AI Labs Are Writing Industrial Policy Now, and the Framing Deserves Scrutiny

OpenAI published what it calls an industrial policy for the Intelligence Age, a document framed around expanding opportunity, sharing prosperity, and building resilient institutions as AI systems grow more capable. The framing is deliberately civic. The subject matter is infrastructure, labor, governance, and national competitiveness. The author is a private company with a direct financial stake in the policy environment it is proposing to shape.

This is worth sitting with before evaluating any specific proposal.

The Industrial Policy Turn

The term “industrial policy” carries ideological weight in the United States, where it has historically been treated with suspicion by market-oriented economists. The argument against it is simple: governments pick winners poorly, distort price signals, and create rent-seeking behavior around public funds. The argument for it is that the market undersupplies strategic national investments, particularly in technologies with long development timelines and significant spillover effects.

The US mostly resolved this tension by doing industrial policy while calling it defense procurement. DARPA funded the internet, GPS, and the foundations of modern computer science. The space race produced materials science, satellite communications, and semiconductor manufacturing advances. The CHIPS and Science Act of 2022, which committed over $50 billion to domestic semiconductor manufacturing, was a relatively explicit admission that strategic industries sometimes need deliberate cultivation.

What OpenAI is doing fits into a longer tradition of the technology sector attempting to shape the policy environment around its core technologies. The semiconductor industry lobbied for CHIPS. The pharmaceutical industry shapes FDA rulemaking. The difference here is that OpenAI is not simply lobbying for favorable treatment; it is proposing a comprehensive framework for how governments should think about AI as infrastructure, as economic policy, and as geopolitical strategy.

The Stargate Context

The backdrop to any OpenAI policy document is the Stargate initiative, announced in January 2025, which committed $500 billion toward US AI infrastructure over four years, with an initial $100 billion from SoftBank, OpenAI, and Oracle. This was announced at the White House, with explicit framing around national competitiveness and job creation.

Stargate is in many ways the operational version of the industrial policy argument: private capital, aligned with government priorities, building the compute infrastructure that advanced AI requires. The policy document can be read as the intellectual scaffolding for why that kind of arrangement should continue and expand.

What makes this structurally interesting is that OpenAI occupies an unusual position in these conversations. It is simultaneously a company that needs favorable policy outcomes, a technical authority whose expertise governments genuinely rely on, and a brand that has become synonymous with AI capability in the public imagination. These roles reinforce each other in ways that are not always transparent.

What AI Industrial Policy Actually Involves

Stripped of framing, AI industrial policy proposals from major labs tend to cluster around a few concrete areas.

Compute and infrastructure. Advanced AI requires data centers at a scale that strains the existing power grid and semiconductor supply chain. Policy proposals in this space typically involve permitting reform for data center construction, energy policy that prioritizes AI workloads, and domestic manufacturing incentives for advanced chips. The Nvidia H100 and its successors are manufactured using ASML’s extreme ultraviolet lithography machines, which are themselves subject to Dutch export controls negotiated under US pressure. The supply chain for AI compute is genuinely strategic.

Export controls. The Bureau of Industry and Security’s chip export rules have gone through multiple revisions since 2022, progressively tightening restrictions on advanced semiconductor exports to China. The AI labs generally support these controls, for reasons that are both principled (keeping advanced AI capability concentrated in democratic states) and convenient (reducing competition from Chinese models trained on equivalent hardware).

Talent and immigration. AI research is a field where a relatively small number of highly skilled researchers produce most of the advances. US universities train many of them, but visa policy determines whether they stay. This has been a persistent policy friction, and AI labs lobby heavily for expanded technical visa categories.

Intellectual property. Training large models on copyrighted data is currently the subject of multiple ongoing lawsuits. How courts and legislators resolve the fair use question will significantly affect the economics of foundation model development. OpenAI has a direct interest in outcomes that treat training data consumption as fair use, and its policy positions reflect this.

The “People-First” Framing

The document’s stated commitment to expanding opportunity and sharing prosperity addresses a real tension: the productivity gains from AI systems will not distribute themselves automatically. There is a meaningful economic literature on skill-biased technological change, and the evidence from past automation waves suggests that labor market transitions are slower and more painful than productivity curves imply.

What concrete proposals actually follow from the people-first framing matters enormously. A document that advocates for AI infrastructure investment, export controls, and IP protection favorable to AI companies, while gesturing at workforce development programs, is doing something different from one that proposes structural mechanisms for distributing AI-generated productivity gains, whether through sovereign wealth funds, profit-sharing requirements, or expanded public services funded by AI-related tax revenues.

Historically, the institutions that made industrial policy work for broad populations were not just the policies themselves but the countervailing power structures around them. The Marshall Plan worked partly because European governments had strong labor movements that pushed productivity gains toward wages. The postwar US semiconductor industry operated in an environment with strong antitrust enforcement that prevented any single firm from monopolizing foundational technologies.

The question of who has countervailing power in the current AI policy environment is not a comfortable one for major labs to address in their own policy documents.

Historical Parallels and Their Limits

The most common historical analogy for AI right now is the railroad era. Railroads were genuinely transformative infrastructure, required massive capital investment, had strong network effects, and became natural monopolies that eventually required regulation. The Sherman Antitrust Act of 1890 and the Interstate Commerce Commission were both responses to the failure of railroad markets to self-regulate.

The analogy is useful but imprecise. Railroads were physically constrained, geographically distributed, and their network effects operated differently than AI model scaling. A more productive comparison might be to the early internet, where the US made deliberate choices, such as the E-rate program, to treat internet access as infrastructure, and where open standards prevented any single company from owning the protocol stack.

The AI equivalent of that question is whether foundation models become infrastructure in a policy sense, regulated as utilities or made available as public goods, or whether they remain proprietary products that governments access through commercial licensing. OpenAI’s policy document presumably does not advocate for the former.

What Would Actually Be Useful

None of this means that AI industrial policy is bad, or that OpenAI’s proposals are wrong. The arguments for domestic AI infrastructure investment, for export controls on advanced compute, and for talent retention are serious. The US-China technology competition is real, and the geopolitical stakes of who controls frontier AI capability are not invented.

But good industrial policy requires more than the beneficiaries proposing it. It requires independent assessment of which interventions actually serve public interests, mechanisms for accountability when investments do not deliver public returns, and structural protections against capture by the industries being supported.

The DARPA model is instructive here. DARPA funded research through competitive grants to universities and contractors, maintained program managers who rotated in and out of industry, and was explicitly oriented toward spillover effects rather than supporting any particular company’s commercial roadmap. The internet it helped create was not owned by any DARPA contractor.

AI industrial policy designed by AI labs will tend toward outcomes that benefit AI labs. That does not make it wrong, but it makes external scrutiny essential. Governments engaging with these proposals should be asking which parts of the agenda are genuinely public goods, which parts are primarily commercial interests dressed in civic language, and what mechanisms ensure that the public investment produces public returns.

OpenAI has produced a document worth reading and engaging with critically. The proposals are substantive enough to deserve that engagement. The political economy of who wrote them is part of the analysis, not a reason to dismiss them.