If you want to buy a DGX Spark, Nvidia's Founders Edition with 128 GB has cost $6,950 in the US since 2 October 2026, according to The Register; it launched at $3,999 and went to $4,699 in February 2026. The 64 GB version, sold only by partners since October 2026, starts at $4,999 according to Nvidia. In total there are eight GB10 systems: the Founders Edition and seven partner variants with the same chip, which differ in SSD, case and support. In Germany, the Founders Edition costs 8,151.50 euros including VAT at the reseller Nelpx (14 November 2026), and partner machines start at 5,699 euros (Geizhals, 12 November 2026). The serious alternatives are a Mac mini, a Mac Studio and a GPU server of your own; which one fits depends not on the price but on the size of the model you want to run permanently.
I ordered the DGX Spark directly from Nvidia in December 2025 and run Qwen 3.8 with 27 billion parameters on it under SGLang, measured at 50.7 tokens per second. Day-to-day experience is in DGX Spark Tested: Worth Buying?. This post is the buying decision: prices with sources, variants, alternatives, successor. Everything on local AI is on my AI page.
What you get when you buy a DGX Spark
The DGX Spark is a 150 by 150 millimeter desktop box weighing 1.2 kilograms, built around the GB10 Grace Blackwell Superchip. Per Nvidia's product page, retrieved 12 November 2026: 20 Arm cores, 128 GB of LPDDR5x unified memory at 273 GB/s, a self-encrypting NVMe drive of up to 4 TB, ConnectX-7 networking at 200 Gbit/s, one 10 gigabit Ethernet port, a 240 watt power supply and a 140 watt TDP for the chip. Nvidia quotes up to 1 PFLOP of FP4 compute. Since October 2026 the platform also comes with 64 GB, but only from the partners.
The number that matters is the memory. CPU and GPU share one pool, and 128 GB holds a single model with 27 to 70 billion parameters plus context without wiring graphics cards together. Nvidia itself lists models up to 200 billion parameters for one device with 128 GB, and up to 100 billion with 64 GB; my range allows for a long context. My system reports around 119 GB as usable. The second figure, 273 GB/s of bandwidth, sets the response speed, where a Mac Studio and dedicated graphics cards beat the Spark.
What you do not get: a training server; the Spark is built for inference and development. The product page names no price and sends buyers to the Nvidia Marketplace and to retail partners. The partner machines come from Acer, ASUS, Dell, Gigabyte, HP, Lenovo and MSI.
DGX Spark price: US list price and prices in Germany
Nvidia raised the Founders Edition list price twice in 2026. In February it went from $3,999 to $4,699, announced in its developer forum and attributed to industry-wide memory supply constraints. On 2 October 2026 it jumped to $6,950, according to The Register, nearly 75 percent above the launch price. The same day, Nvidia's blog announced the 64 GB version, available only from partners, starting at $4,999. Check the current Founders Edition price on the Nvidia Marketplace before you order. For partner machines, Nvidia refers pricing questions to the manufacturers, so take US prices from the OEM's own store.
Prices in Germany: on 14 November 2026 the reseller Nelpx lists the Founders Edition with 4 TB at 6,850 euros net, or 8,151.50 euros including VAT, with one to two weeks' delivery; on the price comparison site Geizhals it started at 8,999 euros on 12 November 2026. In September, Nelpx still asked 5,180 euros net.
For context, what I paid: I ordered the Spark directly from Nvidia in December 2025 for roughly 4,000 to 4,200 euros. That was my price back then, not an offer today; it sits well below today's prices. For your own budget, what counts is your reseller's price on the day.
Founders Edition or partner variant: eight systems
There are eight GB10 systems: the Founders Edition and seven partner variants, all with the same chip and 128 GB of unified memory. The differences are SSD, case and support. Since 23 October 2026, Acer, ASUS, Dell, Gigabyte, HP and MSI also sell a 64 GB version. The table shows the 128 GB versions with prices from Germany.
Geizhals shows prices including German VAT, excluding shipping; resellers change them daily.
My recommendation: SSD size matters more than the brand. A 27B model as an NVFP4 checkpoint takes 21.9 GB on my machine, a 70B model several times that, plus Docker images and further checkpoints. 1 TB works; with 4 TB you never clean up. A 27B model also runs on the 64 GB version; for the upper end of that range with a long context, take the 128 GB. According to Nvidia, hardware support comes from the vendor that sold the system; for a company with procurement, the familiar supplier can decide it.
Three alternatives: Mac mini, Mac Studio, GPU server
The Spark competes less with other GB10 boxes than with three other routes to a local model. Table prices are US entry prices with sources.
The Mac figures come from Apple's specifications and Apple's US store, retrieved 12 November 2026. The Mac mini tops out at 64 GB: a 27B model in 8-bit fits, a 70B model with context does not. The Mac Studio M5 Ultra, with 1.2 TB/s and up to 512 GB, offers more speed and room than the Spark. With 96 GB it costs $5,499, about $1,450 less than the 128 GB Spark at $6,950, but it has less memory than the Spark. 256 or 512 GB require the larger M5 Ultra (36-core CPU, 80-core GPU), which starts at $6,799 before the memory upgrade. In Germany, Apple charges 6,599 and 8,029 euros including VAT. What the Mac can do in detail is in Local AI on a Mac: Mini as AI Server.
The GPU server is the route with the most performance per dollar and the most effort. insiderllm.com: models up to 24 GB run faster on a single graphics card than on any mini PC in this class. Once the model outgrows one card, it gets expensive and loud. A calculation with new parts and electricity is in Your Own AI Server: Hardware and Cost.
Why I bought one and what I measure
My question in December 2025 was not "do I need more compute" but "can I run a 27B model locally and permanently, reachable from all my devices, with no request going through a cloud API". My MacBook Pro M3 Max with 64 GB loads the model, but at 11 tokens per second in LM Studio (8-bit, 27.5 GiB RAM) it was too slow for agent work. Today my AI workforce of fifteen AI employees runs on Claude; the Spark is my test bed for local models.
On the Spark, Qwen 3.8 with 27 billion parameters runs as a 21.9 GB NVFP4 checkpoint under SGLang, following the hasso5703/dgx-spark-qwen38 repo, the official SGLang cookbook recipe since 21 August 2026. I measured 50.7 tokens per second for decode (300 tokens, one stream, from the Mac via Tailscale). For prefill, the setup recipe lists roughly 1,500 tokens per second; on my machine a 125,000-token prompt without cache took 123 seconds, so roughly 1,000 tokens per second. vLLM as an alternative gave 24 to 26 tokens per second. The measurements and the setup are in Run Qwen 3.8 Locally: 27B on a DGX Spark. How Muse Glimmer, Gemma 4, gpt-oss-120b, Nemotron and Mistral Small 4 do on the same machine is covered in Best LLMs for the DGX Spark, Measured.
Two things appear in no spec sheet. First, the unified memory trap: pin more than half of the memory for SGLang (--mem-fraction-static above 0.50) and you risk the host freezing, because CPU and GPU share one pool. Second, the waiting: the first server start takes around 9 minutes because of compilation, then 5 to 7 minutes, and at roughly 2 MB/s at my location a 70 GB download takes hours. Neither argues against buying; both argue for a free first day.
Clustering: what is official and what tinkerers show
If one machine is not enough, you couple several Sparks via the ConnectX-7 ports. Nvidia's product page speaks of up to four systems for models up to 700 billion parameters. The Nvidia clustering documentation describes up to three machines cabled directly or four via a switch, at 200 Gbit/s per QSFP port with approved cables from Amphenol and Luxshare. According to Nvidia, two machines connect with one QSFP cable and no switch, and the Nvidia Sync app sets up the cluster. Notebookcheck reported on 23 February 2026 an eight-machine cluster running Qwen 3.5 with 397 billion parameters at 24 tokens per second, and wrote that officially only a direct link of two machines is intended.
So Nvidia names four machines as the ceiling, Notebookcheck spoke of two in February, and tinkerers show eight: two machines on a direct cable are the simple normal case, up to four you are within the documented range, beyond that you depend on community guides. For me, one machine was the answer. Using DeepSeek as the example, I work out in Running DeepSeek Locally: What V4 Needs which large open models are too big even for four linked machines.
Successor: DGX Spark 2 or RTX Spark, wait or buy
No DGX Spark 2 has been announced as of November 2026; on the product page, the 64 GB version from October is the latest addition. What Nvidia presented at Computex 2026 is called RTX Spark, and it has been on sale since 16 October 2026: laptops from ASUS, Dell, HP, Lenovo, Microsoft and MSI, starting at $2,599 according to Nvidia's press release, with pricing and configurations varying by region. The top configuration has a 20-core Grace CPU, a Blackwell RTX GPU with 6,144 CUDA cores and up to 128 GB of unified memory; a smaller one goes up to 64 GB. Nvidia says compact RTX Spark desktops follow in November. All of them run Windows 11. There is no "DGX Spark laptop"; anyone searching for one means RTX Spark.
Whether RTX Spark is the better choice for you comes down to one question: if you need Linux with CUDA, Docker and SGLang as the DGX Spark ships it, RTX Spark with Windows is a different product and no substitute; Linux tools run there through the Windows Subsystem for Linux. If a notebook with plenty of unified memory under LM Studio or Ollama is enough, look at the first RTX Spark laptop reviews. I would not bet on falling prices: the list price went up twice in 2026, in February and in October, and euro offers in Germany sit well above their September levels.
Frequently asked questions
How much does the DGX Spark cost?
The Founders Edition with 128 GB has cost $6,950 in the US since 2 October 2026, according to The Register, up from $4,699 after Nvidia's February increase; the 64 GB partner version starts at $4,999 according to Nvidia. In Germany, on 14 November 2026, the reseller Nelpx listed the Founders Edition with 4 TB at 8,151.50 euros including VAT, with one to two weeks' delivery, and on 12 November 2026 partner machines with 128 GB and 1 TB started at 5,699 euros on Geizhals. Check the price on the day before you order.
Can I buy a used DGX Spark?
Yes. On 12 November 2026 the German classifieds site Kleinanzeigen had several offers, used or barely used Founders Editions with 4 TB at asking prices between 4,250 and roughly 7,000 euros, plus partner machines such as the ASUS Ascent GX10 with 1 TB from 5,000 euros. Get the invoice: according to Nvidia, hardware support comes from the vendor that sold the system. And compare with new: a new partner machine with 128 GB and 1 TB starts at 5,699 euros in Germany.
Can I buy an RTX Spark yet?
Yes, as a laptop: since 16 October 2026 ASUS, Dell, HP, Lenovo, Microsoft and MSI sell RTX Spark laptops, starting at $2,599 according to Nvidia. Nvidia has announced compact desktops for November 2026. RTX Spark runs Windows 11, the DGX Spark runs Nvidia's DGX OS; for Linux workloads with CUDA and Docker, that is not a one-to-one swap.
Is the DGX Spark faster than an RTX 5090?
Not as long as the model fits in the card's graphics memory. According to Nvidia, the RTX 5090 has 32 GB of GDDR7 on a 512-bit interface, the Spark 128 GB of LPDDR5x on 256 bits at 273 GB/s. insiderllm.com writes that models up to 24 GB run faster on a single graphics card than on any mini PC in this class. The Spark's strength is its memory: only when the model outgrows the card's VRAM does that flip.
Which version should I buy: Nvidia, ASUS or Dell?
All eight systems share the GB10 chip and come with 128 GB of memory. Decide by SSD size, by the price on the day (in Germany on 12 November 2026 the Lenovo ThinkStation PGX with 1 TB was the cheapest, and ASUS is the most widely listed on Geizhals) and by the supplier your purchasing department knows. The Founders Edition is the 4 TB reference from Nvidia.
Is a DGX Spark 2 coming soon?
None is announced, as of November 2026. Nvidia's autumn additions were the 64 GB version of the DGX Spark and RTX Spark with Windows, on sale as laptops since October. For Linux workloads with CUDA and Docker, the DGX Spark remains the machine of choice.
Can I connect two DGX Sparks together?
Yes, via the ConnectX-7 ports at 200 Gbit/s: according to Nvidia, two machines connect directly with one QSFP cable and no switch, and the Nvidia Sync app sets up the cluster. Nvidia documents up to three machines directly or four via a switch. Up to four machines you are inside the documented range, beyond that you are not.
Is a Mac Studio worth it instead of a DGX Spark?
If you need more than 128 GB of memory or more speed, yes: the Mac Studio M5 Ultra offers 1.2 TB/s and up to 512 GB. With 96 GB it starts at $5,499 at Apple's US store; 256 or 512 GB require the larger M5 Ultra from $6,799 plus the memory upgrade (as of 12 November 2026). If your tools need Linux, CUDA and Docker, no. For a single 27B model on a Mac, a Mac mini with 64 GB is enough.
Where to go from here
Before you order, settle three things: which model in which size runs permanently, whether your environment is Linux or macOS, and how many devices access it. A 27B model on a Mac: the Mac mini is the cheaper entry. A 70B model with CUDA: the DGX Spark. Anything bigger: Mac Studio or GPU server.
How the Spark holds up day to day is in DGX Spark Tested: Worth Buying?. The overview of all routes is Local AI: What Actually Works in 2026, and how I use the local model as a coding agent is in Claude Code Alternative Without Cloud. For a small business, AI Server for Small Businesses asks whether you need your own hardware at all.