NVIDIA did not simply unveil a new chip at GTC Taipei 2026. It revealed a roadmap that extends well past the end of the decade, signaling a long-term commitment to the Windows on Arm PC market that no other entrant has been willing to make. The RTX Spark, the company’s first Arm-based SoC designed specifically for Windows, was the headline product, but the slide behind Jensen Huang during the keynote contained the real news: a three-generation roadmap stretching from 2026 to 2030.
Why the Roadmap Matters More Than the Chip
The RTX Spark itself was not a complete surprise. Leaks under the code names N1 and N1X had already circulated widely. The specifications align with those earlier reports: a unified SoC combining a Grace CPU and a Blackwell GPU connected via NVLink-C2C, built on TSMC’s 3nm process, delivering up to 1 petaflop of FP4 AI performance. The chip features a 20-core Grace CPU, 6,144 CUDA cores, and up to 128 GB of unified LPDDR5X memory. On paper, it looks like a direct scaling-down of the DGX Spark data center platform for the notebook form factor.
But the critical question was never whether NVIDIA could build such a chip. It was whether the company would see the project through. The history of Windows on Arm is littered with half-hearted experiments and abandoned platforms. Qualcomm’s Snapdragon X Elite, launched in 2024, initially generated genuine momentum, but software compatibility barriers and OEM hesitation limited its reach. OEMs and software vendors are understandably cautious about committing resources to a platform that might disappear in a few years. Apple succeeded with the M1 in 2020 because it controlled the entire ecosystem, from silicon to operating system. Replicating that model in the fragmented Windows world is an entirely different challenge.
NVIDIA addressed that uncertainty directly by publishing the roadmap. The message to OEMs, developers, and consumers was unmistakable: this is a long-term product line, not a one-off experiment.
The Three Generations of RTX Spark
The roadmap shown during the keynote laid out three distinct generations, each tied to NVIDIA’s broader data center architecture cycles.
Generation 1 (2026): Grace Blackwell RTX Spark. This is the chip announced at GTC Taipei. First laptops and compact desktops from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI are expected to ship in fall 2026. Acer and GIGABYTE will follow. NVIDIA has stated that more than 30 notebook designs and at least 10 desktop designs are in development across these partners.
Generation 2 (2027–2028): Vera Rubin Spark. The CPU architecture moves from Grace to Vera, and the GPU architecture moves from Blackwell to Rubin. Memory transitions from LPDDR5X to LPDDR6, which will bring higher bandwidth and improved power efficiency. Vera is a 88-core Arm CPU that NVIDIA originally developed for data center use. The Spark variant is expected to scale that design down significantly, just as the 72-core Grace data center CPU was scaled down to 20 cores for the first-generation Spark. The exact core count for the Vera Spark variant has not been disclosed.
Generation 3 (2029–2030): Rosa Feynman Spark. The CPU moves to Rosa, and the GPU moves to Feynman. Specific memory specifications were not discussed, but LPDDR6 is the likely continuation. The data center version of Feynman is expected to introduce 3D die-stacking technology, and the key question is whether that capability will extend to the Spark variant. If it does, it would represent a significant leap in performance density for a notebook-class SoC.
The significance of this three-generation commitment cannot be overstated for the PC supply chain. OEMs can now justify building dedicated design teams, optimizing software stacks, and investing in thermal solutions for a platform that will evolve predictably over several years. That level of commitment has been absent from every previous Windows on Arm initiative.
Two Pillars: AI Agents and Gaming
NVIDIA is positioning the RTX Spark around two distinct use cases that leverage its unique architectural advantages.
The first is on-device AI agents. NVIDIA is working directly with Microsoft to build a security framework for running AI agents locally on Windows. This includes adding a new security primitive to the operating system and leveraging NVIDIA’s own runtime environment, called OpenShell, to manage agent permissions and privacy controls. The company claims the RTX Spark can run a 120-billion-parameter large language model locally with a context length of 1 million tokens. For enterprise users who handle sensitive data that cannot be sent to cloud APIs, this capability is transformative.
The second pillar is gaming performance. The integrated GPU, with its 6,144 CUDA cores, is an outlier in the integrated graphics landscape. NVIDIA claims it can deliver 1440p gaming at over 100 fps in AAA titles with ray tracing and DLSS 4.5 enabled. If true, a 14mm-thin notebook could match or exceed the gaming performance of many current-generation discrete GPU laptops. However, the SoC does not support external discrete GPUs, so users who need more graphics horsepower will need to step up to a higher-tier platform.
What ties both pillars together is the CUDA ecosystem. Qualcomm’s Snapdragon X cannot run CUDA applications. Apple’s Metal API is incompatible. Any NVIDIA SoC, by contrast, runs the same CUDA applications that power the data center. For AI developers and researchers, this is the decisive advantage. An application that runs on an NVIDIA DGX system in the cloud will run identically on an RTX Spark notebook. No other Arm-based platform can make that claim.
DGX Station for Windows: The High-End Complement
Alongside the RTX Spark, NVIDIA also announced the DGX Station for Windows, a desktop-class system that targets a different tier of the market. It is built around the GB300 Grace Blackwell Ultra superchip, pairing a 72-core Grace CPU with a Blackwell Ultra GPU connected via NVLink-C2C. The system supports up to 748 GB of coherent memory and delivers 20 petaflops of FP4 AI performance, enough to run trillion-parameter AI models locally.
The positioning is clear. The RTX Spark targets consumers, developers, and professionals who need a capable but portable AI workstation. The DGX Station targets enterprise AI teams who need data center-class performance on a desk. Together, they form a complete stack for local AI development and deployment, from a thin-and-light notebook to a workstation that rivals small cloud instances.
This two-tier strategy also gives NVIDIA a clear upgrade path for RTX Spark users. A developer who starts prototyping on a Spark notebook can scale to a DGX Station for training and then deploy back to Spark for inference, all within the same CUDA software environment.
The Real Test Arrives This Fall
For all the strategic clarity of the roadmap, unanswered questions remain. The first RTX Spark laptops will not ship until fall 2026. Pricing has not been announced, but the bill of materials for a TSMC 3nm die combined with up to 128 GB of LPDDR5X memory suggests that initial models will land at premium price points. It may take a generation or two before more accessible configurations appear.
The competitive response is also unfolding. Qualcomm is preparing next-generation Snapdragon X chips. Intel and AMD are not standing still on x86. And Apple continues to refine its M-series silicon with each generation. NVIDIA is entering a market with entrenched incumbents and a history of platform failures.
But the roadmap changes the calculation. No previous Windows on Arm entrant has committed to three generations of silicon spanning half a decade. That alone gives NVIDIA a credibility that Qualcomm, despite its early lead, has never truly established. The answer to whether NVIDIA can finally make Windows on Arm work will arrive when the first RTX Spark notebooks reach reviewers’ hands later this year. The roadmap, however, has already made the argument that this time, the commitment is real.
