TAIPEI – For most of the past decade, if you wanted the kind of machine that could run a 70-billion-parameter language model locally, hold a 12K video timeline without stuttering, or power an AI agent that actually did things while you slept, you had two choices: spend the better part of $4,000 on a Linux workstation aimed at enterprise developers, or rent the compute from a cloud you did not control. Nvidia changed the terms of that equation at Computex 2026 in Taipei.
The company unveiled the RTX Spark, a superchip it built in close partnership with Microsoft over several years, and announced that laptops and compact desktop PCs carrying the chip would reach consumers in the fall of 2026. The hardware brings together a 20-core Arm-based CPU – derived from Nvidia’s Grace architecture and co-developed with MediaTek – and a Blackwell-generation RTX GPU with 6,144 CUDA cores, all connected through Nvidia’s NVLink chip-to-chip interconnect. The memory pool, shared between the CPU and GPU, runs up to 128 gigabytes of LPDDR5X with bandwidth approaching 300 gigabytes per second. Nvidia rates the full configuration at one petaflop of AI compute in the FP4 numerical format, according to the company’s official announcement.
Those are substantial numbers. But the specifications, impressive as they are, are not the most consequential thing Nvidia announced in Taipei.
What matters more is the posture. Nvidia has spent three decades selling discrete graphics cards into Windows PCs built around Intel and AMD processors. The RTX Spark is the company’s formal declaration that it intends to own the entire silicon stack – CPU, GPU, memory interconnect, and now the operating system scheduling layer – on the consumer Windows platform. That is a structurally different ambition from anything Nvidia has attempted in the PC market before, and it puts the company in direct competition with Qualcomm, which has spent the past two years building out the Snapdragon X ecosystem for Windows on Arm.
Jensen Huang, Nvidia’s chief executive, framed the unveiling in terms that left little room for understatement. He described the RTX Spark as turning personal computers from tools into systems purpose-built for AI agents – machines that can receive instructions in natural language, call external services, evaluate their own output, and refine it, all without sending a request to a cloud server. The company’s own benchmarks suggest the chip can run language models with up to roughly 120 billion parameters locally, handle 90-gigabyte-plus 3D scenes, and generate 4K AI video clips on-device.
Whether those figures hold in practice will depend heavily on which laptop or desktop chassis the chip ends up in. This is not a trivial qualification. Nvidia confirmed that RTX Spark devices will range from single-digit watt operation in the thinnest configurations to a ceiling of approximately 80 watts in the largest desktop enclosures. The chip, in other words, is capable of very different things depending on the thermal solution the OEM chooses to build around it. A 14-millimeter slim laptop running the same silicon as a full-sized compact desktop will be a meaningfully different machine, even if both carry the RTX Spark name. That reality hands OEMs an unusual amount of influence over the platform’s reputation at launch.
The list of OEM partners committed to RTX Spark devices is extensive: ASUS, Dell, HP, Lenovo, Microsoft’s Surface division, and MSI confirmed for the fall launch, with Acer and Gigabyte to follow. Eastern Herald reported earlier this month on the breadth of the partner ecosystem assembled for Computex. Dell’s XPS 16 and Microsoft’s Surface Laptop Ultra were specifically named by Nvidia as early vehicles for the chip.
The software dimension of the RTX Spark announcement is at least as significant as the hardware. Microsoft’s Pavan Davuluri, who oversees Windows and Devices, confirmed that the company had restructured the workload-profile scheduling architecture inside Windows 11 specifically to accommodate the RTX Spark’s unified memory design. The idea is that the Windows scheduler can now shift computational tasks fluidly between the CPU’s 20 cores and the GPU’s CUDA cores depending on what each workload actually requires, rather than treating them as separate pools. Davuluri described this as ensuring maximum efficiency whether a user is checking email or running a local AI agent for code debugging – a formulation that deliberately bridges the gap between casual consumer and power user.
The RTX Spark also qualifies for Microsoft’s Copilot+ PC category, which requires a neural processing unit capable of at least 40 trillion operations per second. The chip’s integrated NPU clears that bar, meaning RTX Spark devices will carry the same AI-feature eligibility – Windows Studio Effects, Cocreator in Paint, live captions, real-time image generation – as the Qualcomm Snapdragon X machines that have been on the market since late 2024.
That framing points to what the RTX Spark contest is really about. Qualcomm was the sole Windows on Arm partner for years, protected by an exclusivity arrangement with Microsoft that expired in 2024. The Snapdragon X series established a credible baseline – good battery life, genuine AI performance, improving x86 emulation – but it arrived without the deep GPU heritage that developers and creators associate with Nvidia. RTX Spark carries the full CUDA ecosystem, DLSS upscaling, RTX ray tracing, and TensorRT inference optimization. For a game developer, a video editor, or an AI researcher who has spent years building workflows around CUDA, that software continuity is not a minor convenience. It is the difference between porting an entire toolchain and simply running existing code on new hardware.
Nvidia also disclosed at Computex that RTX Spark is the first chip in a planned multi-generation roadmap, with future iterations code-named Rubin – which will use LPDDR6 memory – and a subsequent generation called Rosa Feynman. The public commitment to a roadmap is a calculated move. OEM partners and enterprise software developers need confidence that investment in the platform will not be orphaned after one product cycle. Nvidia reported $81.6 billion in quarterly revenue in May, giving it the financial weight to sustain a long-term platform push that would have been implausible for most semiconductor companies.
What RTX Spark does not answer – and what Nvidia has not been asked to answer publicly – is whether consumers will actually use local AI agents in ways that require one petaflop of on-device compute. The use case Huang described at Computex – an agent that debugs code locally, runs autonomously overnight, calls external tools, and refines its own output – is compelling for a narrow class of users who already understand what they would do with such a system. For the broader consumer market, the practical value of that capability remains genuinely unclear. Tom’s Hardware noted that Nvidia enters a market where AI PC enthusiasm has outpaced demonstrated consumer demand, and the Copilot+ experience on Snapdragon X has been commercially modest so far. Financial markets have already priced in substantial expectations for the RTX Spark platform, which makes the fall launch a moment of reckoning for whether the hardware and its use cases can justify the narrative.
The chip is real. The partners are committed. The roadmap is on record. Whether the person who buys a Dell XPS 16 in November actually changes how they work because of the silicon inside it is the question Nvidia has not yet had to answer.
