NVIDIA RTX Spark Laptops: Prices, Specs & First Reviews

By Muzamil Ahmad · Updated 2026-10-16 · Productivity

#productivity#nvidia rtx spark laptops#rtx spark laptop price#rtx spark n1x#surface laptop ultra rtx spark#rtx spark vs macbook#rtx spark review#best rtx spark laptop#rtx spark release date

Save

Today, Friday, October 16, 2026, the laptop NVIDIA has been building toward all year finally goes on sale. RTX Spark laptops from Microsoft, ASUS, Dell, HP, Lenovo, and MSI hit shelves today, starting at $2,599 — machines designed to run large AI models locally instead of renting intelligence from a cloud API by the token.

This is the hardware side of the story we covered yesterday with Nemotron 3.5 Bolt. Bolt is the model; RTX Spark is the machine it was built for. One day after the weights landed, the laptops arrive. And unlike yesterday, when everything was pre-launch promise, today we have something better: the first independent reviews with real numbers.

Here is the honest version: what RTX Spark is, what every maker charges, what the first testers actually measured, and where the launch-day story still has holes.

What Is RTX Spark?

RTX Spark is not a laptop. It is the chip platform inside a wave of laptops — NVIDIA's first full client PC platform, after decades of selling the GPU inside other companies' machines. NVIDIA calls it a reinvention that combines a gaming PC, a creative studio, and a personal AI supercomputer in one superchip.

The idea is simple: put enough memory and AI compute in a thin laptop that large models run on the machine itself. No API bill, no data leaving your desk, no waiting on a datacenter. That is the same local-AI bet as Nemotron 3.5 Bolt, and the two launches are deliberately paired — the model landed October 15, the laptops October 16.

The N1X Chip: What Is Actually Inside

Every RTX Spark laptop is built around the same silicon family, in two tiers:

A few things in that table deserve a slower read.

Unified memory is the whole game. In a normal laptop, the CPU has its own RAM and the GPU has its own VRAM, and models must fit in the smaller VRAM pool. RTX Spark shares one memory pool between CPU and GPU, and lets you hand most of it to the GPU — one reviewer allocated 96GB of a 128GB machine to the GPU. That is why a laptop can run models that would never fit in a 24GB laptop GPU.

Arm, not x86. The Grace CPU is Arm-based, which means these are Windows on Arm machines. Most Windows software now runs on Arm through emulation or native builds, but some niche tools, drivers, and older apps may lag. NVIDIA and Microsoft spent the launch week insisting the ecosystem is ready — 97% of the top 15,000 Python wheels, native Conda, native PyTorch-CUDA. The remaining few percent is where early buyers may still trip.

The petaflop number is NVIDIA's claim. One petaflop of FP4 compute is a peak, lab-conditions figure. What you actually get depends on the model, the engine, the power limit, and whether you are plugged in. Treat it as a ceiling, not a promise.

Same chip family, different shelves. If you want a laptop, you want RTX Spark. If you want a desk-bound developer box, that is DGX Spark.

The Laptop Lineup: Prices by Maker

This is the part most launch coverage fudges. Prices are real now — but they come from a mix of Microsoft's own store, retailer listings, and secondary reports, and a few numbers differ between sources. Every price below is in US dollars, and where reports disagree I say so.

Acer and Gigabyte are confirmed partners with models to follow, but no prices yet.

Microsoft Surface Laptop Ultra: The Flagship

Microsoft's own machine is the launch's center of gravity — and the one with the most complete public pricing. Eight configurations are listed, from the $2,599.99 entry model (18-core CPU, 5,120-core GPU, 24GB memory, 512GB SSD) to the $5,899.99 flagship (20-core CPU, 6,144-core GPU, 128GB memory, 1TB SSD).

The hardware around the chip: a 15-inch 120Hz touchscreen rated at 2,000 nits peak HDR brightness, under 4.5 pounds, under 18mm thick, in Platinum or a new Nightfall color. Microsoft is offering up to $1,000 back on an eligible MacBook Pro trade-in through November 23 in the US and Canada.

Two honest notes. First, Microsoft does not publish battery-life hours for this machine — only that battery life "varies a lot by use." That is a gap, and the first battery tests suggest why. Second, Microsoft's claim of 2.5x the thermal capacity of the Surface Laptop 8 is Microsoft's own comparison, not independent testing.

The Rest of the Lineup

ASUS targets creators with the ProArt line: the P14 starts at $2,599.99, the P16 at $2,799.99. The top P16 with 128GB is where reports split — one retailer listing shows $6,499, another report cites $6,999.99. Treat either as "roughly six and a half to seven thousand," and confirm the exact figure with the seller before buying.

Dell's XPS 16 Creator Edition starts at $3,799.99 — about $1,200 above Microsoft's and ASUS's entry machines — with a 3.2K OLED display and a 99Wh battery. The premium is the screen, the battery, and the XPS badge, not a faster chip.

HP's OmniBook X 14 starts at $2,999.99 and the OmniBook Ultra 16 at $3,199.99, but shipping dates are the murkiest in the lineup — some reports put HP in the October 16 wave, others list November 4. Check the retailer's date, not the headline. MSI's Prestige N16 Flip AI+ starts at $3,299.99, with preorders from October 8; it is the only convertible 2-in-1 in the launch group. Lenovo's Yoga 9n is announced but not yet priced, with a Fall 2026 window.

The Desktop: Surface RTX Spark Dev Box

For developers who don't need portability, Microsoft's Surface RTX Spark Dev Box is a compact desktop with the full chip (20-core CPU, 6,144-core GPU, 128GB memory) for $5,999, shipping in November, sold exclusively through Microsoft.com in the US. It is the cheapest way to get the full 128GB configuration in a box that sits on a desk — though the DGX Spark 64GB, arriving October 23 from $4,999, is the cheaper desk option if 64GB is enough for your models.

First Reviews: What Testers Actually Measured

This is where the launch gets interesting. Yesterday, every performance number was NVIDIA's. Today, independent testers have published measurements — and they are more useful than the launch claims because they show where the platform shines and where it strains.

Local AI: The Numbers That Matter

StorageReview tested a Surface Laptop Ultra for Business with the full chip (20-core CPU, 6,144-core GPU, 128GB memory, up to 96GB allocatable to the GPU). Its headline result: the open models gpt-oss:20b and Qwen3.6-35B-A3B both loaded and generated at 58 to 89 tokens per second. A 35-billion-parameter model running at 89 tokens per second on a laptop is genuinely new — fast enough for a real coding agent, not a demo.

A separate reviewer measured a 120-billion-parameter model at 63 tokens per second and a 27-billion model at 16 tokens per second on a 128GB machine — a reminder that speed depends on the model's architecture (how much of it wakes up per token), not just its size. The same reviewer found the machine roughly tied with a MacBook Pro on writing speed but about 2.5x faster at reading prompts.

The same StorageReview test also found the platform's limits. On anything that fits in 24GB of memory, a laptop with an RTX 5090 GPU was 2.3x faster. RTX Spark wins on models too big for a 24GB GPU. It loses on small ones. That is the honest shape of the tradeoff: memory capacity, not raw speed, is what you are buying.

The Prefill Problem: 170 Seconds Before the First Word

Here is the number NVIDIA did not put in its launch post. Microsoft's own published benchmark for the MAI Code 1.1 Flash model (137B total parameters, 6.8B active, 53GB) on the Windows on Arm build of llama.cpp showed prefill — reading your prompt before generating anything — at 923.5 tokens per second on a 64K-token prompt (about 71 seconds of waiting) and 769.8 tokens per second on a 128K-token prompt (about 170 seconds). Decode, the actual writing, ran at about 63 tokens per second on short prompts. It matters because an agentic coding assistant re-sends the repository, the failing test, and every tool result with each turn — most tokens go in, not out — so cold sessions pay the full wait while warm, cached turns stay fast.

Gaming: Better Than Expected

The gaming results surprised even skeptical testers. Gizbot's review ran Control Resonant at 2880x1800 on Ultra settings with ray tracing off at a stable 50 FPS — without any frame generation. Turning on 2x frame generation pushed it to 100 FPS; turning ray tracing on brought it back to 50 FPS but still playable. 007 First Light hit around 139 FPS at 1440p on high settings with DLSS balanced and 3x frame generation. Cyberpunk 2077 ran around 135 FPS at 1440p with ray tracing on and path tracing off.

HardwareZone's ProArt P16 test was more modest but still strong: Cyberpunk 2077 held a fairly steady 75 FPS at 1440p on mid-range settings, and Gears of War: E-Day ran well. The machine stayed comfortable during long sessions, with the keyboard and palm rest noticeably warm but never painful.

Thermals: One Red Flag

StorageReview's test produced the launch's clearest warning. During a sustained all-core double-precision solve, the laptop cut power three times in one evening — and an earlier review unit had failed the same way and had to be replaced. That is two units showing the same failure mode under the same sustained load. Context matters: a double-precision all-core solve is an extreme workload — not browsing, not coding, not gaming. But it is exactly the kind of sustained load a researcher might run, and two failures is a pattern worth watching, not a fluke to dismiss.

Battery: The Big Unknown

The one early battery datapoint is sobering: about an hour of AI inference on battery in one reviewer's test. General battery life — browsing, video, office work — is still unmeasured in any published review. Microsoft publishes no battery-life hours at all. If you plan to run local models away from a plug, plan around the charger until real battery tests arrive.

The Software and AI Story

Hardware is only half of this launch. NVIDIA spent the week before sale day shipping the software the laptops are meant to run — and this is where RTX Spark looks most finished.

Nemotron 3.5 Bolt is live. The weights we reported yesterday as "scheduled" are now on Hugging Face in official NVIDIA repositories — the 24GB build, a GGUF build, a BF16 build — plus community GGUF and MLX conversions from Unsloth and others. It runs through Ollama, LM Studio, Unsloth, vLLM, and llama.cpp, and is wired into the Hermes and OpenClaw agent frameworks.

NVIDIA PAIR left beta. The free, open-source router spreads AI jobs across the PCs on your home network, and now adds llama.cpp support and direct model downloads from Hugging Face. It routes whole jobs to one machine; it does not pool GPU memory.

DLSS 5's faster model lands today in NBA 2K27, with NVIDIA claiming frame-rate boosts up to 1.4x in that game. The same model reaches Hogwarts Legacy on November 5 and RTX Remix on October 23. NVIDIA says the underlying AI model runs twice as fast with lower GPU overhead — a vendor claim, and the game-specific gains will vary.

Adobe is up to 2x faster — NVIDIA and Adobe rearchitected Photoshop for GPU-accelerated compositing and Premiere for playback, rendering, and effects. The AI optimizations are in beta now; Substance 3D updates follow soon.

The developer stack is ready on day one. The first PyTorch-CUDA build for Windows on Arm, the first natively-built Python 3.15 for Windows on Arm, the first native Conda installer with over 2,000 packages, and 97% of the top 15,000 Python wheels. NVIDIA is also shipping AI playbooks for RTX Spark — guided setups for a first agent, an AI chief of staff, a dev machine, or a creative assistant.

Agents get sandboxing. NVIDIA OpenShell and Microsoft Execution Containers (MXC) give local agents OS-level containment, credential protection, and monitoring. NVIDIA Switchyard routes queries between local and cloud models so you pay for cloud intelligence only when you need it.

Decision models get fast. NVIDIA says decision models — small models that pick an agent's next action without a full reasoning loop — now run up to 200x faster than traditional agents in llama.cpp and Unsloth on RTX Spark and GeForce GPUs. That is a vendor claim, but if it holds, it changes how snappy local agents feel.

RTX Spark vs MacBook Pro: The Real Comparison

Every buyer here is really asking one question: this or a MacBook Pro? They win different races.

Memory bandwidth: About 300 GB/s on the RTX Spark N1X versus 614 GB/s on a MacBook Pro with M5 Max, per the comparisons circulating this week. Bandwidth governs decode — how fast the model writes back to you. The MacBook wins there.

Prompt reading (prefill): Compute-bound, and compute needs watts. The N1X's 45–80W laptop envelope is the bottleneck, which is why cold 128K prompts take ~170 seconds.

Model size: This is RTX Spark's real edge. A 128GB machine can load models a MacBook Pro with 48GB cannot, and can hand 96GB to the GPU. If your model fits in the MacBook's memory, the MacBook is competitive or faster. If it doesn't, the MacBook can't run it at all.

Ecosystem and price: The MacBook has mature local-AI tooling and years of developer habit. RTX Spark has the fresh Windows stack, CUDA, NVIDIA's agent ecosystem — plus Windows games, which the MacBook doesn't match. If your work is local AI with big models, RTX Spark is the stronger machine. If it is general computing with occasional AI, the MacBook's polish and proven software still count for a lot. Either way, these laptops exist to replace cloud models — see our Muse Wonder vs GPT-6 vs Claude comparison, our GPT-6 guide, and our Gemini 4 Argon coverage.

Who Should Buy an RTX Spark Laptop — and Who Should Skip

Buy it if you:

• Run local AI agents or code with large models daily, and want them off the cloud bill. This is the machine's reason to exist.
• Need 48GB+ of unified memory for models that don't fit in a normal laptop GPU. The 48GB configuration is the floor reviewers recommend for serious local AI; 128GB only if your models exceed ~30GB.
• Are a creator who lives in Photoshop, Premiere, or Blender and wants the Adobe speedups plus a real GPU.
• Are a gamer who wants AAA games in a thin laptop and already lives in the DLSS ecosystem.

Skip it (for now) if you:

• Mostly browse, write, stream, and video-call. A $2,599+ machine is vast overkill; a regular productivity laptop does all of that for far less.
• Need guaranteed all-day battery life away from a plug. No verified battery numbers exist yet.
• Depend on a specific Windows app with no Arm build and no emulation path. Check your must-have software before buying.
• Run sustained double-precision compute. The StorageReview thermal failure is a warning for that workload specifically.
• Live outside the launch regions. There is no confirmed Pakistan pricing or availability.

Honest Caveats

Every launch story needs this section, and this one needs it more than most, because the marketing is loud.

• Most performance numbers are still vendor claims. The "up to 2x" Adobe figure, the "200x faster" decision-model figure, the "1 petaflop" figure, and the "5x local LLM downloads" figure all come from NVIDIA, with no published methodology. The independent numbers in this article are the exception, not the rule.
• The review sample is small. One full lab review, two hands-on reviews, and a handful of video tests. Prices, thermals, and battery life will get clearer as more units reach testers.
• The thermal failure may be a pattern. Two StorageReview units, same failure, same workload. Watch for follow-up before running sustained heavy compute.
• Prices disagree between sources. The ASUS P16 128GB is reported at both $6,499 and $6,999.99. HP's shipping dates differ between reports. Confirm with the seller.
• No Pakistan pricing or availability. Nothing in this article establishes what these laptops will cost or when they will reach Pakistan. Do not budget from US prices alone.
• Windows on Arm is much better than it was — and still not perfect. The last few percent of software compatibility is where early adopters get surprised.

Conclusion

RTX Spark is the most interesting PC launch in years — not because of the chip alone, but because it arrives with the software to justify it. The Nemotron 3.5 Bolt weights are live, the agent tools are ready, the developer stack is real, and the first independent numbers are genuinely good: a 35B model at 89 tokens per second, AAA games above 100 FPS, a 120B model running on a laptop at all.

But launch day is not the whole story. The thermal warning, the 170-second cold-start wait, the missing battery numbers, and the thin review sample all say the same thing: a strong first wave with rough edges, not a finished victory. If you buy today, buy the 48GB configuration as your floor, keep the charger nearby for AI work, and check your must-have software for Arm support first.

Prices and availability in this article reflect US reports as of October 16, 2026, and will change. Vendor performance claims are labeled as vendor claims throughout.