← Writing
← Back to writing
Debrief #2: Cactus Compute
Why a founder declined a major tech offer to build AI that runs entirely on your phone.
Written for Axiom Pathways.
Henry Ndubuaku turned down an offer from Nvidia to build Cactus, a framework that runs language models directly on phone hardware — no server round-trip — by writing custom low-level kernels for a problem existing mobile-ML frameworks were never designed to solve. This debrief covers how he and co-founder Roman Shemet turned open-sourcing the project into its own fundraising strategy, recruiting dozens of tech executives as investors after they discovered the GitHub repo themselves, and why Apple’s walled-off Neural Engine access remains the ceiling on how far Cactus can go on iOS specifically.