Business Finland
Are GPU rentals (AWS, Lambda, GCP) eligible in a Business Finland R&D budget?
Compute rental is a real line item in AI R&D projects. Here's how to think about eligibility and how to present it in a budget.
Short answer: Yes — GPU rental from AWS, Lambda Labs, GCP, or similar providers is a legitimate eligible cost in a Business Finland AI R&D budget, when the compute is spent on the R&D work itself: training runs, architecture experiments, evaluation sweeps. It sits in the same category as any other consumed R&D resource. The same rule as API costs applies: compute that’s part of running the finished product afterward is operating cost, not R&D, even if it’s the same GPU instance type.
Compute is often the single largest line item in an AI R&D budget, and it’s also the easiest one to get wrong — either padded with vague monthly estimates that don’t survive scrutiny, or left under-budgeted because founders assume “cloud bill” doesn’t sound like a fundable R&D cost.
What makes GPU spend eligible
The test is identical to any other eligible R&D cost: is this resource consumed to resolve a named technical uncertainty, during the period that uncertainty is being resolved?
Eligible: training runs to test a novel architecture or approach; compute for hyperparameter sweeps or ablation studies answering a specific research question; large evaluation runs comparing approaches before a build decision is made; compute for building and validating a dataset pipeline that’s itself part of the R&D.
Not eligible: GPU instances serving your product’s live inference traffic; compute for routine batch jobs that aren’t testing anything, just running an established pipeline; ongoing infrastructure cost once the R&D question has been answered and the system is in production use.
Budgeting compute so it holds up
Vague monthly cloud-spend estimates read as padding to evaluators, because they can’t be tied back to specific R&D work. A defensible compute budget instead:
- Estimates GPU-hours per experiment or training run, not per month
- Names the work package or uncertainty each block of compute serves
- Uses realistic instance pricing (spot vs on-demand, GPU class) rather than a round number
- Shows compute concentrated in the R&D period, tapering or stopping once the uncertainty is resolved — a budget that shows flat GPU spend running indefinitely past the project’s own end date invites the question of whether it’s really R&D
If the project also has a production deployment phase, keep that compute entirely separate — a different budget line, or explicitly outside the funded scope — so there’s no ambiguity about which euros are R&D and which are operations.
A common mistake
Founders sometimes budget compute at the scale they expect to need once the product is live, because that’s the number they actually have in their head. That’s the wrong number for an R&D application. The R&D compute budget should reflect what it costs to run the experiments needed to answer the project’s technical questions — usually much smaller and much more front-loaded than production-scale compute, and that’s fine. Evaluators aren’t looking for the biggest number; they’re looking for a number that matches the actual R&D work described.
FAQ
Does the cloud provider matter (AWS vs GCP vs a specialized GPU host)? No — the eligibility test is purpose, not vendor. Specialized GPU rental providers (Lambda, CoreWeave, etc.) are treated the same as hyperscaler GPU instances.
Can I include compute for a proof-of-concept build inside the R&D budget? Yes, if the proof-of-concept is itself testing the uncertainty named in the application — that’s a normal and expected part of an R&D project, not a separate production build.
What if I can’t predict exact GPU-hours in advance? Use a reasoned estimate based on planned experiments (e.g. number of training runs × expected hours × instance cost), and be ready to explain the assumption. Budgets are estimates, not invoices.
Should I confirm eligible-cost categories before finalizing a budget? Yes — this describes the reasoning pattern that holds up across applications; exact cost-category rules can shift between funding rounds, so confirm current Business Finland guidance before submitting.
The one-sentence version
GPU rental is eligible R&D cost when it’s spent answering the project’s named technical uncertainty, budgeted per experiment rather than per month, and clearly separated from any production compute the project might eventually need.
Related: Eligible costs in a Business Finland R&D application · Are OpenAI and Anthropic API fees eligible R&D costs under Business Finland? · Business Finland R&D application budget: how to structure it