OpenAI didn’t start with Sam Altman. It didn’t even start with a name. According to evidence unsealed on April 29, 2026, in the Musk v. Altman trial, the AI lab’s earliest structure, mission, and even hardware were shaped by Elon Musk—years before Altman became its public face.
Key Takeaways
- Elon Musk drafted OpenAI’s original mission statement and helped define its nonprofit governance long before Sam Altman joined.
- Nvidia CEO Jensen Huang personally delivered a DGX-1 supercomputer to the founding group in 2015—before OpenAI was formally incorporated.
- Emails show Sam Altman sought to anchor early OpenAI within Y Combinator, raising concerns among co-founders about dependency and control.
- Greg Brockman and Ilya Sutskever expressed internal concerns about Musk’s influence as early as 2016.
- The documents suggest Musk contributed over $30 million in seed funding and infrastructure before stepping away.
Musk Wrote the Mission Before Altman Led the Lab
The most jarring revelation from the trial exhibits? Musk didn’t just fund OpenAI—he authored its philosophical backbone. A redacted 2015 email thread, timestamped June 12, shows Musk circulating a draft titled “OpenAI: Mission & Structure.” In it, he lays out a vision for “non-corporate AI development, resistant to shareholder capture,” with strict limits on executive pay and profit motives. That language—nearly verbatim—appeared in OpenAI’s first public manifesto months later.
At the time, Altman wasn’t leading the project. He wasn’t even a full-time participant. He was CEO of Y Combinator, and his involvement was advisory at best. The driving force was Musk, working from notes exchanged with Greg Brockman, then a Stripe executive, and Ilya Sutskever, fresh off Google’s DeepMind team.
Musk didn’t just write the mission. He structured the initial governance. One document shows a proposed board with himself, Sutskever, Brockman, and two external academics—no YC representatives. That plan was later overturned. By late 2015, Altman had pushed for YC integration, citing resource access and startup credibility.
The Supercomputer That Started It All
Before cloud clusters and Azure deals, OpenAI ran on a single machine—personally delivered by Jensen Huang.
A photo dated August 3, 2015, shows a black Nvidia DGX-1 sitting on a folding table in what appears to be a Palo Alto garage. Handwritten label on the side reads “For the good guys.” According to accompanying internal notes, Huang handed it over during a private meeting with Musk, calling it “the first GPU stack built for AI that thinks.”
The DGX-1 was incredibly rare at the time—Nvidia had only built a dozen prototypes. Most were reserved for internal research or elite partners like Stanford and MIT. Huang’s decision to give one to an unincorporated group of AI idealists was, in industry terms, a moonshot endorsement.
- The machine had 8 Tesla P100 GPUs and 128GB of HBM2 memory.
- It achieved 170 teraflops of mixed-precision performance—unmatched for startups in 2015.
- It trained OpenAI’s first language models, including the precursor to GPT-1.
- Nvidia never invoiced for it. No formal agreement was signed.
The gesture wasn’t just generous. It was strategic. Huang reportedly told Musk: “If general AI happens, it won’t be in a boardroom. It’ll be in a garage with people who don’t care about profit.” That quote appears in a later email from Musk to Sutskever, dated September 14, 2015.
Altman’s YC Play—and the Pushback It Sparked
Sam Altman didn’t enter OpenAI as a neutral party. He came with Y Combinator in his back pocket.
Emails from October 2015 show Altman proposing that OpenAI operate as a “YC-backed nonprofit startup,” with shared office space, legal infrastructure, and recruiting pipelines. He argued that YC’s network could accelerate hiring and deployment. “We don’t need to reinvent HR or payroll,” he wrote. “We need to build models.”
But not everyone agreed. In a November 3 email, Ilya Sutskever replied: “YC has a culture of growth and scaling. We need a culture of restraint and caution. I worry the incentives don’t align.”
Concerns About Control
Brockman echoed that concern in a Slack message (later exported and preserved as evidence): “If YC controls the backend, they control the destiny. That’s not what we signed up for.”
The tension simmered. By early 2016, Musk was openly质疑 (the term appears in a forwarded email, untranslated) the direction. He had envisioned a lean, independent collective—what he called “a monastery for machine minds.” What he saw emerging was what he later described in a January 2016 email as “a startup in a nonprofit costume.”
He wasn’t wrong. By 2017, OpenAI had moved into YC facilities, adopted its HR systems, and begun using its investor outreach network. The nonprofit remained on paper. But the machinery of a high-growth tech company was already in motion.
The Funding Timeline No One Talks About
Musk didn’t just contribute ideas and ideology. He bankrolled the earliest phase.
According to financial records introduced as Exhibit 44-B, Musk transferred $4.5 million in the fall of 2015 via his personal account. Another $25 million followed in early 2016, split between two shell entities: Neural Holdings LLC and Future Fund Inc. Both were dissolved by 2018.
These transfers predate any known investment from Altman, Reid Hoffman, or Peter Thiel. They also predate OpenAI’s formal 501(c)(3) status, meaning Musk was funding an entity that didn’t legally exist.
“That’s either extreme faith or extreme risk tolerance,” said one financial analyst who reviewed the documents, speaking to original report. “Doing that today would raise red flags at the IRS.”
How Competing Labs Approached Early Structure
While OpenAI was wrestling with governance and funding in 2015, other AI research groups were making different bets. DeepMind, acquired by Google in 2014 for $500 million, operated under corporate ownership from day one. Its founders, including Demis Hassabis, accepted that scale required capital—and oversight. But they negotiated an AI ethics board and veto rights over controversial applications, conditions baked into their acquisition agreement.
Meanwhile, Anthropic took a hybrid path. Founded in 2021 by former OpenAI researchers Dario Amodei and Daniela Amodei, it adopted a “long-term benefit” charter and a unique ownership structure: a nonprofit majority stake in a for-profit entity. This model, inspired partly by OpenAI’s early ideals, was designed to resist takeover pressure. By 2025, Amazon had invested $4 billion into Anthropic, but the nonprofit board retained control over core decisions.
Then there’s Meta’s FAIR lab. No nonprofit wrappers. No investor restrictions. FAIR published openly and moved fast. Its Llama models were released with permissive licenses, betting that transparency would build trust. Critics argued this only accelerated commercial mimicry. But internally, Meta treated FAIR as a strategic asset, not a public trust.
Each model reflects a different answer to the same question: can AI safety survive inside systems built for growth? OpenAI’s pivot from Musk’s vision to a capped-profit model in 2019—and its $10 billion Microsoft partnership—mirrors the pressure others have faced. But unlike DeepMind or Anthropic, OpenAI’s shift happened after its foundational tools and culture had already been shaped by outside infrastructure.
The Policy Vacuum Around Early-Stage AI Funding
The Musk v. Altman trial isn’t just about personalities. It exposes a regulatory blind spot: there are no federal rules governing private funding of pre-incorporated AI research collectives. In 2015, when Musk moved $30 million through shell entities, he wasn’t breaking any laws. No disclosure was required. No ethics review was triggered. The IRS doesn’t track donations to informal groups unless they later apply for nonprofit status.
This loophole still exists. In 2024, billionaire investor Nat Friedman funded a group called Frontier Minds with $20 million in undisclosed transfers. The group, working on open-source AI alignment tools, never incorporated. They used cloud credits donated by Google and operated out of a rented house in Berkeley. No filings. No oversight.
Other countries are starting to act. The EU’s AI Act, updated in 2025, now requires any group receiving over €5 million in private funding for “general-purpose AI development” to register and disclose backers—even if unincorporated. Canada’s 2024 Digital Innovation Framework mandates transparency for AI projects using public cloud infrastructure subsidized by government grants.
In the U.S. that kind of scrutiny doesn’t apply. The National Science Foundation funds academic AI work with strict reporting rules. But private funding? It’s the Wild West. The OpenAI story shows how much can happen in that gray zone: machines delivered, code written, models trained—all before a legal entity exists. The policy lag matters. Because once infrastructure is in place, the direction of travel becomes harder to change.
The Bigger Picture: Why Governance Happens in the Shadows
The early OpenAI story isn’t unique because of Musk or Altman. It’s typical. Most tech movements begin in ambiguity. WhatsApp started as a failed status app. Instagram launched without video. The real decisions—about structure, values, control—happen before the press release, before the incorporation date, often in a series of late-night emails and whiteboard sketches.
What makes OpenAI different is that its early choices now carry global weight. The models it trains influence healthcare, education, and national security. Its business model is studied by governments and startups alike. And yet, the pivot from Musk’s vision to Altman’s execution occurred without public input, without regulatory notice, in private messages among a half-dozen people.
This isn’t just about one organization. It’s about how power accumulates in AI. Influence isn’t always claimed. Sometimes it’s embedded—in server access, in hiring pipelines, in the quiet assumption that certain partners are “necessary.” Y Combinator didn’t seize control of OpenAI. It was invited in, for practical reasons. But each integration narrowed the range of possible futures.
We’re now at a moment where early governance choices have real consequences. When OpenAI released GPT-4, it didn’t just ship a model. It shipped a legacy: a technical architecture shaped by 2015 hardware limits, a corporate structure influenced by 2017 HR systems, a strategic direction set by decisions made before the CEO was even on payroll.
The trial won’t settle who “built” OpenAI. But it forces a reckoning: if the most powerful AI lab on Earth was steered by unincorporated agreements, undocumented donations, and unsupervised infrastructure transfers, then the systems meant to guide AI aren’t just weak. They’re missing.
What This Means For You
If you’re building an AI startup, the Musk v. Altman trial isn’t just corporate drama—it’s a case study in how mission drift happens. The documents show how a nonprofit vision, written in good faith, can be Quietly Reshaped by infrastructure dependencies. Relying on YC’s systems seemed practical. But it also handed influence to an organization built for speed, not restraint.
For developers, the takeaway is starker: the tools you use today—APIs, cloud credits, incubator programs—aren’t neutral. They come with embedded incentives. OpenAI started with a garage and a DGX-1. It ended with a $10 billion Microsoft deal. That trajectory wasn’t inevitable. It was chosen, one integration at a time.
So who really built OpenAI? The answer isn’t coded in a model. It’s buried in emails, transfer logs, and a single supercomputer that arrived before the name was picked. The trial won’t undo history. But it might force the AI world to stop pretending that ideals survive untouched inside startup engines.
If OpenAI’s mission was meant to resist capture, what does it mean that its most defining choices were made before its CEO had even committed full-time?
Sources: The Verge, Bloomberg Technology
