Root means root.
Install your stack. Change your kernel. Run containers, fine-tune models, or start from scratch. No GPU sharing. No arbitrary software restrictions.
ssh root@your-spark↗Your own NVIDIA DGX Spark. Full root access.
One machine or a private cluster — the power to build
what comes next, without owning the hardware.
~ $ nvidia-smi
NVIDIA GB10 · 128 GB unified memory
~ $ ▌NOT A SLICE. THE WHOLE MACHINE.
From your first inference to multi-node experiments.
Real hardware, a private fabric, and nothing
between you and your next idea.
Install your stack. Change your kernel. Run containers, fine-tune models, or start from scratch. No GPU sharing. No arbitrary software restrictions.
ssh root@your-spark↗Rent 1–6 dedicated systems. Multi-node clusters use ConnectX-7 Ethernet and verified RoCE networking, with bandwidth dependent on the installed fabric.
Separate Ethernet and RDMA VLANs, enforced on MikroTik switches. Hosts are sanitized between rentals, not just given a new password.
SERIOUS COMPUTE. SENSIBLE PRICING.
No dollar surprises. Configure your cluster
and see exactly what you reserve in rupees.
First Spark at the base rate.
Each additional Spark costs 20% more than the previous one.
Prepay for 1–24 hours.
Stop early. Unused time is refunded.
Choose how much power you need.
Loading pricing…
FROM IDEA TO TERMINAL
Choose your Sparks, reserve your time, and add an SSH public key. Your quote is locked when you book.
Once isolation and provisioning checks pass, connect as root using your private key. The entire machine is yours.
Train, infer, experiment. Stop when you’re done. Access is revoked and hardware is sanitized before its next rental.
GOOD QUESTIONS.
No. Each rental reserves dedicated physical DGX Spark systems, including their CPU, GPU, memory, and local disk. Your root access is not a container shell.
Yes. A two-node and a four-node cluster consume all six machines. They use distinct access and RDMA VLANs enforced by managed switches. Tenant gateways must also deny cross-tenant routing, including via public IPs.
No inter-machine NVLink is exposed on DGX Spark. Clustering uses ConnectX-7 Ethernet with RoCE. NVIDIA specifies up to 200 Gb/s per QSFP port; actual bandwidth depends on the switches, cables, and validated configuration. Multi-node memory is not one transparent shared pool.
Rates are in INR using an administrator-configured USD base price and reference exchange rate. Each additional Spark costs the configured percentage more than the previous Spark. With a 20% premium and a base rate of ₹100, individual machines cost ₹100, ₹120, ₹144 and ₹172.80 per hour; a four-Spark cluster totals ₹536.80. Cluster totals are rounded once to the paise. Reservations are prepaid; unused time is refunded after verified access revocation. Applicable taxes are extra. No automatic card charging is enabled.
Local storage is ephemeral. Back up your work before stopping or before your reservation expires. Hardware stays quarantined until the external reset process has verified sanitization. Root access does not permit abuse, attacks, or access to other tenants or provider infrastructure.
YOUR COMPUTE, YOUR RULES
A dedicated home for your next big idea.
If checkout closes or your connection fails, check the existing payment before paying again. Captured live payments fund your wallet once; test payments never do.
| Date | Amount | Status | Action |
|---|
Reservations debit your wallet upfront. Unused funds are returned after access revocation. This is a usage statement, not a tax invoice.
| Date | Description | Amount |
|---|
| Customer / order | Amount / mode | Status | Action |
|---|
Agent mode is cooperative: tenant root can disable the client or forge health reports. Luckfox mode controls recovery boot externally, but is not firmware attestation. Both require hardware commissioning before live rental.
| Spark | Operation / mode | Progress | Updated |
|---|
| Cluster | Customer | Nodes | State | Rate / hour |
|---|
ACCOUNT SECURITY
Confirm that this email belongs to you. This does not sign you in or change your password.