Last month we gave a whole article to Hetzner's price rise, and the conclusion was that no single provider had lost its mind, a market was giving way. In August every other provider confirmed it. OVHcloud announced a 5% to 10% increase spread between April and September 2026. AWS, Azure and Google Cloud have announced nothing yet, but forecasts put them on the same 5-10% in the second half, with the usual lag of players holding supply contracts signed years earlier. The number that explains it all came from Hetzner: DRAM cost is up to 500% higher than September 2025.
Why it is happening
There is one cause, the AI data centre rush on memory. Manufacturers have diverted an enormous share of capacity to high-bandwidth memory for GPUs, and what is left for ordinary server RAM costs a fraction of what it used to. Hetzner's 128 GB add-on went from 45.88 to 264 euro, a +575% that is not a typo. OVHcloud's projections put RAM prices up 250-300% by the end of 2026 against a year earlier.
Migrating does not fix it, reducing does
We wrote it in July and it still holds: switching provider because it is cheaper today moves the problem by a quarter, because everyone buys the same chips on the same strained market. The one lever you genuinely control is how much memory your application uses.
What we are doing on client projects
- Serious right-sizing: most instances we find are sized against an imagined peak, not a measured one. A two-week memory profile often frees a whole machine tier.
- Connection pool under control: every idle Postgres connection is parked RAM. A pooler like PgBouncer in front of the database cuts usage without touching application code.
- Cache out of process RAM: keeping everything in container memory was free when RAM was free. Today it pays to move the cache onto a shared, sized layer.
- Arm by default: Arm lines rose far less than x86, and an SME's PHP, Node and Postgres run on them without drama. It is the value tier still standing.
- Cost inside monitoring: euro per month per service next to the technical metrics, reviewed quarterly. You do not optimise what you do not watch.
Verdict
The increase is not over and it does not depend on your provider, it depends on how many memory chips the AI data centres pull out of the market. Stop hunting for the cheapest host and start treating RAM as a scarce resource to design around, exactly as you did with bandwidth fifteen years ago. Over the next two quarters, the gap between teams that profiled memory and teams that did not is a three-figure line on the monthly bill.