Hetzner Doubles Prices: The AI Memory Crunch Is Why
If you run anything on Hetzner, you've seen the notice. As of 08:00 CEST on June 15, 2026, new cloud and dedicated server orders jumped an average of 99% in Germany, 158% in US locations, and 78% in Singapore. Some line items more than tripled.
For a host known for "absurdly cheap European iron," a near-doubling is a shock. But the reason behind it will show up in your bills too, whether or not you host on Hetzner or do anything with AI.
The Numbers
Existing machines keep their old price until you reorder or resize. Web hosting, managed servers, IPs, storage boxes, and load balancers were left out. Representative changes from Hetzner's price tables and heise:
| Server | Before | After | Change |
|---|---|---|---|
| CAX11 (ARM, DE/FI) | €4.49/mo | €5.99/mo | +33% |
| CCX13 (dedicated vCPU, DE/FI) | €15.99/mo | €42.99/mo | +169% |
| CPX41 (US region) | €38.99/mo | €120.49/mo | +209% |
Two patterns: ARM took the smallest hit. x86 dedicated-vCPU lines with more memory took the largest. US capacity rose more than European, tracking where new hardware is hardest to get.
This is the third price adjustment in 2026. On April 1, cloud servers rose 30-43%, object storage 30-53%, and memory add-ons by ~575%.
The Real Story: Memory Market
Hetzner cites "extremely high procurement costs for new hardware." That undersells the moment. The component market is in an AI supercycle.
From Tom's Hardware, IEEE Spectrum, and TrendForce:
- DRAM and NAND prices rose 50-200% in H1 2026. DRAM is up ~171% YoY.
- AI data centers will consume ~70% of high-end DRAM output in 2026.
- Samsung, SK hynix, and Micron redirected capacity to HBM and advanced DDR5 for AI. Micron's entire 2026 HBM output is already committed.
- Hard drives are sold out for the year. Analysts expect tight allocation into 2027.
Server memory and storage are most of a machine's BOM. When DRAM nearly doubles and high-capacity drives are on allocation, the cost of new servers rises sharply. Hetzner's 575% memory add-on jump in April makes sense.
It's Not Just Hetzner
Hetzner is a leading indicator. It sells close to cost, so component price spikes reach customers in weeks. AWS, Google Cloud, and Azure buy in volume on long contracts with higher margins, hiding the shock temporarily. The same DRAM and drives go into their racks. The bill arrives later as worse renewal terms, thinner discounts, pricier memory-optimized instances, and instance families that stop getting cheaper. If a near-cost provider went up 99%, the providers selling the same silicon at a markup are not immune.
Is Hetzner Still Cheap?
Mostly yes. Even after this increase, Hetzner remains dramatically cheaper than hyperscalers for raw compute and bandwidth. A comparably shaped box (~2 vCPU, 8 GB):
| Provider | Price (approx) |
|---|---|
| Hetzner (after increase) | ~€5.99/mo (ARM) |
| DigitalOcean | ~$12/mo |
| AWS (on demand) | ~$30/mo |
Before June 15, that Hetzner box was roughly a quarter of AWS price; now it's closer to two thirds. The moat shrank but didn't close. Egress pricing (generous on Hetzner, ~$0.09/GB on hyperscalers) didn't change. For bandwidth-heavy services, egress can still dwarf compute.
What You Should Do Now
-
Protect grandfathered machines. Existing servers keep old price until reorder or rescale. A casual resize reprices the whole machine at the new rate. Before bumping a server up a tier, check the new cost.
-
Treat memory as the cost center. RAM is the line item that exploded. Audit over-provisioned instances. Right-sizing memory pays for itself this quarter.
-
Look hard at ARM. Hetzner's ARM line took a third of the increase of x86 lines. If your stack runs on ARM, you dodge a large part of this and get better price-performance. Same on hyperscalers with Graviton.
-
Re-run cost models for hardware inflation. This is not contained to one host or quarter. Price your colo refresh, cloud renewals, and RAM in your next batch of laptops against a market tight into 2027.
-
Don't over-correct. Migrating providers has large engineering costs. The right move is to measure, right-size, and renegotiate, not flee.
The Takeaway
The AI build-out is now large enough to move the price of components every other workload depends on. You don't have to train a model to pay for the boom. If your service needs memory and disks, you're bidding for the same supply that AI data centers are buying 70% of.
Hetzner is just the first invoice to say so out loud. The rest will follow. Plan your next year of infrastructure spend as if memory is expensive and scarce — because for the foreseeable future, it is.
