In July 2026, the National Law Review published one of the sharpest analyses yet of an emerging policy frontier that founders cannot afford to ignore: the taxation of artificial intelligence. As AI systems absorb tasks once performed by human employees, governments are staring down a fiscal gap. Income tax revenue from labor is at risk. And their answer is not to slow AI adoption but to tax it at three distinct points in the value chain. A robot tax on labor substitution. A token tax on model usage. And a FLOP tax on the compute itself. Each proposal hits a different layer of the AI stack, and each carries different implications for startups building on or with AI.
Robot Tax: Taxing the Replacement of Human Labor
The oldest and most debated of the three proposals, a robot tax targets the direct substitution of human workers by AI systems. The logic is straightforward: when a business replaces an employee earning a salary, the government loses income tax and payroll tax revenue. If the AI system performing that work is treated as capital, it creates a perverse tax incentive favoring automation over hiring. A robot tax would level that playing field by imposing a levy on automated systems that perform tasks historically done by humans.
South Korea and several European Union member states have already explored versions of this idea, though no major economy has enacted one yet. The practical challenge is definitional. Not all AI usage replaces jobs. Some augments workers. Some creates entirely new roles. A blanket robot tax risks penalizing productivity-enhancing automation while doing nothing to protect the workers it aims to compensate. For founders, the risk here is asymmetric. If a robot tax is structured as a per-seat levy on AI agents, startups running large fleets of autonomous agents could face costs that scale with the very efficiency gains they are trying to capture.
Token Tax: Taxing the Usage Itself
A token tax sidesteps the messy labor-substitution debate entirely. Instead of asking whether an AI system replaced a job, it taxes the underlying activity: token consumption. Every API call to a large language model consumes tokens, and providers already bill by the token. A token tax would simply layer a government levy on top of that existing metering infrastructure. This makes it administratively simpler than a robot tax, but it introduces its own distortions.
Less efficient models generate far more tokens for the same output, meaning the tax burden would vary dramatically depending on model choice rather than the value of the work performed. A startup running an older, less optimized model could face several times the tax liability of a competitor using a more efficient one, even if both produce identical results. Token taxes also raise cross-border questions. If a US startup uses a model hosted in Europe and billed in euros, which jurisdiction collects the tax? The complexity multiplies fast.
FLOP Tax: Taxing the Compute Itself
The FLOP tax is the most ambitious and the most controversial. It targets the compute infrastructure itself by imposing a levy based on floating-point operations consumed during training or inference. The European Union has already signaled that compute usage is a relevant proxy for model capability under the EU AI Act. Models above a certain FLOP threshold trigger systemic risk obligations. Extending that logic from safety regulation to taxation is a relatively small conceptual step.
A FLOP tax would be borne by model providers and data center operators, but the cost would pass through to end users in higher API prices and hosting fees. Critics argue that taxing compute is self-defeating. AI development already faces capital constraints and geopolitical competition. Adding a tax on the very resource that defines AI capability could push compute-intensive training to lower-tax jurisdictions, fragmenting the AI supply chain and concentrating development in countries with the most permissive fiscal regimes. For startups, the impact would be indirect but real. If inference compute gets more expensive, every AI product's unit economics shift. Margin calculations that assumed $0.10 per million tokens would need a tax-adjusted line item.
What Founders Need to Do Now
None of these taxes has been enacted at the federal level in the United States as of July 2026. But the legislative groundwork is being laid. The Illinois AI Safety Measures Act, signed into law this month, shows that states are moving ahead of Congress. The US Treasury has commissioned studies on the fiscal impact of AI-driven labor displacement. The OECD is developing a framework for digital services taxes that could extend to AI compute. Founders should act on three fronts.
First, model your unit economics with a tax sensitivity layer. If compute costs increase by 10, 20, or 50 percent, does your product still work at your target price point? Second, track state-level activity. AI taxation will likely start at the state level before any federal framework emerges, just as data privacy did with the CCPA. Third, engage early with trade groups and policymakers. The shape of these taxes is not settled. Founders who articulate how robot, token, and FLOP taxes would affect real businesses can influence the design before it hardens into law. The window for input is open now. It will not stay open forever.