The Case for the Second Wave

Every story being told about AI right now is a pioneer story, and a defensible one. It’s incomplete, though, and the history of the last major platform shift suggests it’s worth holding loosely. The pattern from the last cycle is worth restating, not because it repeats exactly, but because it rhymes. Google was not first in search. Facebook was fourth in social networking. Chewy was built on the wreckage of Webvan and Pets.com. In each case, a first generation of companies taught users new behavior, discovered which business models did not work, and forced the necessary infrastructure into existence, often at their own expense. The generation that followed inherited a solved category and competed on execution.

AI’s first wave is real, and it has accomplished something important. The foundation model labs raised enormous capital, trained a generation of users to open a chat window and type a prompt, and built a genuine business around it. Alongside them, a cohort of 2023 and 2024 startups did the harder, more expensive work of teaching enterprises to trust a model with real workflows. Many of them will not survive as standalone companies, which isn’t necessarily a verdict on the market they were building toward so much as a repeat of what happened with the first generation of search engines and social networks: expensive, necessary, and often unrewarded for the companies that did it first.

Skeptics have good reason to push back here. OpenAI and Anthropic are not AltaVista or Webvan. They control the underlying model, they have raised capital at a scale no browser-era startup ever touched, and they are moving up the stack into agents and vertical products rather than sitting still as a portal. A platform with a technical and capital moat this large may simply keep compounding rather than cede ground the way the first generation of internet companies did, and anyone making the second-wave case needs to take that possibility seriously rather than assume history repeats on schedule. These two companies may very much be the exception rather than the rule.

What has changed since 2023, though, is the cost structure underneath the whole category. Inference prices have fallen more than 280-fold in under two years (from $20 to $0.07 per million tokens for GPT-3.5-level performance, per Stanford HAI's 2025 AI Index Report), which means the unit economics that killed the first wave of thin applications no longer apply in the same way. Procurement teams that would not let AI near a customer record in 2023 now have a standing line item for it. The open question inside most enterprises is no longer whether to use AI, but which vendor to trust with a specific, narrow, high-stakes workflow, a different sales motion than the one the platforms were built for.

The more interesting opportunity, in our view, is forming in a different place: healthcare documentation, insurance claims, legal workflows, and other categories where the depth of the workflow matters as much as the model underneath it. These companies are not trying to out-build the labs on model capability. They are showing up after the behavior has been taught and the cost curve has turned in their favor, betting that domain depth and trust, not the model itself, is what a customer is actually buying. Whether that is enough to win against platforms with the labs’ capital and distribution is the real question, and it is one we would rather test with results over the next few years than answer with certainty today.

‍

Photo by Daniel Boberg / Unsplash

‍