Redistributed

Every transformative technology faces the same question: who gets to build it, and on what terms does everyone else get access? The answer is never purely technical. It's a design choice — shaped by policy, funding, ideology, and power. AI is no exception. But the choices being made right now will lock in patterns that persist for decades.

Understanding the models available — from fully proprietary to fully open, with a range of hybrid options in between — is the first step toward choosing wisely.

The Spectrum: Five Models for Building and Distributing Technology

No technology follows a single model. Most evolve through several. But the starting conditions matter enormously, because early choices create path dependencies that are hard to reverse.

1. Proprietary monopoly. A single entity controls development, access, and pricing. Edison's early electrical grid. AT&T's telephone network before the 1984 breakup. Today: OpenAI's GPT-4, Google's Gemini. The logic is familiar — massive capital investment justifies exclusive control. Innovation can be rapid, but access is dictated by the monopolist's interests.

2. Regulated utility. Private ownership, public oversight. Electricity after the New Deal. Water systems. Telecommunications after deregulation-with-rules. The state doesn't build the infrastructure but sets prices, mandates access, and enforces service obligations. Nobody has seriously proposed this for AI yet — but the template exists.

3. Public infrastructure. Government builds and operates the system. The interstate highway system. ARPANET, which became the internet — $200 million in NSFNET funding established TCP/IP as the dominant global protocol, and the open-architecture design enabled an innovation ecosystem no private network could have planned. Public libraries. The EU's EuroHPC initiative for shared compute. The logic: some infrastructure is too important to leave to market incentives alone. Mariana Mazzucato argues that public investment doesn't just correct market failures — it creates the conditions for entire industries to exist (Mazzucato, 2021). Beyond regulation, she contends, governments must "assume an active role in AI development by designing systems and shaping markets for the common good" (Mazzucato & Gernone, 2024).

4. Open standards and commons. No single entity controls the core technology. TCP/IP and the web protocols. Linux. Wikipedia. The internet worked because its protocols were open — anyone could build on them without permission. Yochai Benkler calls this "commons-based peer production," where decentralised contributors create shared resources that no single actor owns (Benkler, 2006).

5. Hybrid and federated models. Combinations of the above. Pharmaceuticals: publicly funded research, privately developed drugs, compulsory licensing in emergencies. Semiconductors: public R&D investment (CHIPS Act), private manufacturing, export controls shaping global access. Most real systems are hybrids. The question is which elements dominate.

Macro Level: How Nations Choose

At the macro level, governments are making AI infrastructure choices that will shape the next generation of industrial capacity. These choices are rarely framed as choices — they're presented as inevitabilities. They're not.

The American model: private-led, lightly regulated. The US has largely outsourced AI development to private labs backed by venture capital and big tech. Public funding exists — the National AI Research Resource (NAIRR) pilot has supported over 600 research projects and 6,000 students across all 50 states, but Congress has not fully authorised permanent funding. The scale remains modest compared to private investment. The logic: let the market innovate, regulate later. The result: the most capable models are controlled by a handful of corporations, and the public sector that funded foundational research (through DARPA, NSF, university grants) captures almost none of the value. This is the pattern Mazzucato identifies across multiple sectors — socialised risk, privatised reward. Thomas Piketty's central insight applies here with particular force: when returns to capital consistently exceed the rate of economic growth, wealth concentrates (Piketty, 2014). AI infrastructure is capital at its most extreme — billions in compute, data, and talent, generating returns that flow overwhelmingly to the entities that own the infrastructure. Without deliberate policy intervention, AI follows the r > g trajectory toward concentration by default.

The European model: regulate first, build second. The EU passed the AI Act, the world's most comprehensive AI regulation. It's also investing in public compute through EuroHPC — 19 AI Factories across Europe, three supercomputers in the world top 10, EUR 10 billion committed through 2027. But the EU's roughly 57,000 accelerators pale against a single company like Meta planning 600,000 H100 GPUs. Europe has no frontier model provider. The regulatory framework is sophisticated; the industrial capacity to compete at the frontier is not. This creates a dependency: European values encoded in regulation, but European AI users dependent on American and Chinese infrastructure.

The Chinese model: state-directed development. China treats AI as strategic industrial capacity. The state directs investment, sets research priorities, and controls data flows. Baidu, Alibaba, and others build models within this framework. Access is shaped by state priorities, not market dynamics alone. The trade-off: rapid capability development, but tightly coupled to state surveillance and control infrastructure.

Sovereign AI initiatives. A growing number of countries are rejecting dependence on external providers. The UAE developed Falcon and aims to become "the first fully AI-native government by 2027" — but Falcon was trained on AWS, revealing how sovereignty claims can mask infrastructure dependencies. India launched its IndiaAI Mission with roughly $1.2 billion and 58,000 GPUs, pursuing models for its vernacular languages. Singapore funded the SEA-LION project for 11 Southeast Asian languages. France's Mistral offers a European alternative. These initiatives recognise that AI infrastructure shapes national autonomy — but as analysts note, "the key question is not whether a country owns a model, but how it governs, adapts, and applies the models it relies on" (Tony Blair Institute, 2025). Most lack the scale to compete at the frontier alone, creating pressure toward regional cooperation or strategic partnerships.

The macro pattern is clear: every major economy is making AI infrastructure decisions, but most are doing so within frameworks designed for previous technologies. As Langdon Winner argued, technological arrangements embody political choices — and the choice to treat AI development as a private market activity is itself a political decision with distributional consequences (Winner, 1980). Daron Acemoglu frames this as a fork between "machine usefulness" — AI that complements human labour and expands capabilities — and pure automation that replaces it. As he and Simon Johnson document in Power and Progress, shared prosperity from technology was "not the automatic result of new technology but due to countervailing powers like labour unions and legislation" (Acemoglu & Johnson, 2023). The development model shapes which fork we take.

Micro Level: What Organisations Actually Face

At the micro level — individual developers, small companies, public institutions, community organisations — the choice of development model is more constrained but no less consequential.

Build on proprietary APIs. The easiest path. Call OpenAI or Anthropic's API, get frontier capabilities, pay per token. Fast to start, impossible to control. Your costs, your data flows, your product roadmap — all shaped by someone else's decisions. As documented in The API Economy, this creates structural dependency.

Deploy open-weight models. Download Llama or Mistral, run it on your own infrastructure (or rented cloud). More control, more cost, more expertise required. You can fine-tune for your use case. But as Open Source vs. Open Weights explains, you can't retrain, can't fully audit, and you're still building on decisions someone else made. The frontier remains elsewhere.

Contribute to truly open projects. EleutherAI, AI2's OLMo, BigScience's BLOOM — these projects release training data, code, and weights. They're genuinely reproducible. AI2's OLMo 3, released in 2025, includes the first fully open 32B reasoning model — weights, training code, training data, and evaluation suite all under Apache 2.0. These projects are narrowing the gap with proprietary labs while maintaining the transparency that makes independent verification possible. They provide the only path to genuine accountability.

Build locally. Fine-tune a small model on your own data for your specific use case. A hospital training a diagnostic model on its patient records. A municipality using language models for permit processing. A cooperative building tools for its members. This is where the micro level gets interesting — because local deployment can address needs that global models ignore, and it can keep data under local control.

Do nothing and inherit the defaults. The most common choice. Most organisations will use whatever AI is bundled into the software they already use — Microsoft Copilot in Office, Google's AI in Workspace, Adobe's tools. They won't choose a development model; they'll have one chosen for them. This is how infrastructure typically works: the defaults become invisible, and the power dynamics become structural.

What Open Actually Enables

Openness isn't just a licensing preference. It's a systemic condition that enables specific outcomes that closed systems structurally cannot produce.

Verification. You can't audit what you can't inspect. Open training data, open code, and open weights are preconditions for independent evaluation of bias, safety, and reliability. Closed systems ask you to trust the provider. Open systems let you check. As Arvind Narayanan has argued, many AI claims collapse under independent scrutiny — but scrutiny requires access (Narayanan & Kapoor, 2024).

Adaptation. Open models can be fine-tuned for languages, domains, and contexts that commercial providers have no incentive to serve. Swahili-language models, agricultural diagnostic tools, indigenous knowledge preservation — these applications emerge from openness because they're not profitable enough for proprietary development but genuinely valuable.

Competition. Open foundations lower barriers to entry. When the base model is freely available, competition shifts to applications, integration, and service — areas where smaller players can compete. Without open foundations, the market concentrates around whoever controls the base model. The data here is striking: on Chatbot Arena, the performance gap between the best open-weight and proprietary models collapsed from 8% to 1.7% between January 2024 and February 2025. DeepSeek R1, reportedly trained for under $6 million, matched OpenAI's o1 on reasoning benchmarks. The capability gap that justified proprietary control is narrowing faster than the governance gap.

Resilience. Distributed systems are harder to shut down, harder to weaponise, harder to capture. If a single provider controls access, a policy change, a business decision, or a geopolitical event can cut off entire populations. Open infrastructure distributes this risk.

Accountability. Elinor Ostrom's work on commons governance shows that communities can manage shared resources sustainably — but only when they have meaningful participation in rule-setting (Ostrom, 1990). Open AI development creates the conditions for this kind of governance. Proprietary development doesn't.

But openness has a vulnerability: it can be extracted from. Generative AI companies trained their models on the digital commons — open data, Creative Commons content, public code repositories, Wikipedia — and captured the value privately. As the Collective Intelligence Project has documented, foundation models may degrade the very commons they depend on without returning value. This is the enclosure dynamic that Benkler warned about in the broader information economy, now playing out at scale. Openness enables — but without governance, it also enables extraction.

The Middle Options: Where Most of the Action Is

The binary of "fully open" vs. "fully closed" obscures where most real decisions are being made. The interesting territory is the middle ground — and the intermediary steps that might move systems toward greater openness over time. A useful framework comes from the Bertelsmann Foundation and Open Future, whose 2025 "Public AI" white paper maps a "gradient of publicness" — from fully private to fully public — across each layer of the AI stack: compute, data, and models (Tarkowski, Osborne & Sieker, 2025). Different layers can sit at different points on the gradient. A model can use public compute, proprietary data, and open weights simultaneously. The question isn't "open or closed?" but "how public, at which layer, governed by whom?"

Open weights with restricted licenses. Meta's Llama is free to use but not fully open source. You can deploy it, fine-tune it, build products with it — but Meta retains control over the licence terms and can change them. This is better than pure proprietary access but doesn't create genuine commons. It's a corporate gift that can be revoked.

Structured access. Some researchers argue for a middle path: models that aren't fully open but are available to vetted researchers, auditors, and public-interest organisations under controlled conditions. This addresses safety concerns about unrestricted release while still enabling oversight. The risk: who decides who gets access? Structured access can easily become a mechanism for incumbents to control the research agenda.

Public-private partnerships. Government provides compute or funding; private labs provide expertise; results are shared under negotiated terms. The NAIRR model, or the UK AI Safety Institute's evaluations. These work when the terms are genuinely negotiated. They fail when "partnership" means "public subsidy, private capture."

Tiered release. Release a safety-tested version openly, provide full access to researchers under agreement, keep the most capable version behind an API. This is roughly what some labs already do, though the tiers are set by the provider, not negotiated with the public.

Cooperative and federated development. Multiple institutions pool resources to train models collectively, sharing costs and governance. BLOOM was an early example — over 1,000 researchers across 60 countries collaborating through BigScience. It produced a capable model. It also revealed the coordination costs: decision-making was slow, resource allocation was contested, and the model was ultimately smaller than proprietary alternatives. The lesson isn't that cooperation doesn't work. It's that cooperation at this scale requires institutional infrastructure that doesn't yet exist.

Intermediary Steps: What's Viable Now

Grand visions are worth less than workable next steps. Given current conditions — concentrated compute, established business models, emerging regulation — what moves are actually viable?

Mandate transparency as a prerequisite. Before debating whether models should be open, require that all deployed models disclose training data provenance, evaluation results, and known limitations. The EU AI Act moves in this direction. Transparency doesn't redistribute power directly, but it creates the information conditions for accountability. You can't govern what you can't see.

Fund public compute seriously. The gap between public and private compute capacity is the binding constraint on alternative development models. A researcher at a public university cannot compete with a lab backed by billions in cloud credits. Public compute initiatives exist but are dramatically underfunded relative to the task. Scaling them is a concrete policy lever.

Support open model ecosystems, not just open models. A single open model is a starting point. An ecosystem — shared evaluation benchmarks, community fine-tuning infrastructure, interoperable tooling, federated training protocols — is what makes open development sustainable. This requires investment in coordination, not just code.

Create procurement standards. Governments are massive AI customers. Procurement rules that require transparency, interoperability, and data sovereignty would shift market incentives without requiring new regulation. If the public sector demands open standards, providers will supply them.

Build governance infrastructure for shared resources. Ostrom showed that commons succeed when communities have clear boundaries, graduated sanctions, conflict-resolution mechanisms, and the right to self-organise. AI commons need the same. The technical infrastructure for sharing models exists. The governance infrastructure for managing shared AI resources is still primitive — but it's being built. The Mozilla Foundation has published a practical framework translating Ostrom's eight design principles to data commons. The Ada Lovelace Institute is applying them to data governance. Indiana University's Ostrom Workshop has launched a tech governance programme. These are early efforts, but they demonstrate that commons governance for AI is not merely theoretical — the institutional design work has begun.

Invest in local capacity. Not every use case needs a frontier model. Many need a small, well-tuned model running on modest hardware. Investing in the skills, infrastructure, and institutions that enable local AI deployment — community colleges training technicians, regional data centres, cooperative development organisations — may matter more than the frontier race.

The Infrastructure Implication

The development model chosen for AI will shape who benefits from the technology for decades. This is the lesson of every previous infrastructure transition. Electricity took 40 years to reach rural America — not because the technology didn't work, but because the business model didn't serve dispersed populations. Then the Rural Electrification Administration changed the model, and within two decades 90% of farms had power. The lesson is not that distribution is slow — it is that distribution requires political will. The internet was open by design, which enabled an unprecedented wave of innovation — and then was largely captured by platforms that exploited that openness for private concentration.

But concentration is not destiny. As Piketty documents in A Brief History of Equality, the long arc of the past two centuries has bent toward redistribution — through progressive taxation, public education, social insurance, antitrust, and the expansion of political rights (Piketty, 2022). Every one of those mechanisms was politically contested and hard-won. None was inevitable. The question for AI is whether this generation will build the institutional infrastructure to distribute the technology's benefits, or whether the current moment of concentration will harden into a new inequality regime — justified, as Piketty warns in Capital and Ideology, by whatever narrative makes the arrangement seem natural.

AI is at a similar inflection point. The technology works. The question is how it gets wired.

Fully proprietary models maximise short-term capability and minimise accountability. Fully open models maximise verifiability and adaptation but face funding and coordination challenges. The middle options — structured access, public-private partnerships, tiered release, cooperative development — are where the real policy design is happening.

Ursula Franklin distinguished between "holistic" technologies, which leave control with the operator, and "prescriptive" technologies, which embed control in the system's design (Franklin, 1989). Most current AI infrastructure is prescriptive — the choices are made upstream, and users inherit the consequences. Moving toward holistic AI infrastructure means building systems where the people affected by the technology have meaningful input into how it's developed, deployed, and governed.

That's not a technical problem. It's a political one. And the choices being made right now — in procurement offices, standards bodies, legislatures, and boardrooms — will determine which model prevails.

Sources

Benkler, Yochai. The Wealth of Networks. Yale University Press, 2006. Publisher
Franklin, Ursula. The Real World of Technology. Anansi, 1989. Publisher
Mazzucato, Mariana. Mission Economy: A Moonshot Guide to Changing Capitalism. Allen Lane, 2021. Publisher
Narayanan, Arvind & Kapoor, Sayash. AI Snake Oil. Princeton University Press, 2024. Publisher
Ostrom, Elinor. Governing the Commons. Cambridge University Press, 1990. Publisher
Winner, Langdon. "Do Artifacts Have Politics?" Daedalus 109, no. 1 (1980): 121–136. JSTOR
Widder, David Gray et al. "Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI." SSRN, 2023. SSRN
Piketty, Thomas. Capital in the Twenty-First Century. Harvard University Press, 2014. Publisher
Piketty, Thomas. A Brief History of Equality. Harvard University Press, 2022. Publisher
Acemoglu, Daron & Johnson, Simon. Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. PublicAffairs, 2023. Publisher
Mazzucato, Mariana & Gernone, Fausto. "Governments Must Shape AI's Future." Project Syndicate, March 2024. Project Syndicate
Tarkowski, Alek, Osborne, Cailean & Sieker, Felix. "Public AI." Bertelsmann Stiftung / Open Future, 2025. PDF
Tony Blair Institute. "Sovereignty in the Age of AI: Strategic Choices, Structural Dependencies." 2025. TBI
Kapoor, Sayash et al. "On the Societal Impact of Open Foundation Models." 2024. arXiv

General

How Technology Gets Distributed: Models That Shape Who Benefits