Quick answer: Cloud and GPU spend is not claimable or unclaimable as a category. Section 355-205 of the ITAA 1997 allows a notional deduction “to the extent that” expenditure is incurred on registered R&D activities, so the claim is the eligible proportion supported by contemporaneous records, which may include activity-level accounts, resource tags, run logs and a documented apportionment method. You self-assess.
As at 27 July 2026, the Australian Government had announced reforms to the R&D Tax Incentive in the 2026–27 Federal Budget, intended to apply to income years starting on or after 1 July 2028. Until those changes take effect, the program continues to operate under the current legislation.
For an AI company, compute can be a significant cost, and it may arrive as one undifferentiated invoice from a vendor that does not identify which workloads related to experiments. Cloud expenditure is sometimes described broadly as claimable. Put flatly, that is wrong — nothing is claimable as a category. The real question: of the six figures your provider billed you last year, how much can you show was incurred on the activities you registered? Whether the work is R&D at all is covered in whether the AI work is eligible in the first place.
The Word Doing All the Work Is "To the Extent"
Section 355-205(1) of the Income Tax Assessment Act 1997 lets an R&D entity deduct expenditure it incurs "to the extent that the expenditure … is incurred on one or more R&D activities … for which the R&D entity is registered under section 27A of the Industry Research and Development Act 1986" (ITAA 1997). Three consequences follow:
1. There is no category answer.
The provision applies a fraction to expenditure you incurred: whatever share was incurred on registered activities. A GPU hour on a registered experiment and a GPU hour serving paying customers sit on the same invoice and are treated differently.
2. The activity comes first; registration is applied for afterwards.
You define, conduct and document the activity during the income year; the registration application comes after the year ends, within "10 months after the end of the income year" or "a further period allowed by the Board" (s 27D, IR&D Act 1986). Registration is still a logical precondition, since s 355-205 reaches only registered activities. So you cannot attribute compute to an activity with no boundary, and the boundary has to exist while the GPU hours are burning. If the “project” in your tags is a product line rather than an identified core activity with a hypothesis and a defined activity boundary, the tags cannot establish the required link.
3. Renting is not buying.
Section 355-225(1)(b) says s 355-205 does not apply to "expenditure included in the cost of a tangible depreciating asset"; that asset's decline in value may instead be notionally deductible under s 355-305. Which limb applies to a reserved-instance commitment or hardware lease is a question for your tax agent.
An Attribution Method You Can Actually Evidence
business.gov.au expects records showing "which R&D activities the expenditure relates to" and "documents that support any apportionment methods used", notes "the most accurate records are created at the time the activities are conducted", and asks that they be kept five years (Record keeping).
Method note: The account separation, tagging and apportionment practices described below are evidence and governance practices. They are not all express statutory requirements. The statutory eligibility tests remain those in Division 355.
Cloud platforms can provide timestamped cost data at resource level, but companies need to configure their systems to connect that data to the relevant activities:
1. Define the experiment boundary before you tag anything. Write down the core activity: the technical unknown, the hypothesis, the planned experiments, start and end dates, and an identifier. Everything downstream keys off that identifier.
2. Separate the accounts, not just the tags — as a matter of practice. A dedicated AWS account, GCP project or Azure subscription for each identified R&D activity gives a boundary that survives a refactor, a reorganisation and an engineer who forgets. This is a governance recommendation, not a legal requirement — nothing in Division 355 asks for a particular account structure — but it makes the evidence cheap.
3. Tag at resource level, against activity IDs — not team names. Enforce a mandatory tag policy (rdti-activity=CORE-2026-01, rdti-role=experiment|production|shared) through infrastructure-as-code, so untagged resources cannot be created. Then adopt a conservative default: treat untagged spend as unclaimed unless other contemporaneous evidence supports it. That default is a policy choice, not a rule of law — missing tags do not automatically disqualify expenditure, and where tagging is incomplete other contemporaneous records such as run logs, deployment history and engineer time records may still support a defensible apportionment. A conservative policy of excluding untagged expenditure may reduce attribution risk, although other contemporaneous evidence may still support an apportionment.
4. Tie run logs to the activity. This is the bridge between "this was an experiment" and "this was the invoice". Your tracker already records a run ID, start and end time, configuration and result; store the cluster, job or instance ID alongside it. You then have a chain — registered activity → experiment run → compute resource → billed line item — that a reviewer can walk in either direction.
5. Write the apportionment method down while the year is still open. One page per shared cost pool: the driver, why it measures use, and the source of the numbers. A method chosen in advance reads differently from a percentage produced in June.
Worked Example: What the Attribution Is Actually Worth
Cortex Labs is a hypothetical Adelaide AI company: aggregated turnover $2.1 million, base-rate entity at 25%, not controlled by any exempt entity. Its FY2025-26 cloud bill was $312,000 on one account; mid-year it introduced account separation, tagging and run-log linkage.
Cost pool
FY spend
Attributed to registered R&D
Training and evaluation runs, tagged and matched to run logs
$86,000
$86,000
Training runs from the four months before the tag policy went live
$12,000
nil — could not be evidenced
Production inference serving paying customers
$141,000
nil
Shared data platform, apportioned at 31% (measured storage and query volume)
$46,000
$14,260
Shared dev, staging and CI, apportioned at 38% (recorded engineer time)
$27,000
$10,260
Total
$312,000
$110,520
Attributed spend totals $86,000 + $14,260 + $10,260 = $110,520. Assuming the full amount gives rise to eligible notional R&D deductions, the current refundable offset rate for this hypothetical 25% base-rate entity is 43.5% (ATO, Rates of R&D tax incentive offset), producing an illustrative gross offset of approximately $48,076. Figures are illustrative; see refundable vs non-refundable offset.
The second line highlights the importance of contemporaneous evidence. The additional $12,000 of compute expenditure may have been incurred in connection with the R&D activity, but Cortex Labs could not establish a sufficient evidentiary link because the relevant run logs predated the implementation of instance-level tracking. The amount was therefore excluded from the claim. On the same assumptions, this represents an illustrative gross offset difference of approximately $5,220.
Where the Compute Physically Ran: An Open Question
Section 355-210 applies conditions concerning where the R&D activity is conducted. Expenditure on R&D activities conducted overseas generally requires a positive Overseas Finding. Current software and AI guidance does not expressly determine the effect of using overseas cloud infrastructure where the people designing and directing the experiment are in Australia. Server location should therefore not be assumed to determine, or to be irrelevant to, the answer on its own. Record the compute region and where the relevant work was performed, and confirm the position with your registered tax agent.
Third-Party Model APIs and the Core Technology Exclusion
A related issue may arise where expenditure is incurred to acquire, or acquire the right to use, a third-party model or other technology. Section 355-225(2) may exclude expenditure incurred in acquiring technology, or the right to use technology, where the R&D activities are conducted for the purpose of obtaining new knowledge based on that technology, or creating products, processes or services based on that technology. Whether model API expenditure falls within this exclusion depends on the rights acquired, the nature of the technology involved and the purpose of the R&D activities. It should not be determined solely by characterising the API as either a tool or the foundation on which the experiment is built (what does not qualify).
When Compute Is Nearly Your Whole Claim
Where an AI company’s eligible cost base consists largely of founder time and compute, its total notional deductions may fall below $20,000. Where total notional deductions are below A$20,000, the offset base is generally limited to qualifying expenditure incurred to a non-associate RSP for services within a field for which it is registered, together with eligible CRC Program contributions. Other in-house amounts do not automatically form part of that below-threshold offset base — including your own cloud bill.
The RSP route is the carve-out: RSP-conducted eligible R&D activities can be claimed even where the usual $20,000 R&D expenditure threshold is not met, where the provider is not an associate and the services fall within a research field the provider is registered for (s 355-100(2), ITAA 1997). business.gov.au adds the qualification that matters: using an RSP does not guarantee eligibility — you still self-assess (Get help from a research service provider). Mechanics: claiming R&D under $20,000.
The Australian Government announced reforms to the R&D Tax Incentive in the 2026–27 Budget, intended to apply to income years starting on or after 1 July 2028. Until those changes take effect, the program continues to operate under the current legislation. See our dedicated Budget update for the proposed measures and their status.
Frequently Asked Questions
Q: Can I claim AWS, Azure or GCP costs — including GPU training hours — under the R&D Tax Incentive?
A: Cloud and GPU spend can form part of a claim, but not as a category. Section 355-205 allows a notional deduction "to the extent that" expenditure is incurred on registered R&D activities, so what goes in is the evidenced proportion. Being expensive is not the test, and production hosting is not experimental compute. Keep run logs with job IDs matched to billing data, and self-assess.
Q: Can I claim OpenAI, Anthropic or other third-party model API costs as R&D?
A: The same "to the extent" analysis applies to model API spend as to any other cloud resource — nothing follows from the name on the invoice. It turns on whether the underlying activity is registered and eligible, and on the evidence linking the spend to it. Apportionment matters: experiment calls sit differently from production traffic, and billing exports can usually be split by API key or project. Section 355-225(2) may also apply — expenditure incurred in acquiring, or acquiring the right to use, technology may fall within the core-technology exclusion where the statutory conditions in s 355-225(2) are met. Isolate the line, keep run-level evidence, and put the position to your tax agent.
Q: How do I apportion cloud costs between R&D and production?
A: Use a driver that measures use, decide it before the year ends, apply it consistently and document it. Common drivers are tagged resource-level spend, GPU hours by job, measured storage and query volume, and recorded engineer time. business.gov.au expects "documents that support any apportionment methods used". Treating untagged spend as unclaimed is a sound default, though other contemporaneous records may still support an apportionment where tags are missing.
Q: Does it matter if my model trains on overseas servers?
A: Section 355-210 imposes conditions concerning where the R&D activity is conducted. R&D activities conducted outside Australia generally require a positive Overseas Finding under the applicable provisions of the IR&D Act 1986. The location of cloud infrastructure or compute resources should not, by itself, be assumed to determine whether an R&D activity is conducted overseas or in Australia. The relevant facts may include where the experimental work is designed, directed, performed and analysed. Record the compute region, the location of the relevant personnel, and the connection between the infrastructure used and the R&D activity for each run. Confirm the position with your registered tax agent.
Sources & Further Reading
Income Tax Assessment Act 1997, ss 355-100, 355-205, 355-210, 355-225 — legislation.gov.au
Industry Research and Development Act 1986, ss 27A, 27D, 28C — legislation.gov.au
Record keeping for the R&D Tax Incentive — business.gov.au
Rates of R&D tax incentive offset — ato.gov.au
Get help from a research service provider — business.gov.au
Software development sector guide — business.gov.au
AI-related activities and the R&D Tax Incentive — business.gov.au
Talk to Ignition Research before relying on a compute-heavy R&D budget. Defining the experimental activity and establishing contemporaneous attribution records before or during the work can support a clearer link between eligible activities and cloud expenditure. Get in touch.
This article is general information from a Registered Research Service Provider about the R&D Tax Incentive. It is not tax, legal or financial advice; eligibility depends on your circumstances and you should self-assess and seek your own advice.
Thinking about a project like this?
If you're weighing up an AI, software or technical improvement project and can't tell yet whether it's implementation or research, start with a quick read on where it sits.