NVIDIA's RTX Spark laptops went on sale October 16, 2026 from ASUS, Dell, HP, Lenovo, Microsoft and MSI, starting at $2,599. They are the first Windows machines built on NVIDIA's N1X "superchip" – 5120 or 6144 Blackwell RTX cores, an 18- or 20-core Grace CPU bare-die CPU block co-designed with MediaTek, and up to 128GB of unified LPDDR5X feeding both.
Every other site will repeat that press release. Here is what they skip: NVIDIA's 1440p/100-fps gaming claim leans on DLSS 5, early independent testing says raw gaming actually trails comparable RTX 5070 laptops, the price per gigabyte of unified memory is dramatically worse than the desktop GB10 box, and the cheapest 24GB SKU hides most of the AI story you are paying the "superchip" premium for. That is what this guide covers.
Source note: specs, prices and availability come from NVIDIA's launch announcement, the RTX Spark product page and OEM pre-order listings (Oct 7–16, 2026). Benchmark numbers quoted are first reviewing-party measurements, not an independent VelsTech Solutions test – testers also used pre-release samples on early drivers, so treat them as direction, not verdicts.
Background: RTX Spark is the laptop form of a platform we track closely. Read our DGX Spark vs RTX Spark vs Ryzen AI Halo showdown for the desktop picture, then come back – this page is only about the laptops that started shipping this week.
The short version: RTX Spark in 60 seconds
- What: the first Windows-on-Arm laptops with NVIDIA silicon: the N1X superchip (Blackwell RTX GPU + Grace CPU) that shares one pool of up to 128GB of unified LPDDR5X memory between CPU and GPU.
- Price / date: on sale globally since October 16, 2026 from $2,599 (cheapest SKU: 18 cores, 5120 GPU cores, 24GB memory). Top 128GB configurations reach roughly $5,900–$7,000.
- Specs: up to 1 petaflop of FP4 AI compute, about 300 GB/s memory bandwidth on early samples (own measurement by testers; the desktop GB10 plate sits at 273 GB/s), DLSS 5, Reflex 2 with Frame Warp, PCIe Gen 5, HDMI 2.1b + DisplayPort 2.1b, 9th-gen encode / 6th-gen decode engines.
- Headline claim: NVIDIA says it will run frontier-scale local models and agents, render 90GB+ 3D scenes, and play AAA titles at 1440p above 100 fps (that last one assumes DLSS 5).
- Catch: memory-busy testing concurs on local AI, but independent gaming tests so far trail RTX 5070 laptops at similar power, and you pay roughly double the desktop GB10 Spark's per-GB price.
- Who: creators and developers who want CUDA + big unified memory in a portable machine with one machine for local agents, 3D and video – not pure AAA game buyers.
What the RTX Spark actually is
Think DGX Spark's chip shrunk into a laptop chassis. At the heart of every RTX Spark is the N1X superchip: a Blackwell-generation RTX GPU bolted to a Grace-class CPU on one die. The CPU is a custom Arm design NVIDIA co-developed with MediaTek for slim chassis power budgets (18 or 20 cores depending on the tier); the GPU brings NVIDIA's 4th-gen RT cores, 5th-gen Tensor cores and DLSS 5. Because CPU and GPU share the memory pool, the same 24–128GB of LPDDR5X serves your agent context window and your render buffers.
Why now: laptop workloads drifted toward AI anyway (Copilot-class features, local agents, TensorRT-only effects), Windows on Arm finally has decent developer tooling, and the desktop DGX Spark showed people will pay for a quiet 128GB AI box. RTX Spark productizes the same recipe for the machine you actually carry, with NVIDIA's entire consumer stack (DLSS, Reflex, Broadcast, Studio drivers, CUDA) bundled.
Specs at a glance
| Spec | N1X (20-core tier) | N1X (18-core tier) |
|---|---|---|
| GPU | 6144-core Blackwell RTX | 5120-core Blackwell RTX |
| CPU | 20-core Grace-class (10P+10E) | 18-core Grace-class |
| Unified memory | 64–128GB LPDDR5X | 24–64GB LPDDR5X |
| Laptop power range | 45–80W configurable | 45–80W configurable |
| AI compute | Up to 1 petaflop FP4 | Scaled-down FP4 |
| Memory bandwidth | ~300 GB/s (early samples, tester-measured) | Lower with smaller config |
| Uplift tech | DLSS 5, Reflex 2 + Frame Warp, RT cores gen 4, Tensor cores gen 5 | Same feature set |
| Video | 9th-gen NVENC, 6th-gen NVDEC | Same |
| Display / IO | HDMI 2.1b, DisplayPort 2.1b, PCIe Gen 5 | Same |
| OS | Windows 11 (Arm) | Windows 11 (Arm) |
The honest row is power. NVIDIA lists a 140W ceiling for the biggest N1X configuration, but the laptop SKUs ship configurable between 45W and 80W – that is less sustained GPU power than the 90–100W mobile RTX 5070s it is measured against in reviews. Benchmark wins below were recorded at these laptop power targets; expect the desktop Dev Box to be the variant that actually sustains 140W.
The launch lineup: who sells what
Nine models were announced for day one. Prices below come from OEM pre-order listings and launch-day coverage; the top of the range is a 128GB machine, and every OEM sells the cheaper 24GB base config too.
| Model | Screen | Price range |
|---|---|---|
| Microsoft Surface Laptop Ultra | 15-inch 3270×2180 touchscreen | $2,599 – $5,900 |
| ASUS ProArt P14 / P16 | 14-inch or 16-inch 3K/4K OLED | $2,599 – $6,500 |
| Dell XPS 16 (Creator Edition) | 16-inch 3200×2000 OLED | $3,800+ |
| HP OmniBook Ultra / OmniBook X 14 | 16-inch or 14-inch 3K OLED | $3,200 – $5,000+ |
| Lenovo Yoga Pro 9n | 16-inch OLED convertible | $3,000 – $4,500 |
| MSI Prestige N16 Flip AI+ / Prestige N16 | 16-inch | $2,599 – $4,500 |
- Ranges are memory tiers, not SKUs. The steps between $2,999 → $5,900 are almost always memory (24 → 32 → 48 → 64 → 128GB) plus storage. Same logic as our DGX Spark 64GB guide: check which RAM tier a listing quotes before comparing two models.
- The Dev Box matters. Microsoft also sells a Surface RTX Spark Dev Box desktop from $5,999, which pairs the same chip with more sustained power – useful if you want the chip without the thermals trade of a thin chassis.
- Regional availability differs. NVIDIA's list is global, but which OEM actually stocks your country on day one is not guaranteed; launch coverage flagged that pricing and configs vary by region.
Day-one software: what "AI PC" buys you
- Agents: guided installers for GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy agents (no command line); NVIDIA's Nemotron 3.5 Bolt 74B-A9B 24GB model is optimized for local tool-calling; the OpenShell secure runtime and Microsoft Execution Containers sandbox agent permissions.
- Creative: Photoshop and Premiere ship native Arm beta support with TensorRT AI-accelerated effects at launch, Substance 3D coming later; NVIDIA Broadcast 2.3 cleans old recordings; ComfyUI nodes turn 15 fps AI video into 60 fps with HDR without per-clip cloud costs.
- Developers: the first PyTorch-CUDA build for Windows on Arm, native-by-default Python, a native Conda installer and 97% compatibility with the top 15,000 Python packages, plus WSL for Linux-side tooling.
- Games: Cyberpunk 2077, Fortnite, Gears of War: E-Day, Marvel Rivals, World of Warcraft and NBA 2K27 lead day-one support; NBA 2K27's DLSS 5 "3D-Guided Neural Rendering" mode and a new 2× faster upscaler model (up to +1.4× frames) are the showpieces.
One more piece: NVIDIA PAIR, which lets a Spark machine serve AI inference to the Windows, Linux or macOS laptops around your network. On a desk with a small PC per family member or team, that quietly turns the laptop into the household's local inference node – relevant if you mirror our local AI getting-started workflow.
Independent testing so far
Three outlets posted early numbers before and around launch. Read them as first-party measurements on pre-release samples, not settled verdicts:
- Local AI: reviewers running the Surface Laptop Ultra describe it as the strongest AI laptop tested yet – a full 35B-class model runs in the 64GB GPU-visible memory within a laptop, and agent workflows behave like a desktop DGX Spark.
- CPU: early Cinebench/Geekbench-style results put the 20-core Grace CPU ahead of current x86 flagship laptop chips (AMD's Strix Halo, Intel's Panther Lake) in multicore work, while trailing Apple's top M5 Pro/Max cores.
- Games: the weak spot. Reviewers repeatedly note gaming trails 90–100W RTX 5070-class x86 laptops at equal-to-higher price; NVIDIA's 1440p/100+ fps claim is DLSS 5-assisted, not raw. If smooth AAA raw performance is your metric, wait for fuller driver-matured testing.
Pricing: what it actually costs
| Tier | Price | What you get | Effective per-GB of unified memory |
|---|---|---|---|
| Base (all OEMs) | $2,599 | 18 cores, 5120 GPU cores, 24GB | ~$108/GB |
| Middle (Surface/OmniBook/Yoga) | $3,000–$4,000 | 32–48GB tier | ~$85–95/GB |
| Big (64GB tier) | $4,000–$5,000 | 64GB unified, usually top-OEM screen | ~$65–75/GB |
| Max (128GB tier) | $5,900–$7,000 | 20 cores, 6144 GPU cores, 128GB | ~$46–55/GB |
| Desktop alternative | $5,999 (Dev Box) / $4,999 (DGX Spark 64GB) | Higher sustained power | ~$78/GB (64GB desktop) |
- The cliff: the 128GB laptop configuration is the honest price of this platform. Every RAM tier below it exists mostly to slap a sub-$3k sticker in ads, but can't hold a 70B model or a serious context window.
- Concrete: a $2,599 base holds 24GB – enough for 8–13B agent models or 27B at heavy quant, the VRAM calculator territory – while a 128GB tier holds 100B-class models the way DGX Spark did. Same "RAMpocalypse" math we documented for the 64GB desktop in October applies directly.
- Detail: the $5,999 Dev Box with up to 140W sustained vs the $2,599 laptop's 45–80W are the same platform with two very different thermal envelopes – decide by workload hotness, not by name.
How it compares
| Platform | RTX Spark laptop | DGX Spark 64/128GB desktop | Ryzen AI Halo mini PC (192GB) |
|---|---|---|---|
| Chip | N1X (Grace CPU + Blackwell) | GB10 (same family, higher sustained power) | Ryzen AI Max+ 495 |
| Max unified memory | 128GB | 64/128GB | 192GB (160GB VRAM-allocatable) |
| Compute ceiling | ~1 PFLOP FP4 | ~1 PFLOP FP4 | No CUDA |
| Portability | Laptop, 45–80W | Desktop brick, quiet | Desktop brick, larger |
| CUDA | Yes | Yes | No (ROCm/Vulkan paths) |
| Price from | $2,599 | $4,999 | $6,799 |
The trade-off story is old: the laptop is the cheapest ticket into the ecosystem (CUDA + 128GB of unified memory), but the desktop wins on price per gigabyte – and the Halo mini PC buys the biggest total memory of any APU setup, which is why our Evo-X5 Pro explainer reads it the same way.
The India angle
- Base $2,599 ≈ ₹2.3 lakh before shipping, GST and duty – realistically ₹3–3.5 lakh landed via gray-import routes.
- OEM India stores (Lenovo India, ASUS India, HP India) have not confirmed local stock or India pricing as of Oct 17; NVIDIA's "available globally" line references OEM websites and major retailers, who list by region.
- Compare: the native 128GB DGX Spark desktop already lands around ₹5–5.5 lakh via import, and used 24GB RTX 3090 setups still dominate sub-₹60k DIY builds. At current memory prices a 128GB RTX Spark tier might put you somewhere around ₹5.5–7 lakh total landed cost.
Who should buy it – and who should skip it
- Buy if: you need one machine to (a) run local CUDA-accelerated agents with, (b) shuttle into Premiere/Photoshop betas, and (c) edit or render outside cloud SLAs – and the 64GB+ tier is essential.
- Skip if: your goal is raw AAA gaming (RTX 5070 x86 rivals do it cheaper today), your agent models stay under 32B (any 16–24GB laptop covers that), or you want maximum memory per rupee (the Halo minis still double it).
- Consider instead: the $4,999 DGX Spark 64GB desktop for max GPU-hours per dollar, our sub-50k GPU guide with x86 laptops for occasional CUDA tasks, and the Evo-X5 Pro explainer if 100GB+ on an APU is the problem you are actually solving.
What this means if you run local AI
From someone who benchmarks open models on a 12GB RX 6800M weekly: RTX Spark matches the DC's "one die does inference" pitch, but compressed. Agent-in-context is genuinely new for Windows laptops: a 35B model with long context in the bag is currently impossible elsewhere. But the whole platform rides on memory cost. The 128GB tier decides whether this becomes a serious ~$5k workshop tool or a $3k retail mirage, and pre-release gaming numbers suggest the raw-silicon pitch stays blurred until drivers mature. Buy on memory tiers and DLSS capability, not the $2,599 ad line.
Bottom line
RTX Spark is not a laptop with a laptop-killer GPU – it is the GPU-killer ecosystem coming to one (battery-powered) bag: CUDA, DLSS 5, Broadcast and TensorRT, driven by a memory-bandwidth hungry chip, at prices that only make sense from 64GB up. Local AI buyers on a budget or desk-bound CUDA users may find the desktops still better value; travelers with 90GB renders and 120B-ish model detours finally have a machine to buy today. The N1X at $2,599 is a demo-tier product; the platform genuinely arrives at the 128GB tier.
Sources
- NVIDIA press release: "NVIDIA RTX Spark PCs Available Today — A New Beginning for Windows PCs" (Oct 16, 2026)
- NVIDIA RTX Spark product and spec page (N1X configurations, DLSS 5, Reflex 2, HDMI 2.1b/DP 2.1b)
- Tom's Hardware: Surface Laptop Ultra launch, $2,599 start, Dev Box from $5,999 (Oct 7, 2026)
- Tom's Hardware hands-on: "impresses in local AI but gaming isn't its strong suit"
- StorageReview review: 35B models in 64GB of GPU-visible memory, CPU render leads (Oct 16, 2026)
- PCWorld: every launch RTX Spark laptop and its price range, up to $7,000 (Oct 8, 2026)
- Windows Central: RTX Spark week of testing, benchmarks (Oct 16, 2026)
- Our DGX Spark vs RTX Spark vs Ryzen AI Halo showdown and the 64GB DGX Spark explainer
Content and image notice: this is VelsTech Solutions' own summary of the cited sources – no press-release text is reproduced wholesale. Product images are © NVIDIA Corporation, used here for editorial coverage with credit (see each figure caption); NVIDIA, RTX, DLSS, Reflex, G-SYNC, CUDA, DGX, Nemotron and related marks are trademarks of NVIDIA Corporation. Other product names are trademarks of their respective owners.
FAQ
What is an RTX Spark laptop?
A Windows-on-Arm laptop built around NVIDIA's N1X superchip, which combines a Blackwell RTX GPU, an 18- or 20-core Grace-class CPU and up to 128GB of shared LPDDR5X unified memory. Nine models launched globally on October 16, 2026, starting at $2,599.
How much does an RTX Spark laptop cost, and where can I buy it?
From $2,599 for the 24GB base tier to roughly $5,900–7,000 for 128GB top models, sold via ASUS, Dell, HP, Lenovo, Microsoft and MSI direct stores and major retailers. Pricing and configurations vary by region; Microsoft also sells a $5,999 Surface RTX Spark Dev Box desktop.
What exactly is the N1X superchip?
A single-die NVIDIA platform chip pairing a Blackwell RTX GPU (5120 or 6144 CUDA cores) with a custom Arm Grace-class CPU co-designed with MediaTek. It delivers up to 1 petaflop of FP4 AI compute, runs 45–80W in laptops (140W ceiling on the top tier), and supports DLSS 5, Reflex 2 and PCIe Gen 5.
Can RTX Spark run big local AI models like the DGX Spark?
Yes – reviewers measured full 35B-class models running inside a Surface Laptop Ultra's 64GB of GPU-visible memory, and the 128GB tier covers DGX Spark-class ceilings (~100B parameters on one machine). Just remember you pay for that memory in the price tiers, and lower SKUs (24GB) are for 8–13B agents.
Is RTX Spark good for gaming?
Feature-rich but not the performance class. Early independent testing shows raw gaming trails 90–100W RTX 5070 x86 laptops, and NVIDIA's 1440p/100+ fps headline figures lean on DLSS 5. It is a first-generation platform; expect driver-led improvement over time.
Should I buy an RTX Spark laptop in India?
Only if a 64–128GB tier at 64GB-plus price fits your workflow. No India pricing or stock dates were confirmed at launch. Base $2,599 maps to roughly ₹3–3.5 lakh landed via gray import, which for budget tier is much worse value than a 27B-capable gaming laptop.