Single-Node Instances
One VM on one physical host — what every non-2X tier gives you, and why the RAM numbers are what they are.
A single-node instance puts the GPU, CPU, and RAM on the same physical machine. There is no network hop between your GPU and its memory or storage, which is what makes it the default choice for anything that runs as one process: training a model that fits, serving inference, building a dataset, or iterating on a notebook.
Every tier in the catalog except the 2X tiers is single-node.
The single-node catalog
Specs below are what a single instance receives. Prices live on Pricing and Billing.
| Tier | GPU | VRAM | vCPU | RAM | Local disk |
|---|---|---|---|---|---|
| RTX-3090-S | 1× RTX 3090 | 24 GB | 4 | 16 GB | 100 GB |
| RTX-3090-M | 1× RTX 3090 | 24 GB | 8 | 24 GB | 200 GB |
| RTX-5080-S | 1× RTX 5080 | 16 GB | 8 | 16 GB | 100 GB |
| RTX-5080-M | 1× RTX 5080 | 16 GB | 8 | 24 GB | 200 GB |
| RTX-5080-L | 1× RTX 5080 | 16 GB | 12 | 32 GB | 300 GB |
| RTX-5090-S | 1× RTX 5090 | 32 GB | 8 | 32 GB | 150 GB |
| RTX-5090-M | 1× RTX 5090 | 32 GB | 12 | 48 GB | 250 GB |
| RTX-5090-XL | 1× RTX 5090 | 32 GB | 16 | 56 GB | 300 GB |
| RTX-5090-XXL | 1× RTX 5090 | 32 GB | 16 | 56 GB | 600 GB |
Coming Soon: RTX PRO 6000 (S, M, L, XL) and DGX Spark (4XL).
Why the RAM numbers aren't round
Capacity is committed against what a physical host can actually carry. The 3090 nodes expose roughly 16 vCPU / 30.5 GiB RAM / ~350 GB of usable local disk to a single instance, and the 5090 pool ranges up to a 60 GiB RAM pool on some hosts. A tier is only real if it can actually be placed on a host, so the published numbers are capped under that ceiling rather than being the roundest number that sounds good:
RTX-3090-Mis 24 GB, not 32 GB. A 3090 host cannot safely commit 32 GB to one instance and still have room for its own overhead.RTX-5090-XLandXXLare 56 GB, not 64 GB. That leaves headroom under the host's own overhead on the 5090 pool, so the instance can actually be scheduled.RTX-5090-XXLdisk is 600 GB, not 800 GB, which keeps the tier placeable across the 5090 pool.
These are deliberate, verified commits, not rounding. Ask ANK if you need capacity the published table doesn't show — custom configurations exist.
What fits on a single node
| You are doing | Reasonable starting point |
|---|---|
| Fine-tuning a small model, LoRA, embeddings | RTX 3090 S or M |
| Batch inference, API serving, data prep | RTX 5080 S / M / L |
| 7B–8B class fine-tuning, larger batches | RTX 5090 S or M |
| Bigger batches, long-context work, heavy preprocessing alongside training | RTX 5090 XL or XXL |
The binding limit is usually VRAM, not vCPU — the whole card is yours either way, so move up a tier when the model won't fit, and move up a CPU/RAM config when data loading is starving the GPU.
What to keep in mind
- Local disk is ephemeral. It dies with the instance. Keep datasets and checkpoints somewhere restorable — see Snapshots and Backups and Persistent storage plans.
- One failure domain. A host problem takes the whole instance with it. If you need a run to survive that, use multi-node so work is split, or checkpoint aggressively.
- Bigger is not faster. Two smaller instances can beat one big one for throughput-per-rupee when the workload parallelises. Compare against Multi-Node Instances before assuming you need the largest tier.
- Capacity moves. Confirm what's actually provisionable on the dashboard before planning a deadline around a specific config — Availability.