Nvidia’s new RTX Spark packs a petaflop of AI compute into a desktop superchip due this fall

Nvidia used its Computex 2026 keynote to introduce a desktop part aimed at running large artificial-intelligence models on a machine sitting under a desk rather than in a distant data center. The company describes the product, called RTX Spark, as a single-package “superchip” that fuses a processor, a graphics engine and a large shared memory pool, and it says the platform will reach buyers in the fall.

A petaflop of AI compute in a single package

The headline specification is raw throughput. Nvidia says the chip can deliver up to one petaflop of AI performance, a figure that until recently was associated with rack-mounted servers rather than a consumer desktop. A petaflop describes a quadrillion operations per second on the reduced-precision math that modern AI models lean on, and reaching that number in a package meant for a single workstation marks a steep jump in what a desktop is expected to handle. That level of compute is meant to let a machine load and run sizable language and image models entirely on local hardware, without shipping data back and forth to a cloud provider.

Alongside the compute claim, the design carries 128 gigabytes of unified memory, according to the specifications shared at the event. Unified memory means the processor and graphics engine draw from the same pool rather than copying data between separate banks, which matters for models whose size is limited less by processing speed than by how much data can be held in memory at once.

Turning Windows into an “agentic” system

Nvidia framed the launch as more than a faster graphics card. The company pitched a platform intended to turn Windows into what it called an agentic AI operating system, meaning software agents that can carry out multi-step tasks on a user’s behalf would run on the local chip. The argument is that keeping those agents on the device reduces latency and keeps sensitive material off external servers.

That positioning places the product in a broader industry contest over where AI computation happens. Cloud vendors have built their businesses on centralizing that work in large facilities, while a competing camp argues that personal and business data are better processed close to where they are created. A desktop capable of a petaflop is a bid to make the local option credible for heavier workloads. It also reflects a bet that everyday software, not just specialized research tools, will increasingly call on AI models in the background, creating demand for hardware that can answer those calls instantly without a network round trip.

The Arm and Blackwell pairing

The chip combines an Arm-based central processor with a graphics engine built on Nvidia’s Blackwell architecture. Pairing an Arm processor with a high-end graphics core on one package is notable because most Windows desktops still run on x86 processors from Intel or AMD. An Arm design can offer efficiency advantages, but it also depends on software being compiled or translated to run well on that instruction set.

Blackwell is the same generation of graphics architecture Nvidia has used across its recent data-center and consumer lines, which suggests the desktop part is meant to inherit the model-serving features developed for larger systems. Bringing those capabilities down to a single desktop package is the central engineering claim behind the announcement. Fusing the two elements on one piece of silicon, rather than wiring a separate processor to a separate graphics card, is also what allows the shared memory pool to work, since both halves of the chip sit close enough to reach the same data without a slow hand-off.

Why unified memory is the pivotal number

For anyone running large models, memory capacity often decides what is possible at all. A model that does not fit in available memory either cannot run or must be split and slowed. By allocating 128 gigabytes to a shared pool, Nvidia is signaling that the machine is designed to hold models that would overwhelm a typical consumer graphics card, which usually carries far less dedicated memory. The gap is stark: many high-end consumer cards top out well below that figure, so a model that once required stringing several cards together might fit inside a single Spark.

The tradeoff with unified memory is that the processor and graphics engine share a fixed budget, so heavy demand from one side can constrain the other. How the platform manages that balance in practice will shape whether the advertised compute figure translates into smooth performance on real workloads. Independent testing will also have to confirm how much of the 128 gigabytes is available to a model once the operating system and background tasks take their share.

A fall release and unanswered questions

Nvidia set availability for the fall of 2026 but left several practical details for later, including pricing, power draw and the exact configurations that will ship. Those figures will determine whether the product appeals mainly to developers and researchers or reaches a wider set of buyers curious about running AI locally. Power draw is a particular unknown, since a chip pushing a petaflop is likely to demand cooling and electricity closer to a workstation than a slim office machine.

The announcement also raises questions about software support. An Arm-based Windows desktop pushing agentic features depends on applications and AI frameworks being tuned for the hardware, and the strength of that ecosystem on release day will influence how much of the promised petaflop users can actually put to work. Programs written for x86 chips may need to be recompiled or run through a translation layer, and the performance of that translation has been an obstacle for earlier Arm-based Windows machines. Until the platform is in independent hands and running third-party software, the specifications remain the company’s own claims rather than measured results.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

More from Morning Overview