Nvidia RTX Spark laptops go on sale from $2,599

Laptops built on Nvidia's RTX Spark chip went on sale on Friday, October 16, 2026, from six manufacturers, at prices starting at $2,599. Each one carries up to 128GB of memory shared by the processor and the graphics chip. That is enough to run large AI models and AI agents on the laptop itself, with no cloud service, which is the whole pitch.

Prices and configurations vary by region, and the machines sell through the makers' websites and major retailers, Investing.com reports. Microsoft's own model, the Surface Laptop Ultra, opened for preorders on October 8 at $2,599.99.

The laptops on sale

Investing.com and The Next Web list the same launch models:

Maker	Models
ASUS	ProArt P14, ProArt P16
Dell	XPS 16
HP	OmniBook Ultra, OmniBook X 14
Lenovo	Yoga Pro 9n
Microsoft	Surface Laptop Ultra
MSI	Prestige N16, Prestige N16 Flip AI+

Nvidia's original announcement on May 31 said Acer and GIGABYTE would follow later. It also described laptops as slim as 14 millimeters and as light as three pounds, in 14- to 16-inch sizes.

What is inside

RTX Spark is a single "superchip." It pairs a Blackwell RTX graphics chip with a Grace processor, a CPU that MediaTek helped design. Nvidia rates it at up to 1 petaflop of AI performance at FP4, a low-precision number format used for fast AI math.

The shared memory is the key part. On most laptops, the graphics chip has its own small memory pool, which caps the size of an AI model it can hold. Here, up to 128GB is available to both chips. Nvidia's May announcement said the machines can run models with 120 billion parameters and up to 1 million tokens of context. Tokens are the word pieces a model reads.

Nemotron 3.5 Bolt, the model built for it

Nvidia released a new open model for these machines on Thursday, October 15. Nvidia's blog post calls it Nemotron 3.5 Bolt 74B-A9B. It has 74 billion parameters in total, but only 9 billion are active for each step. That design is called a mixture of experts. Compressed to 2-bit numbers, the model fits in 24GB of memory.

Nvidia says the model is fully open, including its training data, weights and training recipes. It runs through common local tools: llama.cpp, Ollama, LM Studio, Unsloth and vLLM. A smaller 4GB version, Nemotron 3.5 Bolt E4B, is due on October 23.

The Next Web adds a reality check. Bolt scores 21 on the Artificial Analysis Intelligence Index, against 58 for Anthropic's Claude Opus 5.5. A local model saves cloud costs, but it is not a frontier model. Nvidia also shipped Nemotron 3.5 Super VL, a larger 120B open model, this week.

Agents and software at launch

The launch agents include GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy, Nvidia says. The Next Web reports they install without a command line.

Two sandboxes limit what an agent can touch: Nvidia's OpenShell and Microsoft's Execution Containers. A router called Switchyard, due later in October, lets the Hermes agent split work between local and cloud models, according to The Next Web.

"We're building Windows to be the platform for hybrid intelligence, where AI can run locally when it makes sense," said Yusuf Mehdi, an executive vice president at Microsoft. Adobe's Photoshop and Premiere support starts in beta, with Substance 3D coming later.

What this means for developers

• Check your Python stack. These laptops run Windows on Arm. Nvidia says the first PyTorch-CUDA build for that platform is ready, and that 97% of the top 15,000 Python packages work. Test the other 3% before you buy for a team.
• Size models to the memory. Bolt needs 24GB at 2-bit precision. A 128GB machine leaves room for a larger model or several at once, but RAM is fixed, so pick the configuration for the biggest model you plan to run.
• Benchmark before you switch off the cloud. Bolt's score of 21 against 58 for Claude Opus 5.5 is a large gap. Route simple, private or high-volume tasks locally and keep hard ones on a cloud model.
• Set agent permissions first. OpenShell and Execution Containers exist to limit what an agent can read and change. Configure them before giving an agent access to email, code or files.
• Expect regional prices. $2,599 is the starting price, and Investing.com says pricing varies by region. Check local prices and stock before you budget.