Nvidia unveils RTX Spark superchip for Windows AI PCs, with autumn 2026 ship date

Nvidia opened Computex 2026 in Taipei on Monday with the introduction of RTX Spark, a new superchip the company describes as the first platform built from silicon to operating system for on-device AI agents. The announcement came in a keynote from chief executive Jensen Huang and was accompanied by parallel statements from Microsoft, which confirmed a co-engineered Windows runtime layer, and from at least seven major OEMs that committed to shipping hardware in autumn 2026.

The headline claim is that a Spark-equipped Windows PC can run a 70-billion-parameter language model locally, with full quantisation, at interactive token rates. That puts the platform in the same on-device inference class as Apple's M4 Max and M5 Ultra Macs, which until now have been the only consumer-grade hardware capable of running models of that size without quality-destroying compression.

What's Actually in the Silicon

RTX Spark is a system-on-chip rather than a discrete GPU. The package combines an Arm-architecture CPU (custom design, derived from but distinct from Nvidia's Grace server core), a Blackwell-derived GPU block with full CUDA support, a dedicated NPU for low-power inference, and unified memory addressable by all three compute domains. Nvidia is positioning it against the system-architectural approach Apple has taken with its M-series chips since 2020.

The desktop variant targets a 100-watt power envelope and ships with up to 128 GB of unified memory in the announced configurations. The laptop variant operates in a 35 to 65 watt envelope with up to 64 GB unified memory. Both share the same instruction set and developer toolchain, which is a deliberate echo of how Apple's Mac lineup is structured.

The Arm pivot matters for two reasons. First, it lets Nvidia design power management at the system level rather than working through x86 firmware constraints. Second, it gives the company a path to compete on battery life with the Apple Silicon Macs that have dominated thin-and-light reviews since the M1 launch. The cost is software compatibility: Windows on Arm has had a difficult decade, although Microsoft's Prism translation layer and the increased native-Arm coverage from Adobe, Autodesk and the major IDE vendors over the past two years has narrowed the gap.

The Microsoft Layer

Microsoft's contribution is what it is calling the Windows AI Runtime, a new system layer that sits above the device drivers and exposes a unified inference API to applications. Apps that target it can run any compatible model on any compatible NPU or GPU without rebuilding for each vendor. Microsoft confirmed that Qualcomm and AMD parts will also support the runtime, which means Spark is the launch platform but not the exclusive one.

Practically, the runtime resolves the most frustrating part of the Windows AI PC story to date. Until now, Copilot+ PC features have required platform-specific implementations, and applications outside the Microsoft 365 stack have not had a clean integration path. The new runtime is meant to address that, although the developer documentation and SDK are not yet public.

Pricing, Availability and OEM Lineup

Nvidia did not announce a chip-level price, which is typical for the segment. OEM device prices indicated in OEM briefings put the entry-level Spark laptops in the US$1,499 to US$1,999 band and Spark desktops in the US$1,999 to US$3,499 band. Both ranges sit at or below the equivalent Apple Silicon configurations with comparable unified memory.

The first wave of devices comes from Asus, Dell, HP, Lenovo, Acer, MSI and Razer. Lenovo confirmed three ThinkPad variants and one ThinkBook. Dell confirmed two XPS configurations and a Precision mobile workstation. Asus committed to seven SKUs across the ROG, Zenbook and ProArt lines. Australian availability dates were not given in the Computex announcement and typically lag US launches by four to eight weeks for the volume tiers.

Where This Sits in the AI Hardware Landscape

Spark is Nvidia's first serious entry into the consumer client market in a decade. The company has been the dominant supplier of data centre AI accelerators since 2022, but its consumer GPU business has been a side business of comparable scale only because gaming graphics carried it. Spark is an attempt to expand that consumer footprint into the AI inference workload that is rapidly becoming a primary use of personal computers.

The realistic competitive picture for an Australian buyer in autumn 2026 will look like this. Apple Silicon Macs remain the on-device inference leader at the high end, with the M5 Ultra Mac Studio capable of larger models than Spark in the desktop tier. Spark-equipped Windows desktops will be priced lower for comparable inference performance up to 70-billion-parameter models. AMD's Strix Halo platform, already shipping, sits between the two on inference but ahead on the integrated graphics for gaming. Qualcomm Snapdragon X Elite second-generation parts (announced for late 2026) will compete on battery life and Arm-native compatibility.

For the buyer who specifically wants to run local models without sending prompts to a cloud API, Spark is the first Windows hardware that makes that viable for the model classes most people actually use. Whether that is enough to overcome the Apple Silicon habit of the past five years will depend on how clean the Microsoft runtime turns out to be and on what the autumn pricing actually settles at.