There is a standard objection to infrastructure-first investing in any technology cycle: "everyone knows infrastructure is the right call in retrospect, but while the cycle is happening, the applications layer has more obvious traction and gets the better valuations." This argument has some empirical support — infrastructure businesses are harder to build early because the customers are other developers, not end users. Developer-to-developer revenue is slower to compound than consumer adoption.
Our counter is not that this observation is wrong. It is that the objection conflates "harder to build" with "worse investment." Infrastructure businesses in a platform cycle typically produce the most durable returns precisely because they are harder to build, because they require deep technical knowledge that is hard to replicate, and because their customers are the most technically sophisticated buyers in the ecosystem — the ones who evaluate switching costs clearly and do not churn on a sales pitch.
The platform cycle pattern
Every major technology platform cycle has the same structure: a foundational capability appears (the transistor, the microprocessor, the internet protocol, the smartphone), a period of uncertainty follows about which applications will dominate, and eventually a small number of application layer winners emerge. The infrastructure companies that serve the platform build during the uncertainty period — before it is clear who wins at the application layer — and find that their customers span every subsequent application winner.
The internet cycle is instructive. By 1998, it was not clear whether e-commerce, search, or online media would be the dominant application category. The companies that built networking infrastructure, hosting, CDN, and database systems served all of them. Akamai, Verisign, and Cisco — infrastructure companies — were the most durable value creators from that period, not the first generation of e-commerce and media companies who mostly did not survive to the current era.
Mobile follows the same pattern. In 2010, it was unclear whether social, commerce, gaming, or productivity would dominate on smartphones. Stripe, Twilio, and AWS — companies that served all application developers regardless of vertical — compounded more durably than most first-generation app companies.
Why AI infrastructure follows this pattern
We are now in the AI equivalent of 2010 for mobile. A foundational capability — LLMs with general reasoning and instruction following — has appeared and is clearly real. The application layer is producing thousands of companies, most of which will not survive the next three years as the market consolidates and the easy use cases get commoditized by incumbents. The infrastructure layer — serving, fine-tuning, optimization, routing, observability — will serve every application that survives, regardless of which vertical wins.
This is not saying all infrastructure companies are good investments. It is saying that the infrastructure layer is the right place to be if you have the technical depth to evaluate which infrastructure bets are real versus which ones are proximate to model providers who will eventually bundle the capability. That is a demanding condition — which is why we staffed Firntal with engineers who have built these systems, not with analysts who read about them.
The application layer does not know which infrastructure it needs yet
There is a second, more specific reason to invest in infrastructure before applications: application layer companies cannot describe their infrastructure requirements clearly until they have been in production for six to eighteen months. The first version of a product is almost never the version that becomes a scaling problem. The infrastructure companies that survive and dominate are the ones that are already building the right system before the application companies know they need it.
In the current cycle, we observe this clearly in inference. When a growing API product scaled from a few hundred users to hundreds of thousands of concurrent users, the serving architecture that worked for their first deployment — a straightforward hosted endpoint on a major cloud — broke down immediately. The techniques for solving that scaling problem — dynamic batching, memory-efficient attention, fractional GPU scheduling — were not things the application company could build internally on a relevant timescale. They needed infrastructure companies that had already solved those problems in production.
The infrastructure companies that had been building for 18 months before this demand appeared were positioned to solve it. The ones that started building after the demand signal became obvious were, by definition, too late to serve the earliest customers.
The objection about infrastructure commoditization
The serious objection to infrastructure investing in AI is: the major cloud providers will eventually build and bundle these capabilities, commoditizing the standalone infrastructure companies. This is a real risk, not a dismissible one. AWS has historically bundled database, message queue, search, and CDN capabilities that once supported standalone companies.
Our position on this is nuanced. Some infrastructure will get bundled — the parts where the performance gap between the cloud-native managed product and the standalone specialist is small and the switching cost from the managed product is low. Managed model endpoints from AWS and Google will satisfy the majority of use cases that do not require frontier optimization.
The infrastructure that will not be bundled effectively is the infrastructure where the performance gap is large and technically maintained — where a dedicated team continuously advances the optimization state of the art, and where the managed cloud product cannot keep pace. Continuous batching, hardware-aware kernel optimization, and multi-adapter serving for fine-tuned models fall into this category today. They require depth of focus that a cloud provider's general-purpose managed service cannot realistically maintain at the frontier.
We are not saying cloud commoditization risk is zero. We are saying that in the current 3–5 year window, the performance gap is large enough that the infrastructure specialists have time to build durable customer relationships and product depth that survives the managed-cloud push. The investment window is real, and it is open now.