Server Auction: how to get dedicated bare metal for less, and what to check before you buy

What is in the auction and why it is cheaper

Data centres refresh hardware on a cycle. Servers that come back from customers or that belong to a prior generation are fully functional, already racked and cabled, and would otherwise sit idle. The auction lists them at a discount to the current line price and waives the setup fee because no build work is needed. You get a dedicated machine — nothing is shared — with the exact CPU, RAM and drives in the listing. Browse the live list on the Server Auction page; the summary below is read from the catalog each time this page loads:

Auction versus new configuration

How prices and stock move

Each listing is a single physical unit. When someone orders it, it disappears from the list; when a similar unit returns, it is relisted, sometimes at a different price. Popular profiles — lots of RAM with NVMe, or any unit with a data-centre GPU — tend to go quickly. The practical consequence is that the auction rewards decisiveness: if a listing meets your requirements and budget, ordering it beats waiting for a marginally better one that may never appear. The total count on the page today is 146.

A five-minute evaluation checklist

1. CPU. Note the model and generation. For databases and inference, core count and memory bandwidth matter more than clock; for single-threaded workloads, clock matters.
2. RAM. Capacity first, then type (ECC is standard on server lines). 64 GB is a comfortable PostgreSQL box; 128 GB+ suits large vector indexes.
3. Disks. NVMe for databases and anything latency-sensitive; large SATA or HDD for storage and backups. Two drives let you run RAID 1.
4. Network. Check the port speed and traffic terms on the listing against your expected egress.
5. Price versus the standard line. Compare with the equivalent AX or EX configuration; the discount should be meaningful, not marginal.

Operating older hardware well

Previous-generation hardware is not unreliable, but you should run it like any bare-metal server: mirror the drives (software RAID 1 is a few commands on Ubuntu 24.04 or Debian 12), monitor SMART data, and schedule off-box backups from day one. If the workload is critical, run a replica on a second auction unit; two discounted servers often cost less than one new one and give you redundancy the single machine cannot. Provisioning offers Ubuntu 24.04, Ubuntu 22.04 and Debian 12 images so the base setup matches what you run on CPU Cloud nodes.

Workloads that fit the auction best

• Primary databases (PostgreSQL, MySQL) where consistent I/O beats the newest CPU.
• Vector stores and search indexes that need RAM more than compute.
• Backup and object storage on high-capacity units.
• Self-hosted inference on GPU listings when traffic is steady enough that a fixed monthly price beats per-minute GPU pods.
• Build farms and batch processing that run around the clock.

Workloads that do not fit are those that need live resizing, very short lifetimes or the latest accelerator features; those belong on per-minute cloud servers or current-generation lines.

From listing to running server

Pick the listing, choose an OS image, add your SSH key and order; the machine provisions quickly because it is already built. Then follow the same first-hour routine as any server: non-root user, key-only SSH, firewall, unattended upgrades, RAID and backups. Billing is monthly from activation, with no setup fee, and the configuration stays yours for as long as you keep paying. More on sizing databases and choosing between lines is in the dedicated servers guides.

Comparing an auction unit with the hosted alternatives

Before ordering, put three numbers side by side: the listing's monthly price, the per-minute cost of a CPU Cloud plan with similar RAM run 24×7 (hourly rate × about 730), and, if the workload is inference, the per-token cost of the same model on the pricing page at your traffic. The auction unit wins when the machine will be busy most of the month and the workload is stateful or performance-sensitive; the cloud plan wins when you need elasticity; the API wins for bursty inference. Writing the three figures down takes ten minutes and prevents the common mistake of renting a server that sits mostly idle.

Finally, check the status page history for the platform you are moving to and keep a current backup elsewhere during the first month, as you would with any new hardware. If the listing is for a database or storage role, restore that backup onto the new machine as the first task and time it; a restore that takes four hours tells you something about disk throughput that no spec sheet will, and it proves the recovery path before you depend on it.