NVIDIA DGX Station vs ChatGPT, Claude & Gemini: When Does Owning AI Become Cheaper Than Renting It?

Owning becomes cheaper than renting only above a fairly high monthly volume, and only for work an open-weight model can do well. In this illustrative model, a DGX Station costs about S$6,177 a month to own and run over three years. That matches what you would pay for about 1.7 billion tokens a month on a mid-tier frontier API such as Claude Sonnet 5.5, GPT-6.1 Sol or Gemini 4 Argon. Against Claude Opus 5.5 it drops to about 0.84 billion tokens. Against a hosted open-weight model such as gpt-oss-120b it rises to about 24 billion tokens. The DGX Station GB300 systems sold by SourceIT, including the HP ZGX Fury G1n, MSI XpertStation WS300 and ASUS ExpertCenter Pro ET900N G3, are all price on request. The model below assumes S$150,000 for the hardware.

What does each option cost per million tokens?

API prices are quoted separately for input tokens (what you send) and output tokens (what the model writes). To compare like with like, the table blends them at 8 input tokens to 1 output token. That matches the 8K-in, 1K-out workload StorageReview used to measure a DGX Station. Prices are as listed on the vendors' pricing pages on 5 Nov 2026 and converted at USD/SGD 1.2774.

Sources: OpenAI, Anthropic, Google, Together and Fireworks pricing pages, accessed 5 Nov 2026. Throughput is from StorageReview's ASUS ET900N G3 review: gpt-oss-120b, 128 concurrent requests, 45,205 total tokens per second. The DGX Station cost is SourceIT's illustrative model using the assumptions below. These are list prices; cached input, batch discounts and negotiated rates lower the API figures. SourceIT has not tested these systems.

What goes into the S$6,177 a month?

Most of it is the hardware. These are the assumptions behind the model, so you can swap in your own numbers:

• Hardware: S$150,000, written off over three years (S$50,000 a year). This is the ballpark used in this series, not a SourceIT quotation. Public US pricing of roughly US$93,000–125,000 for most OEMs works out to about S$119,000–160,000 before GST at 1.2774. At S$120,000 the monthly cost falls to about S$5,343; at S$180,000 it rises to about S$7,010.
• Electricity: S$2,730 a year. That assumes a 1 kW average draw at SP Group's Q4 2026 tariff of 31.16 cents per kWh including GST. Running at the full 1,600 W around the clock would cost about S$4,367 a year, which is the upper bound.
• People: S$20,000 a year. That is a fifth of an engineer earning S$100,000. Morgan McKinley's 2026 guide puts Singapore DevOps engineers with 5 to 10 years' experience at S$80,000–150,000.
• Power protection: S$4,165 once for an APC Smart-UPS SRT 3000VA (listed 5 Nov 2026).
• Not included: the NVIDIA AI Enterprise subscription if you need it, rack space, network upgrades, support after year three and GST.

A per-token price is only cheaper if the answer is as good. The cheapest token is worthless if someone has to redo the work.

Why can't you just compare token prices?

Because the models are not equal. A DGX Station runs open-weight models such as gpt-oss-120b, DeepSeek or Qwen. It does not run Claude, GPT or Gemini, which you can only reach through their makers' services. On the Artificial Analysis Intelligence Index, the top open-weight model on its open-weights list (DeepSeek V4.1 Flash) scores 33. The closed frontier models score 55 to 60. Arena-style human preference rankings show a much smaller gap at the top, so the difference depends on the task. For classification, extraction, summarising your own documents and routine drafting, an open model is often good enough. For hard reasoning, complex coding or high-stakes writing, a frontier model may produce fewer errors per answer, and that saves staff time.

This is why the fair comparison sits in the middle rows of the table. If an open model can do the job, compare the DGX Station with what you would pay a hosted service to run that same open model. That price is very low, which makes ownership hard to justify on cost alone. If the work needs frontier quality, the DGX Station is not a substitute at any volume.

Where does owning clearly come out ahead?

Ownership tends to win when three things are true together:

1. Volume is high and steady. Billions of tokens a month, spread across the working day or run overnight as batch jobs. See The 24/7 AI Factory.
2. An open model meets the quality bar. Test this on your own tasks first.
3. You would otherwise pay mid-tier or frontier rates. Teams already on Sonnet-, Sol- or Argon-class pricing reach break-even far sooner than teams already on a cheap hosted open model. Google's flagship Gemini 4 Argon is due to double from US$2/US$10 to US$4/US$20 per million tokens on 1 Dec 2026, which halves its break-even volume.

There are also reasons to own that are not about price. These include keeping data on premises, predictable monthly cost and not depending on a vendor's model retirements. Those are covered in Sovereign AI: Why Should Your Company's Intelligence Live on Someone Else's Server? and AI Token Inflation.

Where does renting stay cheaper?

For most companies starting out. Below a billion tokens a month, API bills for all but the top models are smaller than the cost of running a DGX Station. Cloud services also give you the strongest models, no hardware to manage and capacity that scales with demand. Per-seat plans are a different comparison. ChatGPT Business and Claude Team list at US$20 per user per month on annual billing, and Microsoft 365 Copilot at US$30. These bundle apps, connectors and admin controls that a bare model server does not have. One DGX Station or 100 AI Subscriptions compares them in detail. If your volume is modest, a GB10 system such as the NVIDIA DGX Spark (S$9,300, listed 5 Nov 2026) may be the better first step. Our earlier 12-month local vs cloud cost model works through a single 70B model in more detail.

How do you run the numbers for your own company?

Start from real usage, not guesses:

• Export three months of token counts from your API dashboards, split into input, cached input and output. Reasoning tokens are billed as output by OpenAI, Anthropic and Google.
• Note which model each workload uses, and whether a cheaper or open model could handle it.
• Re-price the open-model share using the hosted open-model rate, then compare that with the S$6,177-a-month owned cost, adjusted for your own hardware quote, staffing and power.
• Check that the volume fits. 117 billion tokens a month is a full-load benchmark ceiling. Plan for 30% of it, about 35 billion, in practice.

Our post How Many Tokens Must a DGX Station Generate Before It Pays for Itself? shows the break-even maths under different hardware prices and utilisation levels.

Buying in Singapore

SourceIT lists five DGX Station GB300 systems, all price on request: the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC. Browse them alongside GB10 systems in the NVIDIA DGX Spark & DGX Station collection. Send us your monthly token volumes and we can help you check whether ownership pays off. 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

Can a DGX Station run ChatGPT, Claude or Gemini? No. Those models are only available through OpenAI, Anthropic and Google. A DGX Station runs open-weight models such as gpt-oss-120b, DeepSeek, Qwen and Llama.

At what volume is a DGX Station cheaper than the Claude or GPT API? In SourceIT's illustrative model, around 1.7 billion tokens a month against Claude Sonnet 5.5 or GPT-6.1 Sol list prices, and around 0.84 billion against Claude Opus 5.5. This assumes an 8:1 input-to-output mix and that an open-weight model can do the same work.

Is a DGX Station cheaper than renting the same open model through an API? Only at very high volume. Hosted gpt-oss-120b lists at US$0.15 input and US$0.60 output per million tokens. Matching the owned cost takes about 24 billion tokens a month.

What does it cost to run a DGX Station per month? About S$6,177 in SourceIT's illustrative model. That covers S$150,000 of hardware over three years, S$2,730 a year of electricity, a fifth of an engineer and a UPS. It excludes software subscriptions and GST.

Do API prices only go down? No. Many have fallen, but some rise. Google's Gemini 4 Argon is scheduled to double to US$4 input and US$20 output per million tokens on 1 Dec 2026. Check vendor pricing pages before you decide.