Talk notes
AI infrastructure is the sector where engineering acceleration shows up loudest in the data, because the buyer of AI-infra products is also a developer, which means the codebase is also the marketing surface. Q1 2026 was particularly clean. We flagged 14 organizations in the AI-infra sector with the velocity-plus-contributor composite, and 11 of them announced a Series A within the 47-day window.
The dominant pattern is the inference-runtime team. A founding team of 4–7 senior engineers, all with prior big-tech or unicorn-startup pedigree, ships a vendor-fork base (typically vLLM, llama.cpp, or a custom CUDA kernel layer), then accelerates commit velocity sharply for 30–45 days while contributor influx adds 3–5 net-new senior engineers. By day 47, the announcement lands. The pattern is so consistent that we've started using it as a sourcing filter on its own.
The single false alarm of the quarter was an open-core abandonment. An organization shipped a flashy commit-velocity spike, then quietly stopped accepting community PRs, then went private — what looked like a Series A signal was actually a product unwind. The tell, in retrospect, was a sharp drop in issue-engagement coupled with the velocity spike. Real Series A teams stay engaged with the community right up to the announcement, and often after.
The actionable takeaway for sourcing is simple. Filter for AI-infra organizations with a 14-day velocity z-score above 2.5, contributor influx above 3×, and issue-engagement steady or increasing. That filter, applied to public GitHub data, surfaced the 11 winners with 78 percent precision. The remaining 22 percent — the false positives — are still high-quality leads even if they're not raising a Series A this quarter.