AI Markets Have Crystalized
AI markets have evolved significantly over the last 4 years. When GPT-3 came out and scaling laws were openly discussed in the AI literature, it seemed clear that you could extrapolate the rate of progress from GPT-2 to GPT-3 onwards through GPT-4, 5 etc and realize a revolution was going to happen.
4 years ago, I started looking for Generative AI companies to back or help start given this curve. I ended up leading or participating in early rounds in companies like Harvey, Perplexity, Character.AI, BrainTrust, and others. At the time it was clear to just “back all the best people working on all the biggest problems” because very few people were actually starting generative AI companies. OpenAI seemed to be the only clear foundation model company (with Anthropic pre-launch and promising, Llama non-existent, and Google clearly someone who could (and would eventually) aggressively innovate but was stymied at the time by its own internal processes).
As more people outside of the core AI community woke up to this opportunity, or researchers and engineers from the main labs left to start new companies in AI, the world of AI became murkier. I used to say that the more I learn about AI, the less I knew about AI – because it was unclear in many early markets who the likely winners would be and the underlying models and tech were changing so fast. For example in 2022, it was clear code / AI driven software engineering was going to be important, but it was unclear who the winners would be (for example Cursor was not launched until 2023, Codium launched Windsurf in GA about 9 months ago, and Cognition launched Devin in limited release a bit over a year ago).
We have now entered an era where the first set of AI markets have solidified and a likely set of winners have emerged. This does not mean others won’t show up over time to compete in these markets or that current leaders won’t get acquired or eventually die (just as Stripe launched over a decade after PayPal and 4 or so years after Braintree, and Facebook launched a few years after Friendster and Myspace). We will also see a new set of markets crystalize in the coming few years and these markets today seem quite uncertain.
Markets with more clarity
1. Foundation Models- LLMs
There are many types of foundation models including large language models (LLMs), as well as models for voice, images, video, music, chemistry, biology, materials, physics and other areas. Foundation models are often driven by scale (of data, compute, certain types of post training and feedback, etc). Scale means capital, so to win in the LLM market you need high availability of capital now entering the many billions.
In the LLM market, a core set of companies have clearly emerged as the ongoing players of the future. They are often partnered with hyperscalers (Amazon with Anthropic, Google GCP with Gemini, Microsoft Azure with OpenAI and its own efforts) as these companies have an economic incentive (cloud spend on AI via AI adoption) to fund these companies that is independent of whether these companies are good investments or not (they often are). Revenue ramps for foundation model companies are rumored to be in the $0 to many $billions range in just 3 or so years, while cloud spend on “AI” has been reported to have reached a few billion per quarter for some of the main clouds.
The core players in the LLM world are now Anthropic, Google, Meta (via Llama), Microsoft, Mistral, OpenAI, X.AI. Three or four of these companies are the clear winners on various benchmarks and most broadly adopted by developers and enterprise, and are also driving most of the spend in the industry. There are newer entrants like SSI and Thinking Machine Labs driven by brilliant AI researchers, which may either come up with innovative approaches or raise ongoing money to compete, or end up in a few years as acquisitions for companies wishing to enter the market or double down on talent.
... continue reading