Quick answer

A GPU server in an R&D project is planned backwards: research tasks first, then workloads (training, inference, data), and hardware last. Under the SMART Path for SMEs in 2026, equipment for R&D work is not reimbursed at its purchase price but through depreciation charges for the period and share in which it serves the project. The cost has to arise after the application is submitted, and a purchase above PLN 80,000 net goes through the competitiveness principle, with a spec written in parameters rather than model names. On top of that come costs a grant usually doesn't cover: power, cooling, rack space and upkeep through the project's whole durability period.

This is not legal or financial advice. Whether a specific expense is eligible is decided by the call regulations, the funding agreement and the institution that signs it. Below we describe how to prepare technically and which questions to ask.

We write from the perspective of a supplier that has been through this on the technical side. For the MT5 Foundation we selected and delivered an NVIDIA Blackwell GPU cluster that sits at the foundation and supports research on sleep recordings.

Research tasks first, hardware second

Evaluators of R&D programmes check whether costs follow from the work plan. For a GPU server that's good news, because that's exactly how it should be chosen anyway. Before any card is named, answer these questions:

QuestionWhy it mattersExample decision
Training or inference?training needs far more memory for parameters, gradients and optimizer statestraining from scratch, LoRA fine-tuning, or only serving an existing model
Which models, how large?model size and precision set the minimum GPU memorya 30B model in FP8 on one card or a larger model across several
What data, how much?dataset size drives storage, networking and training timemedical signals, images, documents
Where can the data live?sensitive data often rules out the cloudon-site server, colocation or cloud
How many jobs at once?several teams or experiments in parallel means more cards, not a faster cardsplitting cards between projects
What happens after the project?the hardware stays and should have a further useinference, follow-up projects, deploying results

The answers go into the application as cost justification and into the spec as requirements. How to work out memory for a specific model is covered in our guide to how much VRAM a local LLM needs and its second part on professional and data-center GPUs.

At MT5, where the data could live decided the architecture. Sleep study recordings can't go to the cloud, so the pipeline and the model run on hardware at the foundation.

How hardware is claimed in an R&D project

Rules differ between programmes, so below we walk through one current example: the SMART Path (Ścieżka SMART) under FENG, the European Funds for a Modern Economy programme, specifically the SME call FENG.01.01-IP.02-001/26 run by PARP, the Polish Agency for Enterprise Development. The call ran from 26 February to 31 March 2026 with a budget of PLN 700 million. This edition dropped separate modules in favour of cost categories.

The eligibility guide for this call says that under "Depreciation (research equipment and equipment)", depreciation charges are eligible if all of the following hold:

  • the equipment is a fixed asset and is necessary for, and directly used in, the R&D work;
  • charges are calculated under accounting law and the company's accounting policy;
  • only the period in which the equipment actually serves the project's R&D work is eligible;
  • if the equipment also serves other purposes, only the proportional part of the charge is eligible;
  • the purchase was not co-financed by a national or EU grant and has not already been claimed in the project;
  • the equipment was bought at market prices.

In a footnote, the same document states that computer hardware and other devices not used directly in R&D work do not count as "research equipment" (aparatura naukowo-badawcza). The category, however, covers "research equipment and equipment". How a GPU server on which you train models as part of the R&D work should be classified is worth confirming with the institution before you apply, with the justification resting on how the server is directly used in the research tasks.

Other routes in the same guide:

ApproachCost categoryWatch out for
Purchase and depreciationDepreciation (research equipment and equipment)only charges for the period and share of use in the project
Equipment rentalExternal services (operating costs)only to the extent and for the period needed in the project, no buyout option
Software subscriptionExternal servicesonly to the extent and for the period used in R&D work
Indirect costs (administration, overheads)flat 25% of direct costssubcontracting excluded from the base

Two points from the guide's general rules are easy to miss. First, eligible costs can arise no earlier than the day after the application is submitted. A server bought "just in case" before applying won't enter the project. Second, the final eligibility date for expenses is 31 December 2030. That is the end of the 2021-2027 period, so the project and delivery schedule has to fit inside it.

The value left once the project ends (residual value) is not reimbursed. For the company's budget, this means the grant covers part of the hardware cost, and you finance the rest yourself and should have a plan for it.

The competitiveness principle in practice

Poland's guidelines on eligibility of expenditure for 2021-2027 require institutions to bind beneficiaries to the competitiveness principle (zasada konkurencyjności) in the funding agreement. It doesn't apply to contracts up to PLN 80,000 net. A GPU server for work with language models usually exceeds that threshold.

What this means for buying a server:

  • A request in the Competitiveness Database (BK2021). The bid deadline is at least 7 days for supplies, and 30 days when the contract value is EUR 750,000 or more. The deadline has to reflect the contract's complexity, and a multi-GPU server is a complex purchase.
  • Parameters, not brands. The description can't contain trademarks or name a manufacturer where that would favour certain suppliers. An exception is possible when the subject can't be described otherwise, and then "or equivalent" must be added.
  • No splitting. You may not understate the value or split the purchase to get below the threshold. If you buy in parts, the combined value counts.
  • Criteria concern the bid, not the company. Besides price you can score technical parameters, after-sales service, technical support and delivery time. You can't score the supplier's experience or reliability in the award criteria (that belongs in the participation conditions).
  • No related parties. A beneficiary that is not a public contracting authority can't award the contract to an entity linked to it by personal or capital ties.

Public universities, institutes and hospitals are usually contracting authorities under the Public Procurement Law (Pzp). Since 1 January 2026, the Pzp threshold has been PLN 170,000 net (previously PLN 130,000).

How do you write a spec that passes and doesn't exclude good hardware? Describe what the research tasks require:

  • minimum memory per GPU and total GPU memory in the server;
  • supported compute precisions (such as FP8 or FP4), if the plan involves training or serving at low precision;
  • the interconnect between cards, if a model is to be split across several GPUs;
  • CPU, RAM and NVMe capacity for datasets and model weights;
  • network interface;
  • maximum power draw and power requirements, matched to your server room;
  • warranty, service response time, delivery date.

Avoid requirements only one product can meet unless you can justify them with a research task. That is the most common source of questions during audits.

TCO: the cost doesn't end with the invoice

A grant covers part of the purchase cost, but a GPU server costs money for its whole life. For scale: NVIDIA lists a maximum power draw of about 14.3 kW for its eight-GPU DGX B200 system. Smaller servers draw less, but still more than a typical office server room expects.

ItemQuestion to plan for
Powercan the circuits, UPS and switchboard handle peak draw, not average?
Coolingcan the air conditioning remove the heat at full load, including in summer?
Spacerack, depth, weight, noise; your own server room or colocation?
Electricity during and after the projecthow many full-load hours a year, at what price?
Administrationwho updates drivers, the OS and containers, who responds to failures?
Softwarelicences, subscriptions, monitoring tools
Warranty and servicehow long, with what response time, and what after it ends?
Project durabilitymust the hardware stay in place and running for a required period after the project?

Durability is worth checking separately. According to the guidelines, for ERDF projects with infrastructure or productive investments it is 5 years from the final payment, and 3 years for SMEs in projects with a requirement to maintain the investment or jobs, unless state aid rules say otherwise. Your funding agreement tells you which requirement applies.

More broadly, we cover when your own server pays off compared with the cloud or an API in build vs buy in AI, and how we set up and maintain servers on the AI infrastructure page.

Where to put the server: on site, colocation, or rented compute

Each option has different consequences for the project:

  • On site. Full control over data, but all power, cooling and administration costs are yours. A good fit when data can't leave the organisation, as with the medical recordings at MT5.
  • Colocation. Professional power and cooling without building a server room. The hardware is yours, so depreciation works the same way. Check where the data physically sits and who has access to the rack.
  • Rented compute. No large purchase and no risk of the hardware ageing. The SMART guide includes equipment rental under operating costs, but whether a specific cloud GPU service fits has to be confirmed with the institution.
  • AI Factories. Poland is getting two AI Factories co-funded by EuroHPC: PIAST at PCSS in Poznań and Gaia at ACK Cyfronet AGH in Kraków. They are meant to provide compute access to companies too, including SMEs and startups. Access rules and dates are announced by the operators, so check at the source before planning a project around them.

Programmes in 2026: what to check

As of late September 2026, to be verified before any decision:

  • SMART Path (FENG). PARP's SME call closed on 31 March 2026. The 2021-2027 period is drawing to a close, so further calls are announced in the schedules on funduszeeuropejskie.gov.pl and on the PARP and NCBR websites.
  • KPO (Poland's Recovery and Resilience Plan). Under the Recovery and Resilience Facility regulation, the Commission makes payments until 31 December 2026, and KPO investments were due to reach their targets by 31 August 2026. For a new project with a server purchase, that is effectively too late.
  • FENG loans for digital and green transformation. PARP is launching a loan fund: loans of PLN 2.5 to 15 million at 0.5% a year, applied for through Financing Partners. It finances investments, not R&D projects, and whether a GPU server qualifies depends on the purpose of the investment. Details are on the programme page.

Checklist

  1. List the research tasks and derive the workloads from them: training or inference, models, data, parallelism.
  2. Establish where the data may be processed. This often decides between your own server and the cloud.
  3. Read the call regulations and the eligibility guide. Check which cost category the server falls under and whether it is claimed through depreciation.
  4. Confirm with the institution how a GPU server is classified as R&D equipment before you apply.
  5. Don't buy hardware before submitting the application.
  6. Write the spec in parameters, without brands, and where a reference is unavoidable, with "or equivalent".
  7. Prepare the request in the Competitiveness Database with a realistic bid deadline and criteria that concern the bid.
  8. Calculate TCO for the full project and durability period: power, cooling, space, administration, service.
  9. Plan what the hardware will do after the project, because its residual value stays with you.
  10. Build time for delivery, installation and testing into the project schedule.

Sources