blog / AI
AI3 June 20264 min read

Microsoft and NVIDIA just bet the AI PC future on RTX Spark

At Microsoft Build this week, Microsoft and NVIDIA announced RTX Spark: a superchip with 128GB of unified memory and a petaflop of on-device AI compute. Worth paying attention to, even if you're not in the market for new hardware right now.

by Matt Roberts

Something landed at Microsoft Build this week that deserves more attention than most of the coverage is giving it. NVIDIA and Microsoft announced RTX Spark, and the specs are genuinely mind blowing for a laptop device.

The short version: it's a superchip pairing a 20-core ARM CPU (designed with MediaTek) with a Blackwell GPU on a single piece of silicon, up to 128GB of unified memory, and 300 GB/s of memory bandwidth. Microsoft is claiming one petaflop of AI compute, which in practice means you could run a 120 billion parameter model locally, without touching the internet.

That memory figure is the thing I can't stop thinking about, I still remember my father bringing home his new work laptop with a whopping 8mb of RAM - yes, I am THAT old! Copilot+ PCs, which have been Microsoft's AI hardware story for the past year or so, top out at 32GB in most configurations. The jump to 128GB isn't incremental. It changes what's actually viable to run on-device.

For most enterprise use cases today, running serious AI workloads locally means either compromising on model size or routing everything to cloud APIs. Both have costs: smaller models are less capable, and cloud APIs mean your data leaves the building. For anyone working in a regulated environment, that second one is often a non-starter. RTX Spark at 128GB starts to change that calculation.

What Microsoft is actually building

The hardware announcement came alongside something called Aion 1.0, which is Microsoft's new framework for agentic AI on Windows. Local reasoning, tool-calling, agent pipelines that run on the device rather than phoning home for every inference step. AMD, Intel, Qualcomm, and NVIDIA are all in the ecosystem, so this isn't exclusive to RTX Spark. But RTX Spark is clearly the high-end tier, the configuration where you actually have enough headroom to run something serious.

Reading through the Build announcements, Microsoft is trying to make Windows the platform of choice for agentic AI development before anyone else can claim that ground. The Dev Box is part of this: a passive-cooled compact workstation with the full 128GB configuration, aimed at developers building for the Windows AI stack. Get developers building on your platform early and the application ecosystem follows.

The Surface Laptop Ultra

Microsoft's own RTX Spark device is the Surface Laptop Ultra: 15 inches, mini-LED display at 2880 by 1920, peak HDR brightness of 2,000 nits, under 4.5 pounds. Official pricing hasn't been confirmed, but estimates from analyst briefings put the N1X configuration at around $2,899. Autumn 2026, alongside devices from Dell, HP, Lenovo, Asus, MSI, Acer, and Gigabyte.

That OEM breadth matters. A reference design without partners is a prototype. RTX Spark launching with eight manufacturers on board means there's real commercial intent behind it, not just a headline announcement.

What it means for anyone planning hardware refresh

Enterprise procurement cycles are long, and autumn 2026 availability means organisations thinking about 2027 or 2028 refresh cycles should be tracking this now rather than waiting for the devices to land.

The more pressing question is what RTX Spark does to existing guidance on AI workloads. The assumptions underpinning most on-device AI recommendations today, about what needs cloud compute versus what you can run locally, don't hold at this spec level. If you've been telling customers or stakeholders that certain workloads require cloud infrastructure, that answer may need revisiting once 128GB on-device machines are available at scale.

The developer story matters for the channel too. Organisations adopting Windows AI applications in 2027 and 2028 will be running software built on Aion, by developers who had access to RTX Spark hardware early. The hardware shapes what gets built, and what gets built shapes what customers need.

One last thing

I've used Windows for most of my professional life. Twenty-plus years of Active Directory, Group Policy, and more 365 migrations and Intune deployments than I can sensibly count. I know the platform well, and I mean it sincerely when I say this is the most interesting piece of Windows hardware announced in a long time.

But I'm writing this on a MacBook Air, and I'm going to keep doing that for now. The battery life still surprises me after all this time. It never gets warm. I've stopped noticing it in my bag. A petaflop of on-device AI hasn't quite tipped the scales yet.

Ask me again in 2027.

#ai#nvidia#microsoft#hardware#windows#ai-pc
Share:X / TwitterLinkedIn

Related posts

Building my first AI-powered app: what I learned as a non-ML developer
AI

Building my first AI-powered app: what I learned as a non-ML developer

I built this website using Next.js, AWS, and Claude. I'm not a developer by trade. Here's an honest account of what that process was like and what surprised me.

14 May 20253 min read
Claude 3.7 and the rise of agentic AI — this is the inflection point
AI

Claude 3.7 and the rise of agentic AI — this is the inflection point

Anthropic released Claude 3.7 Sonnet in February 2025 with extended thinking mode. Combined with the MCP protocol, something important just shifted.

10 Mar 20253 min read
DeepSeek just changed the economics of AI. What it means for enterprise
AI

DeepSeek just changed the economics of AI. What it means for enterprise

DeepSeek R1 arrived in January 2025 and sent the AI industry into a brief panic. The dust has settled. Here's what actually happened and what it means for enterprise AI strategy.

28 Jan 20253 min read