OpenAIInside OpenAI's path from research lab to $3B ARR platform.

OpenAI began life in 2015 as a non-profit AI research lab with a stated mission of "ensuring AGI benefits all of humanity." For its first six years it published papers, released the occasional eye-catching demo (DALL·E, GPT-3), and lived off donations and a $1B Microsoft commitment. Nothing about that origin predicted the company that, by late 2024, would post $3.7B in revenue and serve 200M weekly active users.

What changed wasn't the research. The research had been steadily compounding for years. What changed was distribution. ChatGPT, launched in November 2022 as a low-priority "research preview," did something that no prior AI product had done, it gave a normal person a useful experience in their first session. Five days later it had a million users. Two months later it had a hundred million. The lab had accidentally shipped the fastest-adopted consumer product in history.

1. 1

   2015

   Founded as a non-profit

   Mission statement: ensure AGI benefits all of humanity.

2. 2

   2019

   Capped-profit + Microsoft $1B

   Structure flipped to enable commercial scaling. Azure becomes the exclusive compute partner.

3. 3

   2022

   ChatGPT research preview

   1M users in 5 days. 100M in 2 months. Fastest consumer product launch in history.

4. 4

   2023

   Board crisis weekend

   Sam Altman fired and re-hired in 5 days. Microsoft's leverage becomes public.

5. 5

   2024

   Corporate restructure

   Non-profit veto removed. $3.7B revenue, $13B+ Microsoft commitment.

OpenAI's monetisation is structured as a three-tier funnel that prints fundamentally different margins. ChatGPT (free + Plus + Team + Enterprise) is the consumer wedge, high volume, mass-market pricing, and the layer where most users meet the brand. The API is where the actual platform business lives: priced per token, used by thousands of SaaS companies, and the layer that turned every founder into an OpenAI distributor without OpenAI having to hire a single sales rep for them. GPTs and the GPT Store sit on top, attempting to capture some of the value those founders are creating.

The genius of this stack is that each tier subsidises the next. ChatGPT generates the user feedback that improves the models. The improved models make the API more attractive. The API ecosystem creates demand for the underlying compute, which justifies the next round of model investment. None of the tiers in isolation is defensible, but the combination compounds.

Most discussions of OpenAI's competitive position focus on the model weights, but raw model capability has commoditised faster than anyone predicted. Llama, Claude, Gemini and Mistral all match or exceed GPT-4 on most public benchmarks. If the moat were the model, OpenAI would already be in trouble.

OpenAI has three years of preference data from real ChatGPT conversations, scaled annotation pipelines staffed by thousands of contractors, and an experimentation system that lets them ship a model variant to a small slice of users and read the result in days. Competitors can copy the architecture; they cannot copy the feedback loop.

This is why the most useful frame for OpenAI is not "AI lab" but "data company that happens to train models." The labelling, the rating, the safety classifiers, that's where the work is.

Every conversation about OpenAI eventually arrives at Microsoft. The original 2019 deal, $1B in compute credits in exchange for an exclusive Azure partnership, has grown into a $13B+ commitment and a structural dependency: OpenAI doesn't own its compute, and Microsoft has commercial rights to OpenAI's models in perpetuity (with limits that have been re-negotiated multiple times).

This is the classic platform tradeoff played at maximum stakes. Microsoft gives OpenAI scale they couldn't otherwise afford. In exchange, Microsoft gets to ship Copilot, Azure OpenAI, and the broader enterprise integration story without having to build the underlying models themselves. Both companies need the relationship to work, and both have publicly hedged. Microsoft has invested in alternate model providers (and built its own MAI-1). OpenAI has built relationships with Oracle and SoftBank for the Stargate compute project.

The November 2023 board crisis (when OpenAI's board fired and then re-hired Sam Altman over a weekend) was, in part, a referendum on this dependency. Microsoft's threat to absorb the entire research staff if Altman wasn't reinstated revealed who actually controlled the strategic optionality. Twelve months later, OpenAI announced a corporate restructure to remove the non-profit's veto over commercial decisions.

Three things are transferable to almost any product team. First: the API wedge. Even if your product is consumer-facing, exposing a clean programmatic interface lets other companies build distribution for you for free. OpenAI shipped ChatGPT and the API the same week, the API was not an afterthought, it was the strategy.

Second: instrument the feedback loop before you scale the product. OpenAI was logging conversation-level signals from day one of ChatGPT's research preview. By the time competitors caught up on model quality, OpenAI had a dataset they could not. If you're shipping ML in your product, the question "how do we know if the output was good?" should be answered before launch, not after.

Third: pricing tiers as user segmentation, not just monetisation. ChatGPT Free → Plus → Team → Enterprise isn't a price ladder, it's four different products with different success metrics. Free optimises for engagement. Plus optimises for retention. Team optimises for seat expansion. Enterprise optimises for security and compliance. Building four products inside one brand is hard, but it's the move that lets a single account team sell from $0 to $7-figure deals without re-positioning.