$1.5 billion does not get you to product market fit, founders lessons from inflection.ai
It was the fastest ticket to unicorn status and back — 12 months.
Inflection.ai launched in March last year, five months after ChatGPT launched in Nov 2022. In June the same year, it followed up with investment announcements totaling up to $1.5 billion. On paper, it had everything going for it: two of the best research scientists in AI at the helm as co-founders, and 30 more members as technical staff, including engineers and researchers, among them the authors of the Chinchilla paper, which was the biggest breakthrough in AI research that year. They shipped well as an agile team, running tight 6-week sprints. They built a user interface and model that users seemed to rave about. In fact, their fourth voice template was so good it reminded users of the voice of the woman from the movie Her. The world’s top investors backed it, including Reid Hoffman, Bill Gates, Jensen Huang, and Microsoft.
Yet, in 12 months, they went from launch to wrapped up. In March 2023 product launch happened; March 2024 acquihire announcement came out.
Is this a warning bell for all others? Is this the beginning of the end of the AI hype cycle? It’s easy to hide behind macro babble like this. Reality is more nuanced.
If you don’t differentiate then you die
Was Inflection.ai climbing the right mountain? Yes. Did it start with a good idea? Definitely yes. The vision of a co-pilot for every profession, as laid out by Reid Hoffman, is still very compelling and has not yet been conquered by anyone.
Dozens of startups are accumulating ARR by building AI companions. Perplexity is making good progress in carrying the torch toward that vision of a personal assistant. They decided not to invest in the model infrastructure like Inflection.ai, which may be the reason they are still around. But merely being on the right mountain is not enough, especially when the competition intensifies. If you don’t stand out, then you don’t stand a chance.
Currently, the difference offered by Inflection, compared to ChatGPT, Claude Opus, and Perplexity, does not warrant a user switching from them to the personal assistant from Inflection.ai. Too much focus was placed on what model, what infrastructure, what language model, etc., without sufficiently addressing why a user would choose Inflection.ai over ChatGPT and the myriad other assistants entering the market.
It turns out that being the most human-like, offering high personalization, or having a stronger Inflection 2.5 model is not enough to convince users to switch to Inflection’s personal assistant, Pi.
ChatGPT went from its launch in November 2022 to 100 million users by November 2023. In the 12-month period for Inflection.ai, i.e., March 2023 to March 2024, the traction amounted to only 6 million users.
Had it pursued a vertical use case, such as mental therapy, it might differentiated enough and stood to have gained more traction.
Just because you could, does not mean you should. Raise boatloads of funds
When asked in an interview, Mustafa said that he raised large sums of money to build a vast amount of compute that is needed to power the personal assistant.
At peak, OpenAI spends $700K every day to run its operations. At the same scale, in a year, Inflection would have spent $255 million. They had a team size of 30–70, assuming they spent ~$50 million to operate the team; that would have still added up to close to $300 million every year. With $1.5 billion of funds in the bank, it would have meant a runway of 5 years. They were operating at 1/15 of the OpenAI scale; the runway would have been between 10–20 years.
The funding amount is influenced by more than the operational budgeting exercise that a startup does. It is driven by how much ownership investors would like to have and maintain, how much a competitor is raising, and how you can raise to choke competitors of resources. When you raise too much, there is also an expectation around the growth rate and implicit expectations around returns that you end up setting. If you raise a million dollars in seed and have a decent-sized investor, then you are setting the expectation of returning at least a hundred million dollars. Now, if you raise a billion at seed, then the expectation is around a hundred times that. That return expectation sets the pace around the growth expectations; if you are not growing at the same pace as the best in class, then you disappoint.
That big pool of money then becomes the noose around your neck. Experienced founders, therefore, stagger their raise based on the next milestones that must be reached
Super fast growth and stickiest retention, not any other antics.
AI was not very interesting a few years ago; in 2016, most folks thought OpenAI was delusional. All of that changed between November 2022 and January 2023. In 2 months, ChatGPT surpassed 100 million users. If people don’t adopt a product en masse, then it is a toy. AI or no AI, distribution is everything. Something about ChatGPT attracted people to try it out and constantly come back to it.
In June 2023, there was another product that unseated ChatGPT, thanks to the distribution strength that Meta had. ‘Threads’ product got to 100 million users in 5 days. But practically today, it is forgotten. Social products or AI product retention is even more of everything. Watch your retention, not just growth.
Before OpenAI got successful with ChatGPT and GPT-3.5, it had different versions of GPT that were quietly used by developers. In the very beginning, OpenAI started with building robots that learn; they did not take off in the product-market fit sense of the way. They tried Dall-E, which had high interest. It is finally GPT-3.5 that took off. Finding product-market fit needs multiple iterations to get there. Most startups go through at least 3 pivots. Hard pivots. Inflection.ai did not give itself the time and opportunity of doing hard pivots, hard to do when you raise a large amount of funding.
Mustafa wrote a New York Times bestseller book, which was provocative enough to get a lot of attention. But thought leadership and past credentials do not guarantee getting to product-market fit
It is the simplest things are worth remembering again and again
Building and selling a startup does not guarantee that you will achieve product-market fit next time. Don’t raise too little, and don’t raise too much. Stagger it. Raise enough for the next goal. Give yourself permission for at least 3 iterations; it takes that long. Marketing to spread the word is important, but the product is equally, if not more, important. Keep the focus on user growth and retention only. Ignore everything else.
