NVIDIA Brings Trillion-Parameter AI to the Desktop With RTX Spark and DGX Station for Windows

NVIDIA has announced two distinct products aimed at pushing frontier AI inference out of the data center and onto enterprise desks and personal workstations: the RTX Spark superchip, co-announced with Microsoft for a new generation of Windows PCs, and the DGX Station for Windows, a workstation-class system built on GB300 Grace Blackwell infrastructure capable of running models up to one trillion parameters locally.

The announcements, made at Computex in Taiwan, mark a structural push by NVIDIA — the world's most valuable company by market capitalisation, built largely on data-centre AI chip dominance — to extend that position into the client and edge compute tier.

RTX Spark: A Unified Superchip for Personal AI Agents

The RTX Spark is a system-on-chip design that integrates CPU and GPU capabilities into a single die, targeting Windows laptops and desktop computers. NVIDIA and Microsoft announced the chip is rated at 1 petaflop of AI performance and supports up to 128 GB of unified memory — a unified memory pool that spans both compute domains and eliminates the bandwidth bottleneck that has historically constrained on-device inference for large models.

Microsoft described PCs running RTX Spark superchips as capable of supporting highly capable AI models and complex workloads. New PC models from Dell, Microsoft, and other OEM partners are scheduled to debut in fall 2026.

The 1 PFLOP figure is worth unpacking. At INT8 precision, 1 petaflop corresponds to roughly 1,000 trillion integer operations per second — enough headroom to run 7B- to 13B-parameter dense models at practical inference latency, and, with aggressive quantisation to INT4 or lower, to push meaningfully into the 30B–70B range. The 128 GB unified memory ceiling is the more consequential number for practitioners: it determines the maximum KV-cache capacity and, more directly, the largest model that can be loaded without offloading weights to slower storage tiers.

The positioning is explicitly around personal agents — persistent, context-aware processes that integrate with user data and local applications rather than routing every query through a cloud endpoint. That architecture implies low-latency, high-privacy inference loops that cannot be cost-effectively served from a remote API at the request frequencies involved in ambient, always-on agent workloads.

DGX Station for Windows: Frontier Inference at the Desk

The DGX Station for Windows is a different product class entirely. Built on NVIDIA's GB300 Grace Blackwell-class AI infrastructure, it is designed to run frontier models at up to one trillion parameters locally — a specification that places it in the same capability tier as the multi-rack GPU clusters most enterprises currently depend on for that scale of inference. NVIDIA's announcement frames it as putting a trillion-parameter AI supercomputer on every enterprise desk.

Grace Blackwell (GB300) combines NVIDIA's Blackwell GPU architecture with Arm-based Grace CPUs over NVLink-C2C interconnect, yielding chip-to-chip bandwidth that far exceeds what PCIe-attached configurations can deliver. That interconnect is what makes trillion-parameter inference feasible in a single-node, desk-side form factor: the memory across GPU and CPU dies is addressable as a coherent pool rather than requiring explicit tensor sharding across discrete VRAM boundaries.

For enterprises running fine-tuned or proprietary frontier models — the use case where cloud latency, data-residency constraints, or per-token API costs make remote inference unattractive — the DGX Station for Windows is an alternative deployment target that did not exist at this capability level before.

The Windows OS environment is notable. Previous DGX Station generations ran Linux. Targeting Windows natively means these systems slot into existing enterprise IT infrastructure without requiring a separate OS management track, which has been a practical friction point for deploying GPU workstations in environments where the desktop estate is managed through Active Directory and Windows-native tooling.

The Historical Pattern Here

We have seen this compression arc before. In the early 2000s, the hardware needed to render a full feature-length CGI sequence — the kind of farm SGI and Sun workstations had exclusive claim on — collapsed into a commodity workstation within a decade, then into a laptop GPU within two. The same pattern played out with genomic sequencing compute, real-time physics simulation, and, more recently, the deep learning training workloads that required purpose-built clusters as recently as 2018. In each case, the timeline from "only in the data centre" to "on your desk" ran faster than the prior cycle.

The inference side of large language models is following the same trajectory, and the RTX Spark and DGX Station for Windows are products of that compression — not its cause. NVIDIA is building into a shift already underway in model efficiency (quantisation, sparse attention, speculative decoding) that is reducing the compute and memory requirements of frontier inference faster than most forecasts assumed even three years ago.

Implications for Enterprise Deployment Architects

For those responsible for AI infrastructure decisions, the two products carve out distinct but complementary positions. RTX Spark targets developer workstations, power-user endpoints, and scenarios where per-device agent intelligence is the design goal — think co-pilot tooling that operates on local files, email, and calendars without phone-home dependencies. The 128 GB unified memory ceiling and 1 PFLOP rating are sufficient for sub-70B models at reasonable throughput, which covers the majority of production-grade open-weight models available today.

DGX Station for Windows targets a different workload profile: organisations that need to run 100B+ parameter models — including emerging mixture-of-experts (MoE) architectures where active parameter counts are lower but total model weight is large — without provisioning cloud GPU instances or managing a rack-mounted on-premises cluster. The Grace Blackwell substrate means it is also a credible fine-tuning node for smaller model variants, not purely an inference endpoint.

Worth flagging: neither announcement included pricing or detailed thermal/power specifications at the time of writing. For the DGX Station for Windows in particular, power envelope and cooling requirements will be determinative for enterprises evaluating whether the system can realistically be deployed at individual desks rather than in a controlled server room. Grace Blackwell-class hardware in rack configurations draws several kilowatts; the desk-side form factor will require either meaningful thermal engineering trade-offs or power specs that narrow the performance ceiling relative to rack deployments.

What Comes Next

The fall 2026 debut window for RTX Spark-based OEM systems gives the ecosystem — ISVs, enterprise IT teams, and agent platform developers — roughly a quarter to build out software stacks that exploit the hardware. Microsoft's involvement from the announcement stage suggests Windows AI and Copilot runtime layers will be optimised for the RTX Spark memory architecture, but the degree to which third-party frameworks (llama.cpp, vLLM, ONNX Runtime) will expose the full unified memory pool cleanly is an open question that benchmarks, not press releases, will settle.

For practitioners tracking where inference gravity is shifting, both announcements reinforce a direction that has been building for several quarters: meaningful AI workloads are migrating from centralised cloud endpoints toward a distributed topology where capable local nodes handle latency-sensitive, privacy-sensitive, or cost-sensitive inference, and cloud remains the tier for burst capacity and the largest training runs. NVIDIA is positioning its silicon across both tiers simultaneously — and, given its current market position, it is better placed than any other vendor to set the terms of that transition.