We are writing our first fund thesis in September 2021, against a backdrop of genuine excitement about what large language models can do. GPT-3 has been available to developers for just over a year. The capability is clearly real; the applications are beginning to appear. The natural investor instinct is to find the applications — the products that will be built on top of this capability — and bet on the ones that look most durable.
We are going to do something different. We are going to bet on the infrastructure layer, before the application wave arrives. I want to explain the reasoning clearly, because it is a contrarian choice and it requires a clear thesis to execute well.
The Railroad Argument
The canonical version of this argument is the railroad analogy. When railroads were being built in the mid-nineteenth century, the investment opportunities were everywhere: the railroad companies themselves, the land along the routes, the towns that would spring up at the depots, the industries that would be transformed by cheap transportation. Most of the fortunes made from the railroad era were not made by picking the right railroad stock. They were made by the suppliers — the steel companies, the timber companies, the companies making spikes and rails and tools — whose products were required by every railroad regardless of which specific routes succeeded.
The AI version of this argument: the specific applications that will succeed when AI capability gets widely deployed are genuinely uncertain. But the infrastructure required to build those applications — the tools for deploying, observing, evaluating, and improving LLM-based systems — will be needed regardless of which applications win. The infrastructure bet has better risk-adjusted characteristics than the application bet.
Why Now Is the Right Timing
The timing argument for infrastructure investment is that the infrastructure companies need to be built before the demand wave arrives, not in response to it. When the application layer explodes — when every enterprise is trying to build AI-powered products and every developer is reaching for LLM APIs — the companies with well-developed infrastructure will capture the demand. The companies that start building infrastructure in response to the demand wave will be too late; they will be competing for a market that is already being served.
We believe the application wave is two to four years away from being large enough to create the kind of enterprise infrastructure demand that supports large infrastructure companies. That timeline gives us the right window to identify and fund the infrastructure companies that will be ready when the demand arrives. We are planting in the spring so we can harvest in the fall — and we believe the harvest will be substantial.
The Selection Criteria
Not all infrastructure bets are equal. The infrastructure companies that will matter are those that solve problems that are genuinely hard, that cannot be solved by the model providers themselves without significant conflict of interest, and that are amenable to specialization. Evaluation and testing infrastructure: hard, not in the model provider's interest to provide well, highly amenable to specialization. Deployment and orchestration infrastructure: hard, not in the model provider's interest, specialization creates value. These are the categories we are targeting in Fund I. The specific companies will vary; the thesis is stable.