ASUS starts selling the ProArt GR1X, a $4,499 mini PC built to run AI at home
The palm-sized Windows desktop uses Nvidia's new RTX Spark chip to run big AI models without the cloud, but it costs as much as a used car and is really meant for developers.
- ASUS put the ProArt GR1X on sale in the US on November 18: a 6-inch-square Windows desktop with Nvidia's RTX Spark chip1,2,5.
- It runs AI models on the box itself instead of in a company's data center, so your data stays home and there's no AI subscription1,11.
- It starts at $4,499 with 64GB of memory; the 128GB model, needed for the biggest models, has no US price yet3,6.
- Reviewers like how fast and cool it runs, but it's a tool for AI developers and small studios, not a family PC4,5.
Should you care?
2/5 Only if you build things with AI
How hot right now
HotLaunch-day reviews and news from several PC hardware sites since November 18, little mainstream coverage
Fire meter checked Nov 19, 2026
Hype check
ASUS and Nvidia pitch it as an always-on home for AI helpers1, but at this price it changes things for a small group of developers, not for everyday computer users.
We'll check back on Feb 17, 2027: did it live up to the hype?
What it costs where you live
Who is this for?
- ✓Developers building AI agentsFor youIt runs large models and Nvidia's CUDA tools on Windows, in a box small enough to leave running on a desk1,11.
- ~Small studios with private client workMaybeKeeping AI work on your own machine helps with confidentiality, but the 128GB model's US price isn't out yet6.
- ✕People tired of paying for ChatGPTSkip itAt $4,499 and up, it would take many years of chatbot subscriptions to break even, and you'd have to set it all up yourself3.
- ✕PC gamersSkip itGames mostly run through a translation layer on this kind of chip, and a regular gaming PC costs far less14.
- ✕Anyone shopping for a regular desktopSkip itFor email, browsing and photos, a normal mini PC does the job for a fraction of the price.
A tiny, very expensive computer that runs AI by itself
ASUS's new ProArt GR1X is a small desktop that can run big AI models on your desk instead of on the internet. It starts at $4,499 and is meant for people who build things with AI.
ASUS has started selling the ProArt GR1X in the US1,5. It's a square box about 6 inches wide and 2 inches tall, weighing a little over 3 pounds2. Inside is Nvidia's new RTX Spark chip, which is built mainly for one job: running artificial intelligence on the computer itself1.
What "AI on your own computer" means
When you ask ChatGPT something, your question travels over the internet to huge computers owned by a company. They work out the answer and send it back. Free or paid, your words are handled on the company's machines.
The GR1X is built to do that work right on your desk. ASUS says the version with the most memory can run AI models with 120 billion "parametersJargon busterThe numbers an AI model learned during training; more parameters usually means a bigger, more capable model that needs more memory."1, a size class that includes openly released models such as OpenAI's GPT-OSS 120B10. ASUS pitches it as a machine you leave on day and night, running AI "agents"1: helpers that carry out multi-step jobs for you, such as tracking down a bug in some code9.
The good and the not-so-good
The upside: your documents and questions don't leave your office, there's no monthly fee for the AI itself, and it keeps working if the internet goes down.
The catch is the price. The cheapest version, with 64GB of memory and 1TB of storage, costs $4,4993,4. The 128GB version, which you need for the biggest models, doesn't have a US price yet6. The models that fit at home are also smaller than the biggest ones companies run in their data centers, and getting them going takes some know-how. And it uses real electricity: its chip can use up to 140 watts when working hard6, about as much as two old-style 75-watt light bulbs, on a machine ASUS expects you to leave running1.
Should you care?
Probably not, unless you build things with AI. It's made for developers and small studios that want powerful, private AI on hand. For everyone else, the chatbot app you already use is far cheaper.
What should you do?
The desktop version of Nvidia's Windows bet, and a quick price fight
The GR1X runs the same RTX Spark chip as ASUS's ProArt laptops, but with more power and cooling. Rivals from MSI and Microsoft are already undercutting or matching it.
RTX Spark is Nvidia's Arm chip for Windows, unveiled at Computex in June with ASUS, Dell, HP, Lenovo, Microsoft and MSI signed up as launch partners9,11. Nvidia has tried Windows on ArmJargon busterThe version of Windows for phone-style Arm chips instead of the usual Intel or AMD ones. before: the Nvidia-powered Surface RT ended in a $900 million write-off for Microsoft in 201311. It's a Windows sibling of the chip in Nvidia's DGX Spark, which runs Linux10,13. The GR1X is ASUS's desktop take, next to its ProArt P16 and P14 laptops1.
Same chip, more breathing room
In laptops, RTX Spark is held to 80 watts. In the GR1X it can use up to 1407. That shows: StorageReview found the extra headroom let it beat RTX Spark laptops on both temperature and speed, and put it at the top of the RTX Spark mini PCs it has tested4. Windows Central also praised how cool and quiet it is, but marked it down for the price and for having no USB-A port or SD card reader5. Nvidia says the graphics are in the same league as a laptop GeForce RTX 50709, which held up in Engadget's testing of the P16 laptop14.
The price fight started on day one
MSI cut its EdgeMesa N AI+, with the same chip, to $4,399 for 64GB and 1TB, $100 under ASUS, and its 128GB model to $7,4998. StorageReview measured the two within about 2% of each other, with ASUS running cooler and offering two SSD slots to MSI's one4. Microsoft's Surface RTX Spark Dev Box comes with 128GB, 2TB and Windows 11 Pro for $6,00012. Nvidia's own 128GB DGX Spark now costs $6,950 after memory prices soared, and AMD's rival Gorgon Halo chips go up to 192GB13.
That makes the missing 128GB US price the number to watch.
The Windows on Arm trade-off
This is not a regular Intel or AMD PC. Most games and some older apps run through Microsoft's PrismJargon busterMicrosoft's translation layer that lets regular Windows apps run on Arm chips, a bit slower. translation layer, which costs speed and can break things14. Engadget's reviewer said he'd wait at least a year before suggesting RTX Spark to ordinary buyers, gamers included14. For AI developers who live in Nvidia's software tools, that matters much less.
By the numbers
How we got here
GB10-class silicon in 1.15 liters: what the spec sheet does and doesn't say
A 20-core Arm CPU and a 6144-core Blackwell GPU share up to 128GB of LPDDR5X over NVLink-C2CJargon busterNvidia's fast link that joins the processor and graphics chip inside one package., at a 140W limit. The 120B claim holds for the 128GB model only.
The chip
The GR1X ships only with the full RTX Spark part, which Nvidia calls N1X: the same silicon family as the GB10 in DGX Spark7,10,13. Engadget reports a cut-down variant with 18 CPU and 5,120 GPU cores for cheaper PCs16.
| Block | Detail |
|---|---|
| CPU | 10x Cortex-X925 at up to 4.0GHzProcessor speed4.0GHz is how fast the chip's clock ticks. It's only comparable within the same chip family; newer chips do more per tick. + 10x Cortex-A725 at up to 2.85GHzProcessor speed2.85GHz is how fast the chip's clock ticks. It's only comparable within the same chip family; newer chips do more per tick., co-designed with MediaTek10 |
| Cache | up to 2MB L2 per X925, 512KB per A725, 16MB L3, 16MB system cache10 |
| GPU | Blackwell, 6144 CUDAJargon busterNvidia's software platform for running AI and other heavy math on its graphics chips. cores, fifth-gen Tensor cores, hardware ray tracing10,16 |
| CPU-GPU link | NVLink-C2CJargon busterNvidia's fast link that joins the processor and graphics chip inside one package., up to 600GBStorage · GoodSpace for your apps, photos and files: 600GB. Plenty for most people./s bidirectional10 |
| AI compute | up to 1 PFLOPJargon busterA petaflop: a thousand trillion calculations per second. at FP4Jargon busterA very compact number format used to run AI models faster; chip makers quote their biggest speed figures in it. (vendor figure)1,10 |
| Memory | 64GBStorage · Weak spotSpace for your apps, photos and files: 64GB. Fills up fast; you'll be deleting things. or 128GBStorage · Weak spotSpace for your apps, photos and files: 128GB. Fills up fast; you'll be deleting things. LPDDR5X-9400, soldered, one shared pool4,7; 256-bit bus, 300GBStorage · OKSpace for your apps, photos and files: 300GB. Fine for most, but photos and games fill it up./s17 |
Will 120B actually fit?
GPT-OSS 120B takes about 80GBStorage · Weak spotSpace for your apps, photos and files: 80GB. Fills up fast; you'll be deleting things.10, so only the 128GB model can hold it with room for the OS and a context window. The 64GB box is realistically for models in the 26 to 35 billion parameter range, which Nvidia argues are now good enough for private agents13. StorageReview measured both N1X desktops serving Qwen3.8 27B to eight concurrent requests at about 40 tokens per second15. We found no published test of a 120B model on a GR1X yet, and no long-run test of ASUS's 24/7 claim.
Memory bandwidth is the real limit. Nvidia's RTX Spark porting guide gives the chip a 256-bit memory interface at 300GB/s17, which matches the math for LPDDR5X-9400 (9,400 MT/s across 256 bits is about 300GB/s). That's roughly 10% more than the DGX Spark's 273GBStorage · OKSpace for your apps, photos and files: 273GB. Fine for most, but photos and games fill it up./s13, but far below the roughly 768GBStorage · GoodSpace for your apps, photos and files: 768GB. Plenty for most people./s Android Authority cites for discrete GDDR cards10. Bandwidth, not FLOPS, usually sets token speed on big models.
Thermals and power
The dual-fan cooler is sized for 140W sustained4,6. In a 30-minute Cinebench 2026 loop, StorageReview logged an 83.3C average and 88.7C peak, against about 97C in two RTX Spark laptops4. Power comes from a 240W USB-CPorts · GoodUSB-C: the oval plug that most phones, laptops and chargers now share. adapter (USB PD 3.1 EPR) plugged into one of the four 20Gbps Type-C ports; the other three carry DisplayPort Alt Mode7.
The software side
Nvidia and Microsoft jointly built sandboxes for running agents such as OpenClaw and Hermes Agent, alongside Nvidia's CUDA software11. Unlike DGX Spark, which has ConnectX-7 networking for clustering up to four boxes13, the GR1X's port list stops at 10GbEJargon busterWired network port that runs at 10 gigabits per second, ten times regular home Ethernet.7, so scaling past 128GB means a different machine.
Still unknown: the US price of the 128GB/2TB model6, and official prices outside the US, which ASUS says vary by region1.
Spec sheet
Hover or tap anything underlined to see what it means.
- Made by
- ASUS, with Nvidia's RTX Spark platform1
- Chip
- Nvidia RTX Spark N1X7
- CPU
- 20-coreProcessor cores · Great20 cores: the chip can work on 20 things at once. Helps with heavy work like video editing; everyday browsing barely notices past 6–8. Grace (Arm)1
- GPU
- Blackwell RTX, 6144 CUDAJargon busterNvidia's software platform for running AI and other heavy math on its graphics chips. cores1
- AI performance
- Up to 1 PFLOPJargon busterA petaflop: a thousand trillion calculations per second. FP4Jargon busterA very compact number format used to run AI models faster; chip makers quote their biggest speed figures in it. (ASUS figure)1
- Memory
- 64GBStorage · Weak spotSpace for your apps, photos and files: 64GB. Fills up fast; you'll be deleting things. or 128GB LPDDR5X-9400 unified memoryJargon busterOne pool of memory shared by the main processor and the graphics chip, so big AI models don't have to be split up., soldered4,7
- Storage
- 1TBStorage · GreatSpace for your apps, photos and files: 1TB (1,024GB). Plenty for most people. or 2TBStorage · GreatSpace for your apps, photos and files: 2TB (2,048GB). Plenty for most people. PCIe 4.0 SSD; 2x M.2 2280 slots (PCIe 5.0/4.0)2,4
- Ports
- 4x USB-CPorts · GoodUSB-C: the oval plug that most phones, laptops and chargers now share. 20Gbps, HDMI 2.1, 10GbEJargon busterWired network port that runs at 10 gigabits per second, ten times regular home Ethernet. LAN; no USB-A or SD slot5,7
- Wireless
- Wi-Fi 7Wi-Fi · GreatWi-Fi 7: the newest and fastest Wi-Fi. You only get the full benefit with a Wi-Fi 7 router, but it works with any., Bluetooth 5.4Bluetooth · GoodBluetooth 5.4: for wireless headphones, watches and keyboards. Recent and power-efficient.2
- Power
- 140W TDPJargon busterThe amount of power, in watts, a chip is designed to use and its cooler is built to handle., 240W USB-CPorts · GoodUSB-C: the oval plug that most phones, laptops and chargers now share. adapter6,7
- Size and weight
- 150 x 150 x 51 mm, 1.48 kg2
- Operating system
- Windows 11 Home or Pro (Arm version)4
- US price
- $4,499 for 64GBStorage · Weak spotSpace for your apps, photos and files: 64GB. Fills up fast; you'll be deleting things./1TBStorage · GreatSpace for your apps, photos and files: 1TB (1,024GB). Plenty for most people. with Windows 11 Home (ASUS store); 64GB/2TBStorage · GreatSpace for your apps, photos and files: 2TB (2,048GB). Plenty for most people. $4,999.99 at Best Buy3,4
- 128GB model
- US price not announced yet6
- Other countries
- ASUS says prices and availability vary by region1
Jargon buster
- Local AI
- AI that runs on your own computer instead of on a company's servers over the internet.
- Parameters
- The numbers an AI model learned during training; more parameters usually means a bigger, more capable model that needs more memory.
- AI agent
- An AI program that carries out multi-step tasks on its own, like opening files, running tools and checking its work.
- Unified memory
- One pool of memory shared by the main processor and the graphics chip, so big AI models don't have to be split up.
- Windows on Arm
- The version of Windows for phone-style Arm chips instead of the usual Intel or AMD ones.
- Prism
- Microsoft's translation layer that lets regular Windows apps run on Arm chips, a bit slower.
- TDP
- The amount of power, in watts, a chip is designed to use and its cooler is built to handle.
- FP4
- A very compact number format used to run AI models faster; chip makers quote their biggest speed figures in it.
- PFLOP
- A petaflop: a thousand trillion calculations per second.
- CUDA
- Nvidia's software platform for running AI and other heavy math on its graphics chips.
- NVLink-C2C
- Nvidia's fast link that joins the processor and graphics chip inside one package.
- 10GbE
- Wired network port that runs at 10 gigabits per second, ten times regular home Ethernet.
Sources17 pages · ASUS, ASUS eShop (US), StorageReview and moreShowHide
The small numbers in the text point to these pages. Everything is written in our own words. How we work · Spotted a mistake? Tell us
- 1ASUS Announces Availability of ProArt GR1X · ASUS
- 2ProArt GR1X: Tech Specs · ASUS
- 3Buy ProArt GR1X · ASUS eShop (US)
- 4ASUS ProArt GR1X Review: RTX Spark N1X With Two M.2 2280 Slots in 1.15 Liters · StorageReview
- 5ASUS ProArt GR1X review: The first RTX Spark mini PCs have landed · Windows Central
- 6Asus releases powerful new mini-PC with 128 GB RAM and Nvidia RTX Spark · Notebookcheck
- 7Hands-on with the Asus ProArt GR1X Mini-PC with Nvidia RTX Spark N1X and 128 GB of RAM · Notebookcheck
- 8MSI undercuts Asus with first Nvidia RTX Spark mini-PC release · Notebookcheck
- 9NVIDIA's RTX Spark is an AI "superchip" that will power Windows laptops and desktops · Engadget
- 10NVIDIA's RTX Spark looks like a PC chip, but it's built like a smartphone · Android Authority
- 11Nvidia chases $200B CPU market with AI agent PCs from Microsoft, Dell, and HP · TechCrunch
- 12Microsoft's Surface RTX Spark Dev Box costs an eye-watering $6,000 · Engadget
- 13Nvidia debuts $4,999 DGX Spark with half the RAM and storage, amid memory crunch · The Register
- 14ASUS ProArt P16 review: A sleek showcase for NVIDIA's RTX Spark · Engadget
- 15Best Mini PCs in 2026: Lab-Tested Leaderboard · StorageReview
- 16NVIDIA's RTX Spark is good for Windows, even if you don't need local AI · Engadget
- 17RTX Spark Porting Guide: System Overview · Nvidia
Read next
Computers · ★ Saved
A free Googlebook update gives back up to 6GB of RAM and fixes about 30 bugs
Googlebooks, Google's new laptops that run on Android, are getting their first big update. It fixes crashes and frees up memory so they feel smoother. Regular Chromebooks aren't affected.
Computers · ★ Saved
Apple's M6 MacBook Pro and iMac, OLED iPad mini and first touchscreen Mac are in stores
Apple's new MacBook Pro, iMac and iPad mini went on sale November 6. The Macs look the same but are faster; the iPad mini got a better screen; and for the first time you can buy a Mac with a touchscreen.
Money & business · ★ Saved
Nvidia's sales double to $115.8 billion in its first report since Jensen Huang left
Nvidia, which makes the chips that power most AI, says its sales doubled in a year. It's the first report since its longtime boss left, and it matters to more people than you'd think.