What the business is
An AI research company in San Francisco, founded in 2015 as a nonprofit lab and now selling access to its large language models, with a staff of around 300.
The bet
That releasing a free chat demo of its LLM — rough, unsourced, knowing little past 2021 — would prove the technology and seed a platform.
How it started
OpenAI built the chat interface in early 2022 as a training tool: reinforcement learning from human feedback needed exactly the constant dialogue a chatbot produces. GPT-3 — 175 billion statistical connections, trained on about two-thirds of the internet, all of Wikipedia and two book datasets — was powerful but hard to steer. Beta testers didn't embrace the bot: 'it wasn't clear to people what they were supposed to talk to the chatbot about', president Greg Brockman said, and OpenAI nearly killed the project. A parallel effort to build expert chatbots for professionals stalled for lack of training data.
What happened
As a Hail Mary — 'I'll admit that I was on the side of, like, I don't know if this is going to work', Brockman admitted — OpenAI put ChatGPT out free on November 30, 2022, announced in a single Altman tweet. It passed 1 million users in five days. CTO Mira Murati called the virality 'definitely surprising'; Altman said he 'would have expected maybe one order of magnitude less of everything'. Google reportedly declared an internal 'code red', and in mid-January 2023 OpenAI opened a waitlist for a commercial version.
How it ended up
By late January 2023 the six-week-old demo carried a reported $29 billion valuation, a tender-offer feeding frenzy among VCs, and an expanded Microsoft partnership of up to $10 billion announced the week of January 23 — alongside documents seen by Fortune showing how unprofitable the business remained. GPT-4 sat in beta testing, expected imminently.
What has to be true
The chat interface was already built and paid for as RLHF infrastructure; releasing it cost OpenAI little but a reputation bet.
Free, open access removed every adoption barrier — no waitlist, no API integration, no prompt skill — which is what let a demo go consumer-viral.
Shipping rough and unsourced was the point of the research-preview framing: it lowered the bar for 'impressive' while Altman publicly managed expectations downward.
A near-cancellation meant the launch carried option value, not conviction: the release was a cheap experiment that happened to catch a generation-defining wave.
What can be applied
Distribution can be a research artifact: the interface built to harvest training feedback became the product — and shipping it free turned a capability demo into a category-defining distribution win.
Aftermath
As of January 25, 2023, ChatGPT was free to the public with a commercial version in the works via waitlist, Microsoft's up-to-$10 billion expanded partnership was just announced, GPT-4 was in beta, and OpenAI — around 300 staff — had moved from research lab to the center of the tech industry's agenda, with its profitability still unproven and its governance under question.
FOLLOW THE EVIDENCE