NVIDIA’s RTX Spark Chip Gives Windows Its Third Shot at Arm

NVIDIA unveiled its RTX Spark superchip on May 31 at Computex 2026, pairing a 20-core Arm CPU with a Blackwell GPU and up to 128GB of shared memory built into thin Windows laptops. It’s NVIDIA’s first system-on-a-chip designed specifically for Windows PCs, built jointly with Microsoft, and it’s also the third attempt in fourteen years to make Windows run properly on Arm instead of Intel or AMD’s x86 chips.

Early pricing chatter puts these machines well above $2,000, with some estimates near $2,900 for the flagship configuration. That positions RTX Spark as gear for AI developers, video editors and power users first, and everyone else a distant second.

A Grace CPU and a Blackwell GPU Share One Pool of Memory

The chip pairs a 20-core NVIDIA Grace CPU, a mix of Cortex-X925 and Cortex-A725 cores co-designed with MediaTek, a market leader in Arm-based chip design, with a Blackwell GPU carrying 6,144 CUDA cores and fifth-generation Tensor Cores built for FP4 precision. NVIDIA rates the combination at up to 128GB of unified memory and one petaflop of AI compute.

NVIDIA says the platform can run 120-billion-parameter language models locally with context windows stretching to a million tokens, edit 12K 4:2:2 video, render 3D scenes larger than 90GB and still play AAA games above 100 frames per second at 1440p. Adobe is rebuilding Photoshop and Premiere around the chip, promising roughly twice the AI and graphics performance of comparable machines.

Several apps are getting native Arm versions built specifically for the platform, including:

• Adobe Photoshop and Premiere, rearchitected around RTX Spark’s unified memory and TensorRT software
• Blender and Maxon’s Cinema 4D and Redshift, for 3D rendering workflows
• DaVinci Resolve, Topaz Photo and CapCut for video and photo editing
• MATLAB, running through Microsoft’s Prism translation layer rather than natively

The laptops themselves are engineered to be as slim as 14 millimeters and as light as 3 pounds, with tandem OLED displays and machined aluminum chassis. Jensen Huang, NVIDIA’s founder and chief executive, framed it as a break from four decades of app-launching. “The PC is being reinvented,” he said at the Computex keynote.

Windows Has Made This Bet Twice Before, and Lost

This isn’t NVIDIA’s first swing at Arm-powered Windows, and it isn’t Microsoft’s either. In 2012, Microsoft launched Windows RT on NVIDIA’s Tegra chips, an experiment that collapsed because it couldn’t run the x86 desktop software that defined the platform. NVIDIA supplied that Tegra silicon for the original Surface RT before quietly exiting the Windows chip business for a decade.

Reuters reported in October 2023 that NVIDIA was already back at work building Arm-based CPUs for Windows, timed to when Microsoft’s chip exclusivity arrangement with Qualcomm was set to expire in 2024. By January 2025, Tom’s Hardware had the codenames: N1 and N1X, developed with MediaTek. That same year, Microsoft’s Copilot+ PC push, built around Qualcomm’s Snapdragon chips, arrived to a muted response. It nudged laptop makers toward standardizing on more RAM, but its headline AI features, including the privacy-troubled Recall tool, never found the audience Microsoft hoped for.

NVIDIA’s own Arm hardware had a trial run first. The company’s DGX Spark workstation, built on the related GB10 Grace Blackwell chip, started shipping in October 2025 as a Linux-based developer box rather than a Windows machine. RTX Spark effectively takes that same silicon lineage and, for the first time, points it at ordinary Windows laptops and desktops.

What Does a Shared Memory Pool Actually Change?

A shared memory pool means the CPU and GPU draw from one pool of RAM instead of splitting it between system memory and separate video memory. That single change is why a thin laptop can hold a 120-billion-parameter model entirely in memory, something no conventional 16GB or 32GB gaming notebook can manage without constantly shuttling data across a much slower PCIe connection.

Microsoft rebuilt parts of Windows to take advantage of it. The company added workload profile scheduling to spread tasks across all 20 cores more efficiently, adopted its Power and Thermal Framework for battery life under heavy load, and rewrote how much system memory the GPU is allowed to see. That last change unlocked a smarter limit on GPU memory, letting the graphics chip pull from more of the system’s total pool than Windows previously allowed.

The comparison to Apple’s approach is direct. NVIDIA claims roughly 1,000 TOPS (trillion operations per second) of AI throughput against the 38 TOPS in Apple’s Neural Engine, though Apple’s chips still lead on memory bandwidth and single-core CPU speed. Here’s how the main unified-memory platforms stack up on paper:

One tradeoff NVIDIA hasn’t advertised as loudly: RTX Spark systems won’t support external dedicated GPUs. Buyers get the integrated Blackwell chip or nothing.

Intel, AMD and Qualcomm Now Share a Common Problem

NVIDIA’s entry threatens the entire premium Windows laptop market at once. Intel and AMD have owned x86 laptop silicon for four decades, and Qualcomm has been the default Arm vendor for Copilot+ PCs since 2024. RTX Spark challenges all three simultaneously, and the market noticed. Shares of Intel fell more than 4% in the days following the keynote, and AMD slipped as well, while Arm Holdings, whose licensing business sits apart from NVIDIA’s custom-designed cores, gained ground.

The stakes go beyond one product line. Analysts covering the launch have framed it as a fight over a PC silicon market worth more than $200 billion through 2027, arriving just as NVIDIA posts record data-center revenue and expands into networking and software on a separate front entirely.

How Much Will an RTX Spark PC Actually Cost?

Nobody official will say. NVIDIA and Microsoft representatives at Microsoft’s Build 2026 conference both declined to confirm pricing when asked directly. “We’re not ready to talk about pricing yet until we get closer to availability,” said Andrew Hill, corporate vice president of Surface, about the Surface Laptop Ultra and the Surface RTX Spark Dev Box, according to PCWorld. Both companies reportedly want to see how memory and storage prices settle before committing to a number.

Away from the podium, estimates vary depending on who’s asked:

• Computex vendor sources, per PCWorld: N1X-based laptops starting around $2,500; N1 models near $2,000
• Morgan Stanley analysts, cited by PC Guide: N1X models needing to price at $2,899; N1 models at $1,799
• NVIDIA and Microsoft, officially: no figure at all until closer to the fall launch

The RTX Spark’s closest sibling offers a preview of how that pricing conversation tends to go. DGX Spark, which Engadget reported launched at $3,999 last October, now sells for $4,699 following a February price increase, according to a review by AI analysis firm IntuitionLabs. Reaction on NVIDIA’s own developer forums wasn’t kind.

Cancelling my pre-order…seriously disappointing for the price tag.

That was one buyer’s reaction after seeing benchmarks, and other forum posters compared the value unfavorably to used GPU rigs and AMD workstation systems. It’s a warning sign for a consumer product aiming at a similar price band with a similar pitch.

NVIDIA’s high-end machines aren’t the only Windows laptops launching this year. At the opposite end of the market, Acer’s Swift Air 14 undercuts the MacBook Air at $699 and skips Copilot+ certification entirely, a reminder that RTX Spark isn’t trying to serve the same buyer at all.

Creators and AI Developers Are the Actual Customers

NVIDIA is chasing a specific crowd: developers and power users adopting AI agents like OpenClaw, which NVIDIA says has hit record activity on developer platforms like GitHub and OpenRouter alongside rivals such as Hermes Agent. “At Nous, we expect tasks to increasingly run on device as personal agents like our Hermes Agent become more capable and ubiquitous,” said Dillon Rolnick, chief executive of Nous Research, a developer of open-source AI agent software.

That local-agent push fits Microsoft’s broader hedging strategy. The company has spent the past year building its own MAI models as a five-year bet on AI independence from OpenAI, and it’s simultaneously locked into a $965 billion cloud partnership with Anthropic through Azure. Betting on NVIDIA’s silicon for on-device agents is one more layer of the same hedge: whichever AI approach wins, Microsoft wants a stake in it.

Beyond AI specifically, video editors and 3D artists stand to gain from the same unified-memory pool, since it removes the out-of-memory errors that plague conventional laptops working on massive project files. NVIDIA’s own product page promises creators can fine-tune and inference the latest models locally, without shipping data to the cloud first.

A Cheaper Spark Is Already Rumored

NVIDIA has already teased a roadmap covering the next three generations of the Spark family, and industry chatter points to stripped-down, cheaper versions of the chip aimed at thinner, more affordable systems down the line. A leaked N1 configuration, reported by outlets tracking Computex disclosures, trims the design to roughly 2,048 CUDA cores, 64GB of memory and an 18 to 45-watt power envelope, though NVIDIA hasn’t confirmed those figures.

Whether that cheaper Spark arrives on schedule may depend on forces outside NVIDIA’s control. AI data-center demand has already driven DRAM prices sharply higher and forced PC makers to ration memory across their lineups, the same shortage squeezing the exact component RTX Spark depends on most. For now, the only confirmed date is fall 2026, when the first laptops from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI are due on shelves.

Frequently Asked Questions

What Is NVIDIA RTX Spark?

RTX Spark is NVIDIA’s first system-on-a-chip built for Windows PCs, combining a 20-core Arm-based Grace CPU with a Blackwell GPU and up to 128GB of shared memory. NVIDIA and Microsoft announced it jointly on May 31, 2026, at Computex in Taipei, and laptops and compact desktops are due from major manufacturers starting in fall 2026.

When Can I Buy an RTX Spark Laptop?

Fall 2026 is the confirmed window, starting with ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, with Acer and GIGABYTE following later. Microsoft has already named its own entries: the Surface Laptop Ultra and the Surface RTX Spark Dev Box, a compact developer system built with 128GB of unified memory inside a 100-watt thermal envelope.

How Is RTX Spark Different From DGX Spark?

DGX Spark is NVIDIA’s earlier, Linux-based workstation aimed squarely at AI researchers, running a custom Ubuntu-based operating system and handling inference on models up to 200 billion parameters. RTX Spark uses closely related silicon but runs Windows instead, built for mainstream laptops and desktops rather than a dedicated development box.

Will RTX Spark Run My Existing Windows Software?

Most of it, through Microsoft’s Prism emulation layer for x86 applications, though some older or niche programs may run poorly or not at all. For gaming specifically, Microsoft says RTX Spark will support native anti-cheat systems from Epic’s Easy Anti-Cheat and BattlEye, plus the Xbox PC app.

Why Won’t NVIDIA or Microsoft Confirm Pricing Yet?

Both companies have said they’re waiting to see how memory and storage costs settle before locking in a price, according to PCWorld’s reporting from Microsoft’s Build 2026 conference. Outside estimates from Computex vendor sources and Morgan Stanley analysts currently range from about $1,800 to $2,900 depending on configuration.

What Is NVIDIA OpenShell?

OpenShell is NVIDIA’s new security runtime, built with Microsoft’s help, meant to let AI agents run locally on a Windows PC without exposing the rest of the system. NVIDIA says agent developers including Hermes Agent and OpenClaw are integrating it directly into their Windows applications.