DGX Spark price, and what the same money rents in cloud GPUs

A DGX Spark costs $6,950 with 128 GB of memory, the price NVIDIA set on 2026-10-02 according to The Register, or about $4,999 with 64 GB from other PC makers. The same 128 GB machine launched at $3,999 in October 2025. At the 2026-11-12 QuantaCloud prices, $6,950 rents an RTX PRO 6000 Blackwell for 2,908 hours or an RTX A6000 for 14,479 hours (today $2.16/GPU-hr and $0.48/GPU-hr). Buy the Spark if you will use it most days for years and its speed is enough. Rent if you need a faster or larger GPU for fewer hours than that.

What a DGX Spark costs today#

The price has gone up twice since launch, and it depends on where you look. NVIDIA's product page carries no price and sends buyers to its Marketplace, so these are dated figures from named sources:

The 128 GB price is 74% above launch (our calculation: 6,950 / 3,999 = 1.74). NVIDIA put the February increase down to constrained memory supply, and The Register blames memory prices for the October one. The lower listings do not help much: the one dealer under $6,000 for the Founders Edition had no stock, and CDW's listing shows the Founders Edition as discontinued there on October 19, 2026. I would plan with $6,950 and check stock before counting on anything cheaper.

What you get for the money#

The Spark is a 1.2 kg desktop box with one chip and a very large pool of slow memory. These are NVIDIA's figures from its product page, read on 2026-11-12:

Two numbers in that table explain the whole machine. 128 GB, shared with the CPU, is more than any single card in QuantaCloud's catalog holds except the H200 NVL. 273 GB/s is the slowest memory of anything on this page: an RTX A6000 reads its 48 GB at 768 GB/s, 2.8 times as fast, and an RTX PRO 6000 Blackwell reads its 96 GB at 1,597 or 1,792 GB/s depending on edition, 5.8 to 6.6 times as fast (our calculation).

What it is good at, and where it is slow#

The Spark's strength is capacity and its weakness is speed, and the independent launch reviews agree on both. The Register's summary was that it "may not be the fastest at any one thing, but can run them all", and LMSYS wrote that "the only downside of this machine lies in memory bandwidth". The measured figures, from tests in October 2025 and February 2026:

So in these tests an RTX PRO 6000 Blackwell generated tokens about 4.3 times as fast, and an RTX 6000 Ada fine-tuned and generated images 2.6 to 3 times as fast (our calculation: 215 / 49.7, 97 / 37, and 90 seconds against 30). The Spark's wins were the jobs that need its memory. The Register fine-tuned FLUX.1 Dev on it in about four hours using a little over 90 GB, which no 48 GB card can hold, and it served four concurrent users at 17 tokens per second each, falling to 4 per user at 64 concurrent requests, which The Register put down to memory bandwidth. LMSYS's verdict on the largest models was that they are best suited to prototyping and experimentation, not production. vLLM's developers put it in a code comment: version 0.33.0 keeps the Spark on its small-GPU batch settings because it "has the memory but not the compute of a large discrete GPU".

The software has moved since launch week. llama.cpp's own benchmark, updated in February 2026, runs gpt-oss-20b at 83.4 tokens per second, against 49.7 in LMSYS's launch test with Ollama. I have not found a newer test that runs the Spark and the rented cards side by side, so read the ratios as the shape of the trade, not as today's exact figures.

DGX Spark vs RTX 5090, and vs Mac Studio#

Against an RTX 5090, the Spark trades speed for capacity. The 5090 has 32 GB of GDDR7 at 1,792 GB/s, and in LMSYS's test it generated 205 tokens per second to the Spark's 49.7, 4.1 times as fast. The Register put it plainly: the Spark will not beat a 5090 at inference, fine-tuning or image generation, and it runs models the 5090 cannot load. The price gap has narrowed. The 5090 launched at $1,999 in January 2025, Tom's Hardware reported it "regularly listed for above $5,000" on 2026-09-05, and one US dealer listed 5090 cards at $5,499.99 to $6,499.99 on 2026-11-12, the ones I checked out of stock. A 5090 also needs a PC around it, which the Spark does not. The RTX PRO 6000 vs RTX 5090 comparison covers the card with three times the 5090's memory.

Against a Mac Studio, the Spark has less memory bandwidth and NVIDIA's software. Apple's specifications list the M5 Max at 460 or 614 GB/s with up to 128 GB, and the M5 Ultra at 1.2 TB/s with up to 512 GB, against the Spark's 273 GB/s and 128 GB. The Register's review made the same point about the earlier Mac Studio: higher memory bandwidth, which translates into faster token generation. What the Spark has is CUDA: The Register's point was that existing NVIDIA code should work out of the box. Anything compiled still has to be built for the Spark's Arm CPU.

DGX Spark clusters#

NVIDIA supports up to four Sparks working together. Two or three connect directly with QSFP cables over the built-in ConnectX-7 ports, and four need an external 200 GbE switch. NVIDIA puts two 128 GB units at models of up to 400 billion parameters and four at up to 700 billion, and says two 64 GB units delivered up to 1.7 times the performance of one in its own Qwen 3.8 27B test. Four of the 128 GB model cost $27,800 at NVIDIA's price (our calculation: 4 x 6,950). ServeTheHome, which runs an eight-node cluster of them, advises buying one 128 GB unit before two 64 GB units.

DGX Station price#

The DGX Station has no NVIDIA list price, and the dealer listings run from about $95,000 to $124,000. It is a different class of machine: a desktop tower with one Blackwell Ultra GPU carrying 252 GB of HBM3e, a 72-core Grace CPU and 496 GB of system memory, 748 GB in all, drawing 1,600 W from a 20 A circuit. Tom's Hardware reported a public listing on 2026-08-23: an Exxact system from $94,930. On 2026-11-12, Central Computer listed ASUS's version at $111,999.99 and $123,569.99 by configuration, and MSI's and Gigabyte's at $99,999.99 on preorder. AI server cost puts that next to GPU servers at the same money. The DGX name also covers NVIDIA's eight-GPU servers, and the DGX B300 guide explains the current one and how it is sold.

What $6,950 rents in cloud GPUs#

The Spark's price buys thousands of hours on cards that are several times faster. These are the live prices for the two cards I would compare it with, and for the RTX 6000 Ada that The Register tested against it:

Prices checked 14 Nov 2026, 09:11 UTC

On 2026-11-12 a 1x RTX A6000 cost from $0.48 an hour and a 1x RTX PRO 6000 Blackwell $2.39 (today $0.48/GPU-hr and $2.16/GPU-hr). At those prices $6,950 equals 14,479 hours of RTX A6000 or 2,908 hours of RTX PRO 6000 (our calculation: 6,950 / 0.48 and 6,950 / 2.39). How long that lasts depends on how many hours a week you would really use a GPU:

The arithmetic for one cell: 20 hours a week x 52 weeks x $2.39 = $2,485.60 a year, and 6,950 / 2,485.60 = 2.8 years. The break-even against the RTX PRO 6000 inside three years is about 19 hours a week (2,908 / 156 weeks = 18.6). Against the RTX A6000 there is no break-even at any normal working pattern: even 40 hours a week takes 7 years.

Hours are not equal work, and that pushes the break-even further out. On LMSYS's gpt-oss-20b test the RTX PRO 6000 generated tokens 4.3 times as fast, so 2,908 rented hours do the work of about 12,600 Spark hours, 1.4 years of running the Spark around the clock (our calculation from one launch-week benchmark: 2,908 x 4.3, then / 8,760). LMSYS tested the Workstation Edition, the Server Edition reads memory 11% slower than the workstation cards, and newer software has raised the Spark's own numbers, so treat 4.3 as an upper bound. The Spark has its own running cost too, a small one: at most, at the full 240 W of its power supply, a year of always-on use is 2,102 kWh, about $382 at the average US residential price of 18.16 cents per kWh in August 2026. The GPU cost calculator runs these sums for your own hours at live prices.

When the Spark wins#

The Spark wins when the meter matters more than the speed. It is always on, so an agent or a private assistant that runs all day costs nothing extra per hour, where the always-on row above costs $4,193 a year on an RTX A6000. It is yours, so prompts, documents and fine-tuning data never leave your desk, and nothing is deleted when you finish for the day. It holds 128 GB, enough for gpt-oss-120b, which llama.cpp's benchmark runs at 58.7 tokens per second, quicker than anyone reads. And small models with batching are where it is strongest: 368 tokens per second in total on an 8B model in LMSYS's test.

When renting wins#

Renting wins on speed, size and idle time. An RTX PRO 6000 was about four times as fast in the tests above, and when the catalog lists it, the H200 NVL has 141 GB at 4.8 TB/s, 17.6 times the Spark's bandwidth. A burst of work costs only its own hours: a 10-hour fine-tune on an RTX A6000 was $4.80 at the 2026-11-12 price (our calculation: 10 x $0.48), and the unused seconds of the last hour are refunded when you stop. Nothing sits idle on your desk, and you do not have to guess what the hardware will cost next year: the Spark's own list price has changed twice in 13 months.

Renting has costs the Spark does not. Stopping a QuantaCloud instance terminates it and deletes its disk, and there are no volumes or snapshots, so models and results have to be copied off or downloaded again. The GPUs run in US regions only, and your data goes to a VM in a data center, which rules it out for some work. How to rent a GPU for AI covers the first launch, and self-hosting an LLM covers the choices between a desk machine and a server.

FAQ#

How much does a DGX Spark cost?

$6,950 for the 128 GB model at NVIDIA's price since 2026-10-02, according to The Register and ServeTheHome, and from about $4,999 for the 64 GB model sold by Acer, ASUS, Dell, Gigabyte, HP and MSI. One US dealer listed the 128 GB Founders Edition in stock at $6,899.99 on 2026-11-12.

Why did the DGX Spark price go up?

Memory. Tom's Hardware reported that NVIDIA raised the price from $3,999 to $4,699 in February 2026 "due to the constrained memory supplies worldwide", and The Register blamed memory prices again when the 128 GB model went to $6,950 in October 2026.

Is the DGX Spark worth it?

For daily, private, always-on use by one person or a small team, yes, if 40 to 85 tokens per second on mixture-of-experts models such as gpt-oss, and a few tokens per second on dense 70B models, is enough. For speed, for serving many users, or for a few hours of heavy work a week, the same money goes further rented: 2,908 hours of a card that tested about four times as fast.

Can a DGX Spark run a 70B or a 120B model?

It loads both, at very different speeds. LMSYS measured 2.7 tokens per second on the dense Llama 3.1 70B at FP8 at launch. llama.cpp's own benchmark, updated in February 2026, measured gpt-oss-120b at 58.7 tokens per second with a short prompt and 42.8 at 32,768 tokens of context, because that model activates 5.1 billion of its 117 billion parameters for each token. NVIDIA's own ceiling is 200 billion parameters for inference on the 128 GB model.

How many DGX Sparks can be clustered?

Up to four, according to NVIDIA's product page. Two or three connect with QSFP cables, and four need a 200 GbE switch. Each unit keeps its own 273 GB/s of memory bandwidth, so a cluster adds capacity more than speed for a single request.

What is the difference between a DGX Spark and a DGX Station?

About $90,000 and a different chip. The Spark is a 240 W box with 128 GB of shared LPDDR5x. The DGX Station is a 1,600 W tower with a Blackwell Ultra GPU, 252 GB of HBM3e and 748 GB of memory in all, listed from $94,930.

My rule: count the GPU hours you would use in a week before you buy. Under about 19 hours a week, the RTX PRO 6000 rented by the hour costs less than a Spark over three years and runs about four times as fast. Above that, or when the work has to stay on your desk and run all day, the Spark earns its $6,950. If you are not sure which side you are on, rent for a month and count.