Marina Chen wakes at 6:47 a.m. in her apartment in Tallinn, Estonia — a city that, by 2031, has become a quiet epicenter of something the world is still learning to name. Not the AI revolution itself, but the bureaucratic aftershock: the global renegotiation of who gets taxed, how much, and where, when the labor force generating the profit is no longer human. She works as a cross-border fiscal liaison — a role that didn’t exist five years ago — mediating between sovereign governments and multinational corporations whose autonomous systems generate revenue in dozens of jurisdictions simultaneously. Her morning begins not with coffee but with a notification from her AI co-analyst, summarizing overnight developments in the OECD’s latest Pillar Three framework negotiations, a successor to the Pillar One and Two agreements that restructured global corporate taxation in the mid-2020s.
The world Marina inhabits is one where industrial production has undergone a metamorphosis so thorough that the term “factory” feels anachronistic. In Shenzhen, Seoul, and Stuttgart, dark factories — fully automated manufacturing plants operating without lighting, heating, or human presence — churn out semiconductors, pharmaceuticals, and precision machinery around the clock. A 2027 McKinsey Global Institute report estimated that by 2030, over 40% of global manufacturing output would originate from facilities with fewer than five human workers per shift, a figure that proved conservative. Autonomous logistics networks, guided by reinforcement learning systems documented in research published in Nature Machine Intelligence (2026), connect these ghost plants to global supply chains with near-zero human intervention. Marina’s job exists precisely because these systems don’t respect borders. A dark factory in Bavaria, owned by a holding company registered in Dublin, running AI models trained on data harvested in São Paulo, shipping products assembled by robotic arms designed in Osaka — where, exactly, is the value created? And who deserves to tax it?
This is the question that dominates her workday. She steps into her office — a shared workspace overlooking Tallinn’s medieval Old Town — and opens a holographic display showing the latest draft of what negotiators are calling the “Computational Value Allocation Treaty,” or CVAT. The treaty attempts to do for AI-generated profits what the OECD’s Base Erosion and Profit Shifting (BEPS) framework tried to do for digital services: establish where revenue is “sourced” when the production chain is almost entirely non-human. An NBER working paper from early 2029 by economists Arnaud Costinot and Iván Werning laid the intellectual groundwork, proposing that taxable value in automated systems should be apportioned based on three factors: where the AI models were trained (data sourcing), where the computational infrastructure physically resides (compute geography), and where the end consumer is located (demand nexus). Marina’s role is to translate these elegant abstractions into enforceable treaty language — a task that is, she often reflects, stubbornly resistant to automation.
Her first meeting of the day is a video call with a Brazilian tax authority representative who is furious. A major European pharmaceutical company has deployed an AI-driven drug discovery platform — what researchers at MIT’s Computer Science and Artificial Intelligence Laboratory have called “zero-human R&D” — that identified and synthesized a new antimalarial compound. The compound was developed using clinical data from Brazilian hospitals, processed on servers in Finland, and is now being sold across sub-Saharan Africa. Brazil argues it deserves a significant share of the taxable revenue because its citizens’ health data was the foundational input. Finland claims compute-infrastructure rights. The pharmaceutical company, headquartered in Basel, insists the intellectual property resides in Switzerland. The African Union, meanwhile, has invoked the demand-nexus principle, arguing that the end-market countries deserve the largest allocation. Marina listens, mediates, and takes notes. The World Economic Forum’s 2030 Global Tax Governance Report warned that precisely these multi-jurisdictional disputes would become the defining fiscal challenge of the decade, and Marina is living proof.
The complexity deepens when she considers white-collar displacement. Law firms that once employed hundreds of associates now operate with skeleton crews of senior partners who oversee AI systems capable of drafting contracts, conducting discovery, and even formulating litigation strategy. A landmark 2028 study published in the Harvard Business Review by researchers from Harvard Law School and the MIT Sloan School of Management found that 67% of tasks previously performed by junior and mid-level attorneys could be fully automated by large language models fine-tuned on legal corpora. The same pattern has swept through finance — algorithmic trading was merely the prologue — and into medicine, where diagnostic AI systems outperform radiologists and pathologists in study after study, as documented in The Lancet Digital Health (2027). Creative industries have not been spared: generative AI produces advertising copy, musical compositions, architectural renderings, and even literary drafts at scale. The human role, as a 2029 IEEE Transactions paper on human-AI collaboration frameworks described it, has shifted from “execution” to “curation, ethical oversight, and emotional contextualization.” Marina’s own work exemplifies this: her AI co-analyst can model seventeen different tax allocation scenarios in seconds, but it cannot sit across from a Brazilian official whose voice trembles with frustration and find the words that build trust.
By midday, Marina shifts her attention to the income side of the equation — not corporate income, but personal. How do ordinary people survive in an economy where traditional employment is contracting? Her own compensation is a hybrid: a base salary from the Estonian government’s International Fiscal Mediation Bureau, supplemented by what is colloquially known as a “data dividend.” Estonia, ever the digital governance pioneer, was among the first nations to implement a framework inspired by proposals from economist Glen Weyl and technologist Eric Posner, whose concept of “data as labor” — articulated in their book Radical Markets and formalized in subsequent academic papers — argued that individuals should be compensated when their personal data is used to train AI systems that generate commercial value. Marina’s health records, her commuting patterns, her language usage — all of it feeds into AI training pipelines, and a quarterly dividend arrives in her digital wallet as compensation. It’s not transformative wealth, but it’s real. The EU’s Data Dividend Directive of 2029, modeled partly on California’s earlier but more limited data privacy frameworks, mandates that companies operating AI systems within the European Economic Area allocate between 1.5% and 4% of AI-attributable revenue to a pooled citizen dividend fund.
Beyond her personal income, Marina benefits from Universal Basic Services — a model that has gained traction as an alternative or complement to Universal Basic Income. Rather than providing a flat cash transfer, UBS guarantees access to essential services: healthcare, education, public transit, housing assistance, and — crucially, in the 2030s — access to compute power. A 2028 report by the University College London’s Institute for Global Prosperity argued that UBS is more cost-effective and socially cohesive than UBI, because it builds shared public infrastructure rather than individualizing the response to automation-driven displacement. Estonia provides Marina with free access to a national AI compute cluster, allowing her to run personal projects, upskill through AI-tutored education, and even launch small entrepreneurial ventures without the capital barrier of purchasing cloud computing time. The financing mechanism behind these services is, fittingly, a compute tax — a levy on the processing power consumed by corporate AI systems. The concept, first rigorously modeled in a 2027 Brookings Institution paper by economists Daron Acemoglu and Simon Johnson, treats computational cycles as a proxy for automated labor and taxes them analogously to payroll taxes. The revenue funds UBS programs, public retraining initiatives, and the data dividend pools.
Some nations have gone further. South Korea implemented a “robot tax” in 2028, building on its earlier reduction of tax incentives for automation investment. The revenue funds a national retraining program and a micro-equity scheme in which displaced workers receive small equity stakes in the companies that automated their positions — a concept explored in a 2029 working paper from the Stanford Institute for Human-Centered AI. Marina has colleagues in Seoul who receive quarterly equity yields from Samsung’s autonomous manufacturing division, a surreal form of compensation that blurs the line between worker and shareholder. Meanwhile, Finland’s UBI pilot — expanded nationally in 2029 after a decade of experimentation — provides every resident with a monthly stipend of €1,100, funded by a combination of compute taxes, carbon levies, and sovereign wealth fund returns. The academic debate between UBI and UBS, documented extensively in the Journal of Economic Perspectives (2028), remains unresolved, but in practice most advanced economies are converging on hybrid models that combine elements of both.
The international dimension of all this is what keeps Marina employed. Compute taxes, data dividends, and robot levies are national instruments, but AI-generated value is borderless. A company training its models on data from thirty countries, running computations in three, and selling products in a hundred creates a tax allocation nightmare that makes the old transfer pricing disputes look quaint. The CVAT negotiations Marina is involved in represent the most ambitious attempt at international fiscal coordination since the Paris Climate Agreement — and they are, if anything, more contentious. Developing nations, organized through the G77+ AI Fiscal Coalition, argue that their citizens’ data is being extracted as a raw resource, much like oil or minerals, and that the current allocation models systematically undervalue data sourcing relative to compute geography and demand nexus. A 2030 UNCTAD report on “Digital Resource Extraction and Fiscal Sovereignty” provided the empirical ammunition: it estimated that low- and middle-income countries contributed approximately 40% of the global training data used by the ten largest AI corporations but received less than 7% of AI-attributable tax revenue under existing frameworks.
Marina’s afternoon is consumed by a working session on this exact imbalance. She facilitates a negotiation between representatives from Nigeria, India, Germany, and the United States, each armed with competing economic models and political imperatives. The American delegation pushes for demand-nexus primacy, which favors large consumer markets. Germany advocates for compute-geography weighting, given its massive investment in domestic AI infrastructure. India and Nigeria champion data-sourcing rights, backed by the UNCTAD data and by a growing body of academic literature — including a widely cited 2029 paper in the American Economic Review — arguing that training data is the most irreplaceable input in the AI value chain and should therefore command the largest share of taxable allocation. Marina’s AI co-analyst generates real-time revenue projections under each proposed weighting scheme, but the final compromises require something no algorithm can yet provide: political judgment, cultural sensitivity, and the willingness to accept imperfect outcomes.
By evening, Marina steps away from the geopolitics of taxation and into the quieter question that haunts her generation: what is all this for? When survival is decoupled from traditional employment, when machines handle not just manual labor but cognitive work, what do humans do with themselves? She enrolls in an evening seminar — part of Estonia’s lifelong learning guarantee — on bioethics and AI-assisted gene therapy, taught by a human professor who uses AI tools to personalize the curriculum for each student. Education in the 2030s, as described in a 2028 UNESCO report on “Lifelong Learning in the Age of Artificial Intelligence,” has shifted from front-loaded credentialing to continuous, modular skill acquisition. Degrees matter less than portfolios of demonstrated capability, and the boundary between “student” and “professional” has dissolved.
Social structure, too, is evolving. Marina participates in her neighborhood’s civic assembly, a monthly gathering where residents deliberate on local policy — zoning, public space design, community investment priorities. The assembly uses what political scientists call “algorithmic governance augmentation”: AI systems model the likely outcomes of proposed policies, simulate demographic impacts, and flag potential unintended consequences, but the decisions themselves are made by human vote. A 2029 paper in Science by researchers at the Oxford Internet Institute found that communities using such hybrid deliberation models reported 23% higher civic satisfaction and 31% greater policy compliance than those relying on either purely human or purely algorithmic governance. Marina finds meaning in these gatherings — in the messy, inefficient, deeply human process of arguing about what kind of neighborhood she wants to live in.
She finds meaning, too, in her work, even though — or perhaps because — it sits at the intersection of the technical and the deeply political. Cross-border AI taxation is not merely an accounting problem. It is a question about global justice: who benefits from the most powerful technology ever created, and how do societies ensure that the gains are shared rather than hoarded? The old international tax system was built for a world of physical goods crossing physical borders. The new one must account for weightless value — data, computation, algorithmic insight — flowing through fiber-optic cables at the speed of light, generating wealth in ways that traditional economic categories struggle to capture. Marina’s daily work is an attempt to build that new architecture, one treaty clause at a time.
As she closes her laptop and watches the Baltic Sea darken beneath a late-autumn sky, Marina reflects on a line from a 2030 editorial in the MIT Technology Review: “The question is no longer whether AI will transform the global economy. The question is whether our institutions — legal, fiscal, democratic — can transform fast enough to keep up.” She doesn’t know the answer. But she knows that the attempt matters, that the negotiations she facilitates are not abstract exercises but the scaffolding of a new social contract — one that must somehow reconcile the borderless nature of artificial intelligence with the stubbornly territorial nature of human governance. Tomorrow, she will wake up and try again.
Frequently Asked Questions
What is the Computational Value Allocation Treaty (CVAT) and how does it determine where AI-generated profits are taxed?
The CVAT is a proposed international treaty that attempts to establish where revenue from AI-driven production should be sourced for taxation. It apportions taxable value based on three factors: where AI models were trained (data sourcing), where computational infrastructure physically resides (compute geography), and where the end consumer is located (demand nexus). It builds on earlier OECD frameworks like BEPS and Pillars One and Two.
Why do dark factories and fully automated supply chains create such complex tax jurisdiction disputes?
Dark factories operate without meaningful human presence and are embedded in global supply chains spanning multiple countries. A single product might involve AI models trained on data from one country, computational infrastructure in another, corporate registration in a third, and consumers in a fourth. This geographic fragmentation of the value chain makes it extremely difficult to determine where value is actually created and which jurisdictions have legitimate taxing rights.
What role does health or personal data play in cross-border AI taxation disputes?
Data used to train AI systems is increasingly treated as a foundational economic input. Countries whose citizens generate that data, such as clinical health records from hospitals, argue they deserve a share of taxable revenue from products developed using it. This creates novel disputes because traditional tax frameworks never anticipated raw data as a value-generating factor comparable to labor, capital, or intellectual property.
How does the three-factor model proposed by Costinot and Werning differ from traditional corporate tax allocation methods?
Traditional methods allocate profits based on factors like physical presence, employee headcount, or where intellectual property is held. The Costinot and Werning model replaces these with AI-specific factors: data sourcing (where training data originates), compute geography (where processing infrastructure is located), and demand nexus (where consumers are). This reflects the reality that automated production chains generate value without significant human labor in any single location.
What is a cross-border fiscal liaison and why has this role emerged?
A cross-border fiscal liaison mediates between sovereign governments and multinational corporations whose AI-driven autonomous systems generate revenue across many jurisdictions simultaneously. The role emerged because AI automation has made traditional tax frameworks inadequate. These specialists translate complex economic theories about automated value creation into enforceable treaty language, bridging the gap between policy, technology, and international law.
Why can't AI itself automate the work of negotiating international tax treaties for automation profits?
Despite advances in AI, translating abstract economic principles into enforceable international treaty language requires navigating competing sovereign interests, political sensitivities, and ambiguous legal concepts that resist algorithmic resolution. Each dispute involves unique contextual factors like data origin, infrastructure location, and consumer markets across different legal systems. This kind of diplomatic and interpretive work remains stubbornly dependent on human judgment and negotiation skills.
探索更多來自 Nenext 的內容
訂閱即可透過電子郵件收到最新文章。