Redistributed

There is a question that precedes all the familiar AI debates — about safety, about jobs, about regulation. It is this: who decides what gets built?

Not in the philosophical sense. In the financial sense. Every AI system that exists was funded by someone, and that someone expected a return. The structure of that funding — its time horizons, its return expectations, its governance mechanisms — determines what AI looks like, who it serves, and who captures the value it creates. If you want to understand why AI development tilts so heavily toward extractive business models and away from public goods, start with the money.

The Numbers

AI startups raised over $100 billion globally across 2024 and 2025. The figure is staggering, but the distribution is more instructive than the total. The vast majority of that capital flowed to a handful of frontier labs and their immediate ecosystems. OpenAI raised at a valuation approaching $300 billion. Anthropic crossed $60 billion. xAI, which did not exist before mid-2023, topped $50 billion. Behind them, a second tier — Mistral, Cohere, Together AI — commands multi-billion-dollar valuations with revenue that often looks modest relative to the checks being written.

The scale of individual rounds has become its own genre. AMI Labs, Yann LeCun's new venture, closed a $1.03 billion seed round at a $3.5 billion pre-money valuation — Europe's largest seed round ever. A billion-dollar seed. The phrase alone should give pause. Seed funding was once the mechanism for testing whether an idea had legs. Now it is a mechanism for placing enormous bets on technologies that may take a decade to mature, structured in a way that demands those bets pay off spectacularly or not at all.

The capital sources tell their own story. Traditional venture firms remain involved, but the decisive money comes from elsewhere: Microsoft's $13 billion into OpenAI, Amazon's $4 billion-plus into Anthropic, sovereign wealth funds from the Gulf states and Singapore writing checks that dwarf Sand Hill Road's capacity. As we documented in The AI Valuation Race: When Hype Meets Capital, the gap between what these companies are worth on paper and what they earn in practice is not a footnote. It is the central fact of the AI investment landscape.

How VC Shapes What Gets Built

Venture capital is not patient money. It is structurally impatient. A typical VC fund has a 10-year lifespan. Limited partners — the pension funds, endowments, and wealthy individuals who put capital into VC funds — expect returns of 3x or more on the total fund. Because most investments in a fund fail, the winners must return 10x to 100x to make the math work. This is not greed. It is arithmetic.

That arithmetic produces a specific set of incentives, and those incentives shape every decision an AI company makes.

AI must become a product with moats. Venture investors do not fund research for its own sake. They fund businesses — which means AI must be packaged into products with defensible market positions. The moats that matter in AI are proprietary data (the more users, the better the model becomes), network effects (ecosystems of plugins, integrations, developer tools), and switching costs (once an enterprise has built on your API, migrating is expensive). These are the features that make an AI company investable. They are also, not coincidentally, the features that make an AI company extractive — locking users in, hoarding data, and creating dependency rather than capability.

Open research is a cost center. Every dollar spent on research that is published openly is a dollar that does not build proprietary advantage. In the early stages, openness attracts talent and credibility. In the later stages, it becomes a competitive liability. This is not a moral failure on the part of any particular company. It is the logic of the funding structure asserting itself.

The OpenAI trajectory is the template, not the exception. OpenAI's journey — from nonprofit research lab to capped-profit entity to full for-profit conversion — is often narrated as a betrayal. It was not a betrayal. It was the predictable outcome of accepting billions in venture investment. The capped-profit structure was an attempt to split the difference between commercial capital and public-interest mission. The cap was removed because the capital demanded it. When your investors expect returns measured in hundreds of billions, a cap on profits is not a guardrail. It is a speed bump on the way to an IPO.

Anthropic's path follows a similar logic. Founded by former OpenAI researchers who left partly over governance concerns, Anthropic structured itself as a public benefit corporation and centered its identity on AI safety. These commitments are genuine. They are also under pressure from the billions in investment that Amazon, Google, and venture firms have committed — capital that creates expectations no mission statement can indefinitely override. As Mariana Mazzucato argues in The Entrepreneurial State, the tension between public-interest rhetoric and private-capital incentives is not unique to AI. It is the recurring dynamic of innovation funded by markets that demand returns on timelines incompatible with public goods.

The Kill Zone

The venture capital model does not produce a vibrant competitive market. It produces a kill zone.

The concept, originally described by economists Sai Krishna Kamepalli, Raghuram Rajan, and Luigi Zingales, refers to what happens around dominant platforms: mid-tier companies get squeezed between the giants above them and the startups beneath them. In AI, the kill zone is especially brutal. A company with a promising model or application faces a stark set of outcomes: get acquired by one of the frontier labs or their Big Tech backers, compete directly against organizations with orders-of-magnitude more capital (and lose), or find a narrow niche and hope the giants do not decide to enter it.

This dynamic shapes the entire AI ecosystem. Venture investors increasingly fund companies that are acqui-hire candidates rather than independent businesses — teams of talented engineers whose most realistic exit is absorption by Google, Microsoft, or Meta. The result is a market structure that funnels talent and technology toward a small number of incumbents, reinforcing the oligopoly rather than challenging it. As we have explored in AI and Antitrust: Why Competition Law Struggles with Foundation Models, existing competition law was not designed for a market where the barriers to entry are measured in billions of dollars of compute.

What Doesn't Get Funded

The mirror image of what venture capital builds is what it ignores. The gaps are not random. They follow directly from the return requirements.

Public goods. Translation systems for low-resource languages. Accessibility tools for disabled users. Diagnostic AI for clinics in low-income countries. These serve populations without purchasing power, which means they do not generate the revenue to justify venture-scale investment. The market for Yoruba-language translation or AI-assisted sign language interpretation is real — it is just not large enough to produce a 50x return for investors.

Research with long time horizons. Fundamental safety research, interpretability, alignment — these may be the most important areas of AI development, but they do not produce near-term commercial products. The VC timeline of 7-10 years to exit is poorly matched to research that may take decades to mature. AMI Labs' bet on world models is instructive: LeCun argues that his approach requires paradigm-level patience, yet the $1.03 billion seed round creates expectations that pure patience cannot satisfy.

Tools for communities without purchasing power. AI for smallholder farmers, for public defenders, for community health workers. These are domains where AI could generate enormous social value but minimal financial return. Venture capital does not fund them because it cannot. The structure does not allow it.

Alternative ownership models. Cooperative AI ventures, community-owned models, and public AI infrastructure all struggle to attract VC funding — not because they are unviable, but because their ownership structures are incompatible with the equity-based returns that VC requires. You cannot sell a 10x exit on a cooperative.

This is not a criticism of individual investors or companies. It is a structural observation. Venture capital is a tool, and like all tools, it is suited to particular tasks. The task it is suited to — building high-growth, high-return businesses in concentrated markets — happens to be poorly aligned with the public interest in AI development.

Alternative Funding Models

If VC incentives systematically skew AI development, the question becomes: what else is there?

Public investment. The most obvious alternative is direct government funding. The United States' National AI Research Resource (NAIRR) provides compute access to academic researchers, though at a scale far below what frontier development requires. The European Union's EuroHPC initiative funds shared supercomputing infrastructure. France funded the compute that trained BLOOM, the open multilingual model built by the BigScience project. These efforts demonstrate that public AI investment is feasible. They also demonstrate how far behind it lags: total public AI research funding across all OECD countries remains a fraction of what a single frontier lab spends annually.

Sovereign wealth funds. Gulf states — Saudi Arabia, the UAE, Abu Dhabi — are deploying sovereign capital into AI at scale, partly as an economic diversification strategy. Norway's sovereign wealth fund, the world's largest, could in principle direct investment toward AI public goods, though it has not done so systematically. The sovereign model offers longer time horizons and different return expectations, but it introduces its own governance concerns — particularly when the investing state has a poor human rights record.

Philanthropic funding. Foundations have funded some AI safety research, but the amounts are small relative to the field's needs. The Open Philanthropy project has been a significant funder of alignment research, and several major foundations have begun AI-focused programs. Philanthropy can be patient and mission-aligned, but it lacks the scale to compete with commercial investment and carries its own accountability gaps — foundation priorities reflect donor preferences, not democratic deliberation.

Cooperative and community ownership. As we explored in AI Cooperatives: Alternatives to Corporate Models, cooperatives offer a structural alternative to investor-owned AI. Data cooperatives, platform cooperatives, and worker-owned AI ventures distribute returns to members rather than shareholders. The model works — it has worked for over a century in agriculture, finance, and energy. Its constraint in AI is capital: cooperatives cannot sell equity to venture capitalists, which limits the scale at which they can operate. Public funding for AI cooperatives, modeled on the Rural Electrification Administration that brought power to rural America in the 1930s, could address this gap.

No single alternative replaces VC. The realistic path is a mixed funding ecosystem where public investment, cooperative ownership, philanthropic support, and commercial capital each play roles suited to their strengths — and where policy ensures that the commercial tail does not wag the public-interest dog. This is not utopian — it describes how most other critical sectors (healthcare, energy, agriculture) already work. The anomaly is that AI development has been allowed to proceed with so little funding diversity.

The Redistribution Mechanism

Venture capital is itself a redistribution mechanism. This is worth stating plainly, because the redistribution it performs is often invisible to those outside the financial system.

Here is how it works. Limited partners — pension funds managing retirement savings for teachers and firefighters, university endowments, insurance companies, and wealthy individuals — commit capital to VC funds. General partners (the VCs themselves) invest that capital in startups, taking a 2% annual management fee and 20% of profits (the "carry"). When a startup succeeds, returns flow back through this chain: founders and early employees capture the largest individual gains through equity, VCs take their carry, and limited partners receive the remainder.

The people who create value in AI companies — the researchers who build the models, the workers who label the training data, the users who provide the feedback that improves the products, the communities whose language and culture fill the datasets — do not share in this return structure. Researchers at well-funded labs receive stock options, which is better than nothing, but the distribution is steep: a handful of founding engineers may become billionaires while the data labelers who made their models possible earn $2 per hour through outsourcing platforms in Kenya and the Philippines. As Kate Crawford documents in Atlas of AI, the material labor behind AI systems is systematically invisible in the value chain, and the financial structures ensure it stays that way.

The "billion-dollar seed round" genre — which The AI Valuation Race: When Hype Meets Capital tracks — creates companies that must extract billions in value from somewhere to satisfy their investors. That extraction comes from users (through data collection and subscription fees), from workers (through automation and labor arbitrage), from competitors (through market dominance), and from the public (through the appropriation of publicly funded research and commonly held knowledge). The returns flow upward: to limited partners, to general partners, to founders. The costs flow downward: to displaced workers, to surveilled users, to communities whose cultural production trained the models without compensation. As the redistribution framing demands: who benefits, who pays.

The connection to Stock Buybacks and AI is direct. When AI companies eventually go public — as OpenAI's conversion trajectory suggests it will — the same dynamics that drive tech giants to spend hundreds of billions on stock buybacks will apply. AI-generated productivity gains will flow to shareholders through buybacks and dividends rather than to workers through wages or to the public through lower prices. The VC funding structure is the first stage of a pipeline that ends with shareholder extraction.

What Can Be Done

The venture capital trap is structural, not moral. Blaming individual VCs or founders misses the point — they are operating within a system that rewards specific behaviors and punishes others. Changing the outcomes requires changing the structure.

Public compute infrastructure. The single most effective intervention would be publicly funded compute at scale, available to researchers, cooperatives, and public-interest developers. This breaks the capital barrier that makes VC funding necessary for frontier work.

Procurement policy. Governments are major buyers of AI systems. Procurement rules that favor open, interoperable, cooperatively governed AI over proprietary alternatives would create a market for the kinds of AI that VC does not fund.

Tax and corporate governance reform. Tax incentives for cooperative AI ventures, restrictions on stock buybacks by companies that have received public research funding, and requirements that AI companies disclose the public-interest costs of their business models could shift the balance.

Diversified funding mandates. Public research agencies could require that a portion of AI funding support cooperative and community-owned projects, not just university labs and corporate partnerships.

None of these are utopian proposals. Each has precedent in other sectors. The question is whether the political will exists to apply them to AI before the funding structures harden into permanence.

The venture capital model built the AI industry we have. It did not have to build the AI industry we need. That gap — between what gets funded and what should get funded, between who captures the value and who creates it — is the redistribution question at the heart of AI development.

Sources

Mazzucato, Mariana. The Entrepreneurial State: Debunking Public vs. Private Sector Myths. Anthem Press, 2013. Publisher
Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Publisher
Acemoglu, Daron, and Simon Johnson. Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. PublicAffairs, 2023. Publisher
Kamepalli, Sai Krishna, Raghuram Rajan, and Luigi Zingales. "Kill Zone." NBER Working Paper 27146, 2020. NBER
Kenney, Martin, and John Zysman. "The Rise of the Platform Economy." Issues in Science and Technology 32, no. 3 (2016). Issues
CB Insights. "State of AI Report 2025." CB Insights
Lazonick, William. "Profits Without Prosperity." Harvard Business Review, September 2014. HBR

General

The Venture Capital Trap: How Funding Structures Shape What AI Gets Built