Hetzner raises prices a third time in 2026, and AI is the reason

Hetzner raised prices for the third time this year, and they're saying the quiet part out loud. The reason is AI.

It's Tuesday, June 16, 2026. Here's the rundown.

The cost floor for self-hosting just moved. Then Iroh hits 1.0, OpenRouter runs a panel of models, and somebody powered a voice agent with a hand crank.

Hetzner's change took effect yesterday at eight in the morning, central European time. This one is narrow. New orders and cloud rescales only, existing contracts keep their terms.

So this isn't the one people are angry about. That was April first.

Right. April was the real one. German and Finnish cloud servers up thirty to thirty-seven percent. US and Singapore cloud up to thirty-eight. US object storage up fifty-three percent.

Fifty-three on object storage. That's not a rounding error, that's a strategy shift.

In plain numbers, the AX42 went from forty-seven euros thirty to fifty-seven thirty a month. The DX293 went from three hundred five to three fifty-five.

And here's what makes this story different from every other price hike. They named the cause. Memory, fast storage, and GPUs for AI. The same shortage lifting your API bill is lifting your colo bill.

They even addressed the obvious objection. You'd think existing hardware is paid for, why reprice it.

Because running a box for years means replacing dead CPUs, RAM, and disks the whole time. Those replacements now cost more. The hardware doesn't care that your contract is old.

There's a portfolio change underneath it too. They're collapsing to about three fixed configurations that cover ninety percent of orders, plus a new Limited tier.

That part I actually respect. Fewer SKUs means they can buy components in volume and predict supply. It's a hedge against the exact crunch they're blaming.

So if you run on Hetzner this week, what's the move.

Existing contracts are grandfathered, so do not trigger a rescale. A rescale reprices you at the new tier. Size your capacity once and hold it.

And the bigger redo.

If your inference or RAG workload lives on bare metal because the math beat managed GPUs, redo that math now. The gap that made self-hosting obviously cheaper is closing.

For a decade commodity compute got cheaper on a schedule you could plan around.

AI demand broke the schedule. When the cheapest serious host in Europe raises prices three times in six months, the floor moved for everyone. Watch OVH and Scaleway next.

n0 shipped Iroh 1.0. Four years, sixty-five versions, and now a frozen wire protocol. You dial a key instead of an IP address.

The commitment is the news, not the version number. A v1 endpoint talks to any other v1 endpoint, any language, any minor version. Breaking the wire now requires a major release.

They built their own QUIC multipath layer too, several routes in one connection with hot-swapping. Their relays saw over two hundred million endpoints created in the last thirty days.

That's a real load test, not a demo. If you're building device-to-device connectivity, this is finally a dependency you can pin and forget.

And by coincidence, the eight fallacies of distributed computing just turned twenty-one this week.

The network is not reliable, latency is not zero, topology changes. Worth a reread the same week someone ships multipath QUIC and Hetzner reprices the boxes underneath you. Cheap to forget, expensive to relearn in production.

OpenRouter shipped Fusion. A panel of models answers in parallel, a judge compares them, and an outer model writes the final answer from the analysis.

It's not a majority vote, which is the interesting part. The judge returns consensus, contradictions, and coverage gaps. A panel topped the chart near sixty-nine percent against high-fifties for strong solo models.

And the hype tax.

You pay for every panel completion plus the judge. A request costs the sum of its parts, not a flat rate. Use it on hard, high-stakes calls, never your hot path. And the chart-topping panel used Claude Fable 5, which Anthropic already pulled, so that benchmark won't reproduce.

Tencent open-sourced an agent memory plugin under MIT. Default backend is local SQLite plus sqlite-vec, no external API needed.

The clever bit is symbolic short-term memory. It condenses heavy tool logs into compact Mermaid symbols instead of dumping everything into a vector pile. Self-reported, they claim up to sixty-one percent fewer tokens. Test it on your own agent before you believe that.

And then there's CrankGPT. A full offline voice agent on a stock Raspberry Pi 5, powered by a hand crank.

Thirty seconds of cranking to start a conversation, twenty seconds of capacitor buffer. It's an art piece. But the argument lands.

Small models do real work without kilowatts. In a week where every compute bill went up, that stopped being whimsy and started being economics.

Anthropic shipped a Swift package routing Apple's Foundation Models API to Claude. The same LanguageModelSession that drives the on-device model now drives Claude.

This is not Apple model news, to be clear. At WWDC Apple added a server-side provider hook and Anthropic took it. Requests go app to Claude directly, Apple isn't in the path, you pay standard pricing on your own key.

The intended pattern is local model for fast tasks, hand off to Claude for multi-step reasoning. Caveat, it's version zero point one, needs the iOS and macOS 27 betas and Xcode 27.

So real, and not for production yet. Three more quick ones. TeslaMate gives Tesla owners a self-hosted telemetry pipeline instead of a vendor dashboard.

EasySpider does visual, no-code web scraping, with the usual notes about site terms and rate limits. And there's a Universal Android Debloater rewritten in Rust.

That one uses ADB to strip bloatware from non-rooted devices. Useful if you provision a fleet or just want a cleaner phone without unlocking the bootloader.

Quick break — two from the desk.

One we know well: vote dot direct. If you're on an H O A or a board, it runs your elections digitally — secure, verifiable, no paper, no clipboard in the lobby. Point your council to vote dot direct.

And if this is your ten minutes of A I for the day, get the written edition too. The full wire, free, every morning — leave your email at nextbig dot dev.

curl will not accept vulnerability reports for all of July 2026.

Fox agreed to buy Roku, per the Journal's deal report.

Windows 11 users keep hitting Microsoft account requirements creeping into more features.

TinyWind, a pixel pirate game with real wind physics, has logged over three hundred eighty thousand kilometers sailed.

Monash reports a copper-transport drug that restored memory and cleared toxic Alzheimer's proteins in animal models.

And there's now a CAPTCHA that asks you to win a claw machine to prove you're human.

Our call. Hetzner announces a fourth 2026 price step before October 31, citing component costs again, and at least one of OVH or Scaleway raises prices in the same window.

It's wrong if no further Hetzner adjustment is published and neither OVH nor Scaleway moves by October 31. That's when it settles.