Did Microsoft just cannibalize a company they were a lead investor in? Seems that way.

Last year, Inflection raised $1.525B to create what they called the “first emotionally intelligent AI” – Pi. Unlike GPT, Pi is supposed to be a kind and supportive companion that can offer advice (in fewer words: a therapist). And they did it. Earlier this month, Pi exceeded 1M active daily users.

And then last week, both co-founders – Mustafa Suleyman and Karén Simonyan – were hired to lead Microsoft AI, taking most of the Inflection team with them.

In the wake of this massive upheaval, Inflection is shifting its focus away from Pi and moving to an AI studio approach (whatever that is), hosted of course on Microsoft Azure.

If you’re wondering, ‘What the heck just happened to one of the hottest AI startups showing tremendous potential with tons of capital?’ you’re in good company. We have a few theories as to why things changed so quickly for Inflection and Pi, and what that means for other AI application layer companies.

But first, what is the application layer?

When you think of AI, you probably think of GPT-4 or Claude. These are some of the top large language models (LLMs) available currently, and they’ve been trained on vast amounts of data and work pretty well. So creating your own LLM is a waste of time and money –  instead, you can build on top of it.

Building on top of these LLMs is the “application layer", and we’re already seeing thousands of startups pop up to serve niche purposes – like Pi, which was built in part on top of GPT-3.5.

So what happened to Inflection?

We have a few theories.

Theory 1: Pi’s business model was fundamentally flawed

Pi worked on a freemium model like ChatGPT – meaning it was free to use but you had to upgrade to a paid model to unlock premium features. One hang up there is that they hadn’t even announced a paid version yet.

If you’re going off of Lenny Rachitsky – founder of the #1 business newsletter on Substack – 5% is a good conversion rate of free to paid users. So if Pi was making $20 a month (ChatGPT-4’s going rate) from 50,000 paid users (5% of their 1M daily active users), they’d have a $12M a year business.

That’s not enough to cover the costs of the 95% of people who will never pay up. Which brings us to theory number two.

Theory 2: The lack of enterprise opportunity stunted its possibilities

Pi is differentiated by its intuitive, emotional responses. That’s nice – but it may not be enough for consumers to pay for it on an ongoing basis. And without a sustainable consumer model, Pi would need an enterprise business to sustain investor interest.

With most technology, the consumer model pays for some of it, but the enterprise side is where they make the bulk of their money. And there’s not much need or demand for a therapy bot in an enterprise setting. That cuts Pi off from enterprise revenue, which most startups need to grow.

Theory 3: Microsoft was a life raft for soft entrepreneurs

Mustafa Suleyman is a leader for humane and responsible AI, with a super impressive track record, but his background is in mental health and social support. It could just be that he and his team don’t have the desire to actually run a startup.

They may have looked at the circumstances – and asked themselves if they wanted to spend their time raising money, getting enterprise customers, and everything else that goes into building a sustainable company. They may have wondered ‘where's the rich uncle that can take this off our hands?’ and they found the richest uncle of all.

And by the way, they did this once before – selling DeepMind to Google in 2014 for £400 million – delivering a great return for investors. Mustafa is already rich, so the brain damage and emotional toll caused by running an AI startup may have looked like too high a price.

Theory 4: Inflection’s Pi was crushed by Character.ai

As it turns out, Pi was not the first to focus on an emotionally intelligent AI. Character.ai launched nine months before and they’re already netting an annual revenue of $16.7M - nearly $5M more than we guestimated Pi could’ve pulled. They’ve been more capital efficient, and seem to have a user experience that has struck a chord with people looking for a more personal connection with AI.

Similar to other “winner-take-all” categories (social media, crypto etc), for AI companionship – at least at this moment, when the market is still forming – there is only room for one consumer winner. That is Character.ai.

And so the management team and investors agreed – no amount of capital or time would close that gap. So they raised the white flag.

The solution: Peace with honor

The AI winners will need big brains and boat loads of capital. Microsoft has the capital and needs more big brains.

Their CEO, Satya Nadella, has described their AI efforts as “frantic” following the success of OpenAI. For a moment last November, it seemed like they might’ve been able to bring the brains behind that success in-house when the OpenAI board fired their CEO – inciting a mutiny where hundreds of employees pledged to join Microsoft if he wasn’t reinstated. When Sam Altman was inevitably rehired at OpenAI, that chance slipped away. Taking in Inflection’s team is one of the next best moves Satya and co. could make in order to bolster their AI talent bench.

What’s happening now is the rebirth of Microsoft. They’re taking big bets by investing in a tremendous amount of leading AI scientists. And maybe Mustafa has decided to bet his career on the future of Microsoft AI, too. In the end, everybody wins: Mustafa doesn't have to run a startup, and Microsoft just acquired some of the very few valuable brains in AI.

What this all means for the application layer

The application layer is still deeply valuable, and the quick demise of Pi doesn’t invalidate that. The fact that Pi acquired 1M daily active users in less than a year means something. It took Netflix nearly 3.5 years to hit a million subscribers, Twitter 2 years, and Pinterest 20 months. Here are the lessons to take away:

1. You better have a value prop that enterprise will pay for

Pi’s user growth proved that lots of people want a therapist- or coach-like relationship with AI. But without an enterprise application, the consumers’ interest in an AI companion isn’t enough on its own to sustain a business.

2. Make sure you’re solving a new problem

Pi had a good use case: A lot of us can't afford therapists and coaches, so rich people tend to be the only ones who get access to them. But the problem was already solved by Character.ai nearly a year before. Don’t build an AI tool for the sake of getting in the game. Make sure there’s a clear gap in the market and that AI can fill it.

3. AI entrepreneurship is going to be a LOT harder than you think

Once the hype is over, AI entrepreneurship is just not for the faint of heart. It’s really hard to build independent companies that can survive. The guts of AI will likely come from two or three players – like OpenAI and Anthropic – and we'll have to build these experiences on top of their models. But the cold shower in AI-land is the go-to-market: What’s the plan to acquire customers (consumer or enterprise) in a capital-efficient way? Hundreds of YC-graduated AI companies are going to learn this the hard way.