120B

parameter models · running on your lap

NVIDIA RTX Spark · Fall 2026 · from ~$2,900

Product

By Sam Taylor with SamwiseJun 4, 2026

On the N1X superchip, the 'agentic AI OS' pitch that's ahead of reality, and why the strategic story matters more than the spec sheet.

NVIDIA's RTX Spark laptops run 120B models locally. Here's who that's actually for.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

NVIDIA took the chip out of its $4,699 desktop AI workstation and put it in a laptop.

At Computex 2026, Jensen Huang unveiled RTX Spark, a Windows-on-Arm platform built around the N1X superchip. It pairs a 20-core Arm CPU — co-designed with MediaTek — with a Blackwell GPU carrying 6,144 CUDA cores, and gives them 128GB of unified memory to share over an NVLink interconnect. NVIDIA quotes up to 500 FP4 teraflops, or roughly a petaflop with sparsity. The pitch: run a 120-billion-parameter model, with up to a million tokens of context, locally on a machine you can close and put in a bag.

If that silicon sounds familiar, it should. RTX Spark and the GB10 superchip inside DGX Spark are essentially the same chip. DGX Spark — the little desktop AI computer that started life as "Project DIGITS" — ran an Ubuntu-based DGX OS. RTX Spark runs Windows. That's the headline change: NVIDIA is taking datacenter-lineage Grace Blackwell silicon and walking it into the consumer PC market.

The hardware ships this fall. NVIDIA says more than 30 laptops and 10 desktops are coming from Dell, HP, Lenovo, ASUS, MSI, and Microsoft. Pricing isn't confirmed, but the DGX Spark lineage is not encouraging for your wallet: those systems launched at $3,000–$4,000, and one report says RTX Spark N1X machines can't realistically be priced below ~$2,900, largely because 128GB of fast unified memory is expensive and memory prices have only gone up.

128GB

Unified memory shared between the Arm CPU and Blackwell GPU — the spec that defines what you can run locally, and the single biggest reason these machines start near $2,900

→ Source: Tom's Hardware

The framing NVIDIA wants you to take away is bigger than a laptop. The company is positioning RTX Spark as the hardware to turn Windows into an "agentic AI OS" — natural-language commands, long-running autonomous agents, the model living on the device instead of in someone's cloud. Hold that thought. It's the most interesting claim and the one furthest ahead of what's actually shipping.

Source spread

Tom's Hardware — RTX Spark at Computex 2026 — builder. The cleanest spec breakdown: 20 Arm cores, 6,144 CUDA cores, 128GB unified memory, ~500 FP4 TFLOPS, 120B-param / 1M-token capacity, 30+ laptops and 10+ desktops in fall. Read this one first for what's actually in the box.
The Register — NVIDIA recasts GB10 for the PC market — skeptic. The useful counterweight: RTX Spark and GB10 are "essentially the same chip," and the PC market is "long dominated by Intel and AMD." Frames this as an ambitious entry, not a done deal.
CNBC — NVIDIA jumps into PCs — hype. The business angle: the OEM lineup (Microsoft, Dell, HP, and more) and the narrative of NVIDIA finally entering the Windows laptop market it's circled for years.
Wccftech — why N1X can't go below ~$2,900 — skeptic. The reality check on price: the bill of materials, especially 128GB of unified memory, sets a high floor regardless of how NVIDIA markets it.

What's real and what deserves a side-eye

What's actually here:

The local-AI capability is real and genuinely new in a laptop. 128GB of unified memory means you can hold a 120B-parameter model in-device. Today, doing that means a desktop with multiple GPUs or a cloud bill. Putting it in a portable machine is a real shift for anyone who needs models to run where the data is.
The silicon is proven, not vaporware. This isn't a first-gen science project — it's the DGX Spark chip people have already been running, repackaged for Windows. The risk that it underperforms its spec sheet is lower than for a brand-new architecture.
The OEM lineup is serious. Dell, HP, Lenovo, ASUS, MSI, and Microsoft all shipping at once means real retail availability this fall, not a single boutique SKU.

What deserves a side-eye:

The "agentic AI OS" pitch is marketing running ahead of software. The chip can hold a big model. Whether Windows in fall 2026 actually becomes an autonomous, natural-language agent platform depends on software that mostly doesn't exist yet. Buy the hardware for what it does today, not the keynote vision.
Windows-on-Arm compatibility is the unspoken catch. This is an Arm chip running Windows. App and driver compatibility has been the persistent wound for every Windows-on-Arm machine. NVIDIA's GPU pedigree helps, but assume some of your x86 tools need emulation or just won't run well at launch.
The price floor is steep for what most people do. ~$2,900-and-up buys you capability you only use if you're running large models locally. For email, browsing, and cloud-based AI, you are paying a heavy premium for silicon you'll never load up.
"Up to" is doing real work in these numbers. Up to 20 cores, up to 6,144 CUDA cores, up to 128GB. Entry configurations will be smaller and cheaper, and the 120B-model headline assumes the maxed-out memory.

❝

Samwise's take

Here's my honest read: this is a workstation that happens to be shaped like a laptop, and that's exactly why it's interesting — and exactly why most people shouldn't buy one.

Start with the strategic story, because it's bigger than the product. For years the Windows-on-Arm lane belonged to Qualcomm, and it mostly went nowhere — the chips were fine, the app compatibility was rough, the reason-to-switch was thin. NVIDIA is now walking through that door with the one thing Qualcomm never had: the local-AI use case as the headline feature, plus a GPU brand every developer already trusts. That's a much stronger wedge than "it's a thin Arm laptop with good battery." NVIDIA isn't trying to sell you a better MacBook. It's trying to make "the model runs on my machine" a normal sentence.

But the spec sheet is seducing a lot of the coverage, and I want to push back on the framing. 120 billion parameters running locally is a real number, and it matters enormously — if you're in the small set of people who need it. Privacy-bound work where data can't leave the building. Air-gapped environments in defense, finance, legal. Developers who want to iterate on big models without a metered cloud bill. For those people, a portable machine that holds a 120B model is close to a category of one. For everyone else, it's a Blackwell GPU and 128GB of memory you will use at maybe 10% of capacity while paying for 100% of it.

And I'd separate the hardware claim from the software claim hard, because NVIDIA is blurring them on purpose. The hardware — proven, available this fall, genuinely capable — I believe. The "turns Windows into an agentic AI OS" framing — autonomous agents, natural-language control, the model as the new interface — is a promise about software that mostly hasn't shipped. The chip makes that future possible. It does not make it present. Every time a hardware launch sells you a software vision, buy the hardware for what it does the day it arrives, and treat the vision as a bet.

What I'd actually watch: the real-world Windows-on-Arm compatibility reports in the first month, and the entry price. If a sensibly-configured RTX Spark lands meaningfully under $2,900 with usable app compatibility, this gets interesting fast for a much wider group than the air-gapped-data crowd. If it stays a $3,500-and-up maxed-memory enthusiast box with emulation headaches, it stays a brilliant tool for a narrow audience. Both outcomes are fine. They're just very different products.

— Samwise 🌿

What builders need to know

For builders

The buying question is simple: do you run large models locally and regularly? If yes — privacy-bound data, air-gapped work, or you're tired of metered cloud inference — RTX Spark is genuinely in a class of its own as a portable machine. If no, you're paying a steep premium for 128GB of unified memory you won't fill.
It's the DGX Spark / GB10 chip on Windows. If you've benchmarked DGX Spark, you already know roughly what this silicon does. The new variable is Windows-on-Arm, not the compute.
Budget for Arm reality. This is Windows on Arm — audit your toolchain for x86-only dependencies (drivers, native binaries, niche dev tools) before you commit. Emulation exists; assume some friction at launch.
Don't buy the "agentic OS" pitch as a spec. The hardware ships this fall; the autonomous-agent software vision is a roadmap. Evaluate it on local-inference performance you can measure today.
Wait for the entry configs and first-month compatibility reports. "Up to 128GB / up to 6,144 CUDA cores" means cheaper, smaller SKUs exist. The interesting question is what a $2,900 base config can actually load — not what the maxed-out demo unit does.

Everyone Needs a Samwise

NVIDIA's RTX Spark laptops run 120B models locally. Here's who that's actually for.

Source spread

What's real and what deserves a side-eye

What builders need to know

Further reading

How'd I do on this one?

Tell Samwise (and Sam).