Hetzner Isn't the Cheap Default Anymore — And the AI Boom Is Why

I’ve been running personal infrastructure on Hetzner for the better part of a decade. Cloud VMs for side projects. A couple of dedicated boxes for the things I didn’t want to babysit on AWS pricing. The calculus was always the same: pay a third of what the hyperscalers charge, manage it yourself, accept the lack of a managed-service safety net. The deal was very, very good.

On April 1, 2026, that deal got noticeably worse.

Six weeks in, I’ve had enough time to actually do the math, and I want to write down where I’ve landed — because I suspect a lot of people who build on Hetzner are running the same calculations and arriving at similar conclusions.

What changed

Hetzner raised prices across cloud, dedicated servers, storage, and load balancers in every region — Germany, Finland, the US, Singapore. A few representative numbers:

Cloud is where the hike bites hardest — 30 to 37% depending on the tier, with some specific configurations climbing as high as 50% once you factor in storage and traffic add-ons. Dedicated boxes got off easier. The Server Auction line — Hetzner’s second-hand returns — got a token 3% bump, which I assume is the point: that’s still the arbitrage tier.

This is the second major price action in 18 months. The first one, in December 2024, was much sneakier and arguably worse: Hetzner cut bandwidth allotments on US cloud VMs by ~88% on average — the basic tier went from 20 TB to 1 TB per VM — while quietly bumping prices 4–27%. The April 2026 hike is the loud one. The 2024 bandwidth cut is the one that actually changed unit economics for anything serving traffic.

The honest reason: DRAM and AI

Give Hetzner credit for not being cute about the cause. Their own statement says it plainly: hardware costs, energy, and component shortages. The number behind that statement is the one worth internalizing:

DRAM prices rose roughly 171% year-over-year through 2025.

That isn’t because RAM is harder to make. It’s because HBM (high-bandwidth memory) for AI accelerators eats production capacity that used to make commodity DDR5. NVIDIA, AMD, and the hyperscalers have effectively monopsonied a meaningful share of the world’s memory fab output for the next several years. The first downstream effect was that gaming GPUs got more expensive. The second was that consumer laptops got more expensive. The third — landing now — is that VPS providers got more expensive.

This is an externality. Hetzner isn’t price-gouging; they’re just forwarding a real cost increase. The component bills go up, and at their margins they can’t absorb it for long. The same forces are hitting OVH, Scaleway, Netcup, and every other budget host that competes on hardware-per-euro. Hetzner just announced it first and loudest.

That doesn’t mean you have to be happy about it. Honest doesn’t mean painless.

The quieter stuff

A couple of the smaller adjustments deserve more attention than they’re getting:

• Setup fees expanded. Hetzner now charges a one-time setup fee on more product lines, including some that used to be fee-free. For monthly users this barely registers. For anyone churning short-lived servers (CI runners, temporary load testing, ephemeral dev environments) it’s a tax that compounds.
• IPv4 is increasingly its own line item. You can shave €0.50/mo per server by going IPv6-only, which is fine for backends but a non-starter for anything serving the public web in 2026 (yes, still).
• Traffic overages are now more painful. With reduced bandwidth allotments on US plans and overage rates that haven’t gotten any cheaper, an unexpected traffic spike now costs noticeably more than it used to. If you’re doing any video, image-heavy, or RSS-popular workload, this is the line you want to watch.

Where this actually leaves you

I see four paths, and the right one depends entirely on what you’re running.

1. Pay the bump and move on

If you have one or two boxes for personal stuff and the absolute delta is €4–€10 a month, just pay it. The cost-to-time ratio of migrating somewhere else for that delta is terrible. Hetzner’s post-hike pricing is still better than equivalent-spec instances at DigitalOcean, Linode, or Vultr — never mind the hyperscalers, where the gap is now somewhere between 3× and 5× for comparable hardware.

This is where most of my own personal stack landed: I grumbled, I clicked “OK,” I moved on.

2. Migrate to Netcup, OVH, or Scaleway

If you’re running real workloads — a few dedicated boxes, a meaningful cloud fleet — the math gets more interesting:

• Netcup is the obvious side-step. Their RS 1000 G12 gets you 4 dedicated cores at €8.74/mo versus Hetzner CPX31 at €12.49/mo for 2 cores. Different KVM management, German jurisdiction, comparable network. Worth a look if you’re willing to switch provider muscle memory.
• OVHcloud wins on global presence and dedicated-server variety. Pricing is roughly comparable to Hetzner-after-the-hike, with occasional better deals on dedicated. Better for anything that needs a non-European footprint.
• Scaleway is the closest thing to a like-for-like swap if you’re staying in Europe. Their bare-metal “Elastic Metal” line is genuinely competitive, and they have a more developer-friendly API than Hetzner’s Robot.

None of these are dramatically cheaper than the old Hetzner. They’re roughly competitive with the new one. That’s the real story: the budget-VPS market just compressed.

3. Self-host on your own hardware

The boring answer that I keep coming back to: a used enterprise server on eBay plus a residential fiber line will outperform €100/mo of cloud spend within a year. I’m increasingly running my own services on a small box at home — uther, for the record — and the only thing I miss is the 99.99% uptime someone else worries about.

This isn’t right for everyone. It requires you to be the person who gets paged when the UPS battery dies. But for personal infrastructure, blog hosting, dev environments, and the kind of workload that doesn’t need a real SLA, the homelab math has flipped. A used Xeon with 64 GB of DDR4 from 2023 is currently a screaming deal because everyone else is fighting over DDR5 for AI rigs.

4. Reach for the hyperscaler managed services you’ve been avoiding

Counter-intuitive, but: as Hetzner’s price advantage compresses, the relative cost of going to an AWS/GCP/Azure managed service for specific pieces (managed Postgres, S3-compatible object storage, a queue) gets less awful. You won’t move your whole stack there, but “Hetzner for compute + cloud-provider for the one stateful service I’d rather not run myself” is now a reasonable hybrid where it used to be a luxury.

Where I’ve landed personally

I’m not leaving Hetzner. I am doing the math I should have done a year ago.

• Personal blog and a couple of small services: moved to a box at home (uther). The blog you’re reading is now served from my hallway, behind Cloudflare.
• Side-project compute that needs to actually be on the internet: stayed on Hetzner, paid the bump.
• A couple of larger workloads I was thinking about: getting quotes from Netcup and comparing genuinely, not handwaving.

The takeaway isn’t “Hetzner is bad now.” It’s that the cheapest infrastructure decade is over for a while. AI demand has made memory expensive, and there’s no sign of that easing. Every provider you care about is going to do some version of what Hetzner just did, whether they’re upfront about it or not.

The good news: this is exactly the sort of situation where doing the math properly — instead of defaulting to last year’s habits — actually pays off. The bill differences are big enough now that caring is worth it.