Hetzner has raised prices three times in 2026. The February price-increase announcement — up to 37% on cloud servers, effective April 1 — got the headlines, but it was the first of three separate adjustments in five months. If you run infrastructure on monthly-billed EU servers, this is worth twenty minutes of your attention, because the cause is structural and it is not going away in 2026.
This is not a "switch to us" post. We run AS60982 and we buy RAM in the same market Hetzner does — more on that below, because it's the part most competitor blog posts conveniently skip. This is a breakdown of what actually changed, why it changed, and how to evaluate alternatives without trading one problem for another.
Hetzner's 2026 price increases: three in five months
Here is the timeline, compiled from Tom's Hardware and Web Hosting Today:
| Announced | Effective | What changed | Who it hit |
|---|---|---|---|
| Feb 23, 2026 | Apr 1, 2026 | Cloud servers in Germany/Finland +30–37%, US/Singapore up to +38%, dedicated servers +3–21%, storage products +30% | New orders and existing customers |
| Apr 29, 2026 | Immediately | Dedicated server setup fees raised, citing RAM and NVMe component costs | New orders |
| May 27, 2026 | Jun 15, 2026 | Dedicated and cloud portfolios restructured into -1/-2/-3 hardware tiers, plus a lower-cost "Ltd" tier; price adjustments on the new lineup | New orders only; running servers kept their terms |
Two things stand out. First, the April 1 round applied to existing customers, not just new orders — servers people had budgeted at one price cost up to a third more the next month. Second, the pace: three adjustments in five months is not a pricing refresh, it's a company repricing in real time as its input costs move. In the Hacker News discussion, users cited Hetzner's own notification putting DRAM cost increases at up to 500% since September 2025, with memory add-ons for bare metal repriced 575% overnight.
Why: the AI DRAM squeeze
The root cause is memory. DRAM contract prices rose roughly 171% year over year — a steeper climb than gold over the same period — driven by AI datacenter buildouts. Industry estimates put AI-related demand at close to 20% of the entire industry's DRAM wafer capacity in 2026, and as much as 70% of high-end memory output. Manufacturers have shifted fabs toward HBM and datacenter parts because that's where the margin is, and everyone else — hosting providers included — bids for what's left.
The uncomfortable part: industry projections put meaningful supply relief no earlier than 2028. This is not a quarter of turbulence. It's the new baseline for anyone whose product contains DIMMs, which is everyone in this business.
To be fair to Hetzner: they were transparent about the cause, they remain cheap by industry standards, and repricing was arguably the honest move versus quietly degrading hardware. The lesson for buyers isn't "Hetzner bad." It's that monthly list prices are no longer a stable planning input anywhere.
The billing detail most people miss
While we're reading Hetzner documentation, one nuance matters more than any percentage: how hourly billing is metered. Per Hetzner's own billing docs, partial hours are rounded up, and billing runs until you delete the server — not until you power it off. A stopped server still bills, capped at the monthly price. That cap is a genuinely customer-friendly touch, but the combination means a "cheap hourly server" you forget to delete converges on the full monthly rate while doing nothing.
This matters because "hourly billing" is not one thing. Round-up-to-the-hour metering, bill-until-deletion semantics, and per-second metering produce very different invoices for the same workload. Which brings us to the honest part.
The part where we admit we buy the same RAM
Every EU host is exposed to this. Including us. When we provision new hypervisors for our VPS fleet or build out bare metal, the DIMMs come from the same three manufacturers fighting the same AI demand curve. Any provider telling you their prices are immune to a 171% input cost increase is either subsidizing losses temporarily or planning to tell you something different next quarter.
So the argument for evaluating alternatives is not price permanence. Nobody can promise you that in 2026. The argument is billing-model flexibility: if list prices are going to move with the DRAM market, the defensible position is paying only for the hours you actually consume, so a price change hits your utilized hours instead of your idle ones. A 20% increase on 720 hours of a mostly-idle staging box hurts; the same increase on the 60 hours it actually ran is noise.
How to evaluate an hourly-billed EU alternative
If the price increases have you comparison shopping — and judging by the size of that Hacker News thread, a lot of people are — here is the checklist we'd use, in order of how much money each item represents:
- Billing granularity. Ask three questions: what's the metering unit (second, minute, rounded-up hour)? Does a powered-off server bill? When exactly does billing stop — deletion, or stop? On our platform the answer is: pay per hour, billed to the second, from €0.006/hr. The distinction between per-second and rounded-up-hour metering is invisible on one server and very visible on a CI fleet cycling forty ephemeral runners a day.
- EU jurisdiction, actually. Not "EU region of a US company" — where is the legal entity, and which law governs your data? GDPR compliance under EU jurisdiction is a checkbox for some workloads and a hard requirement for others. Check before migrating, not after your DPO does.
- Network ownership. Does the provider run its own AS or resell someone else's transit? A provider operating its own network — we run AS60982 with 13 EU POPs and direct peering at 15 internet exchanges — controls its own routing, capacity, and peering disputes. A reseller inherits all three from upstream, along with upstream's pricing changes.
whoison the announced prefixes takes thirty seconds and tells you a lot. If you hold your own address space, also check whether BYOIP is supported, so a future migration doesn't mean renumbering. - DDoS mitigation: included or invoiced? Some providers include basic scrubbing, some sell it as a tier, some null-route you and call it mitigation. We include always-on 1 Tbps DDoS mitigation with detection under 8 seconds, free, on everything, because retrofitting protection during an attack is the most expensive way to buy it. Whatever provider you evaluate: get the mitigation capacity, the detection time, and the price in writing.
- Provisioning speed. Hourly billing is only useful if creating and destroying servers is fast enough to do casually. Our median time to root SSH is 47 seconds. If deploys take twenty minutes, you'll leave servers running "just in case," and the billing model stops mattering.
Where hourly billing actually wins
Hourly billing is not automatically cheaper — a 24/7 production database should be on a monthly plan, full stop. We wrote up the full break-even math separately, but the short version: at €0.006/hr against a €4/mo monthly plan, the break-even is about 667 hours — roughly 92% of a month. Below that utilization, hourly wins. The workloads that live far below it:
- CI/CD runners. A build fleet that runs 6 hours a day on workdays consumes ~130 hours/month:
130 × €0.006 = €0.78. Billed to the second, a single 15-minute pipeline job costs €0.0015. Under rounded-up hourly metering, the same job bills a full hour — 4× the price for the identical workload. Multiply by every job, every day. - Staging and review environments. Business hours only: 10 hours × 21 workdays = 210 hours, or €1.26/month per environment instead of €4. Destroy on merge, recreate on the next PR in about a minute.
- Batch and cron workloads. A nightly 2-hour ETL job is 60 hours a month: €0.36. Paying a full month for two hours a night is paying a 92% idle tax.
- Seasonal and event-driven capacity. Load-test fleets before a launch, extra capacity for a weekend event, restream relays for a one-off broadcast. Spin up, run the event, delete. For recurring streaming workloads the math flips — a 24/7 ingest server belongs on a monthly streaming VPS, not an hourly one.
The pattern: hourly billing converts servers from things you own into things you invoke. That's also your hedge against the DRAM market — when compute is something you invoke for 60 hours instead of rent for 720, input-cost repricing touches a twelfth of your bill.
Bottom line
Hetzner's 2026 increases are a symptom, not the disease. The disease is a memory market where AI buyers outbid everyone else until at least 2028, and it will reach every provider's price list — including ours — in some form. You can't dodge the input costs. You can choose a billing model where you only pay them for hours that produce something.
Audit your fleet's actual utilization before you migrate anything. If a server is busy more than ~90% of the month, monthly billing is correct and boring. Everything below that line is a candidate for per-second metering — and in a market that reprices every quarter, the servers you can delete are the ones that can't surprise you.