Falling into Microsoft

Taken from https://jeffwang.substack.com/

Last year, there were already industry titans and CEO celebrities in the world of AI. If you were able to train your own models and release them to the wild, you automatically became a tier 1 AI company. By demonstrating new models you were able to show that you had a specialized AI team and that you were actually on the forefront of pushing the AI industry forward. In fact, this was similar to being a “Layer 1 chain” in crypto. If you were able to become the foundation, people could build off of it. How could you prove this though? Well it seemed initially that all you had to do was open source your model and make a big deal about benchmarks, and that was enough to gain virality amongst the AI enthusiasts.?

But similar to crypto, this strategy of “build it and revenue will come” did not always work. And quite the opposite of crypto, where it only takes a few lines of code to fork Ethereum and launch a chain, with AI it took massive amounts of capital to purchase GPUs, hire the right talent, and obtain the right data to build these foundational models. Maybe crypto has the best “quick start” model after all, just fork a chain and get a lot of attention and you’ll be worth billions! I really hope that’s not the main takeaway from this.?

Last year, the strategy of spending money to train models was fine when every startup with this talent could raise money to cover the costs. These companies were literally raising hundreds of millions dollars every few months. Mind you, this was not “earning” the revenue, but simply by making headlines you were able to get VCs knocking on your door to take their cash. Reality needed to hit at some point though. After all, spending money is one thing, but not having any return on investment is another.?

I highlight this because it should be a wake up call to companies that have not discovered their business model yet. To give a few examples of this that were quite shocking last month:

Stability AI. A lot of the AI photos I generate come from Stable Diffusion. These rely on foundational models that were open sourced and created by Stability AI. I was blown away by the ability to host models and run them locally on my machine, and follow the community on building tools and fine tuned models. It’s simply amazing. However, Stability spent most of its money training these models and never focused on how to grow its revenue. Lately, it started to turn their models into a licensed model, but it was not clear who was paying for this. Also, it started developing front end applications that it could start charging subscriptions to, but this growth takes a lot of effort and time. Stability raised $101M back in October 2022 at a $1B valuation, and later it raised another $25M in June 2023 and got an additional $50M in Financing in November of 2023. It would then have trouble raising another round, and the top executives and researchers including the CEO would leave the company.?

Inflection AI. LinkedIn co-founder Reid Hoffman and some of the top AI researchers in the space, including the co-founder of Deepmind Mustafa Suleyman, would come up with a “personal AI” chatbot that would be tailored for each individual and be “safe”. They raised a whopping $1.3B (most likely a lot of this as Nvidia chips), and purchased 22,000 H100s to train their next model. It does seem that after version 2.5 came out, it was already over at the company as people departed to Microsoft right away.?

The entire situation is a bit weird. How would it be possible for the main players of a massively funded startup to be able to walk away? Reportedly the investors will be made whole, and that Microsoft will license the hardware and technology so that the company will still continue on. Reid Hoffman being on both the boards makes the situation a bit peculiar. But anyways, let’s get back to the thoughts…

Inflection targeted too narrow of a use case, which was a personal chatbot that would hope to be a person’s agent to do all their everyday tasks eventually. By going for trust, Inflection hoped that it would be the go-to chatbot for all purposes. However as it soon discovered after training, the technology simply isn’t there yet. And by not having any other in-between plan, or even an API, it simply could not sway users away from ChatGPT, Gemini or Anthropic.?

So two tier 1 companies, Stability and Inflection, are not looking so good 1 year later. And with Anthropic proving itself against OpenAI, perhaps this will become a 2 or 3 player game if Google can get back market share in the chatbots arena.?

I wanted to title this “failing into Microsoft”, however, getting acquired by Microsoft is hardly failing. Falling into Microsoft might be a better fit after all, however, Microsoft now pretty much owns OpenAI and Inflection. Out of all the top private market players, that’s 2 out of 3 of the biggest competitors. Microsoft also has a major partnership with Mistral as well. Is this fair? It’s somehow resisted straight acquisitions of these companies, presumably to avoid any antitrust or legal risk. Last year I wrote an article about how Microsoft will come out as the winner of the entire AI race, and unfortunately things seem to be on track.?