One AI Supercomputer, Multiple Departments: How Should Companies Charge for AI Usage?

Meter usage at a single gateway, show each department its share for a few months before billing it, then charge a blend: a small fixed fee for reserved capacity plus a rate per million tokens set from realistic utilisation. The rate is the tricky part. In our illustrative model, a DGX Station GB300 costs about S$5,700 to S$7,000 a month to own and run. Spread over the output tokens departments actually use, that works out to anywhere from about S$0.90 to S$5.40 per million output tokens, depending on whether the system is half-used or barely used. Set the rate too high and teams drift back to personal AI subscriptions. Set it too low and finance sees an unrecovered cost.

What cost are you actually recovering?

A monthly cost pool. Ours uses stated assumptions you should replace with your own figures:

Sources: SP Group tariff (Q4 2026); Morgan McKinley Singapore Salary Guide 2026; hardware figure is an assumption. SourceIT quotes DGX Station GB300 systems on request. Public US pricing of roughly US$93,000 to US$125,000 for most OEMs is about S$119,000 to S$160,000 before GST at USD/SGD 1.2774 (5 Nov 2026). Arithmetic is ours and illustrative. We have not tested these systems.

How much does utilisation change the internal rate?

Enormously. StorageReview measured 5,023 output tokens per second with gpt-oss-120b on the ASUS ET900N G3 at 128 concurrent long-prompt requests (8,192 tokens in, 1,024 out). Running flat out for a 30-day month, that is about 13 billion output tokens. Nobody uses a shared system flat out, so the rate depends on how much of that capacity departments really consume:

• 10% utilisation (about 1.3 billion output tokens a month): S$4.40–5.40 per million output tokens
• 25% utilisation (about 3.3 billion): S$1.76–2.16 per million
• 50% utilisation (about 6.5 billion): S$0.88–1.08 per million

Each of those output tokens arrives with roughly eight input tokens in this workload, so the rate covers both. For comparison, hosted gpt-oss-120b on Together lists at US$0.15 input and US$0.60 output per million tokens (about S$0.19 and S$0.77), while Claude Sonnet 5.5 lists at US$2 and US$10 (about S$2.55 and S$12.77), as listed on 5 Nov 2026. The internal rate only beats a hosted open model at high utilisation, which is why the case for owning usually rests on data control and predictability as well as price. Our post How Many Tokens Must a DGX Station Generate Before It Pays for Itself? works through the break-even in detail.

An internal AI rate is a promise about utilisation. If departments don't use the capacity, someone still pays for it.

Which charging model should you use?

Four common approaches, often combined:

1. Fixed allocation. Split the monthly pool by an agreed share. In a 50/20/10/20 split between customer service, finance, legal and engineering, the S$5,727 low-case pool becomes S$2,864, S$1,145, S$573 and S$1,145. This is simple and predictable, but there is no incentive to use the system efficiently.
2. Metered usage. Charge per million tokens at the internal rate. This is fair and familiar to anyone who has used a commercial API, but monthly bills swing, and early adopters pay for the platform while others wait.
3. Reserved capacity. Sell MIG slices. The GB300 supports up to seven isolated instances, so an always-on workload such as a customer-service model can own a slice outright. If the pool were split evenly over seven slices, each would carry S$818–1,004 a month. In practice, price by slice size, since a 2g instance has twice the compute of a 1g instance.
4. Base fee plus usage. A small reserved fee covers fixed cost, and a per-token rate covers variable use. This is usually where companies end up.

Many firms start with showback, reporting each department's usage and notional cost without billing it, for one or two quarters. That gives real usage data to set the rate from, rather than a guess. Our post Can Seven AI Teams Share One DGX Station? explains how MIG slices are sized and isolated.

What tools do the metering?

The gateway does most of it. Every request should pass through one front door that knows which team sent it:

• LiteLLM proxy issues virtual keys per user or team, tracks spend per key automatically (a database is required), and enforces max_budget with a budget_duration such as 30 days. Once a key exceeds its budget, requests fail.
• Kong AI Gateway can limit by tokens or by cost per consumer group or model, though its token rate limiting is an Enterprise feature.
• Azure API Management's llm-token-limit policy sets hourly to yearly token quotas, which is useful if part of your AI traffic goes to Azure-hosted models.
• NVIDIA Run:ai works at the GPU level. Projects and departments get "deserved" quotas and can borrow idle capacity with preemptible jobs.

The FinOps Foundation's guidance on generative AI recommends "token proxies" in front of APIs to allocate the cost of shared resources, plus limits on individual or team consumption. It also notes that inference often accounts for 80–90% of total cost of ownership. A gateway in front of both the Station and any commercial APIs gives finance one view of total AI spend, which sits alongside the subscription tracking we described in Who's Actually Tracking Your IT Subscriptions?

What goes wrong with AI chargeback?

• Charging too early. Departments trying AI for the first time see a bill and stop, leaving the Station idle and the rate higher for everyone else.
• Ignoring idle capacity. Night-time capacity is free to use but still costs money. Offer it at a discount for batch jobs; see The 24/7 AI Factory.
• Treating all tokens as equal. Long retrieval prompts, agent loops and reasoning models consume far more compute per answer. Meter input and output separately.
• No priority rules. Decide in advance whose work is pre-empted when the system is busy, and write it into the rate card.
• Forgetting shared services. A company-wide document assistant, such as the one in From 10,000 Documents to One Private AI Brain, may be better funded centrally than charged per query.

For the platform side, see Can NVIDIA DGX Station Become the Foundation of Your Company's Private AI Cloud? and The Most Expensive Computer Your Employees Might Never Touch.

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. For departments that want their own dedicated box instead of a shared slice, GB10 systems such as the HP ZGX Nano G1n at S$11,499.00 and the NVIDIA DGX Spark at S$9,300.00 are in stock (listed 5 Nov 2026). Browse the AI workstations and AI servers collection. 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

What is a fair internal price per million tokens? It depends on utilisation. In our model the monthly cost pool is S$5,727–7,031, which works out to S$0.88–1.08 per million output tokens at 50% utilisation and S$4.40–5.40 at 10%.

Should we bill departments from day one? Usually not. A quarter or two of showback, reporting usage without billing, gives real data to set rates from and avoids discouraging early adoption.

Can MIG slices be charged as reserved capacity? Yes. The GB300 supports up to seven isolated GPU instances. Price them by size, since larger profiles carry more compute and memory.

What software tracks AI usage per department? An LLM gateway such as LiteLLM tracks spend per key or team and enforces budgets. Kong and Azure API Management offer token limits, and NVIDIA Run:ai manages GPU quotas.

Does the cost pool include software licences? Ours does not include NVIDIA AI Enterprise, which is a separate subscription if you use it, or support extensions beyond the warranty. Add both if they apply.