Written by Dmitrijs Kravcenko, SSE Riga Associate Professor

AI taxation is back in the policy debate, and it is likely to become more prominent in countries closest to the AI frontier. The underlying question whether tax systems built around human labour remain fit-for-purpose in an economy where a growing share of value may be generated by AI?

The current debate in US is framed around a “token tax”, but that label compresses several different ideas. Any serious AI tax has to answer basic fiscal questions about the taxable base, who ultimately bears the burden, can the state measure and administer the tax, how easily can firms loophole around it, and whether the purpose is revenue, redistribution, externality pricing, or a slowing down of labour substitution? Figuring these out is quite important because the chosen tax base determines whether the measure reaches economic rent, infrastructure cost, consumer use, or ordinary productivity-enhancing adoption.

In this context, there are three broad categories of ideas floating around. The first category is output levies, which tax AI usage directly, whether per token, per API call, or as a share of the revenue that usage generates. Here, Anthropic’s Dario Amodei proposes something close to a revenue charge, while Mark Cuban to a flat usage tax on large providers. Greg Casar recently offered a hybrid version that would look at both tokens and compute, with revenues earmarked specifically for jobs programmes rather than general expenditure.

The attraction of this way of taxation is obvious - usage is clearly visible, demand is growing, and the most profitable part of the market is concentrated among a small number of providers. There are, however, significant weaknesses as well. For one, tokens are not a unit of labour, value, or of social cost. They are an artefact of tokenizer design, varying across models, languages, code, caching, compression, and provider strategy. A flat token charge therefore risks taxing low-margin or recreational use alongside economic rent, while encouraging firms to optimise, route, cache, or reclassify activity in ways that shrink the measured base without reducing the underlying AI activity.

The second category is input levies, which tax compute or energy. This is closer to infrastructure pricing or Pigouvian taxation, since kilowatt-hours and GPU-hours are measurable and linked to real externalities, such as grid pressure, local electricity prices, cooling demand, land use, and environmental cost. Senator Elizabeth Warren’s proposed energy excise sits in this category, for example. The design problem is that AI-specific energy use has to be separated from general data-centre activity, the cost may be passed on to consumers, and national levies can still push compute into lower-tax jurisdictions.

The third category is substitution levies, which tax automation that replaces workers. This is the oldest version of the debate, revived most recently by Senator Bernie Sanders. The theoretical basis here is that tax systems often make labour expensive through payroll taxes, income taxes, and social contributions, while machines, software, and compute are treated as capital inputs. Essentially, this means that because AI is exempt from social taxes, employers will be incentivized by the tax code to replace human workers as soon as productivity gains on humans will become even approximate, which, at the current pace of AI capability growth may be sooner than we would like to think. Recent history, however, is less encouraging - where robot or automation taxes are justified, they tend to be small, second-best, and declining over time, with more welfare gain coming from broader reforms to income, capital, and consumption taxation than from the automation levy itself. South Korean robot tax is the only real empirical example of this, although EU has debated similar things at length also.

In addition to tax or levies, these conversation now also include public ownership claims. Sanders’ recent American AI Sovereign Wealth Fund proposal would impose a one-time 50% tax paid in stock by major AI companies, creating a public ownership stake rather than taxing tokens, compute, or automation flows. Sam Altman of Open AI seems to support this, either as a public buy-out or a donation to pre-seed such a fund. Whether or not such a proposal is ultimately feasible from a political AND a markets perspective, it does reframe AI taxation as a question of who owns future AI rents, not only how to tax current AI activity.

For us in Latvia, the question of imposing a domestic AI tax is not question at all – there is nothing to tax. We can, however, think about adoption, infrastructure, and tax-base resilience IF an AI tax of any kind is adopted in the US. If the United States adopted a flat provider-level token tax, the market and public response might be changes in routing, caching, localisation, billing entities, and inference locations. Latvia could benefit if it can offer reliable low-latency infrastructure, competitively priced renewable power, available data-centre capacity, cybersecurity credibility, EU-market access, and regulatory predictability. We could, in essence offer to become a compute tax haven of sorts.

This may not seem as farfetched as it does at first sight. Latvia has a relatively high renewable-energy share, at 43.2% of final energy consumption in 2023, with 54.3% renewable electricity in final electricity consumption. AI-relevant data-centre capacity is also emerging, including Delska’s 10 MW AI and high-performance-computing facility in Riga and Tet’s EUR 30 million DC7 project in Salaspils. The European Commission’s proposed Cloud and AI Development Act, which aims to expand EU cloud and AI capacity, improve access to energy, land, water, and financing, and at least triple EU data-centre capacity within five to seven years, would also, no doubt, contribute with all kinds of financial and non-financial support too.

Still, if the US were to start taxing AI compute AND the market would respond by moving said compute abroad instead of, say, into space as per SpaceX’s IPO vision, none of this would automatically create a tax base worth the trouble. As long as intellectual property, model ownership, customer contracts, and billing entities remain abroad, most taxable profit will remain abroad as well. We could, however, still benefit through a narrower, but not inconsequential, infrastructure dividend - data-centre investment, electricity demand that contributes to grid costs, network upgrades, skilled technical employment, heat-reuse projects, cybersecurity services, and a stronger position in regional digital infrastructure. That is assuming that Latvia captures more action than just routing server traffic. Achieving that is in large part a policy issue with the key question being whether the incentives for building up a physical footprint of AI supply chains can be balanced in a way that supports local infrastructure development, clean power efficiently, and creates local without ending up burning electricity for someone else’s AI with no further added value.