How Did Sam Altman Turn a Nonprofit Lab Into an AI Empire?

A Lab With a Mission and a Pay Problem

In December 2015, Sam Altman, Elon Musk, and a group of researchers launched OpenAI as a nonprofit with a stated mission to ensure artificial general intelligence benefits all of humanity. The founding blog post promised openness and a $1 billion funding commitment. The structure was a deliberate contrast to Google DeepMind, which had been acquired by Google in 2014.

The contrast was not merely symbolic. DeepMind had been folded into a company with a mature advertising business, a global cloud, and a compensation system built on stock. OpenAI was betting that a mission-first organization could attract the same caliber of researcher without matching the same package. The founding post framed openness as a core value, and the $1 billion pledge was meant to signal that the venture was serious rather than academic. In practice, the pledge was a commitment to raise and spend over time, not a single wire transfer, and the nonprofit form meant there was no equity to hand out.

That distinction matters because the market for AI talent was already heating up. DeepMind had shown what a well-funded industrial lab could do, and Google’s acquisition gave its researchers a currency — liquid stock — that a nonprofit simply could not mint. A nonprofit cannot easily pay top AI researchers with equity. Cash salaries could be competitive at the top, but they could not replicate the upside of a stake in a company whose value might multiply. The pay problem was structural, not a matter of generosity.

The mechanism is worth spelling out, because it explains much of what followed. Equity compensation works by granting employees a claim on future appreciation. At a public company, that claim is liquid and priced daily; at a startup, it is illiquid but carries the prospect of a step-change in value at an acquisition or listing. A nonprofit has no shares to grant, so its only lever is salary and benefits, which are paid out of donations and grants. That makes the nonprofit’s compensation ceiling a function of its fundraising, not of the value it creates. A researcher who builds a model worth billions cannot capture any of that value through ownership; the upside accrues to the mission. For some researchers that is precisely the appeal. For many others, especially those with offers from Google, it is a real cost.

By 2019, OpenAI had created a capped-profit arm, OpenAI LP, to raise outside capital while keeping the nonprofit board in control. Microsoft invested $1 billion that year. The structure was unusual: investors could earn returns up to a cap, with excess profits flowing back to the nonprofit. The design was an attempt to square a circle — to let capital in without letting it own the mission. The cap was meant to answer the obvious objection: if investors can only earn so much, why would they invest at all? The answer was that the cap was high enough to be attractive and the mission was compelling enough to matter.

It is worth being precise about what the capped-profit form did and did not solve. It gave the lab a way to accept large checks and to offer employees a participation mechanism that a pure nonprofit could not, while keeping ultimate authority with a board whose fiduciary duty ran to the mission rather than to shareholders. But it also created two classes of interest inside one organization: the nonprofit, which was chartered to prioritize safety and broad benefit, and the limited partners, who expected a return within the cap. Every subsequent decision — how fast to ship, how much to spend on compute, when to release a model — had to be legible to both. The structure did not eliminate the tension between mission and money; it housed it.

Microsoft, Compute, and the Scale Problem

Microsoft’s investment was not just cash. It included Azure compute credits, which mattered because training large models is a capital-intensive race. OpenAI’s GPT-3, released in 2020, required massive compute and produced a research API that developers could pay to use. The commercial arm began generating revenue while the nonprofit board retained formal authority.

The compute component is the part that is easy to underrate. In the years after 2019, the binding constraint on frontier AI was not ideas or even researchers; it was access to clusters of specialized chips and the electricity and networking to run them. A dollar of cash could buy talent, but compute credits bought time — the ability to run experiments at a scale that academic labs could not match. GPT-3 was the visible result: a model whose capabilities emerged from scale as much as from architecture. The research API turned that capability into a product, and developers paying to use it gave the commercial arm its first real revenue stream. The board still held formal authority, but the center of gravity was shifting toward the entity that could sign contracts and pay bills.

This is a general pattern in capital-intensive technology, and it is worth naming. When the scarce input is a physical resource — chips, power, floor space — the party that controls access to that resource gains leverage that no governance document anticipates. The nonprofit board could hire and fire executives, but it could not conjure a cluster. Microsoft could. That asymmetry did not show up in the org chart, but it shaped every negotiation that followed.

By January 2023, Microsoft had extended its partnership with a reported $10 billion investment. OpenAI’s models were integrated into Microsoft products including Bing and GitHub Copilot. The relationship gave OpenAI distribution and Microsoft a stake in the leading model developer. Revenue grew, but so did the governance tension between mission and money.

That tension had a concrete mechanism. The nonprofit board’s authority was designed to protect the mission, but the commercial arm’s obligations ran to investors and customers. Every expansion of the partnership increased the value of the commercial entity and, by extension, the stakes involved when the board exercised its authority. Distribution through Bing and GitHub Copilot made OpenAI’s models part of Microsoft’s product surface, which meant that a governance dispute at OpenAI would be felt inside Microsoft’s own roadmap. The two organizations were becoming entangled faster than the governance documents could adapt.

Entanglement of this kind is hard to reverse. Once a partner’s products depend on your models, the partner has a strong interest in your stability, and that interest can be expressed as support for management rather than for the board. The relationship that began as an investment became a form of implicit governance insurance for whoever ran the commercial arm. None of this was hidden; it was simply not the kind of thing a charter can encode.

The November 2023 Board Crisis

On November 17, 2023, OpenAI’s nonprofit board fired Altman as CEO, citing a loss of confidence. The announcement triggered a staff revolt: nearly all employees signed a letter threatening to leave unless Altman was reinstated. Microsoft CEO Satya Nadella offered Altman and co-founder Greg Brockman roles at Microsoft.

Within five days, Altman returned. The board was reconstituted, and Altman consolidated power. The episode exposed the fragility of a structure where a nonprofit board controlled a commercial entity with billions in outside capital. It also showed Altman’s leverage: the staff and Microsoft were aligned with him, not the board.

The sequence is worth reconstructing because it revealed where power actually sat. The board acted on a Friday. By the weekend, the staff letter had made the threat of an exodus explicit, and Nadella’s offer gave Altman a credible alternative that also signaled Microsoft’s alignment. The board’s formal authority was real, but it depended on the cooperation of the people who built and sold the products. When that cooperation evaporated, the board’s position became untenable. The reconstituted board and Altman’s consolidated role were the settlement terms of a dispute the governance structure was never designed to handle.

Counter-evidence to the simplest reading is worth noting. The board did have a legitimate mandate: it was created precisely to check the commercial arm when mission and money diverged. Its decision was not obviously wrong on the merits; the loss-of-confidence rationale reflected a real disagreement about direction and pace. The failure was less about judgment than about leverage. A board can fire a CEO, but it cannot fire the staff, the customers, or the partner whose cloud the models run on. The crisis demonstrated that formal control and practical control had come apart.

There is a further nuance. A staff revolt is not automatic; it happens when employees believe their own interests and the CEO’s are aligned, and when the alternative employer is credible. In this case both conditions held. The letter was effective because the signatories were the people who would have to execute any plan the board adopted, and because Microsoft’s offer made the threat believable rather than rhetorical. Governance theories that treat the board as the ultimate locus of authority tend to assume that everyone else will stay put. The OpenAI episode is a reminder that they do not have to.

The For-Profit Pivot and the Valuation Race

In 2024 and 2025, OpenAI moved toward a more conventional for-profit structure. Reports described plans to convert the capped-profit arm into a public benefit corporation, with the nonprofit retaining a significant equity stake. In October 2025, OpenAI completed a share sale that valued it at roughly $500 billion, according to reporting by Reuters and others. Microsoft’s stake was restructured.

The public benefit corporation form is a middle path rather than a clean break. It preserves a mission obligation in the corporate charter while allowing the kind of equity compensation and investor terms that a capped-profit structure struggled to accommodate. The nonprofit retaining a significant equity stake keeps the mission-aligned entity as a major owner, which is a different arrangement from a straightforward spin-out. The share sale at roughly $500 billion, reported by Reuters and others, was less an IPO than a private repricing: existing and new investors transacting at a valuation that reflected expectations about future revenue. Microsoft’s restructured stake changed the shape of its exposure without ending the partnership.

A public benefit corporation is not a magic fix, and it is worth being clear about what it does. It writes a mission purpose into the charter and, in many jurisdictions, requires directors to weigh that purpose alongside shareholder interests. That gives the mission a legal footing it lacked inside a capped-profit vehicle. But it also normalizes the company’s relationship with capital: equity compensation becomes straightforward, investor terms become conventional, and the cap that once limited returns disappears. The mission survives as an obligation in the charter, not as a ceiling on outcomes.

The valuation reflected both revenue growth and the strategic importance of frontier models. OpenAI’s consumer products, including ChatGPT, reached hundreds of millions of weekly users. Enterprise deals and API revenue became meaningful, though the company continued to spend heavily on compute. The capped-profit structure that once limited investor returns was being unwound.

Hundreds of millions of weekly users is a distribution fact with financial consequences: consumer scale creates a data and feedback loop, and it gives the company a brand that enterprise buyers already know. But the same scale drives compute costs. Training and serving models at that volume is expensive, and the company’s spending on compute remained heavy even as revenue grew. The unwinding of the capped-profit structure was, in part, an acknowledgment that the cap had done its job — attracting capital under mission constraints — and that the next phase required more conventional terms.

It is also worth noting what the valuation does not tell you. A private share sale prices the company at a moment, based on the expectations of a relatively small set of investors. It is not a verdict on the mission, and it is not a guarantee that the governance questions are settled. What it does show is that the market believes the commercial entity will generate substantial cash flows, and that belief is now embedded in the terms under which employees and investors participate. The pay problem that started this story has been substantially solved — but by making the company look much more like the companies it was founded to differ from.

What Altman Actually Controls

Altman’s power does not come from majority ownership. He holds little or no equity in OpenAI, by his own public statements. His influence comes from his role as CEO, his relationships with investors and Microsoft, and his ability to set strategy. The 2023 crisis demonstrated that staff loyalty and partner alignment can outweigh formal board authority.

This is the unusual part of the story. In most companies, control follows ownership; a founder with a small stake can be outvoted. At OpenAI, control has followed relationships and the ability to mobilize the people who do the work and the partners who provide the compute and distribution. That is a real form of power, but it is also a contingent one. It depends on continued staff loyalty, on a partner whose interests remain aligned, and on a board that defers to the CEO in practice. None of those is guaranteed by a charter.

The open question is whether that arrangement scales. As OpenAI raises more capital and competes with Google, Anthropic, and Meta, governance will matter. Altman has built an empire on a structure designed to avoid one. Whether it holds depends on whether the mission and the money can stay aligned.

For the reader, the lesson is not that mission-driven structures fail. It is that structures are tested by the pressures they were designed to resist. OpenAI’s capped-profit design anticipated the tension between capital and mission and tried to contain it; the 2023 crisis showed that containment depends on people, not just paperwork. The for-profit pivot and the roughly $500 billion valuation are the next test. If the mission survives the money, it will be because the people with leverage chose to keep it — not because a document required them to.