If your agency runs machine learning models, trains AI tools, processes large datasets, or develops software on cloud infrastructure, you have probably wondered whether those AWS, GCP, or GPU rental costs qualify for R&D tax credits. The short answer is yes, but only under specific conditions. HMRC has tightened the rules on what counts as qualifying expenditure for cloud compute costs, so you need to get the detail right.

As ICAEW qualified accountants working exclusively with agency founders, we see a lot of claims where cloud costs are either missed entirely or included incorrectly. Both mistakes cost you money. This guide covers what qualifies, what does not, and how to structure your claim so it stands up to HMRC scrutiny.

What Qualifies as R&D for Cloud Compute Costs?

HMRC defines R&D for tax purposes as work that seeks to achieve an advance in science or technology. For a digital or creative agency, that typically means developing new software, algorithms, or data processing methods that solve a technical uncertainty. It does not mean routine website builds, standard integrations, or using off-the-shelf tools in the way they were designed to be used.

Cloud compute costs qualify where they are directly attributable to qualifying R&D activities. If you are training a proprietary machine learning model on AWS SageMaker, that qualifies. If you are running A/B tests on a standard WordPress hosting setup, it does not.

The Three Conditions for Cloud Compute Costs

For a cloud compute cost to be eligible, it must meet all three of these conditions:

  • Directly attributable to R&D, the cost must be incurred as part of a specific R&D project, not general business computing.
  • Consumed in the R&D process, you must have used the compute resource, not just reserved it. Reserved instances sitting idle do not qualify.
  • Not capital in nature, ongoing consumption costs (pay-as-you-go, spot instances, on-demand) qualify. Upfront reserved instance purchases may be treated differently.

This is where many claims go wrong. An agency might have a £40,000 monthly AWS bill, but only £12,000 of that relates to active R&D compute. Claiming the full £40,000 is a red flag for HMRC.

Which Cloud Services Qualify for R&D Claims?

The most common qualifying cloud services for agency R&D claims include:

  • AWS, EC2 instances used for model training, SageMaker for ML pipelines, Lambda functions for novel data processing, EMR for large-scale data transformation.
  • GCP, Compute Engine for training runs, Vertex AI for ML workflows, BigQuery for novel data analysis at scale, Cloud Functions for custom processing.
  • Azure, similar treatment for Azure ML, Batch AI, and custom compute workloads.
  • GPU rental, dedicated GPU instances from AWS (p3/p4/p5 instances), GCP (A100, V100), or third-party providers like Lambda Labs, Paperspace, or CoreWeave. These are often the largest single R&D cost for AI-focused agencies.

Storage costs (S3, Cloud Storage, EBS volumes) can qualify if the data stored is used exclusively for R&D. But data storage for general business operations, client deliverables, or backups does not qualify.

What Does Not Qualify

HMRC is clear that the following cloud costs do not qualify for R&D tax credits:

  • Hosting, any cloud costs for hosting live client websites, applications, or services. Even if the software was developed through R&D, the ongoing hosting is not R&D.
  • General business computing, email, CRM hosting, collaboration tools, accounting software.
  • Content delivery networks, CloudFront, Cloudflare, or similar CDN costs.
  • Reserved instances, upfront payments for reserved compute capacity, unless you can demonstrate they were used exclusively for R&D during the period.
  • Licence costs, third-party software licences embedded in cloud services, unless they are integral to the R&D and not available separately.

A common mistake we see is agencies claiming the full AWS bill because "we do R&D on AWS." That is like claiming your entire office rent because you do R&D in that office. HMRC expects you to apportion costs accurately.

How to Calculate the R&D Cloud Compute Claim

There are two approaches to calculating the qualifying cloud compute costs for your R&D claim. The method you choose depends on how your cloud infrastructure is set up.

Method 1: Tagging and Cost Allocation

The best approach is to tag your cloud resources at the project level. On AWS, use cost allocation tags. On GCP, use labels. Tag every resource that is used for R&D with a project code or identifier. Then pull the cost report for those tagged resources at year-end.

This gives you an auditable trail. If HMRC asks "how did you arrive at £24,700 for R&D cloud compute?" you can show them the tagged resources and the cost breakdown.

If you are not already tagging resources, start today. Retrospective tagging is difficult and less reliable.

Method 2: Time-Based Apportionment

If your R&D and non-R&D workloads run on the same infrastructure without tagging, you can use a time-based apportionment. Track the hours each resource runs for R&D versus non-R&D activities. Apply that ratio to the total cost.

For example, if a GPU instance runs 120 hours per month, and 80 of those hours are for R&D model training, then 66.7% of that instance's cost qualifies. You need logs or monitoring data to support this. HMRC will ask for it.

Worked Example

Let us take a 15-person AI agency based in Shoreditch. They develop custom NLP models for clients. Their annual cloud costs break down as follows:

  • AWS SageMaker and EC2 GPU instances for model training: £63,400, fully qualifying. Tagged as R&D.
  • AWS S3 storage for training datasets: £8,200, qualifying, as the data is used exclusively for R&D.
  • AWS EC2 instances for client API hosting: £31,500, not qualifying. This is live service hosting.
  • AWS CloudFront CDN: £4,300, not qualifying.
  • GCP BigQuery for novel data analysis: £14,700, qualifying. Tagged as R&D.
  • GCP Compute Engine for general dev/test: £9,100, partially qualifying. 60% of usage was for R&D prototyping, supported by logs.

Total qualifying cloud compute costs: £63,400 + £8,200 + £14,700 + (£9,100 x 60%) = £91,760.

Under the SME R&D scheme, that £91,760 is enhanced by 186%, giving a qualifying expenditure of £170,873. At 19% corporation tax, that saves the agency £32,466 in tax. If the agency is loss-making, they can surrender the loss for a cash credit of up to 14.5% of the enhanced expenditure.

That is a real return for proper record-keeping.

GPU Rental and Third-Party Cloud Providers

GPU rental from third-party providers like Lambda Labs, Paperspace, CoreWeave, or Vast.ai qualifies on the same basis. The key is that you are consuming compute for R&D. The provider does not matter. What matters is the use.

One nuance: if you rent dedicated GPU servers on a monthly contract, HMRC may view that as a capital lease rather than revenue expenditure. In practice, most monthly GPU rentals are treated as revenue costs because there is no transfer of ownership. But if you sign a 12-month contract with no ability to cancel, the treatment gets more complex. Speak to your accountant before committing to long-term GPU rental contracts if you plan to claim R&D.

What HMRC Looks For in Cloud Compute R&D Claims

HMRC has become more sophisticated in reviewing cloud costs in R&D claims. They look for three things specifically:

  • Technical narrative, you need to describe the technical uncertainty you were trying to resolve and why cloud compute was necessary to resolve it. "We trained a model on AWS" is not enough. "We needed to process 2TB of unstructured data using a novel neural network architecture, which required GPU instances with 80GB memory per node" is better.
  • Cost breakdown, a simple total will not do. HMRC wants to see how you arrived at the figure. Tagging reports, resource logs, or time-based apportionment calculations.
  • Exclusion of non-qualifying costs, show that you have removed hosting, CDN, and general business compute. A claim that includes everything looks like you do not understand the rules.

If your claim includes cloud compute costs over £50,000, expect HMRC to ask for supporting evidence before they process it. That is standard. Have your documentation ready.

Claiming Cloud Compute Costs Under RDEC

If your agency is large enough to fall under the RDEC (Research and Development Expenditure Credit) scheme, the treatment is slightly different. RDEC gives a 20% taxable credit on qualifying expenditure, including cloud compute. The same rules apply about what qualifies and what does not. The difference is in how the credit is calculated and reported.

Most agencies with fewer than 500 employees and either turnover under €100M or a balance sheet under €86M will use the SME scheme. If you are part of a larger group or have received notified state aid, you may be on RDEC. Check with your accountant.

Common Mistakes with Cloud Compute R&D Claims

We see the same errors repeatedly. Avoid these:

  • Claiming the full cloud bill. HMRC will disallow the non-R&D portion and may open a full enquiry.
  • No supporting documentation. A spreadsheet with one line saying "AWS costs £40k" will not survive an enquiry.
  • Claiming hosting costs. Live client hosting is not R&D, even if the software was developed through R&D.
  • Ignoring GPU rental. If you rent GPUs from a third party, include those costs. They are often the largest qualifying cost.
  • Not tagging resources. Start tagging now. It makes the claim easier and more defensible.

If you are unsure whether your cloud compute costs qualify, ask your accountant before submitting the claim. A rejected claim is worse than no claim at all, because it flags your agency for HMRC scrutiny in future years.

Getting Your R&D Cloud Compute Claim Right

Cloud compute costs can be a significant part of your R&D claim, but only if you have the documentation to support them. Tag your resources, keep logs of usage, and separate R&D compute from hosting and general business computing. If you do that, the claim is straightforward. If you do not, you are leaving money on the table or inviting an HMRC enquiry.

If your agency is spending more than £10,000 per year on cloud compute for development, AI training, or data processing, it is worth reviewing whether those costs qualify. Our ICAEW qualified team specialises in R&D claims for agency founders. We can help you structure the claim and prepare the supporting evidence.

For agencies working across multiple sectors, we also have dedicated pages for digital agencies, creative agencies, and SEO agencies where cloud compute is often a major R&D cost.