Visionologies
FINANCE•4 min read

When the Algorithm Decides the Rate, Who Owes the Tax?

Digital fintechs running AI-driven credit pricing, dynamic FX, or algorithmic compliance face a tax question their human-decision predecessors never had to answer in this form: jurisdiction. If a Lisbon-licensed neobank runs an underwriting model trained on European data but executes inference on AWS Frankfurt, with the loan booked in Luxembourg and disbursed to a customer in Aveiro — where, exactly, is the value created?

The OECD's BEPS 2.0 framework and the EU's Digital Services Tax debates were drafted before AI inference cost-allocation became a meaningful financial line item. They allocated taxing rights across distinct human-decided activities. They struggle when a single API call spans four jurisdictions in eighty milliseconds.

The CFO question that follows is concrete: should the inference compute spend — which can run €0.05 to €0.50 per credit decision, multiplied across millions of decisions a year — be capitalised as part of the model's intangible IP, or expensed as a cloud service? In Portugal, the AT has yet to issue formal guidance specifically on AI inference. By analogy with existing rulings on cloud computing classification under CIRC, current best practice for most fintechs is to treat inference as a recurring service expense rather than an asset.

That is unlikely to remain the prudent treatment. Two regimes are likely to emerge: a transition period where AI inference costs are routine OPEX and broadly expensable, and a maturing regime — likely 2027 or beyond — where the trained model itself, the proprietary training corpus, and possibly the inference workload, fall under intangible-asset reporting and transfer-pricing scrutiny. Treasury teams that want to avoid retroactive exposure should be modelling both regimes in parallel today. Platforms that wait for guidance to crystallise will discover their cost structures were misclassified for several reporting cycles.

arrow_back All insights