The DGX Spark question is not which chip. It is where to put it.

Not on the sticker. NVIDIA’s Founders Edition listed at $4,699 in February 2026 and a year with us is $6,109.79. Renting wins once you count what a bought machine still needs — a routed public address, power and cooling that do not stop, and somebody to touch it — and it wins immediately if you would otherwise be waiting on supply. Buying wins if you already have a rack and a spare pair of hands.

Yes. It is plan D105x: one whole NVIDIA DGX Spark, single tenant, in New York. $6,109.79 for twelve months — $509.15 a month, paid once at the start. It is built to order rather than kept in stock, which is why it never appears in our live-stock listings — that silence is about the shelf, not about whether we sell it.

On a twelve-month term, paid once at the start: $6,109.79 for the year, which is $509.15 a month. That is the $599 monthly list rate with the 15% annual discount applied. There is no monthly, hourly or daily option on this line. There is no setup fee and the renewal figure is the same figure.

Not from us. Providers renting GB10 capacity by the hour were advertising roughly $0.55 to $0.75 an hour in 2026, and for a short evaluation that is the right shape. Come to us when you want the machine to still be there next month, with your data on it and at the same address.

Yes, and it is worth asking rather than clicking through the configurator four times. Each machine has a ConnectX-7 with a 200 GbE pair, NVIDIA supports up to four connected together for models of up to 700 billion parameters, and ours sit in the same rack in the same hall. Tell us the shape on chat and we raise one build for the set.

A GB10 Grace Blackwell Superchip, 20 Arm cores (10 Cortex-X925 + 10 Cortex-A725 Arm), 128 GB of LPDDR5 shared by the CPU and the GPU, and a NVMe-SSD 1 TB. NVIDIA also sells a 4 TB variant; ours is the 1 TB one. It runs NVIDIA DGX OS with the CUDA stack already installed.

A 1 Gbps port, unmetered and unshared, with a routed IPv4 and IPv6. The box itself carries a 10 GbE port; we sell it on a gigabit that nothing counts and nobody else is on, which for serving a model is the number that matters. The ConnectX-7 is used for machine-to-machine links inside the rack.

Root over SSH on your own routed address, from the first login. There is no BMC in a DGX Spark, so unlike our x86 servers there is no iLO or iDRAC-style out-of-band console; power cycles and anything physical are done by our engineers, on request, at any hour.

Not inside the machine. A DGX Spark has a single power supply and a single input, and nothing NVIDIA ships changes that. Redundancy lives upstream, in the facility — our feeds, our distribution, our generators — and a failed unit is replaced within four hours.

Days rather than hours. Our racked servers go live about half an hour after an order clears; a DGX Spark is bought and built against the order, so the honest window is several business days and longer when supply is tight. Ask on chat before you pay and we will give you the figure as it stands that morning.

Both come up often. Listing hardware you own so other people can rent it by the month is what the GPU marketplace exists for; putting a machine you own into our hall is colocation, which sales will price for you.

Data Room, Inc. of Long Island City, New York, trading as Server Room, hosting since 2004, with Data Room SRL in Bucharest. The machine is deployed in New York, so the purchase runs through the US entity wherever you are. The network is AS19624 — look it up at bgp.tools.