The Cyprus IP Box for AI & machine-learning companies
AI and machine-learning software qualifies for the Cyprus IP Box, so companies that build their own models, training pipelines and inference software can tax that profit at an effective rate as low as ~3% in 2026.
Why AI & ML companies qualify
Under the Cyprus IP Box, the code behind an AI product — model architecture, training and data pipelines, inference and serving software — is copyrighted software, a core qualifying asset. You do not need a patent. Because AI teams develop their models in-house, their nexus ratio is high, which delivers the full benefit and the lowest effective rate. Any patented invention behind the model qualifies separately.
- Model weights, training pipelines and inference code you develop in-house count as qualifying R&D.
- Copyrighted software qualifies — no patent required.
- API usage fees, subscriptions and licence income all qualify.
- In-house development pushes your nexus ratio toward 100%.
Effective tax rate for an owner-developed SaaS
~3%
15% corporate tax on just 20% of qualifying profit (2026).
Estimate your rate →Which AI income qualifies
Income earned from your qualifying AI software, less the direct costs of earning it.
API & usage-based revenue
Fees customers pay to call your model or AI API.
SaaS subscriptions
Recurring access to an AI product you built and own.
Licence fees & royalties
Payments to use your models, weights or software.
Pure resale or unrelated trading income does not qualify, and marketing IP (brands, trademarks) is always excluded.
Built for AI builders
Foundation & fine-tuned models
Proprietary LLMs, vision and speech models you train.
ML platforms & MLOps
Training, orchestration and serving software you own.
Applied AI & agents
Vertical AI products, copilots and autonomous agents.
Data & inference pipelines
In-house pipelines that create defensible IP.
What AI assets qualify as copyrighted software
Most AI and ML assets qualify for the Cyprus IP Box as copyrighted software, and you do not need a patent to claim the benefit. Software protected by copyright is a qualifying intangible asset in its own right, so the code and models you develop can already earn the reduced effective rate.
In practice this covers the model architecture and its source code, the training and data pipelines that prepare and feed your datasets, the inference and serving software that runs the model in production, and the fine-tuned weights you develop from your own work. Each of these is an original expression fixed in code, which is exactly what copyright protects.
Where an underlying invention is genuinely novel, it can also be patented, and that patented invention qualifies separately as its own IP Box asset. But for the great majority of AI and ML teams the copyrighted-software route is the direct path, and no patent is required to secure the Cyprus IP Box treatment.
- Model architecture and source code, protected by copyright as original software.
- Training and data pipelines that build, clean and feed your datasets.
- Inference and serving software, including APIs that expose the model.
- Fine-tuned weights you develop from your own training runs.
- Any genuinely novel underlying invention can be patented and qualifies separately.
- No patent is needed: copyrighted software alone secures the IP Box benefit.
Open-source and third-party components and your nexus ratio
Using open-source and third-party components does not disqualify you, but it does shape your nexus ratio, which sets how much of your AI profit gets the reduced rate. The OECD nexus approach rewards the R&D you actually carry out, so your own development pushes the ratio towards 100%.
Qualifying expenditure includes your in-house engineering and, importantly, work performed by unrelated contractors, which counts even when those contractors are based abroad. That means a Cyprus company running a distributed ML team can still reach a very high nexus ratio provided the outsourced work is with unrelated parties.
What reduces the ratio is acquiring models or IP and R&D outsourced to related parties, both of which sit outside qualifying expenditure. To soften this, the rules add a 30% uplift to your qualifying costs, capped at your total actual expenditure, so a limited amount of acquisition or related-party R&D need not sink the benefit.
- Your own development and in-house engineering push the nexus ratio towards 100%.
- Unrelated contractors count as qualifying spend, even when they work abroad.
- Acquired models and related-party R&D fall outside qualifying expenditure and lower the ratio.
- The 30% uplift (capped at total expenditure) softens the impact of acquisition costs.
- Building on open-source is fine; the ratio simply tracks the value you add yourself.
Which AI income qualifies
Income that flows from your qualifying AI software qualifies for the Cyprus IP Box, while income for services that merely surround the software does not. The test is whether the revenue is earned from the IP itself or from human effort sold alongside it.
Qualifying income includes API and usage fees, subscriptions to the AI product, licence income from letting others use your model, and IP embedded in a wider product or device. Income that does not qualify includes pure consulting and support fees, and marketing intangibles such as brands, trademarks and customer lists, which are excluded from the regime entirely.
Worked example: an AI company books €1,000,000 of qualifying profit and, having done its own development, reaches a 100% nexus ratio. The IP Box exempts 80% of that profit, leaving €200,000 taxable. At the 15% corporate rate that is €30,000 of tax, an effective rate of 3% on the qualifying AI profit.
- Qualifies: API and usage fees, subscriptions, licence income, embedded IP.
- Does not qualify: pure consulting and support, and marketing intangibles.
- Worked example: €1,000,000 profit, 100% nexus, 80% exempt leaves €200,000 taxable.
- €200,000 at 15% is €30,000 of tax, an effective rate of 3%.
Documentation for AI R&D
Robust documentation is what turns an eligible AI project into a defensible IP Box claim, so build the evidence trail as you develop rather than reconstructing it later. The nexus approach requires you to link income to specific assets and to the qualifying costs behind them, and AI R&D generates exactly the records needed to do this.
Keep experiment logs and training-run records that show how each model was developed, dataset provenance that establishes what data was used and its source, and commit history from your version control that time-stamps the engineering effort. Together these demonstrate that the software is original and that the R&D was genuinely yours.
Finally, track costs per asset so each model or ML product carries its own qualifying and non-qualifying expenditure. Per-asset cost tracking is the backbone of the nexus calculation and lets you support the ratio you claim if the Tax Department asks.
- Experiment logs and training-run records tying models to development effort.
- Dataset provenance showing what data was used and where it came from.
- Commit history from version control time-stamping the engineering work.
- Per-asset cost tracking underpinning the nexus ratio for each model.
AI & machine-learning FAQ
Talk to a Cyprus IP Box specialist.
Book a free, no-obligation assessment. We'll confirm whether you qualify, estimate your effective rate, and give you a fixed quote — confidentially, usually within one business day.
Reviewed by a Cyprus-admitted advocate · Last updated 21 June 2026.