DGX Station GB300 Storage Planning: How Much NVMe Capacity Do AI Teams Actually Need?
It depends far more on what the team does than on the model it runs. An inference-only team can live inside a few hundred gigabytes of model files. A RAG team adds caches, container images and a vector index. A team that runs full fine-tunes is in a different league: each resumable checkpoint of a 70B model is roughly 0.84–1.1 TB under a common rule of thumb, which is more than the usable data space on most DGX Station configurations as shipped. The workable plan is three tiers: fast local NVMe for what the GPU is reading now, a NAS for the model library, datasets and checkpoints, and tape or cloud for archives. The worksheet below gives the line items and figures to start from. SourceIT has not measured a Station's disk usage; the figures come from model repositories, NVIDIA documentation and a published rule of thumb, and our calculations are labelled CALC.
How much capacity does each kind of team need?
| Line item | Inference only | RAG assistant | Fine-tuning | Basis |
|---|---|---|---|---|
| Model library (3 current models) | About 216 GB (gpt-oss-120b, Nemotron 3 Super NVFP4, Llama 3.3 70B FP8) | Same, plus small embedding and rerank models | Base models in BF16 for training: 70B is about 141 GB each | Hugging Face repository sizes; CALC |
| Very large MoE models | GLM-5.2 NVFP4 433 GB; DeepSeek-R1 FP8 689 GB | Rarely needed | Rarely needed | StorageReview; Hugging Face |
| Containers and caches | Tens of GB of images | 200 GB free or more for NVIDIA's RAG Blueprint (NIM models 100–150 GB, images 20–30 GB) | Tens of GB of images | NVIDIA RAG Blueprint support matrix |
| Vector index | None | Depends on corpus and chunking; measure on a pilot | None | No published per-document figure |
| Checkpoints | None | None | Full fine-tune at 12–16 bytes per parameter: 8B 96–128 GB, 32B 384–512 GB, 70B 840–1,120 GB per checkpoint | Rule of thumb from ZeRO accounting; CALC |
| LoRA adapters | None | None | Small next to the base model; size depends on rank and target modules | Not published as a fixed figure |
| Datasets and staging | Prompt logs | Source documents (keep on the NAS) | Training data plus a working copy | Workload-specific |
Source: Hugging Face model repositories and Llama 3.1 blog (23 Jul 2024); StorageReview MSI XpertStation WS300 review (24 Aug 2026); NVIDIA RAG Blueprint support matrix (accessed 9 Nov 2026); ZeRO paper (Rajbhandari et al.). The 12–16 bytes per parameter checkpoint figure is a rule of thumb derived from optimizer-state accounting, not a published specification. CALC = our arithmetic; facts as of 9 Nov 2026.
How much local NVMe do you actually get?
Less than the headline. NVIDIA's reference design has four M.2 slots, with two PCIe Gen5 drives for the operating system in software RAID 1. That mirrored pair is protection against a drive failure, not extra space. The OEMs fill the other slots differently. On HP's configuration of four 2 TB drives with RAID 1 for the OS and RAID 0 for data, about 4 TB is usable for data (CALC). Dell's 16 TB option gives roughly 4 TB mirrored plus 8 TB unmirrored, about 12 TB usable (CALC). Our DGX Station SSD architecture post lists every OEM's layout, so we won't repeat them here.
Two consequences follow. A RAID 0 or unmirrored data volume loses everything if one drive fails, so nothing should live there that isn't also on the NAS. And a single 70B full-fine-tune checkpoint can fill a 4 TB data volume after three or four saves.
Which data belongs on which tier?
Think of local NVMe as the GPU's working area. It holds the model being served, the active dataset and the most recent checkpoint. Everything else lives one tier down.
- Hot, local NVMe: the current model weights, today's training data, the latest checkpoint and container caches.
- Warm, NAS: the full model library, source documents for RAG, all datasets, and checkpoints you might resume from. An all-flash NAS suits teams that reload models often; a hard-drive NAS suits bulk datasets and checkpoints.
- Cold, tape or cloud: finished training runs, final models and datasets you must keep but rarely read. LTO-9 holds 18 TB native per cartridge.
Write the retention rules down before the disks fill. A common pattern is to keep the last two or three checkpoints of a run plus the best one by evaluation score, and to move superseded models to the NAS after a set period. Model weights downloaded from a public repository can be fetched again; your fine-tuned adapters, checkpoints, vector databases and configuration files cannot, so those are what the backup must cover.
Rule of thumb: size local NVMe for what the GPU needs this week, the NAS for what the team needs this year, and the archive for what the company must keep.
How quickly can you move data between tiers?
Local storage is fast. StorageReview measured 13.1 GiB/s reads (about 14.1 GB/s) from a single Gen5 NVMe drive on the MSI WS300 using NVIDIA's GPUDirect Storage test tool. At that rate a 689 GB DeepSeek-R1 FP8 checkpoint loads in under a minute (CALC). Over the Station's 10GbE port, the same file takes about nine minutes at line rate (CALC), and real NAS throughput will be lower. That gap is why hot data is staged locally, and why teams that reload large models often move up to 25GbE or faster links. Our NAS integration guide covers the mounts, link speeds and the GPUDirect Storage caveats.
What should you check before adding drives?
Check the OEM's supported-drive list and warranty terms before fitting your own SSDs; some OEMs ship self-encrypting drives, and replacing them can change support. Match the endurance rating to the job. Checkpointing writes large files repeatedly, so drives used as a training scratch area need a suitable drive-writes-per-day rating; our enterprise vs consumer SSD post explains the ratings. AI demand is also pushing SSD prices up, as our AI storage crisis post explains, so budget the NAS tier at today's quotation rather than last year's.
For the archive tier, our LTO tape post sets out when tape beats disk. For model choice and memory sizing, see the vLLM deployment guide and the fine-tuning guide.
Buying in Singapore
The six DGX Station GB300 systems we list are price on request: the ASUS ExpertCenter Pro ET900N G3, Dell Pro Max with GB300, HP ZGX Fury G1n, MSI XpertStation WS300, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC, each supplied through the OEM's channel. For the tiers around them (listed 9 Nov 2026): the ASUSTOR Flashstor 12 Pro Gen2 all-flash NAS is S$2,312.50; the ASUSTOR Lockerstor 12R Pro Gen2 rackmount NAS is S$6,437.50; Seagate IronWolf Pro drives start at S$567.50 for the 8 TB variant; and HPE LTO-9 cartridges (18 TB native, 45 TB compressed) are S$135 each. If you plan to add an M.2 drive, the Micron 7450 PRO 960GB is S$2,178, but check it against your OEM's supported list first. Prices change; ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
How much storage does a DGX Station GB300 come with? It depends on the OEM configuration, from 4 TB to 16 TB raw across four M.2 slots. Two drives are usually mirrored for the operating system, so usable data space is lower.
How big is a fine-tuning checkpoint? A resumable full-fine-tune checkpoint is roughly 12–16 bytes per parameter by a common rule of thumb, so about 0.84–1.1 TB for a 70B model. LoRA adapters are far smaller.
Is RAID 1 on the Station a backup? No. It protects against one drive failing, not against deletion, corruption or ransomware. Keep copies of checkpoints, adapters and configuration on separate storage.
Do I need a NAS next to a DGX Station? For most teams, yes. Local NVMe is a working area; the model library, datasets and checkpoints belong on shared storage that is backed up.
How fast can the Station load a large model? StorageReview measured about 14 GB/s from one local Gen5 NVMe drive, so a 689 GB checkpoint loads in under a minute. Over 10GbE the same file takes about nine minutes at line rate.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
