems with built-in identity and policy execution. The Microsoft eXecution Containers (MXC) form the policy layer, defining and instrumenting isolation and containment while relying on native Windows operating system constructs to apply these policies. Alongside MXC, NVIDIA and Microsoft collaborate to deliver a native Windows experience for personal agents, including new security primitives and NVIDIA OpenShell to run agents securely on primary devices. OpenShell lets users define exactly what an agent can and cannot do, and can strip personal information from any queries that do get routed to the cloud.

The hardware lineup

NVIDIA says the first RTX Spark laptops will arrive in Q4 2026 from partners including Asus, Dell, HP, and Lenovo, alongside a Microsoft Surface device. The breadth of the OEM lineup is notable:

• Microsoft Surface Laptop Ultra -- the flagship laptop, positioned as a portable, premium device for creators and developers
• Surface RTX Spark Dev Box -- a compact desktop aimed squarely at AI engineers who need sustained local compute
• Dell XPS 16 Creator Edition -- Dell's first RTX Spark laptop, announced by Michael Dell personally
• ASUS, HP OmniBook, Lenovo, MSI -- covering ultraportable, workstation, and gaming form factors
• Acer and GIGABYTE -- to follow after the initial wave

Pricing is expected to start around $1,699, a premium over entry-level Copilot+ PCs but competitive with high-end MacBook Pros. Configurations will range from ultraportable 13-inch models with 16GB of memory to beefy 16-inch workstations with 64GB and optional discrete RTX 5000-series GPU for double the AI throughput.

The developer machine

The most interesting product for practitioners may not be the laptop. The way developers build software is fundamentally changing. AI models are growing in capability and complexity, agentic workflows demand sustained compute, and every iteration can incur cloud costs even when the work does not require state-of-the-art models. Surface RTX Spark Dev Box changes that equation. It is a purpose-built Windows AI developer box that puts up to 1 petaflop of AI compute directly on the desk. By bringing powerful AI compute to the edge, developers can reserve frontier model calls for truly frontier problems and handle the rest on their own hardware.

Surface RTX Spark Dev Box ships with Windows 11 Pro pre-configured for developers at the image level, bringing a purposeful set of defaults, preinstalled tools and tuned settings so the development environment is the default from first sign-in. The setup keeps developers in their workflow: dark theme, taskbar simplified for development, Widgets removed, Do Not Disturb on. Developer Mode is enabled. PowerShell 7 is the default shell. Under the hood, WSL 2 is configured with GPU passthrough and CUDA support. VS Code, GitHub Copilot, Git, Python and Node.js are installed.

Who wins and who loses

The market reaction was immediate and unambiguous. The news sent shockwaves through the semiconductor sector. Intel shares fell 5.1% to $114.68, while Dell surged 32.9% to $421.30, hitting a new 52-week high. That spread tells the story: NVIDIA entering the PC chip market is a direct threat to Intel's core business, while OEMs like Dell are seen as clear beneficiaries of a premium hardware cycle.

Qualcomm is also in a difficult position. Its Snapdragon X chips were supposed to be the Windows-on-Arm answer, but the first wave of Copilot+ PCs leaned heavily on NPUs and battery-efficient inference. That made sense for features like background blur, image generation, recall-style indexing, and small-model assistance. But it did not fully satisfy developers, creators, or enterprises who wanted to run heavier models, test agents locally, or keep sensitive workflows off external cloud infrastructure.

For Apple, the competitive picture is more nuanced. RTX Spark's unified memory architecture mirrors Apple Silicon's design, and the pitch -- powerful local AI on a slim laptop -- is essentially Apple's value proposition brought to Windows. The move attacks several settled assumptions at once: that x86 compatibility is the safest default, that AI PCs are defined by NPUs, that Windows on Arm is a secondary tier, and that local agents are still a future problem.

The software ecosystem is already moving

Hardware without software is a press release. The ecosystem response here is unusually fast. Adobe is rearchitecting Photoshop and Premiere from the ground up for RTX Spark to deliver 2x faster AI and graphics performance. Georgi Gerganov, founder of llama.cpp, confirmed optimized support. The creator of ComfyUI noted that RTX Spark's large unified memory pool will enable complex multimodal workflows that were previously impractical on a laptop. Over 100 Windows software providers are listed as embracing the platform at launch.

NVIDIA also confirmed that Vera Rubin NVL72 systems are now in production and available through AWS, Google Cloud, Azure, and other providers in H2 2026, with inference throughput 5x that of Blackwell at 10x lower cost per token. The strategic logic is explicit: a bi-modal compute environment is emerging by late 2026, with cloud inference getting dramatically cheaper as Vera Rubin systems come online and local inference becoming genuinely capable for the first time on laptop-class hardware. Both trends favor builders.

The real risks

The biggest open question is Windows-on-Arm software compatibility. The key test will be software compatibility. Windows on Arm has historically struggled with x86 application emulation, though Microsoft has steadily improved its Prism translation layer. NVIDIA's brand carries high expectations -- a mediocre Arm laptop is disappointing; a mediocre NVIDIA RTX-branded Arm laptop would be a narrative problem.

There are also hardware tradeoffs worth understanding. Because they share the same pool of LPDDR5X memory, the GB10 GPU enjoys just 273 GB/s of raw bandwidth, far less than that offered by more traditional laptops with dedicated GPUs that have their own pools of GDDR memory. That makes RTX Spark excellent for large-model inference but not necessarily the best gaming or GPU-compute platform compared to a discrete GPU setup. The pitch is local AI first, everything else second.

For practitioners building agentic pipelines today, the practical advice is clear: do not wait for RTX Spark to ship your pipeline. Fall 2026 is four to six months away. Build for cloud now, and build the abstraction layer that will let you swap in local inference when the hardware lands. The official NVIDIA announcement and the NVIDIA developer blog have the full technical details on MXC, OpenShell, and the local agent stack. The Microsoft Devices Blog covers the Surface RTX Spark Dev Box specs in depth.