Distributed GPU Compute Sharing via libp2p for Low-Power Browser AI Inference — Browser Engine, Privacy, Web, Sovereign AI, and Post-Cloud Architecture (Kathon)

Abstract

Local AI inference for browser tasks—including vision-language processing, speech recognition, and neural translation—requires significant computational resources that may exceed the capabilities of low-power devices such as smartphones, tablets, and older laptops. This paper presents the design of a distributed GPU compute sharing system for the Kathon cryptographic browser that enables peer-to-peer AI inference acceleration across trusted devices using libp2p networking. The system partitions neural network inference workloads across participating peers using tensor parallelism, with encrypted communication channels, verifiable computation proofs, and incentive mechanisms based on the .aioss cryptographic ledger. We address key technical challenges: heterogeneous device discovery with capability advertisement, dynamic workload partitioning for variable peer availability, encrypted inference that prevents input reconstruction, and fault tolerance through redundant computation. Simulated benchmarks across a 16-peer testbed demonstrate 3.8x speedup for Whisper transcription and 4.2x speedup for Qwen 2.5 VL inference on low-power client devices. A security analysis confirms that encrypted inference provides semantic security against honest-but-curious peers. The system enables Kathon to deliver AI features on devices that lack the local compute capacity for real-time inference. Part of The Anticloud research corpus by Lois-Kleinner Alpasan (ORCID: 0009-0009-2233-6107). This work explores browser engine, privacy in the context of sovereign AI infrastructure, post-cloud computing architectures, and transparent, blackbox-free systems.