NVIDIA DGX Spark • Built to order in New York

One GB10 Grace Blackwell machine, and nothing else on it

Twenty Arm cores, 128 GB of unified LPDDR5 and the GB10's Blackwell GPU, racked in New York with root on the metal. No hypervisor, no neighbour, no vCPU accounting. $599 a month, and that figure does not move.

Configure from $599/mo

Root on the metal Unmetered gigabit port Built to order in 4-24h Engineers on chat 24/7

Single tenant Bare metal Root access

The whole machine, not a share of one

The GB10 puts the Arm cores and the Blackwell GPU on one package behind a single pool of memory. Partition that across tenants and you have thrown away the reason to run the part at all, so we do not. You get the board, the memory and the accelerator, on your own power and your own port.

$599/mo No setup fee Monthly terms

The price is the price

$599 a month, billed monthly, cancellable monthly. No setup charge, no introductory rate that lapses, and no step-up at renewal. Three, six and twelve-month cycles take 5, 10 and 15 percent off if you would rather commit.

24/7 Hardware cover Real engineers

A person racked it and a person answers

Support is included on every machine rather than sold back to you: engineers on live chat around the clock, and failed hardware replaced within four hours. The people who answer are the people who work in the halls the machine sits in.

How the machine reaches you

We do not keep a pile of these on a shelf, and we would rather say so than pretend otherwise.

Built for the order

Every DGX Spark we ship is assembled against the order that paid for it, in our New York hall. The configurator quotes a 4 to 24 hour window and that is the window we work to.

DGX OS, already installed

It arrives running NVIDIA's own DGX OS with the CUDA stack in place, so the first thing you do on it is your work rather than a driver hunt. Full root from the first login.

Nothing counts the traffic

A gigabit port, unmetered. Not a large allowance, not a burst credit, not a fair-use figure that turns into an invoice. One IPv4 and one IPv6 address come with it.

What engineers put on it

A single coherent memory pool next to a Blackwell GPU suits some jobs far better than a rented slice of a bigger box.

Fine-tuning without a cluster

LoRA and full fine-tunes on open-weight models that fit the unified pool. One box you own outright, running for as long as the job takes, with no scheduler queue in front of it.

Inference you keep in-house

Serve a model on hardware nobody else touches, with the weights on your own disk. Sensible when the data cannot leave your control and a shared endpoint is not an option.

Porting to Blackwell first

Get your CUDA code building and running against Blackwell and an Arm host before you commit a budget to racks of it. Cheaper to find the surprises here.

Datasets that hate the PCIe hop

Where the CPU and the GPU share one address space, the copy across the bus stops being the thing you optimise. Useful for iterative work on a set that fits in 128 GB.

Reference specification

The build the configurator opens on. Everything here is what you get for $599 a month.

Accelerator

NVIDIA GB10 Grace Blackwell Superchip

CPU

20 Arm cores - 10x Cortex-X925, 10x Cortex-A725

Memory

128 GB LPDDR5, unified across CPU and GPU

Storage

1 TB NVMe SSD

Network

1 Gbps unmetered, unshared

Addresses

1x IPv4 and 1x IPv6 included

Operating system

NVIDIA DGX OS, preinstalled

Location

New York, US

Availability

Built to order, 4 to 24 hours

Price

$599 per month, no setup fee

Prices are per month and exclude sales tax where it applies. Need something this line does not cover? Have one built to your specification.

DGX Spark, asked and answered

The questions engineers actually send us before ordering one.

Is this a whole machine or a share of one?

A whole machine. One DGX Spark, one tenant, full root, nothing virtualised between you and the GB10. Its memory is a single pool shared by the Arm cores and the Blackwell GPU, and none of it is carved up for anybody else.

How long does it take to get one?

We build each one against the order rather than holding stock, and the configurator quotes a 4 to 24 hour window. If your timing is tight, ask us on chat before you order and we will tell you where the build queue actually stands.

What operating system does it run?

NVIDIA DGX OS, preinstalled, with the CUDA stack ready. This line ships with DGX OS rather than the wider image menu our x86 servers offer, because it is the platform NVIDIA supports the GB10 on. You have full root and can change what you like afterwards.

What does it cost to run once traffic is counted?

$599 a month, and traffic is not counted. The gigabit port is unmetered and unshared, there is no setup fee, and the renewal figure is the same figure. Longer billing cycles reduce it; nothing increases it.

Can I have more than one?

Yes, and they will sit on the same network in the same hall. Tell us how many and what you intend to run across them and we will build the order rather than have you click through the configurator once per box.