The Complete History of OpenAI and ChatGPT: From a Nonprofit Dream to a $730 Billion Corporate Machine

Let me tell you a story. In 2015, a group of very smart, very rich people stood up and said: "AI is dangerous, so we're going to build it ourselves — but we're going to do it safely, as a nonprofit, with no profit motive distorting our judgment." That was the pitch. That was the founding vision of OpenAI.

Ten years later, OpenAI is a $730 billion for-profit corporation, burning through $17 billion a year, competing with the company that gave it $13 billion to survive, and the word "safely" has been deleted from its mission statement.

That's a story worth telling.

This is not a puff piece. This is not a hit piece either. This is a chronological walk through every major milestone, product launch, controversy, boardroom coup, funding round, and turning point in the history of one of the most consequential technology companies ever built — with honest opinions on what it all means.

The Founding — The Nonprofit That Never Really Was

On December 11, 2015, OpenAI was announced to the world with a $1 billion pledge from a list of co-founders that reads like a Silicon Valley Hall of Fame: Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, Wojciech Zaremba, John Schulman, Reid Hoffman, Peter Thiel, and Jessica Livingston.

The stated mission: "to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return."

That last phrase is doing a lot of work.

The founding mythology matters here. Elon Musk — who at this point was publicly sounding alarms about AI existential risk — helped create OpenAI specifically because he feared what Google was doing with AI. If the big corporations were going to build AGI anyway, the argument went, better to have a safety-focused nonprofit doing it in the open, where the results belonged to everyone.

That story about being "open" stuck around for a while. It's worth remembering because it shaped how a lot of people — researchers, journalists, the public — came to trust OpenAI.

The $1 billion pledged was not all cash upfront. Musk and Altman committed to contributing over time. The actual cash available to start was much smaller. But the signal was loud: this was serious, it had serious backing, and it was going to be different.

First Products — Teaching Machines to Play Games

In April 2016, OpenAI released its first real product: OpenAI Gym. Gym was an open-source toolkit for reinforcement learning research — essentially a standardized collection of environments (Atari games, physics simulations, robotic control tasks) that researchers could use to train and benchmark AI agents.

This was genuinely useful work. Reinforcement learning is a branch of machine learning where an AI learns by trial and error — take an action, get a reward or punishment, adjust behavior accordingly. The problem before Gym was that everyone building reinforcement learning systems was also spending enormous effort just setting up the testing environments. Gym standardized that. Researchers could now benchmark against common tasks and actually compare results.

In December 2016, OpenAI followed with Universe — a platform for training AI agents on a far broader range of real-world software environments, including browser-based games and desktop applications. The ambition was to train agents that could use a computer the way a human does.

These were research tools. They were also the building blocks of something much more powerful.

The Transformer Arrives (Not From OpenAI, But It Changes Everything)

I want to pause here and give credit where it belongs, because this detail gets glossed over constantly.

In June 2017, researchers at Google Brain published a paper called "Attention Is All You Need." That paper introduced the Transformer architecture — the neural network design that underlies GPT, ChatGPT, Claude, Gemini, and essentially every large language model that matters today.

OpenAI did not invent Transformers. Google did. This matters because a lot of the breathless coverage of OpenAI's technical genius skips over the fact that the foundational architecture they built their empire on came from a competitor's research paper — a paper that was published openly for the entire field to build on.

What OpenAI did was scale it. That's genuinely impressive engineering. But the credit for the underlying idea belongs in Mountain View, not San Francisco.

Meanwhile in 2017, OpenAI was training OpenAI Five — a reinforcement learning system designed to beat professional players at Dota 2, one of the most complex real-time strategy games ever made. This project ran through 2019 and eventually resulted in a system that defeated the world champion Dota 2 team OG in April 2019. It was a genuine demonstration of what reinforcement learning could do at scale.

Also in 2017: OpenAI allocated roughly $7.9 million specifically for cloud computing — about a quarter of its total functional expenses that year. This number is important context. The cost of compute required to train frontier AI models was already becoming a dominant constraint. That constraint would define everything that came after.

GPT-1 and the Beginning of the Language Model Era

In June 2018, OpenAI published Generative Pre-trained Transformer 1 (GPT-1) — a language model with 117 million parameters trained on BookCorpus, a dataset of unpublished books. The paper, "Improving Language Understanding by Generative Pre-Training," introduced the idea of pre-training a large language model on raw text and then fine-tuning it for specific tasks.

This was not ChatGPT. GPT-1 was a research paper, not a product. The outputs were crude. The model struggled with coherence over more than a few sentences. But the approach — train a huge model on massive amounts of text, then adapt it — turned out to be the right direction. Everyone in NLP (natural language processing) took notice.

Also in 2018: Elon Musk resigned from OpenAI's board, citing potential conflicts of interest with Tesla's own AI development for autonomous vehicles. He remained a donor. The official story was that Tesla's work on Autopilot could put him in conflict with OpenAI's research directions.

The less official story, revealed later through emails and lawsuits, is considerably messier. Musk had grown increasingly frustrated with what he saw as OpenAI's slow progress and had attempted to take control of the organization as its CEO. Altman and the board declined. Musk left. He later claimed the departure was mutual. The emails suggest otherwise.

The Musk situation is worth flagging now because it becomes a decade-long subplot.

The Capped-Profit Pivot, the GPT-2 Saga, and Microsoft's Bet

Three big things happened in 2019.

First: The GPT-2 Controversy. In February 2019, OpenAI announced it had trained GPT-2 — a far more capable language model with 1.5 billion parameters — but declined to release the full model publicly. Their stated reason: fear that GPT-2 could be used to generate convincing fake news, spam, and disinformation at scale.

The AI research community had opinions about this decision.

Some researchers praised it as responsible disclosure. Most were skeptical. The staged release policy — releasing the model in progressively larger versions over several months while monitoring for misuse — read to many as a PR strategy that generated enormous attention for OpenAI while providing little actual safety benefit. By November 2019, OpenAI released the full 1.5 billion parameter model anyway, noting they had found "no strong evidence of misuse so far."

What GPT-2 really demonstrated was that OpenAI had figured out how to capture public and media attention through safety theater. That playbook did not go away.

Second: The Capped-Profit Structure. In March 2019, OpenAI restructured, creating OpenAI LP — a "capped-profit" for-profit subsidiary under the control of the nonprofit. The justification was practical: you cannot attract top AI talent or raise the billions of dollars needed to train frontier models as a pure nonprofit. The capped-profit structure limited early investors to a 100x return on their investment, theoretically preserving the mission.

Initial investors were told to "think of investments in the spirit of donations." The operating agreement warned the company "might never turn a profit."

You can see where this is going.

Third: Microsoft's $1 Billion Investment. In July 2019, Microsoft invested $1 billion in OpenAI and the two companies entered a multi-year partnership to develop AI supercomputing technology on Azure. OpenAI's models would run on Microsoft's cloud. Microsoft would get to market and sell OpenAI's technologies.

This partnership felt like a lifeline. It was. OpenAI at this point was burning through money training large models on compute that cost tens of millions of dollars per run. Microsoft's Azure infrastructure gave them the scale they needed. In exchange, Microsoft got early access to some of the most powerful AI technology ever built.

It seemed like a fair trade. In hindsight, it was the deal that eventually broke everything.

GPT-3 Drops the World's Jaw

In June 2020, OpenAI released GPT-3 — 175 billion parameters, trained on a massive corpus of internet text, books, and code. The jump from GPT-2 to GPT-3 was not incremental. It was categorical.

GPT-3 could write coherent long-form essays. It could answer questions that seemed to require reasoning. It could write code. It could summarize documents, translate languages, and hold something resembling a conversation. The outputs weren't perfect — they hallucinated facts, confused details, occasionally went in completely wrong directions — but the quality was shocking relative to everything that had come before.

Developers who got access to the early beta were stunned. Twitter threads went viral showing GPT-3 writing convincing short stories, generating working code from natural language descriptions, and completing prompts in a dozen different styles. The hype was real, and for once, justified.

OpenAI released GPT-3 via API access only — no open weights, no public model download. The API approach was a deliberate business decision. OpenAI controlled the model, controlled access, and could charge for usage. The "Open" in OpenAI was becoming an artifact of history.

The foundational research insight behind GPT-3 was scaling. OpenAI published "Scaling Laws for Neural Language Models" in January 2020, demonstrating that model performance on language tasks scaled predictably with model size, training data, and compute. If you made the model bigger and gave it more data, it got better. Reliably. Predictably. That finding was explosive because it meant that whoever could spend the most on compute would build the best models.

Compute was now the moat. And Microsoft owned the cloud.

Codex, DALL-E, and the Birth of the Product Era

By 2021, OpenAI shifted from pure research into product mode.

In January, they released DALL-E — a model that could generate images from text descriptions. Type "an avocado shaped like an armchair" and get a photorealistic image of exactly that. The outputs were imperfect by today's standards, but DALL-E demonstrated that the same Transformer approach that worked for text could be extended to images. The world's artists started to pay attention, and not in a good way.

In August, Codex dropped — a version of GPT-3 fine-tuned specifically on code from GitHub. Codex could take natural language descriptions and produce working Python, JavaScript, and other programming language code. The demo was jaw-dropping. You could type "write a function that takes a list of numbers and returns the sum of all even numbers" and get working code back.

Two months later, GitHub and OpenAI jointly launched GitHub Copilot — a coding assistant built on Codex, integrated directly into VS Code and other editors, suggesting code completions in real time as you typed.

GitHub Copilot was the first mass-market AI product built on OpenAI's models. It was also the first shot fired in what would become a war between Microsoft and OpenAI that plays out years later. Keep that in mind.

This is also where the copyright questions started getting loud. Copilot was trained on public GitHub repositories, including code licensed under GPL and other copylicenses that arguably required attribution or derivative work restrictions. Developers noticed Copilot reproducing verbatim snippets of copyrighted code without attribution. The lawsuits would follow.

GPT-3.5, DALL-E 2, and the Product That Changed Everything

DALL-E 2 launched in April 2022 with significantly improved image quality. Text-to-image AI went from a curiosity to a culture war overnight. Artists, illustrators, and photographers looked at DALL-E 2 — and Stable Diffusion, and Midjourney — and recognized an existential threat to their livelihoods. That debate has not resolved.

In August 2022, InstructGPT demonstrated a critical breakthrough: using human feedback to fine-tune language models to follow instructions more reliably. This technique, called RLHF (Reinforcement Learning from Human Feedback), made GPT-3 derivatives significantly more useful and less prone to producing harmful or incoherent outputs. It was the technical foundation for everything that followed.

Then on November 30, 2022, OpenAI launched ChatGPT.

ChatGPT was not the most technically sophisticated model OpenAI had built. It was GPT-3.5 — an improvement on GPT-3 — with RLHF fine-tuning and a simple chat interface. The innovation was access and design: anyone could go to chat.openai.com, make a free account, and talk to an AI in plain English.

Five days after launch, ChatGPT had one million users. Two months later, it hit 100 million. That was the fastest any consumer product in history had reached that milestone. Nothing had ever grown that fast. Not Instagram. Not TikTok. Nothing.

For context: ChatGPT's traffic spike in those first months was large enough that it contributed to measurable strain on web infrastructure. The AI hosting industry started to feel the pressure — and if you want to understand just how badly AI companies have been destroying the economics of web hosting since then, go read our article on The RAMpocalypse.

The world had changed. Everyone could feel it. The question was where it led.

The Year of Hype, Hearings, and the Boardroom Coup

2023 was the year OpenAI went from "interesting AI company" to "the company reshaping civilization."

January 2023: Microsoft announced a multi-billion dollar extension to its OpenAI investment — later confirmed at $10 billion total across the partnership. Microsoft started integrating GPT-4 into Bing, Office, and Azure. Satya Nadella called it the biggest bet in Microsoft's history.

February 2023: ChatGPT Plus launched at $20/month — the first paid tier. OpenAI was not yet profitable, burning through compute costs at a rate that would eventually reach $8+ billion per year. But subscription revenue provided a stable foundation.

March 2023: GPT-4 launched. The jump from GPT-3.5 to GPT-4 was again categorical. GPT-4 passed bar exams in the 90th percentile. It scored in the 99th percentile on the SAT Math section. It could analyze images, reason through multi-step problems, and write code that actually worked most of the time. The benchmarks were startling.

The API for GPT-4 opened a gold rush of AI startup activity. Thousands of companies built products on top of GPT-4 overnight. The AI venture capital market caught fire.

March 2023: Italy became the first country to ban ChatGPT, citing GDPR violations and data privacy concerns. OpenAI made concessions, and the ban was lifted in April after about a month.

May 2023: Sam Altman testified before the U.S. Senate, calling for AI regulation. That session is now infamous for two reasons. First, senators asked questions that revealed startling technical illiteracy about how AI systems actually work. Second, Altman's performance was widely praised as statesmanlike — calling for regulation of his own company in ways that, critics noted, would primarily benefit established players like OpenAI while raising barriers to entry for competitors. Regulatory capture has a lot of faces.

May 2023: OpenAI lost $540 million developing ChatGPT and GPT-4 in 2022, according to reporting at the time. The company was growing revenue fast, but spending faster.

July 2023: OpenAI launched custom GPTs and plugins, expanding the platform into an app ecosystem.

November 2023: The Altman firing. This is the most dramatic episode in OpenAI's history, and possibly in tech history.

On November 17, 2023, the OpenAI board fired Sam Altman as CEO, releasing a statement that he had "not been consistently candid in his communications with the board." The board at the time consisted of Ilya Sutskever (chief scientist), Adam D'Angelo (Quora CEO), Tasha McCauley (tech entrepreneur), and Helen Toner (Georgetown AI safety researcher). None of them were commercial executives. This was the nonprofit board governing a company that had become one of the most commercially significant in the world.

The firing blindsided almost everyone. Altman was on a call learning about it when it happened.

The next 96 hours were chaos. Greg Brockman, OpenAI's president, resigned. Nearly all of OpenAI's 770 employees signed a letter threatening to resign and follow Altman to Microsoft, where he had already been offered a position to lead a new AI initiative. Microsoft — which had invested over $13 billion in OpenAI — made clear it was not happy.

By November 22, Sam Altman was reinstated as CEO. The board was replaced. The nonprofit oversight structure remained, technically, but its practical authority had been exposed as hollow. A CEO could be fired and rehired in five days based almost entirely on the threat that the company would collapse without him.

Microsoft was subsequently given a non-voting observer seat on the board — a sign of just how much leverage the company had accumulated.

The specific reasons for the firing were never fully disclosed publicly. The board members who fired Altman largely left and have not spoken in detail about it. Various theories circulated: a breakthrough in AI capabilities that scared board members, concerns about Altman's honesty in communications, disagreements about safety priorities versus commercial speed. The honest answer is we still don't fully know.

What the incident revealed was clear: the nonprofit governance structure was not built to manage a company at this scale. The board that was supposed to provide independent safety oversight had just demonstrated it could be overridden by the collective threat of mass employee resignation and a billionaire investor making a phone call.

Departures, Sora, GPT-4o, and the Safety Exodus

2024 was the year the safety researchers started leaving.

January 2024: Arizona State University became the first university to purchase ChatGPT Enterprise — a signal that OpenAI was moving aggressively into the education market.

February 2024: Elon Musk filed a lawsuit against OpenAI and Sam Altman, alleging breach of contract and claiming OpenAI had abandoned its nonprofit mission. The lawsuit was eventually dropped, refiled, and turned into a recurring legal saga that continues through 2026. Musk had also, by this point, launched his own AI company called xAI with a competing chatbot called Grok. His financial interest in attacking OpenAI's reputation was not subtle.

February 2024: OpenAI released Sora — a text-to-video generation model. You type a description, you get a short video. The demos were extraordinary. The concern among filmmakers and visual artists was immediate. OpenAI did not rush Sora to public release — it spent months in limited access, partly for technical reasons, partly because the content moderation challenges with video generation are significantly more complex than with images or text.

May 2024: OpenAI released GPT-4o — the "o" standing for "omni." GPT-4o could natively process and generate text, audio, and images in a single model, enabling genuinely conversational interactions with extremely low latency. The demo where GPT-4o assisted a user through a math problem in real-time voice conversation, adjusting its explanation based on the user's responses, was genuinely impressive.

May 2024: Ilya Sutskever left OpenAI. This was significant. Sutskever was not just any employee — he was one of the co-founders, the chief scientist, and the person whose name had appeared on the board that fired Altman. He founded a new company called Safe Superintelligence Inc. (SSI) with a stated focus purely on AI safety research.

The month after Sutskever left, Jan Leike — head of OpenAI's Alignment team — resigned, posting a blunt statement on social media that OpenAI had subordinated safety culture to product speed. He wrote that safety culture had "taken a back seat to shiny products" and that he disagreed with leadership "about the core trade-offs between safety and products."

Throughout 2024, the safety team hemorrhaged talent. Multiple researchers who had spent years building OpenAI's alignment research programs left for Anthropic, DeepMind, academia, or their own ventures. The pattern was consistent enough that it was impossible to ignore.

September 2024: Mira Murati, OpenAI's Chief Technology Officer, resigned along with several other senior executives. The wave of departures was accelerating.

October 2024: OpenAI raised $6.6 billion in a funding round at an $157 billion valuation — at the time, the largest venture funding round in history. The investor list included Thrive Capital, Microsoft, Nvidia, SoftBank, and others. The capital gave OpenAI runway but did not reduce the burn rate.

Also in October 2024: OpenAI launched o1 — a "reasoning model" that spent more time thinking through problems before answering, showing dramatically improved performance on math and logic benchmarks. And OpenAI launched a ChatGPT web search feature, competing directly with Google.

November 2024: Multiple news organizations filed copyright infringement lawsuits against OpenAI, following the New York Times' lawsuit filed in December 2023 alleging that OpenAI had trained its models on millions of NYT articles without permission or compensation. These cases remain ongoing through 2026.

The debate over training data copyright is genuinely unresolved. If you want to use a book to teach yourself something, that's fine. If a company uses millions of books to train a commercial AI product and generates billions in revenue from that training, does the author deserve compensation? Courts are still working through this. The legal outcomes will reshape the AI industry.

Restructuring, Stargate, GPT-5, and the $500B Bet

2025 was the year OpenAI stopped pretending it was anything other than a for-profit corporation.

January 2025: President Donald Trump announced the Stargate Project — a $500 billion joint venture between OpenAI, SoftBank, Oracle, and MGX to build AI data center infrastructure across the United States. This was the largest AI infrastructure commitment in history by an enormous margin. It was also a political statement: this was American AI, backed by American government favor.

January 2025: OpenAI acquired Torch, a healthcare technology startup, for approximately $60 million, signaling an expansion into medical AI.

February 2025: Elon Musk, through a group of investors, made a $97.4 billion bid to acquire OpenAI outright. The board rejected it. Musk subsequently escalated his legal campaign against the company.

March 2025: OpenAI raised $40 billion in a Series F funding round — at the time, the largest private investment in AI history. The round was led by SoftBank, which committed $22.5 billion, with the remainder coming from other institutional investors. The condition: OpenAI needed to complete its transition from a capped-profit structure to a standard for-profit Public Benefit Corporation by early 2026. If it didn't, SoftBank could pull back half the investment.

August 2025: GPT-5 launched — a model that represented meaningful progress in reasoning, coding, and complex task completion. Unlike the transition from GPT-3 to GPT-4, GPT-5 was notable partly because it used less training compute than GPT-4.5 while delivering better performance — a sign that the era of "just throw more compute at it" scaling was reaching diminishing returns, and algorithmic efficiency was becoming the new battleground.

October 2025: OpenAI completed its restructuring as a Public Benefit Corporation (PBC). Microsoft received a 27% equity stake in the for-profit entity and model access rights through 2032. OpenAI committed to a substantial Azure cloud purchase agreement as part of the deal.

The mission statement that had once said OpenAI would develop AI "safely and beneficially" was updated. The word "safely" was removed.

That deletion was not accidental. Nothing in a legal document of this importance is accidental. The people writing those documents know exactly what they're removing and what it means.

November 2025: California Attorney General Rob Bonta approved OpenAI's restructuring after extracting concessions, including a commitment that charitable assets would be preserved and safety would remain a priority. The legal settlement did not address whether the company's actual behavior aligned with those commitments.

December 2025: Disney announced a $1 billion investment in OpenAI and a three-year licensing deal for Sora video generation.

December 2025: SoftBank added another $22.5 billion to its OpenAI investment. The total external funding raised in 2025 exceeded $60 billion.

By the end of 2025, OpenAI's annualized revenue had crossed $20 billion. Its weekly active users on ChatGPT reached 800 million. It had burned through $8.5 billion in cash that year. It projected burning $17 billion in 2026.

$730 Billion, 900 Million Users, and the Wars Begin

As of the writing of this article in March 2026, OpenAI is the fastest-growing technology company in history. Its weekly active user count has reached 910 million. Annualized revenue has crossed $25 billion. It is reportedly targeting a $1 trillion valuation ahead of a planned 2026 IPO filing.

It is also at war with almost everyone.

The Amazon deal: In early 2026, OpenAI signed a landmark $50 billion partnership with Amazon, making AWS the exclusive third-party cloud provider for OpenAI Frontier — its enterprise platform. Microsoft, which had an existing agreement establishing Azure as the primary cloud infrastructure for OpenAI's commercial operations, viewed this as a breach of contract and began weighing legal action. The company that built OpenAI is now potentially suing it.

The GitHub angle: Reports from March 2026 indicate OpenAI engineers built an internal code-hosting platform to escape their own frustration with GitHub's ongoing reliability issues — four separate outages in February 2026 alone — and the company is considering commercializing it as a GitHub competitor. Microsoft owns GitHub. Microsoft also holds a 27% stake in OpenAI. We wrote extensively about what it would actually take to build a GitHub competitor and what this move signals about OpenAI's strategy — go read ChatGPT Is Building Its Own Code Repository for the full technical and strategic breakdown.

The Pentagon contracts: OpenAI has contracts with the Department of Defense permitting use of its AI "for all lawful purposes" within DoD policy. Its Chief Product Officer Kevin Weil was commissioned as a lieutenant colonel in the U.S. military. The company that was founded to ensure AI "benefits all of humanity" is now an active defense contractor.

The money: OpenAI projects $115 billion in total spending through 2029. It expects to be cash-flow positive only by 2030 at the earliest. It is spending significantly more than it earns, funded by investors who believe the bet will pay off when AGI — or something close enough to matter commercially — arrives. Whether that bet is rational depends entirely on whether you believe frontier AI models will continue to get dramatically more capable on the current trajectory.

The Products That Actually Changed the World

Let me pause the timeline for a second and be direct about what OpenAI actually built and why it matters.

ChatGPT is the most rapidly adopted software product in human history. Whatever you think about OpenAI as a company, the product gave hundreds of millions of people access to a tool that can write code, explain complex topics, draft documents, debug errors, and assist with an extraordinary range of cognitive tasks. Developers who use ChatGPT effectively can complete certain kinds of work in a fraction of the time. That's real. That matters.

GitHub Copilot, powered by OpenAI's Codex, fundamentally changed how a lot of developers write code. We have a full breakdown of how to use AI coding tools effectively without destroying your own skills at How to Use AI Coding Tools to Speed Up Development.

DALL-E and Sora democratized visual creation and raised genuinely important questions about the future of creative work and intellectual property.

The GPT API enabled thousands of companies and developers to build products that would have been impossible or unaffordable without access to frontier models. If you've used an AI-powered app in the last two years, there's a reasonable chance it was built on OpenAI's API.

The o1 reasoning model and its successors marked a shift from "predict the next word really well" toward "actually think through problems step by step" — a meaningful architectural evolution with real implications for how capable these systems become.

These are genuine contributions. The products work. The technology is impressive. The capability gains from 2020 to 2026 have been extraordinary.

None of that erases the questions about how OpenAI got here and where it's going.

The Things OpenAI Would Rather You Forget

The safety team is gone. The researchers who built OpenAI's alignment programs — the people whose job was to figure out how to make powerful AI systems behave in ways that are safe and consistent with human values — have largely left. Ilya Sutskever founded a separate safety company. Jan Leike went to Anthropic. Multiple alignment researchers departed across 2023 and 2024. The people who replaced them are oriented toward shipping products. OpenAI's current leadership has not prioritized rebuilding the safety research programs with the same seriousness.

"Open" is a lie. The GPT-4 and GPT-5 model weights are not published. The training data is not disclosed. The company publishes some research papers, but the most commercially sensitive work stays internal. The name is a historical artifact. OpenAI is one of the most closed major AI labs operating today.

The Kenya moderation workers. In 2021, OpenAI sent text data to Sama, a San Francisco-based company employing workers in Kenya, to label and filter toxic content for ChatGPT's training. The content included detailed descriptions of violence, including sexual violence. The workers Time magazine interviewed described themselves as mentally scarred. The people who did the invisible, traumatizing work of making ChatGPT safe to use were paid low wages in Kenya while OpenAI raised billions of dollars in San Francisco.

The GPT-2 "safety" release. Looking back, the staged release of GPT-2 in 2019 generated enormous positive press for OpenAI as a responsible actor, built months of hype that drove signups and API interest, and ultimately resulted in a model that was going to be released anyway. If you're cynical about it — and there are reasonable grounds to be — it was one of the most effective PR campaigns in tech history dressed up as safety research.

The mission drift. OpenAI was founded as a nonprofit to ensure AGI benefits humanity. It became a capped-profit company to attract investment. It became a Public Benefit Corporation to complete the restructuring deal. At each step, the financial structures required to operate at scale required compromises to the founding ideals. The people who warned this would happen — including, ironically, Elon Musk in his lawsuit filings — were right about the direction even if wrong about the motives.

Where This Goes

The honest answer is that nobody knows.

OpenAI is betting everything on continued capability gains leading to products valuable enough to justify the spending. They need ChatGPT and its successors to become so deeply embedded in how businesses operate that switching costs are prohibitive. They need AI agents — systems that can autonomously complete multi-step tasks — to become the enterprise software layer of the 2030s. And they need to get there before Google, Meta, Anthropic, or a Chinese lab beats them to AGI or near-AGI capabilities.

That's a real bet. It might pay off. The capability gains of the last five years have been real and substantial. If GPT-6 or GPT-7 represents the same kind of leap that GPT-3 to GPT-4 did, the investment thesis looks reasonable.

But the structural problems are also real. The company burns more cash than it earns. Its relationship with Microsoft — its largest investor and primary cloud provider — is in active deterioration. Its safety research leadership is gone. Its governance structure has been exposed as incapable of providing meaningful oversight. And the regulatory environment in the U.S. and Europe is becoming more complex by the month.

OpenAI is simultaneously the most impressive technology company of the 2020s and one of the most concerning. Both of those things are true. Holding both in your head at the same time is the only honest way to look at it.

What started as a nonprofit safety research lab is now a $730 billion corporation fighting with its biggest investor, competing with its primary backer's most valuable platform, burning through $17 billion a year, and projecting cash-flow-positive operations by 2030 — while betting the whole thing on capabilities that nobody fully understands yet.

That's the story. Make of it what you will.