Dedicated Instant Custom GPU Unmetered Domains Pricing Data Centers Migrate
Marketplace Configure server

NVIDIA DGX Spark · New York

The DGX Spark, both ways round.

Two completely different people read this page. About half want to rent a GB10 machine from us. The other half already own one and want it earning. Pick your door and the rest of the page follows.

I want one Rent a whole DGX Spark A GB10 machine in our New York building with nobody else on it. $6,109.79 for twelve months, charged once. Five business days to build. The price and the terms I have one Rent yours out instead A Founders Edition, an ASUS Ascent GX10 or a ZGX nano sitting idle earns nothing. List it and other people rent it from you, at a price you set. Let your own machine earn

Turn your GPUs into passive monthly revenue

Got idle server or desktop GPU setups? List them on the Primcast marketplace today and earn steady monthly rents from AI teams, developers, and enterprises needing production-grade compute.

Go to Marketplace

Renting one, end to end

Five steps, and the awkward ones are not hidden in step four.

  1. 1

    You configure it and pay for the year

    $6,109.79, charged once, which works out at $509.15 a month against a list rate of $599. Twelve months is the only cycle this line is sold on, and it is on this page rather than in an email afterwards: we buy the machine when the order arrives, and that is what makes the price possible. No setup charge.

    Why not a month at a time? Because we would have to buy a machine on the chance you stay. If a shorter arrangement matters, ask on chat before you pay — we will answer honestly rather than take the order and work it out later.

  2. 2

    We source it, rack it and give it an address

    Five business days from a confirmed order, and longer when supply is tight. Nothing is on a shelf and we do not pretend otherwise — NVIDIA raised the price of these once already when memory got scarce, and finding them has been the hard part ever since. If you are working to a date, tell us before you order.

  3. 3

    You log in as root, on day one

    NVIDIA DGX OS is already installed — Ubuntu underneath, with NVIDIA's driver, CUDA and container tooling in place. One customer per machine, no hypervisor, nothing of ours running on it. Access is root over SSH.

    Worth knowing before you order: a DGX Spark carries no BMC, so unlike our x86 servers there is no out-of-band console. If you lock yourself out of the network, our engineers power-cycle it for you on request rather than you doing it from a browser.

  4. 4

    It stays there, at the same address

    A gigabit line with no meter on it, one IPv4 and one IPv6. That is the whole difference between this and an hourly instance: yours can serve, and it answers on the same address next month with your work still on the disk.

  5. 5

    When something breaks, we replace it

    Within four hours, included rather than sold as an upgrade. The unit itself has a single power supply — that is how NVIDIA built it and no host can change it. The redundancy is upstream in the building, and the four-hour swap covers the rest.

The number

Asked of the checkout, not typed into a page.

Two things move it. Everything else on this line is fixed, including the term.

  • No setup charge on this line at all
  • The price does not move during the term
  • Traffic is never counted or billed
  • You see it itemised again before paying

Buying instead? NVIDIA lists the Founders Edition at $4,699, up from $3,999 in February 2026 when memory supply tightened. A year here costs more than that, and what the difference pays for is the room, the line, the address and the four-hour replacement — plus having one at all next week.

Your configuration12-month term
MachinesUp to four wire together over ConnectX-7
1
Response coverParts are replaced in four hours on every tier

$6,109.79

$509.15/mo each · Charged once, for the year.

Order this

Is this the right machine for you?

Four honest answers, including the two that say no.

Yes — you are building for this chip

Software that has to run on a GB10 at a customer site. Rent the same machine for the length of the project rather than buy one to find out, and keep your build tree on it between runs.

Yes — the data cannot leave

Client files or records under a contract that says they stay put. One tenant, one machine, one address, nothing shared. Whether the model fits in the shared memory is the question after this one.

No — you want it for a weekend

The shortest term here is a year. For a few hours of experiment an hourly provider is genuinely the better answer, and we will say so on chat rather than sell you twelve months.

No — you actually want a big card

If the job wants raw card memory, or several cards in one chassis, this is the wrong shape. The cards we fit and multi-card builds are the pages for that.

What is in it

NVIDIA's published specification, not a measurement of ours. There is no tokens-per-second figure here: it moves by a multiple with the model, the quantisation, the runtime and the batch size, so a single number would be a guess wearing a promise.

Chip
NVIDIA GB10 Grace Blackwell Superchip
Processor
20 Arm cores: 10x Cortex-X925 + 10x Cortex-A725
Memory
128 GB LPDDR5x, coherent and shared by the processor and the graphics side
Bandwidth
273 GB/s across a 256-bit interface
AI throughput
Up to 1 PFLOP at FP4, sparsity included
Storage
1 TB NVMe M.2
Network
1 Gbps unmetered, one IPv4 and one IPv6
Operating system
NVIDIA DGX OS, built on Ubuntu
Together
Up to four wire together over their ConnectX-7 pairs
Size and power
150 x 150 x 50.5 mm, 1.2 kg, 240 W supply
Where it lives
New York — our buildings
Term
Twelve months, charged once. No setup fee.

Something this line does not cover? Have one built to your specification. Every other machine's price is on pricing, and the marketplace is where hardware other people own is listed.

Asked, in this order, in live chat

The ten questions this page exists to answer.

Can I rent a DGX Spark from you?
Yes. One NVIDIA GB10 machine, single tenant, in our New York building, ordered from the configurator on this page. It is built for your order rather than taken off a shelf, so nothing about it depends on what happens to be racked the day you read this.
What does it cost, and can I pay monthly?
It is $6,109.79 for twelve months, charged once, which works out at $509.15 a month against a list rate of $599. Twelve months is the only cycle this line is sold on: we buy the machine when the order arrives, and the term is what makes that possible. There is no setup charge, and the price does not move during the term.
How long before I can log in?
Five business days from a confirmed order, and longer when supply is tight. That is what we quote in writing rather than a best case. Yours is sourced, racked, addressed and handed over in that window. If you are working to a date, ask on chat before ordering and we will tell you where things genuinely stand.
Can I take one by the hour, or for a few days?
No. The shortest term here is the twelve months above, and we would rather say so than take a short booking we cannot honour. Hourly sellers exist and are the right answer for a weekend experiment. This is for the case where the same machine has to be at the same address next month.
Can I have more than one, wired together?
Yes, up to four. They pair over their ConnectX-7 links in one rack, which is how NVIDIA supports four systems working on models of up to 700 billion parameters. Set the count in the configurator, or ask on chat for a set rather than placing four separate orders.
Is it really the whole machine?
Yes. One customer per machine, root access from the first login, no hypervisor, no partitioning, and nothing of ours running on it. The 128 GB is one coherent pool that the Arm cores and the Blackwell side both address, and all of it is yours. It is not an instance and it is not wiped when you stop using it.
Do I get out-of-band console access or IPKVM?
No, and it is worth knowing before you order. A DGX Spark carries no BMC, so unlike the x86 servers we rent there is no remote console and no browser-based power control. Access is root over SSH. If you lock yourself out of the network, our engineers power-cycle the machine for you on request.
What is on it when I first log in?
NVIDIA DGX OS, which is built on Ubuntu, with NVIDIA's own driver, CUDA and container tooling already installed. That is the stack NVIDIA supports on this chip, which is why this line does not offer the long menu of operating systems our other servers do. After handover it is yours to change.
Is the storage 1 TB or 4 TB?
The machine we fit carries 1 TB of NVMe. NVIDIA's own page shows a 4 TB variant and it exists, so if the larger disk matters to you, say so before you order and we will quote it rather than let you find out afterwards.
I already own a DGX Spark. Can I earn from it here?
Yes, and it is a different product entirely. You keep the hardware and root, list it at a price you set, and other people rent it from you — GB10 machines including the ASUS Ascent GX10 and the ZGX nano are already among the things people list. Start there rather than on this page.

Still deciding? Whether your model fits in the shared memory is the question most people ask next, and the cards we fit is the page to read if what you actually want is a graphics card rather than this machine.