Rob Cheung

Rob Cheung is the co-founder and chief executive of Zo Computer, which provides an always-on cloud computer equipped with an AI agent. He is building personal cloud computing around persistent environments where individuals can store files, develop software, host services, and automate work without administering their own infrastructure.

Cheung met future co-founder Ben Guo on the early Venmo team in 2013. He helped build the personal-assistant service Fin in 2015, then became Substack’s founding engineer in 2017. Machine-learning challenges at Substack sharpened his interest in model-inference infrastructure, and he reunited with Guo in 2023 to found Substrate, a platform for composing applications from multiple AI models. They began developing Zo in April 2025, released a beta that July, and launched publicly in November.

• Multimodal computation graphs. Substrate represents text generation, speech transcription, image creation, embeddings, code execution, and structured outputs as connected operations. Making those relationships explicit helps developers inspect failures, evaluate intermediate decisions, and extend systems without treating a foundation model as the entire application. Its execution engine optimizes batching, caching, concurrency, networking, and workload placement across the graph; dependable structured JSON decoding helps components exchange usable results.

• A personal computer that stays online. Zo extends Cheung’s infrastructure work into a persistent environment for individualized software, files, hosted projects, and automations. Reliability is central: a production infrastructure comparison reported that consolidating model integrations and routing reduced Zo’s retry rate from 7.5% to 0.34%.

• Principal-agent alignment. Cheung argues that agent builders overemphasize harnesses, context, and execution loops while neglecting the person delegating the work. His principal-side critique calls for stronger mechanisms to specify expectations, monitor behavior, and address information asymmetry. He has also explored semantic combinators: reusable semantic functions for assembling agents, skills, loops, and graphs.