OpenAI, public interest and the problem of tech governance
OpenAI’s evolution reveals why governing transformative AI increasingly demands institutions beyond markets and governments alone.
In a nutshell
- Commercial success strengthened AI innovation but weakened governance
- Frontier AI shapes economies and demands accountability
- Public-private AI partnerships appear inevitable
- For comprehensive insights, tune into our AI-powered podcast here
Over the past few months, the lawsuit Elon Musk filed against OpenAI gave the public a rare glimpse into the engine room of the company that, more than any other, defined the new artificial intelligence era. Much of the coverage treated the dispute as a Silicon Valley feud: Elon Musk against Sam Altman, founder against founder, ego against ego. After weeks of testimony, the case ended in anticlimax: The jury found against Mr. Musk on timing grounds, concluding that he had brought his claims too late.
But the procedural resolution left the substantive question unanswered. Mr. Musk’s claim was that OpenAI’s charitable mission had been captured by commercial restructuring. OpenAI’s answer was that the mission could not survive without capital, infrastructure, products and scale.
The trial exposed this question without settling it. Can a public-interest mission survive once frontier AI – the most advanced AI systems, requiring enormous computing power – depends on private capital, cloud infrastructure, product revenue and strategic partners?
That unresolved question is the core of the story. Frontier AI now sits outside the institutional categories built for earlier technologies: too capital-intensive for nonprofit governance alone, too public for ordinary capitalism and too strategically important for regulation alone.
OpenAI as the first frontier AI testbed
OpenAI occupies a unique place in the short history of frontier AI. It launched the first major mass-market large language model (LLM) product, ChatGPT. It faced the first globally visible governance crisis when its board fired, and then reinstated, its CEO, Sam Altman. And it became the first frontier AI company whose founding mission was tested in court.
OpenAI is therefore more than a leading AI company; it is the testbed through which the world is discovering the institutional contradictions of frontier AI. In 2015, it launched as a nonprofit AI research company whose goal was to advance digital intelligence for humanity as a whole, “unconstrained by a need to generate financial return.”
That model suited an era when the main fear was concentrated control over AI research. A nonprofit lab could publish, collaborate, set norms and act as a public-interest counterweight to Google DeepMind and other corporate labs. But as frontier AI became a race for computational power, the problem changed. Staying at the frontier required billions of dollars for cloud compute, talent and AI supercomputers.
In 2019, OpenAI created a capped-profit structure: a for-profit arm that could raise investment and compensate employees with equity, while limiting investor returns and remaining formally controlled by the nonprofit. Microsoft’s $1 billion investment that year made the trade-off explicit: OpenAI could pursue its public-interest purpose only by tying itself to one of the very hyperscalers it had been created to counterbalance.
This became the pattern. Each institutional adjustment protected one part of OpenAI’s mission while weakening another. The capped-profit structure brought capital but introduced investor expectations. Microsoft supplied compute but created infrastructure dependence. ChatGPT expanded public access but turned OpenAI into a product platform with users, revenue and competitive pressure. The board’s failed attempt to remove Mr. Altman in 2023 showed how difficult it had become for formal mission control to override the practical power of employees, investors, customers and partners.
ChatGPT made AI governance public
ChatGPT changed OpenAI’s governance dilemma by reshaping the company’s social role. Released in November 2022 as a research preview, it transformed OpenAI from a frontier AI lab into a mass platform used for writing, coding, learning, search-like inquiry and enterprise work.
The governance issue is not only what the system can do technically, but what role it plays socially, economically and in terms of national security. Search engines return links, snippets and rankings but do not determine directly which one is the “right” one. AI systems, however, synthesize answers. They increasingly stand between users and knowledge, presenting information through a fluent interface that many people experience less as a tool and more as an authority.
This gives frontier AI a dual role. It is becoming both a mediator of knowledge and a general-purpose layer of economic power. It shapes how people encounter information, how professionals work, how organizations make decisions and how nations may deliver services and wage war.
Anthropic’s Claude Mythos Preview illustrates the point from another angle. Anthropic said Project Glasswing partners would receive access to Claude Mythos Preview to identify and fix vulnerabilities in critical software systems, while broader access would remain restricted because the same capabilities could also increase cyber risk.
The decision to release or withhold a model is not only a product decision. It can affect the security environment.
Events in June 2026 showed that such judgment does not stay with companies for long. On June 2, President Donald Trump signed an executive order creating a framework that allows AI companies to voluntarily permit the federal government to vet their models for security risks up to 30 days before release.
Ten days later, the voluntary framework shifted to mandatory intervention. The U.S. government applied export controls to Anthropic’s Fable and Mythos models, requiring restrictions on foreign-national access. Anthropic suspended access for all users because it could not verify nationality in real time. The restrictions were eased by the end of the month, after negotiations over safeguards. The call Anthropic had made voluntarily over Mythos Preview – who may use a frontier model, and under what safeguards – was now being made for it, by the state. The Trump administration also asked OpenAI to delay the broader release of GPT-5.6 for similar reasons.
Whether such responsibility can be left entirely to companies is, in Washington’s eyes, no longer an open question.
The AGI clause as a symbolic turning point
The “AGI (artificial general intelligence) clause” may be the clearest example of OpenAI’s institutional transformation.
The clause was part of the OpenAI-Microsoft commercial relationship. Its purpose was to protect the most powerful future version of OpenAI’s technology from ordinary commercial capture. Reuters reported that, under the then-current terms, Microsoft would lose access to OpenAI’s most advanced models once OpenAI achieved AGI. The clause was the institutional expression of OpenAI’s founding promise: Commercial partners could support the journey, but the destination would remain governed by the public-interest mission.
By October 2025, that safeguard had been relaxed. Microsoft said that once OpenAI declared AGI, an independent expert panel would verify the declaration. Microsoft’s rights to OpenAI models and products were extended through 2032 and included post-AGI models, subject to safety guardrails.
The April 2026 amendment went further. Microsoft’s license to OpenAI IP for models and products runs through 2032. OpenAI’s revenue-share payments to Microsoft continue through 2030, independent of OpenAI’s technological progress, at the same percentage but subject to a total cap.
For the economics of the partnership, AGI no longer served as a breakpoint. The contractual escape hatch from commercialization was replaced by fixed dates and capped payment obligations. A philosophical safeguard became a commercial uncertainty to be managed, then a calendar-based contract term.
The same logic is now visible in OpenAI’s move toward infrastructure-scale financing. Stargate, OpenAI’s infrastructure project, ties the company to data centers, energy and chips at industrial scale. Moreover, in June 2026, OpenAI and Anthropic each confidentially filed for an initial public offering. A mission once meant to be “unconstrained by a need to generate financial return” is preparing to answer to public shareholders.
Legitimacy is becoming contested
These blurred lines between research and commercialization, and between private product decisions and public consequences, are already creating legitimacy problems. In May 2026, former Google CEO Eric Schmidt was booed while discussing AI during a University of Arizona commencement speech. Several other commencement speeches praising AI also drew boos from graduates anxious about the technology’s impact on jobs.
More by innovation expert Uri Gabai
- The battle for AI sovereignty
- From Pax Americana to Pax Silica
- A new century’s dawn amid a democratic polycrisis
The episode is a signal. A technology that shapes work, knowledge, creativity, security and public services requires more than customer adoption and investor confidence.
This is the weakness of ordinary capitalism in frontier AI. It can scale the technology, but it cannot by itself answer why a small number of private institutions should mediate knowledge, reshape labor markets and help define national capability.
Once legitimacy becomes contested, the question is no longer whether frontier AI should be governed, but which institutional model can govern it without destroying its capacity to innovate.
Scenarios
Most likely: Public-private AI compact
The closest working analogy is the defense-industrial base. The state does not usually run the factories. Private firms execute, compete and innovate, while the government acts as the dominant customer, sets security and reliability standards, screens ownership and investment, controls exports and retains residual authority over the most sensitive systems.
Frontier AI may be drifting toward a similar arrangement: public compute programs, security reviews of model releases, export controls on advanced chips and other critical AI inputs, data-center and energy planning, and the state as an anchor buyer.
A frontier-AI compact would formalize that arrangement. It could include critical-infrastructure designation for frontier labs and the compute beneath them; enforceable duties and audit rights instead of voluntary commitments; special public-interest authority over defined strategic decisions, short of majority ownership; and financing terms tied to obligations, because whoever underwrites the compute shapes what the lab can do.
The analogy has limits. A primary defense contractor ultimately serves the state, whereas a frontier AI lab serves enterprises, consumers and governments at once. The compact therefore has to govern a mass-market, knowledge-shaping technology, not merely a strategic supplier.
This is not hypothetical. OpenAI in early June proposed that it could give the federal government a 5 percent equity stake, as a sort of insurance policy, an idea now being considered for other AI giants as well. One American politician is calling for the government to become 50 percent owners of AI leaders.
Some nations are already asserting control over strategically critical AI − not through regulation but through procurement leverage and security designations, a dynamic described as strategic capture. The Pentagon’s labeling Anthropic as a “supply-chain risk” is a clear example.
But this type of capture is precisely what a compact would discipline. The real choice is no longer between private autonomy and state control, but between capture improvised case by case and a compact defined by charter, with enforceable obligations on both sides.
Somewhat likely in the near term: Fully private frontier AI
This is broadly the current model in democratic countries, but it is likely to change in the coming years. This model has strengths: speed, talent, product discipline, capital-market access and operational culture. Its weakness is legitimacy. A fully private model leaves key decisions inside corporate structures: deployment thresholds, safety practices, compute allocation, pricing, access and product design, as well as macroeconomic and societal outcomes.
Least likely: Full nationalization
With an increase in negative public sentiment, states could decide that frontier AI is too strategic to remain privately governed. This would assert public authority over a technology with broad social consequences, but it would also create risks: politicization, bureaucratic rigidity, security overreach, weak product discipline and slower innovation.
Contact us today for tailored geopolitical insights and industry-specific advisory services.
Sign up for our newsletter
Receive insights from our experts every week in your inbox.
