Buying, leasing or renting GPUs for private AI in Australia

Rent first. GPU instances in an Australian cloud region let you measure real load before you sign for hardware. Buy or lease only when the measured load keeps the GPUs busy for a large share of every hour across their useful life.

The break-even point is a utilisation ratio. You can calculate it from two quotes: the hardware price per GPU and the hourly cloud rate per GPU. This article shows the arithmetic and lists the GPU instances each provider offers in Australia as of October 2026. It also sets out the contract terms to check.

Three ways to get GPUs

You can buy servers and run them in your own data centre or in a colocation facility. You carry the capital cost, the spare parts and the risk that a newer GPU generation makes the old one less useful.

You can lease the same servers from a vendor or a financier. A lease spreads the cost over the term. It does not keep the asset off the balance sheet. AASB 16 requires a lessee to recognise assets and liabilities for all leases with a term of more than 12 months. The only exception is an asset of low value.

You can rent GPU instances from a cloud provider. On-demand rent has no commitment. AWS Savings Plans give a lower rate for a commitment of using a specified amount of compute power. The term is one or three years.

GPU instances in Australian regions

Each table row comes from the provider’s own region list as of October 2026. A family in the list does not guarantee capacity in your account. AWS notes that an instance type in a Region is not always supported in all of the Availability Zones of that Region. Every provider also applies quotas.

Sources for the table: the AWS instance types by Region page and the AWS accelerated computing specifications. For Azure, the products by region table and the size pages for NCads H100 v5, ND H100 v5 and ND H200 v5. For Google Cloud, the GPU regions and zones page and the accelerator-optimised machine family page.

Two gaps stand out. The AWS list shows no G6e family in Sydney, and G6e is the AWS family with the NVIDIA L40S GPU. AWS Melbourne shows no NVIDIA GPU family at all. A design that needs NVIDIA GPUs in two Australian AWS regions for disaster recovery does not work on the current lists.

AWS also sells Capacity Blocks for ML, which reserve GPU instances for a fixed period. As of October 2026, the Capacity Blocks table lists p5.4xlarge, p5.48xlarge, p5e.48xlarge and p5en.48xlarge in Asia Pacific (Sydney). You can reserve a block with a reservation start time up to eight weeks in the future. AWS does not allow you to cancel a Capacity Block after purchase. Blocks suit a fixed job such as a fine-tuning run or a load test before a hardware purchase.

Break-even utilisation

Owned hardware costs the same whether it is busy or idle. Rented hardware costs money only while it runs. The break-even is the number of rented hours per year that costs the same as one owned GPU for one year.

Write every cost in hours of on-demand rent for one GPU. This removes the currency and the price level from the comparison.

K  = hardware price per GPU / on-demand rate per GPU-hour   (hours of rent)
L  = useful life of the hardware in years
O  = yearly running cost per GPU (power, colocation, support, staff share) / rate
Break-even utilisation = (K / L + O) / 8,760 hours

Take K from your own quotes. Divide the server price by the number of GPUs in it. Then divide by the per GPU hourly rate on the AWS On-Demand pricing, Azure Linux virtual machine pricing or Google Cloud pricing page. The table below assumes a useful life of 4 years and running costs equal to 1,000 hours of rent per year. Both are assumptions for the example. Replace them with your own figures.

Compare the break-even with the load you expect. A chat assistant that staff use 10 hours a day, 5 days a week, runs 50 x 52 = 2,600 hours a year. That is 2,600 / 8,760 = 29.7% of the year. At that load, ownership pays back only when K is low or when batch work fills the nights. A committed cloud rate lowers the hourly rate, which raises K and moves the break-even up. The full comparison, including hosted APIs, is in private AI versus API cost.

Power, cooling and colocation

The H100 SXM has a maximum thermal design power of up to 700W (configurable). An 8 GPU DGX H100 system lists system power usage of ~10.2kW max. The DGX H100 user guide gives a weight of 130.45 kg max and an 8U rackmount form factor. Its operating temperature range is 5 to 30 degrees Celsius.

Energy at maximum draw, one DGX H100:  10.2 kW x 8,760 h        = 89,352 kWh per year
Facility energy at an assumed PUE 1.4: 89,352 x 1.4             = 125,093 kWh per year
Per GPU:                               125,093 / 8              = about 15,600 kWh per year
Floor load, four systems in one rack:  4 x 130.45 kg            = about 522 kg

PUE (power usage effectiveness) is the energy entering the facility divided by the energy used by the IT equipment. Ask the facility for its own figure and for the tariff you will pay, then multiply. The draw at maximum is a ceiling. The real draw depends on load, but the power contract and the cooling must cover the peak.

PCIe servers are easier to place. The H100 NVL draws 350-400W (configurable) and the L40S lists 350W max power consumption. A server with four L40S cards draws 4 x 350 W = 1.4 kW for the GPUs, plus the CPUs, memory and fans.

For colocation, ask each provider for the contracted kW per rack, the cooling type and the floor loading. The Hosting Certification Framework lists certified facilities from providers that include AirTrunk, Equinix, NEXTDC, Macquarie Technology Group and Canberra Data Centres. The framework is currently undergoing reform, and the list names the certified facilities of each provider. Check that the exact facility in your quote is on it.

Comparison table

Contract risks

Check these terms before you sign:

1. Cloud commitments. A savings plan or reservation commits you to pay for the term whether you use it or not. Capacity Blocks cannot be cancelled.
2. Capacity. On-demand GPU capacity in Sydney is not reserved for you. Ask the provider for quota increases in writing before the go-live date.
3. Hardware delivery. Put the delivery date, the model numbers and a remedy for late delivery in the purchase order.
4. Support and spares. Name the response time for a failed GPU, who holds spare parts in Australia and how long firmware updates continue.
5. Lease end. Set out the return, buy-out and refresh options and the early termination fee.
6. Data on returned drives. Servers hold cached documents, embeddings and logs. APP 11 says an entity has obligations to destroy or de-identify personal information in certain circumstances. Require certified destruction of drives before any server leaves your control.
7. Colocation terms. For an APRA regulated entity, a colocation provider that supports a critical operation can be a material service provider. CPS 230 requires a formal legally binding agreement for all material arrangements. For a critical operation, APRA needs notice not more than 20 business days after entering into or materially changing an agreement.
8. Billing currency. Check the currency in every quote. A price in US dollars moves with the exchange rate for the life of the contract.

What to decide next

Measure your load on rented GPUs first, with the sizing method in sizing GPU hardware for LLM inference. Calculate K from real quotes and compare the break-even with the measured utilisation. Then choose the topology in private LLM deployment architecture, because an on-prem build needs a facility and a VPC build needs regional capacity.

Our AI strategy and readiness audit sets out the load, the costs and the options. Our on-prem AI deployment service builds on the option you choose.