How Many Tokens Must a DGX Station Generate Before It Pays for Itself?

It depends on which bill the DGX Station replaces. In SourceIT's illustrative model (S$150,000 of hardware over three years, plus power, part of an engineer and a UPS), ownership costs about S$6,177 a month. To break even it must replace about 1.7 billion tokens a month of Sonnet-class API spend, or about 60 billion tokens over three years. Against Claude Opus 5.5 list prices that falls to about 0.84 billion a month. Against a hosted open-weight model such as gpt-oss-120b it rises to about 24 billion a month, roughly 20% of the system's measured peak capacity. All figures assume an 8:1 input-to-output token mix and that an open-weight model does the job well enough. The DGX Station GB300 systems SourceIT sells are price on request, so the hardware figure here is an assumption, not a quotation.

What is the break-even formula?

Break-even volume is the cost of owning per month divided by the API price you avoid per token:

Monthly tokens to break even = monthly owned cost ÷ API price per token avoided

Two inputs drive it. The first is the owned cost: hardware spread over its useful life, plus electricity, staff time and support. The second is the price of the tokens you would otherwise buy. Because API prices differ by a factor of almost a hundred between the cheapest and dearest models, the second input matters far more than the hardware price.

How many tokens does it take at different hardware prices?

Sources and assumptions: API list prices from Together/Fireworks (gpt-oss-120b US$0.15/US$0.60), OpenAI (Terra US$1/US$5; Sol US$2/US$10), Anthropic (Sonnet 5.5 US$2/US$10; Opus 5.5 US$4/US$20) and Google (Gemini 4 Argon US$2/US$10 until 30 Nov 2026), accessed 5 Nov 2026. These are blended at 8 input tokens to 1 output token and converted at USD/SGD 1.2774. The owned cost covers the hardware over three years, 1 kW average power at 31.16 cents per kWh (S$2,730 a year), S$20,000 a year of engineer time and a S$4,165 UPS spread over three years. "% of peak" compares against about 117 billion tokens a month, which is StorageReview's measured 45,205 tokens per second (gpt-oss-120b, 8K in / 1K out, 128 concurrent, ASUS ET900N G3) held for 30 days. Illustrative only.

The hardware price moves the break-even point by a few billion tokens. Choosing which API price you compare against moves it by a factor of thirty.

What does a token cost on your own hardware?

Turn the formula round and the owned cost per token depends entirely on how busy the system is. At S$6,177 a month:

• 10% utilisation (about 11.7 billion tokens a month): about S$0.53 per million tokens. That is twice the hosted gpt-oss-120b price, but well below Terra or Sonnet-class prices.
• 30% utilisation (about 35 billion tokens): about S$0.18 per million. That is below the hosted open-model price.
• 50% utilisation (about 59 billion tokens): about S$0.11 per million.

These are benchmark-based ceilings. Real workloads with longer prompts, larger models or uneven demand will see less throughput, and some hours will be idle. For a 120B-class model, 30% of the measured peak is a reasonable planning figure. How Many AI Tokens Can DGX Station Actually Generate? explains why peak benchmarks overstate everyday throughput, and The 24/7 AI Factory covers filling the idle hours.

What can push break-even further out?

• API discounts. Cached input is charged at 0.05 to 0.1 times the normal input rate on most current models, and batch processing is 50% off at OpenAI and Anthropic. A workload that is half batch roughly halves the avoided price, which doubles the break-even volume.
• A bigger or slower model. Larger open models run slower. StorageReview measured 954 output tokens per second (8,587 total) on DeepSeek V4 Flash at FP8 against 5,023 output on gpt-oss-120b at the same 8K/1K shape. Fewer tokens per second means a higher owned cost per token.
• More staff time. If running the system takes half an engineer rather than a fifth, add about S$2,500 a month at S$100,000 a year.
• Software licences. NVIDIA AI Enterprise is a separate subscription. Open-source stacks such as vLLM cost nothing to license but need more in-house skill.

What can bring break-even closer?

• Output-heavy work. At a 1:1 input-to-output mix, Sonnet-class list prices blend to about S$7.66 per million tokens rather than S$3.69. That halves the break-even volume to about 0.8 billion a month.
• Agent workloads. Agents re-read context at every step, so token counts climb quickly. See The Hidden Cost of AI Agents.
• Price rises. Google's Gemini 4 Argon is scheduled to double to US$4/US$20 per million tokens on 1 Dec 2026, which halves its break-even volume.
• Longer life. Spreading the hardware over four years instead of three cuts about S$1,042 a month at S$150,000, though support costs after year three need pricing in.

Is a full DGX Station the right comparison?

Not always. If your volume is far below break-even, a smaller system may pay back sooner. StorageReview measured an NVIDIA RTX PRO 6000 Blackwell Workstation Edition (S$29,100, listed 5 Nov 2026) at 3,064 output tokens per second on gpt-oss-120b at 128 concurrent requests (512 in / 512 out). The DGX Station managed 9,280 on the same test, about three times as much. The NVIDIA DGX Spark managed 500. The Station also has far more memory for larger models and longer contexts. NVIDIA GB10 vs RTX PRO 6000 Blackwell and GB10 vs GB300 cover the step-ups.

How should you use these numbers?

1. Measure three months of real token use by workload: input, cached input, output and reasoning.
2. Mark which workloads an open-weight model could take over, and test quality on real tasks.
3. Apply the formula with your own hardware quotation, staffing and power figures.
4. Check that the required volume is below 30% of measured capacity for the model you plan to run.

For the full three-year view including staff productivity, see A S$150,000 AI Workstation: How Much Productivity Must It Generate?. For a comparison with renting GPUs by the hour, see Buy DGX Station or Keep Paying for Cloud GPUs?.

Buying in Singapore

SourceIT lists the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC, all price on request. Smaller systems are in the AI Workstations & AI Servers collection. With your quotation in hand, you can plug the real hardware price into the calculator above. 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 many tokens does a DGX Station need to pay for itself? In SourceIT's illustrative model, about 1.7 billion tokens a month replacing Claude Sonnet 5.5 or GPT-6.1 Sol list-price spend, about 0.84 billion replacing Claude Opus 5.5, and about 24 billion replacing hosted gpt-oss-120b. That assumes an 8:1 input-to-output mix.

What is the break-even formula? Monthly owned cost divided by the API price per token you avoid. Owned cost includes hardware over its life, electricity, staff time and support.

What does a token cost on a DGX Station? It depends on utilisation. At about S$6,177 a month, it works out to about S$0.53 per million tokens at 10% of measured peak, about S$0.18 at 30% and about S$0.11 at 50%, running gpt-oss-120b.

Do API batch discounts change the answer? Yes. OpenAI and Anthropic batch APIs are 50% cheaper, which roughly doubles the volume a DGX Station must replace to break even.

Is the S$150,000 figure SourceIT's price? No. It is the planning assumption used in this series. SourceIT quotes DGX Station systems on request. Public US pricing of US$93,000–125,000 for most OEMs is about S$119,000–160,000 before GST.