NVIDIA RTX Spark Desktops Are Here: A $4,400 Box That Runs 120B-Parameter AI Models at Home
NVIDIA spent most of 2026 telling developers that the future of AI agents is local - models running on your own machine, not rented by the token from a cloud API.
On November 18, 2026, that pitch became something you can actually buy. RTX Spark compact desktops from ASUS, Dell and MSI went on sale in the US, starting at $4,399.99, and they are built around one idea: a small box on your desk that can run a 120-billion-parameter language model without sending your data anywhere.
This is not another "AI PC" sticker on a laptop. The chip inside is a superchip - a 20-core Grace CPU paired with a Blackwell GPU carrying 6,144 CUDA cores, sharing up to 128GB of unified memory, rated at up to 1 petaflop of AI performance. NVIDIA describes these machines as "systems for always-on agents."
Whether that justifies the price depends entirely on what you run on them. Let's look at what they are, who they are for, and where the honest limits sit.
The short version
RTX Spark desktops are compact Windows machines built to run large AI models and always-on agents locally. The first three on sale are the ASUS ProArt GR1X, the MSI EdgeMesa N AI+ and the Dell Creator Edition Desktop, with Acer, GIGABYTE and HP mini PCs still to come, plus Microsoft's Surface RTX Spark Dev Box.
The headline numbers come from NVIDIA: up to 128GB of unified LPDDR5X memory, up to 1 petaflop of AI performance (FP4), and enough headroom to run a 120B-parameter model on-device. The cheapest listed model is the MSI EdgeMesa N AI+ at $4,399.99 with 64GB of memory and a 1TB SSD.
These are NVIDIA's claims - no independent benchmarks of the desktop hardware have been published yet.
What is inside
Every RTX Spark desktop uses the same chip, the RTX Spark N1X superchip, at full desktop power - the ASUS spec sheet lists 140W sustained, where the laptops run the same silicon in a 45-80W envelope.
The parts that matter:
- 20-core NVIDIA Grace CPU (Arm-based, co-designed with MediaTek), linked to the GPU by
NVLink-C2Cat 600 GB/s. - Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores - roughly RTX 5070 laptop-class graphics.
- Up to 128GB of unified LPDDR5X memory, shared between CPU and GPU, so the graphics unit can use nearly the whole pool instead of being capped by a separate VRAM limit.
- Up to 1 petaflop of AI performance at FP4 precision - note the FP4 qualifier; it is not directly comparable to the INT8 TOPS figures quoted for NPUs.
- 10-gigabit Ethernet for moving large model files, datasets and media to networked storage.
The unified memory is the real story. On a normal PC, a model has to fit in the GPU's VRAM - typically 8-24GB on consumer cards - while the rest of the system RAM sits idle. Here there is one pool of up to 128GB. That is what lets a 120B-parameter model load at all on a machine this size.
NVIDIA also claims the platform can render 3D scenes larger than 90GB in real time, edit 4:2:2 video in real time, and hold a million-token context window for certain 120B models. Again: vendor claims, not independent tests.
One important caveat on speed. Fitting a 120B model into memory prevents out-of-memory crashes, but how fast it generates text is bound by memory bandwidth, which NVIDIA has not published a firm figure for - coverage puts it roughly in the 273-307 GB/s range, well below an Apple M5 Max at up to 614 GB/s. An early llama.cpp-based test of an RTX Spark laptop reported about 65 tokens per second on gpt-oss-120b. Usable, not blazing.
Capacity is the selling point; raw generation speed is the open question.
Why local AI matters - the two real reasons
It changes the cost math. Every prompt you send to a hosted frontier model is metered. For a developer who runs agents all day - code review, triage, research, drafting - the per-token bill adds up, and it never stops. A $4,400 box is a one-off purchase. If you run a capable open model locally for a year or two, the machine can pay for itself against API spend, depending on volume. That is the calculation NVIDIA is pitching, and for heavy local users it is a real one - but only if the local model is actually good enough for your work, which brings us to the second reason.
Your data stays on your desk. A model running through llama.cpp or Ollama on this machine never sends your prompts to a model provider. For codebases under NDA, unreleased product plans, client data or anything you simply do not want leaving the building, that is not a nice-to-have - it is the whole point. Running locally makes private inference possible. It does not make the machine private by itself: the moment you connect a cloud agent, a sync service or a telemetry-heavy app, data leaves again. Local inference is a tool for privacy, not a guarantee of it.
The four boxes: ASUS, MSI, Dell and Microsoft
- MSI EdgeMesa N AI+ - 64GB, 1TB SSD - $4,399.99 - On sale
- ASUS ProArt GR1X - 64GB, 1TB SSD - $4,499 - On sale
- ASUS ProArt GR1X - 64GB, 2TB SSD - $4,999.99 - On sale (Best Buy)
- MSI EdgeMesa N AI+ - 128GB, 4TB SSD - $7,499.99 - On sale
- Dell Creator Edition Desktop - 128GB - Not published - Orders opened Nov 18
- Microsoft Surface RTX Spark Dev Box - 128GB - $5,999.99 - Pre-order since Oct 7, shipping in November
A note on prices, because launch-day listings disagreed: some outlets reported the MSI 64GB model at $4,699 and the 128GB model at $7,999, and both MSI configurations have shown thin or no stock. The figures above are the lowest listed prices found on the makers' own US stores on November 18. Treat any price as "check the store before you buy" - they move, and they vary by retailer. Dell has not published a price at all for its Creator Edition Desktop, and no prices have been announced outside the US yet.
ASUS ProArt GR1X is the most fully specified of the launch machines - a 1.1-litre mini PC aimed at creators, with the full 140W chip and up to 300 GB/s memory bandwidth on its spec sheet. If you want the best-documented box, this is the one to watch.
MSI EdgeMesa N AI+ is the cheapest way in, and the only one with a published 128GB price - but that jump from 64GB to 128GB costs about $3,100, which is the single most expensive decision in this whole launch. Buy the memory you will actually need; you cannot upgrade it later.
Dell Creator Edition Desktop exists in NVIDIA's announcement and Dell's own post, with orders opened on November 18 and a 128GB configuration - but without a published price, it is hard to recommend until Dell's store shows one.
Microsoft Surface RTX Spark Dev Box at $5,999.99 is the developer-focused box, on pre-order since October 7 and due to ship in November. It is the most expensive 128GB option of the four, and its value is the Surface/Windows integration and the dev-box positioning rather than raw price.
Coming next: compact desktops from Acer, GIGABYTE and HP, plus the DGX Spark HPC Developer Kit - four DGX Spark systems clustered for 512GB of memory and models up to 700 billion parameters - aimed at researchers rather than individual developers.
Who should actually buy one
Developers running local agents all day. If you already live in Ollama, llama.cpp or LM Studio and you are hitting VRAM limits on a consumer GPU, this is the first compact box that removes the limit instead of moving it. The always-on agent angle is real: a machine that sits on your desk running agents overnight, triaging, indexing, drafting - without a cloud bill ticking.
AI researchers and small teams. The DGX Spark cluster kit aside, a single 128GB box is a reasonable shared inference node for a small team prototyping with 120B models, or for fine-tuning experiments that do not need a datacentre.
Creators working with heavy local media. Real-time 4:2:2 video editing and 90GB+ 3D scenes are NVIDIA's claims, and the creator positioning of the ASUS ProArt line fits - but wait for independent creative-workload tests before buying on that basis alone.
Probably not you, if: you mostly chat with Claude or ChatGPT a few times a day (the cloud is cheaper and the models are better); you need a general-purpose PC (this is Windows on Arm - most x86 software still runs through emulation); or you want a gaming machine (NVIDIA claims 1440p AAA gaming above 100 fps with DLSS, but nobody has checked that on the desktop hardware yet).
The honest limits
The price is the product's biggest problem. $4,399 is the cheapest way in, and the configuration most people actually want - 128GB - starts around $6,000-$7,500. The RTX Spark laptops start at $2,599 and include a screen, keyboard and battery. The desktop asks you to pay more and bring your own monitor and peripherals.
Windows on Arm. This is an Arm machine. Most x86 software runs through emulation, and some tools will not run at all or will run slower than you expect. Check your toolchain before you buy, not after.
No independent benchmarks yet. Every performance number in this article is NVIDIA's or an OEM's claim. Memory bandwidth, sustained load behaviour, token-generation speed on retail hardware - all untested by third parties at launch.
Thin stock and moving prices. Launch-day listings already disagreed by hundreds of dollars, and configurations have shown thin or no stock. This is a first-batch product.
The 120B promise has a speed asterisk. Holding a 120B model in memory is not the same as running it fast. Bandwidth-limited generation is the likely reality, and the honest comparison is against a Mac Studio or a cloud API, not against a datacentre GPU.
What to do if you are tempted
- Check your toolchain on Arm first. List the five tools you use daily and confirm each one runs natively or acceptably emulated on Windows on Arm.
- Decide the memory before the model. 64GB runs agents and smaller models comfortably; 120B-class models need the 128GB configuration, and the $3,100 jump is the decision, not the $4,400 sticker.
- Price it against a year of API spend. Add up what you actually paid for hosted inference in the last 12 months. If the box costs more than two years of that, the economics do not work yet.
- Wait for one independent benchmark. A single
llama.cpporvLLMtest on a retail desktop unit will tell you more than every launch slide combined. If you are not in a rush, a week of patience is worth $4,400 of certainty.
Pricing and availability
RTX Spark compact desktops from ASUS, Dell and MSI went on sale in the US on November 18, 2026. Prices start at $4,399.99 (MSI EdgeMesa N AI+, 64GB/1TB). Acer, GIGABYTE and HP systems are still to come, and no prices have been announced outside the US.
The Microsoft Surface RTX Spark Dev Box ($5,999.99, 128GB) has been on pre-order since October 7 and is due to ship in November. RTX Spark laptops, which launched October 16, 2026, start at $2,599.
The bottom line
RTX Spark desktops are the first compact machines that make a serious local-AI workstation a real purchase instead of a datacentre rental - 128GB of unified memory in a 1.1-litre box is a genuine engineering achievement, and for developers who run agents all day, the cost and privacy math can work.
But the price is steep, the memory upgrade is steeper, the platform is Arm, and not a single independent benchmark existed at launch. This is a specialist tool at a specialist price, not a PC for everyone.
Verdict: If you already run local models daily and you are VRAM-limited, the 128GB box is worth a serious look - after you check your toolchain on Arm and see one independent benchmark. Everyone else should watch the first tests and the Acer/GIGABYTE/HP follow-ups before spending $4,400.
Sources: NVIDIA blog, "NVIDIA RTX Spark Compact Desktops Out Now" (Nov 18, 2026); The Threshold, "RTX Spark desktops: prices, specs and where to buy" (prices checked Nov 19, 2026); Codema, "RTX Spark desktops: ASUS, Dell, MSI on sale now"; HWBusters, "RTX Spark Mini PCs Get the Full 140W Chip" (Nov 18, 2026); HardwareZone, "NVIDIA RTX Spark Compact Desktop for AI agents"; Gadget Review, "NVIDIA RTX Spark Compact Desktops Are Here, Starting at $4,499". Prices are US store listings as reported on November 18-19, 2026 and vary by retailer; Dell has not published a price. Performance figures are NVIDIA's claims - no hands-on testing was performed and no independent desktop benchmarks were available at writing.
Comments
No comments yet. Start the discussion.