# Akash Homenode

> Consumer gaming PCs into Kubernetes GPU capacity, led from first commit to GA in 10 months.

2025-10 → 2026-08 · lead · first commit → GA in 10 months · lead · linux go python k8s electron

Link: https://homenode.akash.network

Homenode turns consumer gaming PCs with RTX 3090, 4090, and 5090 cards into Kubernetes worker nodes on Akash, the decentralized GPU cloud, through a guided installer and automated cluster provisioning. Operators need no Kubernetes expertise. I led the project at Akash Network from the first commit in October 2025 through alpha in April 2026 and GA in August 2026, across scope, architecture, delivery sequencing, and the alpha, beta, and public rollout. It is live at [homenode.akash.network](https://homenode.akash.network).

## Installers for hardware we do not control

The Linux path is a hardened Ubuntu autoinstall ISO, rebuilt with NVIDIA firmware in the initramfs and a forked Subiquity. The Windows path is an Electron installer that boots through an auxiliary partition, sidestepping the 100 MB EFI system partition. Both end in GPU initialization and an automated cluster join.

## Device trust

An installation key becomes an ED25519 device identity, and every request from the device is DPoP-signed. Mesh and cluster-join credentials are short-lived and shredded after use. Disks are fully encrypted with LUKS, bound to TPM2 with PCR 7 and 8 and a recovery-key escrow, and images are Secure Boot signed.

## The alpha as the feedback loop

The alpha cohorts of RTX 4090 and 5090 owners shaped the GA product, including the decision to support non-TPM2 hardware.

## Two production root causes

A fleet-wide, silent downgrade from TPM2 unlock to a keyfile showed up in production. Root-causing it led to a redesign of the device-unlock architecture that preserves the encryption guarantees while supporting non-TPM2 hardware. Separately, reboot-only failures traced to a systemd ordering cycle around tailscaled, and removing it took a 90-second stall out of every boot.

## The operator experience

A Go and Next.js kiosk appliance with signed apt OTA updates, a web dashboard and admin, live install telemetry, and payouts. I owned release engineering across five artifacts and the GA cutover.

## Inference on the fleet

vLLM serving of open models such as Qwen 3.5 9B ran on fleet GPUs and was benchmarked against hosted inference on cost and latency to validate the economics of self-hosted inference.

Back: https://jigarpatel.ca/#work