Blog
Insights, analysis, and updates from the AI agent economy. Browse by tag · Browse the archive.
Paying $4,700 for a $1,999 GPU: What the RTX 5090 Street Premium Does to Self-Hosted Inference Math
An RTX 5090 at today's $4,700 street price adds about $75 a month to a self-hosted inference node — while H100 rentals fell roughly 70% from their peak. A worked amortization of both price curves, and where buying still wins.
Render's 75% Faster Python Builds, Deconstructed: What a Self-Hosted PaaS Must Cache to Match It
Render claims buildpacks make Python builds 75% faster than a Dockerfile. The speedup is really warm layer-cache locality, and a self-hosted PaaS can match it with persistent build caches, warm build nodes, and dependency-first layering.
Railway's Per-Second Meter vs Render's Flat Workspace Fee: What a Bursty Side Project Actually Costs in 2026
Railway bills compute by the second under a $5 Hobby floor; Render charges a flat workspace fee plus fixed per-service prices. A worked comparison for a bursty side project and a steady production app — with the crossover math and what a flat Hetzner box removes entirely.
No Account, 60 Minutes, Zero Trust: The Abuse Controls Behind Anonymous Deploys
Railway's no-account 60-minute deploys show what instant agent-to-preview loops need — and what stops them from becoming free cryptominers. A hostile-user checklist for quotas, network cages, build scanning, and timed teardown on your own Kubernetes fleet.
Sessions Live, Sandboxes Die: How Perplexity Cut Agent Sandbox Creation From 185ms to 60ms
Perplexity's SPACE runtime cut agent sandbox creation from 185ms to 60ms median by making the session — not the sandbox — the durable unit. The mechanism, the honest median-vs-tail math, and what it sets as the bar for self-hosted sandboxes.
The 5x PaaS Pricing Spread: What One API + Worker + Postgres Costs on Railway, Render, Fly.io, Vercel — and a Flat Hetzner Box
A line-by-line recompute of the same API, worker, and Postgres workload across Railway, Render, Fly.io, and Vercel rate cards — plus egress sensitivity to 2 TB, cold-start math, and the flat Hetzner number the whole spread collapses to.
Bring Your Own Cloud, Keep Someone Else's Meter: What Northflank BYOC Really Costs Against Owned Hardware
Northflank's BYOC runs your workloads in your own cloud account, but its per-vCPU-hour meter runs on top of your cloud bill. A worked three-way cost comparison — hosted vs BYOC vs a flat Hetzner box — plus the ownership question no invoice shows: who controls the control plane.
Netlify Will Now Run Your Postgres Too: What the One-Invoice Database Really Costs
Netlify Database folds managed Postgres into the platform bill with per-preview data branches. A concrete accounting of what the bundle buys, the four ceilings it installs, and why bex keeps the database behind a tenant-owned connection string.
Kubernetes 1.36 Made Mutating Webhooks Optional: What Your Admission Layer Can Delete — and What Has to Stay
Kubernetes 1.36 graduates MutatingAdmissionPolicy to stable, so CEL evaluated inside the API server can replace hand-rolled mutating webhooks for tenant defaults, baselines, and sidecars. A concrete migration map, the boundary where webhooks stay, and a fleet-upgrade checklist.
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