NVIDIA DGX Spark: Price, Specs and Rent vs Buy (2026)

NVIDIA DGX Spark is a 1.2 kg desktop computer built around the GB10 Grace Blackwell Superchip, with 128 GB of unified memory and a Founders Edition price that NVIDIA raised from $3,999 to $4,699 in February 2026 and that ServeTheHome reports has since moved to about $6,950. It is a development box for models up to roughly 200 billion parameters, not a replacement for a datacenter GPU, and the break-even against renting one by the hour is long.

This guide covers what is inside the box, what it runs, what it costs to own, and when a rented datacenter GPU is the better call. For the wider map of chips, see Datacenter GPUs in 2026.

TL;DR

DGX Spark pairs a 20-core Arm CPU with a Blackwell GPU and 128 GB of LPDDR5x memory shared between them, at 273 GB/s. NVIDIA claims up to 1 PFLOP at FP4.
Founders Edition MSRP was raised from $3,999 to $4,699 on NVIDIA's developer forum in February 2026. ServeTheHome reported on October 3, 2026 that the 128 GB model is moving to about $6,950 with various OEMs charging different amounts, alongside a 64 GB model at $4,999 from October 23, 2026. We did not find an NVIDIA page stating the October figure. Partner systems from Acer, ASUS, Dell, Gigabyte, HP, Lenovo and MSI are priced by each partner.
NVIDIA says one unit runs inference on models up to 200 billion parameters and fine-tunes models up to 70 billion. Two linked units reach 400 billion.
The memory is the point, and the bandwidth is the limit. 273 GB/s is a small fraction of the HBM bandwidth on a datacenter card, so token generation on big models is slow compared with a rented H100 or B200.
Verdict: buy it if you want a quiet, always-on local box for prototyping and privacy. Rent by the hour for training runs, throughput serving, or anything that has to finish fast.

DGX Spark specs

All figures below are from NVIDIA's DGX Spark product page unless noted. Performance figures are NVIDIA's claims, not our measurements.

Spec	DGX Spark (GB10)
GPU	Blackwell, 5th-generation Tensor Cores, 4th-generation RT Cores
CPU	20-core Arm: 10 Cortex-X925 and 10 Cortex-A725
Memory	128 GB LPDDR5x unified (a 64 GB variant is offered through OEM partners only)
Memory bandwidth	273 GB/s over a 256-bit interface
AI compute	Up to 1 PFLOP at FP4 (NVIDIA's claim)
Storage	Up to 4 TB NVMe M.2 with self-encryption
Networking	10 GbE RJ-45, ConnectX-7 at 200 Gbps, Wi-Fi 7
Power	240 W power supply, 140 W GB10 TDP
Size and weight	150 x 150 x 50.5 mm, 1.2 kg
Clustering	Up to four systems over ConnectX

The product launched with the press release NVIDIA DGX Spark Arrives for the World's AI Developers dated October 13, 2025, with ordering opening October 15, 2025.

Architecture: one pool of memory

The GB10 Superchip puts the CPU and GPU in one package with coherent access to the same 128 GB. There is no separate VRAM. That is the design choice that matters: a 70B model at 8-bit weights needs about 70 GB for weights alone, which does not fit on any single consumer card but fits comfortably in Spark's pool. For context, NVIDIA's RTX 5090 ships with 32 GB of GDDR7, and the RTX PRO 6000 Blackwell Workstation Edition carries 96 GB.

The trade is bandwidth. Decoding tokens is mostly a memory-bandwidth problem: each generated token reads the model's active weights once. At 273 GB/s, a dense model with tens of gigabytes of weights moves slowly. A datacenter H200 is specified at 4.8 TB/s, roughly 17 times Spark's figure (4,800 / 273). Mixture-of-experts models help, because only a fraction of the weights is read per token, which is why Spark is a better fit for sparse models than for large dense ones. See our mixture-of-experts and HBM glossary entries for the mechanics.

The FP4 figure deserves a caveat. NVIDIA states "up to 1 PFLOP" at FP4 and does not label it dense or sparse on the product page. Treat it as a ceiling for 4-bit inference kernels, and see FP4 and NVFP4 vs MXFP4 for what the format gives you in practice.

What DGX Spark runs

NVIDIA's own sizing, from the product page:

Configuration	Model size NVIDIA lists
1x 64 GB	Up to 100B parameters
1x 128 GB	Up to 200B parameters
2 units linked (256 GB combined)	Up to 400B parameters
4 units linked (512 GB combined)	Up to 700B parameters

Fine-tuning is supported for models up to 70 billion parameters. These limits assume low-precision weights; the sizes are what fit in memory, not a statement about speed.

Good fits:

Local prototyping. Build and debug an agent, a retrieval pipeline or a fine-tune on your desk, using the same CUDA software stack as the datacenter.
Private data. Nothing leaves the room.
Large sparse models at modest speed. A model that cannot load on a 32 GB card can load here.
Always-on dev server. 140 W TDP and a 240 W supply is a laptop-class power draw.

Poor fits:

Pre-training or long fine-tunes. Compute and bandwidth are small next to a datacenter GPU, so a job that takes hours on a rented H100 can take days here.
Serving many users. Aggregate throughput scales with bandwidth and batch size, both of which Spark lacks.
Anything with a deadline. You cannot burst to ten GPUs on a desk.

What it costs to own

NVIDIA raised the Founders Edition price in February 2026. The announcement on the developer forum reads: "The MSRP for DGX Spark (Founders Edition) has been adjusted from $3,999 to $4,699 due to memory supply constraints" (NVIDIA developer forum, as of October 8, 2026). That is the last NVIDIA-posted figure we could open; ServeTheHome (October 3, 2026) reports the 128 GB model is now about $6,950. Some retailers still showed the old $3,999 figure in that thread, which looks like older stock. Partner systems are priced by the partner; NVIDIA directs OEM pricing questions to them.

Ownership costs beyond the sticker:

Power. At the 240 W supply rating, running around the clock is about 5.8 kWh a day (0.24 kW x 24 h). Multiply by your local rate.
Storage. The 4 TB ceiling is a configuration option; larger datasets live on a network share.
Depreciation. Memory pricing moved the list price up in 2026, and a newer generation will eventually replace the GB10.
Idle time. Hardware you own costs the same whether it computes or sits idle.

Rent vs buy: the break-even

Break-even in GPU-hours is the purchase price divided by the hourly price of what you would rent instead. Two prices are in play: the $4,699 NVIDIA posted in February 2026, and the roughly $6,950 ServeTheHome reported in October 2026.

break-even hours = price in dollars / (live hourly price in dollars)

For every $1 of hourly price, that is 4,699 GPU-hours (about 196 days of continuous use, 4,699 / 24) at the February price, or 6,950 GPU-hours (about 290 days, 6,950 / 24) at the October figure. Read the live hourly price from the box at the bottom of this page and divide. Three things shift the answer:

Your utilization. A box that computes 4 hours a day takes six times longer to pay back than one running around the clock.
What you compare against. A datacenter GPU does more work per hour than Spark. If an H100-class card finishes your job in a fraction of the time, the fair comparison is cost per finished job, not cost per hour.
Electricity and your time. Owned hardware adds a power bill and setup time; rented hardware adds neither once the instance is up.

The rule of thumb that falls out: if you run a GPU job a few times a month, rent. If you iterate on a model every day for a year, a desk box starts to make sense, and it still does not replace renting when a run needs an H100, H200 or B200 for a few hours.

When to rent a datacenter GPU instead

Rent when any of these is true:

The model does not fit in 128 GB, or fits but decodes too slowly. Datacenter parts carry HBM bandwidth in the terabytes per second (H100: 80 GB at 3.35 TB/s; H200: 141 GB at 4.8 TB/s).
You are training, not just fine-tuning a small model.
You need results for a deadline and can pay for a few hours instead of owning for years.
You need more than four units' worth of memory for a short period.

A common pattern is to prototype on the desk, then rent the same container on a datacenter GPU for the real run. For the chips behind that step, see the B200 guide, the H100, H200 SXM vs NVL vs PCIe guide, and the RTX PRO 6000 page for a workstation-class card you can rent. The larger sibling to Spark is the DGX Station GB300.

Rent today

Aquanode manages and optimizes GPUs for training and inference workloads, and you can rent the GPUs in the box below on demand. Prices update live. Compare them against the break-even formula above.

What's next

NVIDIA's partner list for GB10 systems has grown since launch, and the company lists a 64 GB variant through OEM partners as coming soon. NVIDIA has not published a successor to GB10 on the DGX Spark page. For what is coming in the datacenter, see NVIDIA Rubin.

FAQ

How much does DGX Spark cost?

NVIDIA raised the Founders Edition MSRP from $3,999 to $4,699 in February 2026 (NVIDIA developer forum). ServeTheHome reported on October 3, 2026 that the 128 GB model is moving to about $6,950. Partner systems are priced by each partner.

How much memory does DGX Spark have?

128 GB of LPDDR5x unified memory shared by the CPU and GPU, at 273 GB/s. A 64 GB variant is sold through OEM partners.

What is the largest model DGX Spark can run?

NVIDIA lists up to 200 billion parameters on one 128 GB unit and up to 400 billion on two linked units, for inference. Fine-tuning is supported up to 70 billion parameters.

Is DGX Spark faster than an RTX 5090?

They are built for different jobs. The RTX 5090 has 32 GB of GDDR7; Spark has 128 GB of unified memory but much lower bandwidth than a discrete card. Spark fits bigger models; a discrete card trades capacity for much higher memory bandwidth. NVIDIA does not publish a head-to-head benchmark, so test your own model.

Can DGX Spark replace renting a datacenter GPU?

For prototyping, yes. For training or high-throughput serving, no: bandwidth and compute are far below an H100, H200 or B200, and you cannot scale out on demand.

Is DGX Spark worth buying?

If you will use it most days for months, it can pay back against hourly rental. Divide the price you would pay ($4,699 in February, about $6,950 per ServeTheHome in October) by the live hourly price to get your break-even in GPU-hours.