We wrote the first version of this thesis in the summer of 2020, shortly after closing Fund I. The core claim was that enterprise AI adoption was running ahead of the infrastructure layer, and that the companies building that infrastructure at the primitive level would be the most durable businesses created in the current technology cycle. Four years in, we think that claim is correct — and more validated by portfolio company performance than we had any right to expect at the time we wrote it.
What has changed is the specific areas where we see the strongest investment opportunity. The orchestration and observability categories have grown substantially — there are now multiple credible companies in each, and the market has broadly validated the thesis. The governance and compliance category has moved from "early adopter" to "mainstream requirement" faster than our models projected. And two new areas have emerged that weren't legible as investable categories in 2020: prompt management infrastructure and human-in-the-loop workflow tooling.
What we got right
Infrastructure stickiness compounds faster than we modeled. The data gravity thesis — that the cost of migrating off an infrastructure product grows proportionally to how much workflow logic or data you've run through it — has proven out more dramatically than our investment memos projected. Portfolio companies in the orchestration and CDC categories are seeing net revenue retention above 140%, driven primarily by increased usage depth rather than seat expansion. This validates the thesis but also means we should be even more aggressive about backing primitive-layer bets earlier.
Developer-first distribution works at enterprise scale if the enterprise requirements are there from day one. The portfolio companies that designed enterprise compliance requirements in from the beginning — SSO, RBAC, audit logs, data residency controls — navigate enterprise procurement reviews faster and with lower churn than those that retrofitted them at Series A. This has hardened into an explicit evaluation criterion: we treat enterprise-readiness as a proxy for long-term architectural thinking, not just a sales requirement.
Where we're looking in the second half of 2024
The categories we're spending most of our research time on are prompt and context management infrastructure, human-in-the-loop workflow primitives, and enterprise-grade document intelligence. In each case, the pattern is the same: a category that existed in a crude form at enterprise scale two years ago, that AI has made both more important and more tractable to solve correctly, and that doesn't yet have a dominant infrastructure player that has won the enterprise segment. That's the window.