Fly.io Halved L40S to $1.25/hr — Then Scheduled GPUs for Deletion on August 1
Fly.io cut its L40S GPU to $1.25/hr and then deprecated all GPUs after August 1, 2026. A line-by-line market comparison shows why rented GPUs look cheap until they disappear — and when owned Hetzner hardware wins on cost and control.
Fly.io Sprites at $0.07 Per CPU-Hour: Why Zero When Idle Still Cost $35,000 for 200 Sandboxes
Fly.io Sprites charges $0.07 per CPU-hour only while active — idle is free. Yet a 200-sandbox benchmark put Sprites at $35,000 while Northflank hit $7,200 and a Hetzner fleet costs under €900. The math of idle ratios explains why.
Fly.io Is Killing GPUs on August 1: What Two Retreats in Three Years Say About Renting vs Owning Your Accelerators
Fly.io's A10/L40S/A100 GPUs go unavailable August 1, 2026 — its second GPU retreat since 2024 — while its jobs page sits at zero. A concrete migration math of where inference workloads land on rented GPU clouds vs owned bare-metal nodes, and why an owned GPU pool can't be sunset.
Hetzner's €889 RTX PRO 6000: What Self-Hosted AI Inference Actually Costs When You Already Own the Fleet
Hetzner's GEX131 with a 96GB RTX PRO 6000 costs €889/month flat — no per-token meter, no pause button. A worked comparison of what that card actually serves via vLLM, where it undercuts hosted API pricing, and where bursty load makes the monthly commit the expensive choice.
The Self-Hosted LLM Gateway Pattern: How Mantis and LiteLLM Turn Model Routing Into a $10 Container
Self-hosted gateways like Mantis and LiteLLM run model routing, fallback, and cost tracking as a $10 container beside your apps — at 100M tokens the 5.5% managed fee costs 3x more and still routes your prompts through a third party.
Omni's One-Postgres Workplace Search: Running BM25 and Vectors in a Single Database as a Tenant Workload
Omni indexes wikis, docs, and drives with ParadeDB BM25 and pgvector in one Postgres, deployable with docker compose up. A walkthrough of the one-database pattern and why it runs as just another tenant on a self-hosted PaaS.
Hetzner Is Building an LLM Inference API: What €889/Month of GPU and a Free Experimental Endpoint Do to the Self-Hosted AI Math
Hetzner is testing an OpenAI-compatible inference API alongside its €889/month RTX PRO 6000 GEX131 — what the experiment changes for self-hosted AI rent-vs-own math, EU sovereignty, and running inference as a regular tenant workload.
Self-Hosting Firecracker Sandboxes for AI Agents: The Real Math Behind the 60–80% Cost Drop
E2B, Modal, and Daytona meter AI-agent sandboxes at $0.17–0.24 per hour; one $115 Hetzner box running Firecracker delivers them for cents. A worked cost model shows where the 60–80% savings actually start — and why the popular 500-sandbox-hour threshold assumes your ops time is free.
Crossplane vs Cluster API: Two CNCF Answers to "Should Infrastructure Be an API?" — and Where a Self-Hosted PaaS Draws the Line
Crossplane models everything from databases to GPU inference fleets as one declarative API; Cluster API keeps a written list of what it refuses to manage beyond cluster lifecycle. A side-by-side of what each CNCF project actually optimizes for, why they compose as layers rather than compete, and why a self-hosted PaaS builds on Cluster API while skipping multi-cloud abstraction.