Auction servers in 2026 — still the best $/perf or a trap?

Been running auction boxes from Contabo and OVHcloud for exactly 12 months now. Heres the data.

Contabo DE-01 (E5-2680v4, 64GB, 2x480GB SATA)
disk: 412 MB/s read, 398 MB/s write
network: 215 Mbit down / 198 Mbit up
geekbench 6 single: 847
geekbench 6 multi: 7823
price: 18.90 eur/mo

OVHcloud NL-02 (E3-1270v6, 32GB, 2x240GB SSD)
disk: 523 MB/s read, 511 MB/s write
network: 412 Mbit down / 389 Mbit up
geekbench 6 single: 1243
geekbench 6 multi: 4512
price: 14.50 eur/mo

Vultr UK-01 (Ryzen 5 3600, 64GB, 2x1TB NVMe)
disk: 1567 MB/s read, 1489 MB/s write
network: 891 Mbit down / 756 Mbit up
geekbench 6 single: 2156
geekbench 6 multi: 8921
price: 29.00 eur/mo

Decent disk IO on the old xeons, meh network on Contabo. The ryzen box is obviously best perf but 2x the cost.

Year total: 748 eur spent, zero hardware failures, one 4-hour downtime on OVHcloud for power mai

Nice data SingaporeRep. For that money you get way more flexibility than renting fresh hardware.

My current auction fleet:

• Contabo DE-01: E5-2650v3, 128GB RAM, 4x4TB HGST NAS drives. Running ZFS raidz1. 10.4TB usable. 31 eur/mo
• Hetzner FI-02: E3-1230v5, 32GB, 2x8TB WD Red. mdadm raid1. 24 eur/mo — https://www.hetzner.com/sb
• RackNerd PL-01: E5-2683v4, 256GB, 8x2TB SAS. Raidz2 for cold storage. 45 eur/mo

The sweet spot is still 4TB+ drives in cheap chassis. New NAS from anywhere costs 3x this.

Only issue: Contabo shipped one server with a failed backplane. Took 8 days to replace. For that money you get what you get.

HDD smart data on the HGSTs shows 38000 power on hours. Still spinning fine. Ymmv on auction drive lottery.

My angle: power draw matters more than benchmark numbers if youre actually racking these. Measured at the pdu:

• E5-2650v3 dual socket idle: 87W
• E5-2680v4 dual socket idle: 94W
• E3-1270v6 single socket idle: 43W
• Ryzen 5 3600: 62W

The v3/v4 xeons are cheap to buy, expensive to run. Over 3 years at 0.25 eur/kWh that 50W gap is 330 eur difference. Wipes out most auction savings versus a newer gen.

Shipping pallets: Contabo uses standard 48x40 wood pallets, OVHcloud ships on cheap pressed wood that crumbles. Learned that the hard way.

Rack units: all these are 2U except the E3 boxes which are 1U. Density is bad. 8 auction servers = 16U minimum. Versus 4U of modern stuff.

Practical advice: if you have cheap power or its just one box, ignore me. If youre filling a 42U, do the math.

Network numbers only tell half the story. Heres routes from my OVHcloud NL-02 to some test targets:

to he.net freemont:
1  185.233.x.x     0.5ms
2  62.115.x.x      1.2ms  OVHcloud core
3  80.249.x.x      4.8ms  AMS-IX
4  184.105.x.x    142ms   route via HE adds 12ms vs my Contabo DE box

to Vultr UK-01 same target:
1  10.0.x.x        0.3ms
2  185.60.x.x      2.1ms
3  195.66.x.x      8.4ms  LINX
4  184.105.x.x    128ms

to Contabo DE-01:
1  185.233.x.x     0.5ms
2  62.115.x.x      1.2ms
3  80.249.x.x      4.8ms
4  80.81.x.x       11ms   DE-CIX
5  184.105.x.x    156ms   route via DE-CIX adds 28ms vs OVHcloud path

OVHcloud has better transatlantic routing. Contabo dumps you through DE-CIX with higher latency to US west coast. For EU-EU both are fine, under 15ms to most targets.

The 198 Mbit up SingaporeRep saw? Thats because OVHcloud has 10G uplinks shared 20:1

Which kernel version on the ryzen box

50W gap is real

5.15.0-105-generic. Had to upgrade from the stock 5.4 to get proper cpuidle on the 3600. Without it idle was 78W instead of 62W.

At Singapore commercial rates of 0.35 SGD/kWh that gap is worse. But my Contabo box is in Nuremberg, not here. Power bill goes to Contabo, not me. If youre colocating, Carls math is correct. If youre renting, its Contabos problem.

Auction servers are still the best $/perf for renters. The trap is buying them thinking you got a deal, then paying for the power yourself.

To clarify: DE-CIX is fine for Europe. My complaint is specifically transatlantic and to Asia. From Stockholm:

• OVHcloud NL-02 to my office: 18ms
• Contabo DE-01 to my office: 34ms
• Hetzner FI-02 to my office: 22ms

OVHcloud wins on latency to Nordics because they peer at NetNod too. Contabo only hits DE-CIX and then its a long haul north.

For a CDN backend or game server this matters. For ZFS cold storage like steveipw, irrelevant.

Measured mine at the wall: 91W idle, 187W under scrub. The HGSTs add 28W spinup each but thats brief.

Still, 330 eur over 3 years is real money. My use case is backup target though. Writes once a day, reads on restore. Power cost per GB stored is tiny versus S3 or B2.

Newer idea: Hetzner auction has E5-2680v4 now for 27 eur with 256GB. Same CPU as SingaporeReps Contabo but more RAM and cheaper per GB. If I was buying today Id skip Contabo for Hetzner on price alone.

My electricity bill is 0.28 USD/kWh here in Tokyo. I looked at auction servers last year and ran the same math Carl did. Even with cheap hardware, running it here destroys the savings.

I rent a Vultr cloud instance in Singapore instead. 12 USD/mo for 2GB. Not comparable performance but I dont pay for cooling or power or the space.

Auction servers only make sense if you have cheap power or someone else pays. Japan, California, most of Europe: no.