ChatGPT launched in November 2022 and reached 100 million users in two months. I want to record our thinking from early 2023 about what it means — for the market, for our thesis, and for what comes next. These views will age and some of them will be wrong; that is the point of writing them down.
The first and most important thing ChatGPT demonstrated is that general-purpose language model capability has crossed a threshold where it is genuinely useful for a very wide range of tasks, to a very wide range of people, without any specialized knowledge of how the underlying system works. This is qualitatively different from the previous state of affairs, where using LLM capabilities required significant technical sophistication. The consumer-grade interface hiding the engineering underneath is not a trivial achievement.
What This Means for Infrastructure
Our thesis was always about the infrastructure layer, not the model itself. ChatGPT and the GPT-4 release in March 2023 validate the thesis dramatically: the model capability is now clearly sufficient for a wide range of applications, which means the limiting factor in deploying AI at scale is no longer model capability but deployment infrastructure. Latency management, cost optimization, reliability, evaluation, customization — these are engineering problems, not research problems. The engineering infrastructure to solve them is largely underdeveloped.
The practical implication for us: the demand signal for every company in our portfolio just increased by an order of magnitude. The teams at Langbase, Superagent, and Mem.ai are all seeing dramatically more inbound interest than they were three months ago. This is welcome. It also puts more pressure on them to build production infrastructure faster than the market is moving, and it attracts more well-funded competitors.
What Comes After the Demo Wave
The natural progression of any technology wave: a demo phase, a product phase, and an infrastructure phase. We are somewhere between the demo and product phases for LLM applications. The demo phase is dominated by people discovering what the technology can do and building impressive but often superficial demonstrations. The product phase begins when teams start discovering what makes these applications reliable in practice — and discovering how much infrastructure they need to get there. The infrastructure phase begins when that demand is large enough to support dedicated infrastructure companies.
We believe we are entering the product phase now, and that the infrastructure phase will follow within 12 to 18 months. That puts the investment timing for AI infrastructure at approximately right for seed-stage checks today — building products for demand that will materialize clearly within the fund's investment horizon.