Ashburn sits next to the transatlantic cable landings and within about 15 ms of New York, Toronto, Chicago and Atlanta. How it pairs with a European primary, why one TCP flow across the Atlantic cannot fill a big port, and what the GDPR says about putting EU data there.
Poland is the EU's fifth most populous country, and every Baltic capital and Kyiv sit within 9 ms of floor from Warsaw. For game servers, regional services and patch-day delivery that distance is the whole argument. The latency table, the sizing rules and when Warsaw is the wrong pick.
Marseille faces south and east: it is where 2Africa, PEACE and AAE-1 come ashore. For North Africa it is 650 to 800 km closer than Frankfurt; for the Gulf it is the European end of the cable. The latency table, the video and regional workloads that fit, and the limits.
For users in Romania, Bulgaria, Moldova and Serbia, a server in Frankfurt adds 15 to 20 ms to every round trip. Bucharest removes it at the same price, keeps the data in the EU, and is where our 96 TB storage servers live. The latency table, the workloads and the sizing.
Amsterdam and Frankfurt are under 4 ms apart, so for most of Europe the choice is a wash. At the edges it is not: the UK, Ireland, the Benelux and the Nordics lean Amsterdam. The latency table, the VPN and delivery workloads that suit it, and a sizing rule for each.
Frankfurt is the default European location, and usually the right one: one of the world's largest internet exchanges, and no major capital in the west or centre of the continent more than about 25 ms away. The latency table, the workloads that fit, and the cases where another city wins.
Unmetered says the bytes are not counted; it does not say the port is yours at the busy hour. The five contract clauses to check, how to size a tier from your peak, what each tier costs per delivered terabyte, and the 13 cities where the server can live at the same price.
Generating a token means reading the model's active weights once, so a CPU server's speed ceiling is its memory bandwidth divided by the model's size. The arithmetic, the catch in prompt processing, why two sockets are not twice as fast, and the workloads a big-memory EPYC box genuinely suits.
Most database performance problems are a working set that no longer fits in memory, or a commit that waits on a slow flush. How to measure both, how much RAM and what kind of disk to buy, where the CPU matters, and how to lay out replicas and backups.
One rented machine, a free hypervisor, and as many virtual machines as the hardware will hold. How to budget RAM, lay out ZFS, get networking right on a data-centre switch port — the part that differs from a homelab — and keep the backups off the box.
What you can rent is not this year's Xeon against this year's EPYC but an eight-year spread of generations at prices seven times apart. Per-core speed, memory bandwidth, PCIe lanes and price per core — and which of those differences your workload can actually feel.
The port is the easy part — it is lit in an afternoon. Then the machine behind it has to produce 12.5 gigabytes every second, and each layer between the disk and the wire gets a chance to be the reason it does not. The arithmetic, in the order it fails.