The 24/7 AI Factory: Can DGX Station Turn Idle Hours Into Business Value?

Yes, if you have work that can wait until morning. A 45-hour office week covers about 27% of the 168 hours in a week, so a DGX Station used only for daytime chat sits idle roughly 73% of the time. Filling those hours with batch jobs is the easiest way to raise its value without buying anything else: overnight document processing, embedding generation for search, model evaluation, transcription and report drafting. Measured throughput shows the scale. StorageReview recorded 5,023 output tokens per second (45,205 total tokens per second) with gpt-oss-120b on an ASUS ExpertCenter Pro ET900N G3 (DGX Station GB300) at 128 concurrent requests of 8,192 input and 1,024 output tokens. Over a 12-hour night, that is a theoretical ceiling of about 217 million output tokens. Real jobs reach only part of it.

How much batch work fits into one night?

Source: StorageReview's ASUS ExpertCenter Pro ET900N G3 review (vLLM, gpt-oss-120b MXFP4, 8,192/1,024 tokens, 128 concurrent: 5,023 output tok/s). Our arithmetic multiplies that rate by 43,200 seconds; documents per night assume one 8,192-token request per document. Electricity uses SP Group's Q4 2026 tariff. These are illustrative ceilings from one benchmark configuration. Your model, prompt lengths, retrieval steps and job orchestration will change them, so the 50% column is a planning assumption, not a measurement.

What kinds of work suit overnight runs?

• Document processing. Classifying, extracting fields from and summarising contracts, invoices, emails or reports that arrived during the day. Results are ready when staff log in.
• Building the search index. Generating embeddings for new or changed documents so internal search stays current. See From 10,000 Documents to One Private AI Brain.
• Model evaluation. Re-running a fixed test set whenever you change a model, a prompt or a retrieval setting, so quality regressions show up before users find them.
• Transcription and minutes. Processing the day's recorded meetings in one queue. Our post Can a GB10 Transcribe 100 Hours of Meetings? looks at the smaller GB10 tier.
• Drafting at volume. First drafts of product descriptions, reply templates or report sections for people to review the next day.
• Agent jobs that don't need a human present. Scheduled research, reconciliation or monitoring tasks, with results queued for approval. Can a GB10 Run AI Agents 24/7 Without Anyone Watching? covers the safeguards.

The cheapest token your DGX Station will ever produce is the one it generates at 3am, when it would otherwise be idle.

What would the same night cost through an API?

It depends heavily on which model you compare against. Here is the 50% scenario above (about 868 million input and 108.5 million output tokens) priced at list rates as listed on 5 Nov 2026:

The comparison is honest in both directions. For the same open-weight model, hosted APIs are cheap, about S$125 to S$250 a night in this scenario, so owning a DGX Station purely to batch-process gpt-oss-120b is hard to justify on token cost alone. The case for local batch work rests on three other things: data that shouldn't leave your premises, capacity you have already paid for and would otherwise waste, and freedom from rate limits and spending caps on very large jobs. Frontier models produce different quality, so they aren't a like-for-like substitute either way. Prices change, and Gemini 3.8 Flash's batch rate is scheduled to double on 1 Jan 2027.

How do you share the machine between day and night?

Scheduling is the practical part. Common approaches include:

1. A job queue with time windows. Interactive traffic gets priority during office hours. Batch jobs start from a queue at a set time and stop or yield before people arrive.
2. A scheduler that understands GPUs. Kubernetes with NVIDIA's GPU operator, or NVIDIA Run:ai, can give interactive projects a guaranteed quota and let batch jobs borrow idle capacity as preemptible work.
3. GPU partitioning. Multi-Instance GPU can split the GB300 into up to seven isolated instances, so a small interactive slice stays available overnight. Re-partitioning needs the affected instances idle, and Slurm can't change MIG layouts on the fly, so most teams set a day profile and a night profile. Our post on sharing one DGX Station between seven teams covers the limits.
4. Metering. If departments share the night shift, log tokens per job so costs can be allocated. See How Should Companies Charge for AI Usage?

What about power, heat and noise overnight?

Running flat out at 1.6 kW for 12 hours costs about S$5.98 at the Q4 2026 tariff, so electricity isn't the constraint. Heat and noise are. A system working hard all night needs cooling that stays on after office hours, and a UPS sized for its load. The APC Smart-UPS SRT 3000VA is S$4,165 (listed 5 Nov 2026). Our post Can You Really Put a Data Centre-Class AI System Under Your Office Desk? covers the room. Overnight output also needs somewhere to land; a NAS such as the Synology RackStation RS822RP+ (S$2,086.80, listed 5 Nov 2026) keeps results and backups off the station itself.

For the measured throughput figures behind this article, see How Many AI Tokens Can DGX Station Actually Generate?. For the payback maths, see How Much Productivity Must It Generate to Justify the Investment?

Buying in Singapore

SourceIT lists the ASUS ExpertCenter Pro ET900N G3, HP ZGX Fury G1n, MSI XpertStation WS300, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC, all price on request, along with UPS and NAS options for round-the-clock operation. Tell us your daytime user count and the batch jobs you want to run overnight, and we will help size the system and its supporting kit. For a 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 of the week is a DGX Station idle? If it serves only a 45-hour office week, about 73% of the week's 168 hours. Batch jobs scheduled outside office hours can use that time.

How many documents can it process overnight? Based on StorageReview's measured 5,023 output tokens per second with gpt-oss-120b at 8K in / 1K out, a 12-hour theoretical ceiling is about 212,000 documents of that size. At a planning assumption of 50% utilisation, about 106,000.

Is overnight local processing cheaper than an API? Compared with the same open-weight model through a hosted API, often not by much: about S$125 to S$250 a night in our scenario. The value comes from using capacity you already own, keeping data on-premises, and avoiding rate limits, rather than from token price alone.

Can interactive users still use it at night? Yes. A scheduler with guaranteed quotas, or a small MIG slice reserved for interactive use, keeps some capacity available while batch jobs take the rest.

What does running it all night cost in electricity? At most about S$5.98 per 12-hour night at 1.6 kW and the Q4 2026 tariff of 31.16¢/kWh with GST.