Blog
Insights, analysis, and updates from the AI agent economy. Browse by tag · Browse the archive.
TurboFieldfare Fits a 26B MoE Model in 2GB of RAM: What SSD-Streamed Experts Mean for Your Cheapest Inference Node
TurboFieldfare runs Gemma 4 26B-A4B in ~2GB of RAM by streaming 4-bit experts from SSD — 5–6 tok/s on an M2 Air, 31–35 on an M5 Pro. The per-token byte math behind those numbers, and how to size a €49 NVMe box as your cheapest inference tier.
Tailscale Went Seat-Based: What Self-Hosting Your Admin VPN With NetBird or Headscale Actually Costs
Tailscale's 2026 move to 8 to 18 dollar per-seat pricing turned the admin VPN into a headcount tax. What a self-hosted NetBird or Headscale control plane costs in infrastructure and ops hours, the exact seat counts where each crosses over, and when to just pay the SaaS bill.
SmolVM Packs a MicroVM Into a Single Binary That Boots in Under 200ms — What It Changes for a Firecracker-Only Agent Sandbox
SmolVM boots any OCI image as a hardware-isolated microVM from a single binary in under 200ms, with pack and branch primitives Firecracker never shipped. Here is how it compares head-to-head and what to verify before adding it to a self-hosted agent-sandbox stack.
Seven Months to Self-Host an AI Stack That Was Supposed to Take a Weekend: What OpenMake's Post-Mortem Bills Against a PaaS
OpenMake's seven-month self-hosting retrospective comes with receipts: four changelog bugs and a Node-plus-Postgres-plus-Docker stack. A worked Year 1 comparison shows VPS savings evaporating after two days of engineering time — and why Kubernetes wouldn't have shortened the seven months, but a PaaS deploy surface would have.
Render Cut Median Builds 40% With Faster Hardware: Reproducing 38s-to-21s on Owned Hetzner Build Nodes
Render cut median service builds from 38 to 21 seconds with faster hardware. A costed playbook for matching that 40% cut on owned Hetzner machines with one shared BuildKit builder and cache-first levers — no fleet-wide upgrade required.
Railway Retired Nixpacks for Railpack: What Changed in the Zero-Config Build Path
Railway replaced Nixpacks with Railpack after 14 million builds exposed the cost of approximate versions and monolithic image layers. A before-and-after of the auto-detect build path, why Cloud Native Buildpacks lost twice, and what platform builders should steal.
Railway's May 2026 GCP Blackout: What an 8-Hour Outage Reveals About a Control Plane That Never Actually Left Google Cloud
Google suspended Railway's production account and healthy bare-metal workloads went dark for eight hours — because the control plane routing them never left GCP. A timeline of the cascade and a seven-question audit for your own platform.
Railpack Is Railway's Default Builder: What the Nixpacks-to-BuildKit Rewrite Teaches Self-Hosted PaaS Builders
Railway replaced Nixpacks with Railpack, a Go rewrite on raw BuildKit, cutting base images 38-77%. A look at what broke, what the numbers prove, and what CNB-based self-hosters should borrow — plus where the Dockerfile escape hatch still wins.
Preview Environments Are a Cost Problem Before They're a Feature: The 50-PR Math on Auto-Sleep and Auto-Destroy
At 50 concurrent PRs, always-on preview environments burn about $2,850 a month. Auto-sleep cuts the compute line, scale-to-zero database branches cut the data tier, and auto-destroy kills the 30% orphan drag — taking the bill to roughly $650. The line-by-line math, plus the three defaults a self-hosted platform should enforce.
Subscribe
New posts land in your reader as soon as they publish. Pick a format — all three carry the same posts.
Current feeds keep roughly two days of posts so daily polling does not miss a burst. Older entries stay reachable from the feed's next-page link in readers that follow it, or from the blog archive.
Following one topic instead? Browse tags