ChatGPT: A Case Study in PMF

TL;DR: ChatGPT is the cleanest specimen of Andreessen’s “the market pulls the product out of the startup” we are ever likely to get. OpenAI shipped what they called a “low-key research preview” on November 30, 2022. By February 2023 it had 100 million monthly active users — the fastest consumer adoption in recorded history (impromptu-hoffman). The product was rough. The market did not care.

The accident

Almost every retrospective gets the story wrong on one detail: ChatGPT was never supposed to be a product launch. Internally it was a research preview — the OpenAI team expected a blog post, some Twitter chatter, maybe a few thousand curious devs poking at it. They spent the launch weekend rationing GPU capacity and slapping up “ChatGPT is at capacity right now” pages because the actual reception broke their planning by an order of magnitude in every direction.

This is what Andreessen meant when he wrote that in a great market, “customers are knocking down your door to get the product; the main goal is to actually answer the phone” (only-thing-that-matters). OpenAI’s launch was, almost literally, OpenAI scrambling to answer the phone for sixty straight days. The company runs on what one engineer called “twitter vibes” — when something goes viral, leadership reads it the same day and considers it (see below). That’s the cultural fingerprint of an organization being hauled along by its market rather than steering it.

The two-and-a-half-year gap

Here’s the part the standard narrative skips, and it’s the most important part. GPT-3 had been publicly available via API since June 2020 — almost two and a half years before ChatGPT. The model under ChatGPT (GPT-3.5) was a fine-tuned variant of a model OpenAI had been shipping for a year. So why did GPT-3 attract tens of thousands of developers and ChatGPT pull in a hundred million normies in two months?

Because GPT-3 had a packaging problem, not a capability problem. As Max Woolf observed in 2020, GPT-3 was a “sentence completion engine” — you had to learn prompt engineering, hide your prompts to make demos look magical, and wrestle with an API (tempering-expectations-gpt3). The 175-billion parameter model was a Ferrari engine bolted to a wheelbarrow.

Then OpenAI did three things, in this order:

InstructGPT (early 2022) — used RLHF to make the model follow instructions instead of just completing text. Suddenly you could ask it things instead of priming it with three labeled examples (brex-prompt-engineering).
The chat UI — collapsed the entire prompt-engineering ritual into a text box that worked like every other chat product on Earth.
Free, no signup gate beyond an email — no API key, no credit card, no rate plan.

The model was the same. The packaging was a different product. This is the atomic-concepts insight in a single bullet: GPT-3’s atomic unit was the prompt (engineering literacy required). ChatGPT’s atomic unit was the conversation (English literacy required). One required a CS degree. The other required a pulse. That’s the difference between a 30K-developer market and what a Pew poll later put at hundreds of millions of US adults.

The Sean Ellis test, retroactively

If you’d run Vohra’s PMF survey on ChatGPT in early 2023, the “very disappointed” number is somewhere unreasonable — probably 80%+. But the more interesting question is who the very disappointed segment was. According to the most rigorous study to date — based on internal OpenAI consumer-plan data from May 2024 to June 2025 — the largest single category of consumer ChatGPT messages is non-work, practical guidance and writing help, not coding, not enterprise productivity, not “agents” (how-people-use-chatgpt).

Translated: ChatGPT’s killer use case was never the one OpenAI was building for. It was thinking out loud with a smart friend. OpenAI’s revealed PMF was something close to “personal cognitive offload,” which is not a product category that existed before, and which is roughly the size of literacy itself.

This inverts the chasm. Most products achieve PMF with technical early adopters and then bleed out trying to cross to the mainstream. ChatGPT achieved mainstream PMF first and then had to reverse-engineer how to serve developers (the API) and enterprises (ChatGPT Enterprise, custom GPTs). The chasm was crossed in the wrong direction. For most teams that’s strategic suicide. For OpenAI it was a windfall — every market segment showed up at once and they got to pick which ones to build for next.

OpenAI’s strategic answer

Sam Altman has been startlingly explicit about the implication: ChatGPT is the strategy. In a leaked May 2023 meeting with developers, Altman said OpenAI “would not release more products beyond ChatGPT” — that great platform companies need a killer app, and being the best customer of your own API was how you made the API better. ChatGPT was that killer app (openai-plans-altman).

In the same meeting he was unusually blunt about plugins: “the usage of plugins, other than browsing, suggests that they don’t have PMF yet.” Then he said the line that should be tattooed on every founder considering an “X for ChatGPT” startup: “A lot of people thought they wanted their apps to be inside ChatGPT but what they really wanted was ChatGPT in their apps.” Two months out from launch, looking at the usage data, Altman called it.

This is the aggregation-theory playbook. Ben Thompson nailed the strategic frame: “the route to becoming a platform is to first be a massively popular product. Acquiring developers and users is not a chicken-and-egg problem; you must get the users first” (openai-windows-play). OpenAI gets to be the platform because they own the demand, not because they own the technology.

By early 2025, ChatGPT had grown from ~100M to more than 300 million weekly active users, per Altman’s own end-of-year reflections — and that number has only kept climbing since (altman-reflections). Revenue followed the same shape: from a $28M annual run rate in 2022 to a >$1B run rate by mid-2023 — a 35x year-over-year jump that was the financial fingerprint of explosive PMF (openai-1b-revenue). That’s Power in the canonical Helmer sense — surplus margins protected by something competitors can’t easily replicate. In OpenAI’s case, the moat is brand + scale economies + a cornered-resource in researcher talent.

What this case actually teaches the framework

1. Andreessen was right about market. ChatGPT was not a great team building a great product into a mediocre market. It was a competent team accidentally serving a market so vast and so latent that the product almost didn’t matter. Andreessen’s claim — “in a great market, the team is remarkably easy to upgrade on the fly” — got tested in November 2023 when OpenAI lost half its leadership for 96 hours and the user count kept climbing. The market did not care about the org chart.

2. Vohra’s engine still applies, even at warp speed. OpenAI’s product roadmap since launch reads exactly like Vohra’s split: half the work is deepening what lovers love (faster models, better reasoning, voice mode, memory), half is removing barriers for fence-sitters (mobile apps, sharing, custom GPTs, search, image generation) (superhuman-pmf-engine). The pace is faster. The pattern is identical.

3. PMF is not a moat. It is the opportunity to build a moat. Two years after launch, every frontier-lab benchmark winner has been someone other than OpenAI at one point or another (Anthropic, Google, Meta’s open Llama models, DeepSeek). Capability leadership has proven shockingly fragile. ChatGPT’s actual moat was never the model — it was the brand, the verb, and the user habit. Distribution beat capability. This is the single most important lesson the AI industry has been refusing to learn for three years (google-no-moat, distribution).

4. Relationship effects > network effects for consumer AI. Victor Stepanov’s argument, written in 2024 from the position of a growth-marketing veteran at Every: for consumer AI, the moat isn’t classic network effects — it’s the memory, personalization, and trust that accumulate between user and model (never-go-viral-ai). This is a category of switching-costs that didn’t exist before LLMs: per-user accumulated context. Every conversation deposits memory and preferences that competitors literally cannot replicate without time. ChatGPT’s memory feature isn’t quality-of-life. It’s a deliberate moat.

The Altman quote that explains everything

In one of the most underrated tweets in modern startup history, Altman wrote:

i am still amazed how most startup investors are great at understanding that startups can grow exponentially but don’t understand that markets can too. “the TAM is too small” has cost startup investors more money than any other often-repeated phrase i know of.

The whole ChatGPT story sits inside that observation. Nobody — not OpenAI — had a market sizing for “AI assistants for everyone with a keyboard” in 2022, because the category did not exist. The market materialized when the product did, and it kept materializing for years afterward. This is the invisible-asymptotes story in reverse: instead of hitting a ceiling no one saw, ChatGPT broke through a floor no one knew was there.

The unanswered question

The interesting open question is whether ChatGPT-the-product can keep outrunning the invisible asymptote that catches every consumer product eventually. Andrew Chen’s thesis, written before ChatGPT existed, was that the era of billion-user consumer apps was over — easy growth channels saturated, ad costs ballooning, viral mechanics nerfed (end-of-billion-user-startup). ChatGPT looks like the violent counterexample, but it grew on a one-time demand release: the world realizing AI worked. The next 300 million users will be harder than the first 300 million.

The smart bet is that ChatGPT’s path looks more like Google’s than Instagram’s: a once-in-a-generation product that becomes infrastructure, then defends that position against incumbents (Microsoft, Google, Meta) and against the slow erosion of capability moats. The case for OpenAI is the same case Helmer makes for scale-economies and branding as Power types — both apply, both compound, and the verb test (“did you ChatGPT it?”) is as durable a brand asset as has ever existed in software.