Remember when “AI PC” meant a keyboard got a new button? That was the pitch not long ago: a dedicated Copilot key, a modest neural processor, and a lot of marketing about TOPS numbers that nobody outside a spec sheet could translate into something useful. You bought a laptop. It still ran a chat window that phoned home to a data center. The local AI part was mostly background noise removal on video calls.
So forgive the raised eyebrow when Microsoft and NVIDIA show up with something called the RTX Spark superchip and claim they’re reinventing the Windows PC for the age of personal AI. I’ve heard versions of this sentence before. The difference this time is the hardware actually looks like it was designed by people who have tried to run a model locally and gotten annoyed.
What was actually announced
On May 31, 2026, NVIDIA and Microsoft unveiled a new line of Windows machines built around the RTX Spark superchip. The headline specs:
- A 1-petaflop RTX Blackwell GPU
- Up to 128GB of unified memory
- A 20-core Grace CPU, pitched on efficiency
- Desktop, laptop, and workstation form factors
- Full CUDA and RTX ecosystem support, plus Windows-native agents
Target audience, per both companies: developers, creators, and power users who want to run capable agents locally and securely. Jensen Huang called it “the new PC” during the GTC Taipei 2026 keynote. Pavan Davuluri, who runs Windows and Devices at Microsoft, framed it as a new chapter for Windows PCs. The machines debut in October 2026, and NVIDIA is already running local-AI demos with Microsoft and partners at IFA 2026.
The spec that actually matters
Everyone is going to fixate on the petaflop. Fine, it’s a big round number. But the figure I care about is 128GB of unified memory, because memory capacity is the wall most people hit when they try to run serious models on their own hardware. A consumer graphics card with limited VRAM will happily run a small quantized model and then fall over the moment you ask it to hold a long context, keep a second model resident for embeddings, and still leave room for the application you’re actually using.
Unified memory changes the shape of that problem. If the CPU and GPU share one pool, you stop playing the shuffle game where weights get copied back and forth and your throughput dies in the transfer. For anyone building agents — things that need to keep state, call tools, re-read documents, and loop — capacity and bandwidth matter more than peak compute. You can have a petaflop and still be bottlenecked by a model that doesn’t fit.
That’s also why the “securely and locally” framing isn’t just privacy theater. Agents that touch your files, your email, your codebase are exactly the workloads people are most nervous about sending to someone else’s servers. A machine that can hold a capable model in local memory is a different proposition than a thin client with a Copilot key.
The part nobody wants to talk about
October 2026. That’s the ship date, announced at the end of May. Five months is a long time in this space, and announcement-to-availability gaps are where enthusiasm goes to die. In five months the cloud providers will have cut inference prices again, the open-weight models will have shifted, and whatever benchmark NVIDIA shows on stage will be measured against something newer.
There’s also no pricing in any of this. None. A 1-petaflop GPU with 128GB of unified memory and a 20-core CPU is not going to be a $900 laptop, and “developers, creators, and power users” is the kind of audience description that usually precedes a workstation price tag. Until a number exists, “the new PC” is a category of one that most people won’t be buying.
What I’ll be testing
When these land, the review doesn’t start with synthetic benchmarks. It starts with: can I run a useful agent for eight hours without the laptop becoming a space heater? Does the Windows-native agent layer actually expose local models to third-party tools, or is it a walled garden with a nice demo? Does CUDA support mean my existing scripts run unchanged, or does it mean a porting weekend? And what happens on battery, because a laptop that only hits its numbers while plugged in is a desktop with a handle.
I’m cautiously interested, which for me is a high bar. The hardware direction is right: memory capacity over marketing numbers, local execution over round trips, real developer tooling over a chat sidebar. That’s a more honest version of the AI PC than the first wave delivered.
But a keynote is not a product, and a petaflop on a slide is not a petaflop on your desk. Ask me again in October.
đź•’ Published: