Do AI & machine-learning products qualify for the Cyprus IP Box? (2026)

The short answer: yes, AI usually qualifies

AI and machine-learning products are, in almost all cases, software - and software protected by copyright is a core qualifying asset under the Cyprus IP Box. That means you do not need a patent to benefit. If your company develops the code, models and systems that power an AI product, the profits from exploiting that software can qualify for the regime's 80% deduction.

In practice, an AI company's stack is a bundle of copyrighted works: the model architecture and source code, the data and training pipelines, the inference and serving software, and the trained weights you develop in-house. Each is an original expression fixed in code and therefore capable of copyright protection, which is exactly what the IP Box rewards.

The result in 2026 is an effective tax rate of around 3% on qualifying profits (15% corporate income tax applied to the 20% of profit that remains taxable after the 80% deduction). For accounting periods before 2026, when the headline rate was 12.5%, the equivalent effective rate was about 2.5%.

What counts as the qualifying AI asset

The qualifying asset is the software your company creates and owns, viewed as a whole. For an AI or ML business, several components typically fall inside that boundary:

• Model architecture and source code - the neural-network design, training scripts and the codebase that defines how the system learns and runs.
• Data and training pipelines - the software that ingests, cleans, labels and feeds data into training, including feature engineering and orchestration code.
• Inference and serving software - the APIs, runtime, optimisation and deployment layer that turns a trained model into a usable product.
• Trained weights you develop - the parameters produced by your own training runs, as the direct output of your development effort.
• Underlying patented inventions - if a genuinely novel technique is patented, that patent is a separate qualifying asset in its own right, alongside the copyrighted software.

Which AI income qualifies (and which does not)

The regime rewards income earned from exploiting the qualifying software - not every euro your company collects. Drawing the line correctly is where most of the tax value is won or lost, so revenue streams should be mapped to the asset that generates them.

Qualifying income typically includes:

• API and usage fees - revenue billed per call, per token or per compute unit for access to your model.
• Subscriptions - recurring SaaS fees where customers pay to use the AI software.
• Licence fees - amounts received for licensing the model, code or weights to third parties.
• Embedded IP - the portion of a product's price attributable to the qualifying software built into it.
• Non-qualifying: pure consulting, setup, implementation and support fees, plus marketing intangibles such as brands and customer lists - these sit outside the IP Box.

Why in-house development gives a high nexus ratio

The benefit is not automatic on all qualifying income - it is scaled by the nexus ratio, an OECD-driven fraction that rewards companies for doing their own R&D. The ratio is (your own qualifying expenditure plus unrelated-party R&D, plus a 30% uplift) divided by total expenditure on the asset, capped at 100%.

For a typical AI startup that writes its own code, runs its own experiments and trains its own models with its own team, most spend sits in the numerator - so the ratio lands at or near 100% and almost all qualifying income enjoys the deduction. This is the natural profile of an in-house AI builder.

Two things drag the ratio down: research outsourced to related parties (group companies), and acquired IP - buying a model or codebase rather than building it. The 30% uplift softens the impact of some outsourcing and acquisition, but it cannot lift the ratio above 100%. Structuring development so the Cyprus company itself is the one incurring the R&D cost is the single most important planning point.

Open-source and third-party components

Very few AI products are built entirely from scratch. Teams fine-tune open-source base models, pull in libraries and frameworks, and rent GPU compute. None of that disqualifies you - but it does affect the nexus arithmetic, so it is worth understanding the effect.

Using open-source frameworks and tools that you build on top of is generally fine: the qualifying asset is the software you create, and your own development of architecture, fine-tuning, pipelines and serving code is your qualifying expenditure. Renting cloud compute for your own training runs is part of that in-house effort.

The sensitive case is acquiring a substantial third-party model or codebase and building your product around it. Because acquisition costs count against the nexus ratio, a product resting heavily on bought-in IP will see a lower ratio and a smaller deduction than one developed in-house. The more of the value you create yourself, the better the outcome.

Patents on underlying inventions

Copyright over software is the main route for AI, but it is not the only one. If your research produces a genuinely novel and non-obvious technical invention - a new training method, hardware-software technique or algorithmic breakthrough - that invention may be patentable, and a granted patent is a qualifying asset separate from the copyrighted code.

This can matter for deep-tech AI companies whose value lies in a specific inventive step rather than the product as a whole. You do not have to choose: the software qualifies as copyright regardless, and any underlying patents qualify in addition. For most application-layer AI businesses, though, the copyrighted-software route alone is sufficient and far simpler to evidence.

Documentation for AI R&D

The IP Box is a documentation regime as much as a tax one. To claim it you must be able to show, per asset, what you developed, what you spent and what income it generated. AI teams are well placed here because good ML practice already produces most of the evidence.

Keep contemporaneous records that tie development effort to the qualifying asset and support your nexus ratio:

• Experiment logs and training-run records showing the work performed in-house.
• Version control history and commit records evidencing your own code development.
• R&D cost tracking that separates own spend, unrelated-party R&D and any acquired IP.
• Income mapping that links each revenue stream to the qualifying software.
• A defensible split of any mixed income between qualifying software and non-qualifying services.

Where AI companies fit in the wider regime

For the vast majority of AI and ML businesses operating from Cyprus, the IP Box is a strong fit: the product is software, the development is in-house, and the income is licence, subscription or usage-based - exactly the profile the regime is built to reward. The combination of an effective rate near 3% and a high nexus ratio is difficult to match elsewhere in the EU.

Two wider points are worth noting. First, the regime sits within a full EU and OECD-compliant framework, so it is a substance-based incentive, not a loophole - real people doing real development in Cyprus. Second, groups with consolidated revenue below EUR 750 million fall outside the OECD Pillar Two global minimum tax, so most startups and scale-ups keep the full benefit of the ~3% effective rate.

This guide is indicative and general in nature; it is not tax advice. The right structure depends on how your company develops, owns and monetises its AI, and should be confirmed with a Cyprus tax adviser before you rely on it.