GPUShare — Self-hosted GPU sharing platform

Case snapshot

Problem

Make an intermittently available home GPU useful to a trusted group without exposing hardware controls or hiding usage cost.

My role

Solo engineer across product design, distributed architecture, hardware integration, billing logic, and installers.

What I changed

Separated public account middleware from the private hardware backend.
Preserved streaming, tool calls, and local-or-cloud model routing.
Replaced manual setup with GPU-aware cross-platform installers.

Result

The core platform and setup workflow work, with corrected usage accounting and documented limits; no hosted service or adoption is claimed.

Idle hardware suggested a shared service

A desktop GPU bought for gaming, rendering, and local inference would sit unused for much of the day. GPUShare explores how a trusted group could use that capacity while paying the electricity it consumes.

The host is not managed infrastructure

The machine can be offline or reclaimed for gaming, and operations such as model loading, Blender execution, and energy monitoring should never be public API surfaces.

The architecture follows the trust boundary

Public middleware handles accounts, billing, and caching. An internal-token-protected backend runs Ollama, Blender, GPU monitoring, and uploads behind an outbound Cloudflare Tunnel with no router port exposed.

Installability became part of the system

GPU-aware macOS, Linux, and Windows scripts recommend model sizes, estimate wattage, generate secrets, check dependencies, and verify service health. OpenAI-compatible streaming and tool calls allow existing clients to connect.

Correct accounting was more valuable than more features

Per-type ledger totals replaced misleading chart arithmetic, and the account view was aggregated to reduce requests. The result is substantial capability with honest uptime and model-quality caveats, not a cloud competitor.