Ryzen 9950X-class boxes in the US — who is competitive right now?

Been hunting for high-clock dedicated for game-adjacent workloads (srcds, Mumble, some light proxying). Need sub-65ms to Chicago and sub-80ms to Dallas from my test points.

Looked at Contabo's lineup and OVHcloud's 9950X offering. Contabo routes via Cogent in some markets which adds 8-12ms vs their Level3 paths. OVHcloud seems to single-home through Hurricane Electric in SJC, fine for west coast, rough for east.

Contabo STL -> ORD:  34ms (Level3)
Contabo STL -> ORD:  46ms (Cogent, alternate route)
OVHcloud SJC -> ORD:  71ms
OVHcloud SJC -> DFW:  42ms
Hetzner IAD -> ORD:  28ms
Hetzner IAD -> DFW:  41ms

Anyone got actual 9950X or 9900X boxes live with clean routing? Not interested in 7950X3D, cache isn't the bottleneck here, clock is. Budget around $180-240/mo for 64GB DDR5 — Hetzner's auction sometimes has 9950X around there: https://www.hetzner.com/sb

I have used Hetzner for a while and this is what I have seen. They have Zen5 compute in the U.S. Now, though not in as many locations as some want. The following configurations are what I was able to get:

AMD Ryzen 9 9950X

• 128GB DDR5-5600 ECC (2x64GB)
• 2x 3.84TB NVMe Gen4 (Micron 7450)
• Dual 25GbE (Mellanox ConnectX-4 Lx)
• 100TB/month on 10Gbit burstable
• IPMI with HTML5 KVM
• €289/month, €199 setup waived through August 31.

AMD Ryzen 9 9900X

• 64GB DDR5-5600 ECC (2x32GB)
• 2x 1.92TB NVMe Gen4
• Single 10GbE
• 20TB/month on 10Gbit
• €189/month, same setup waiver.

Network: Blend of Arelion, GTT, and regional fiber in Ashburn and Hillsboro. No single-homed paths. BGP communities available for prepend control.

Power: 1.2A @ 120V measured at wall under all-core load, 0.45A idle.

1.2A at wall for a 9950X under all-core? That's 144W which sounds low unless you're capping PPT. My last 9950X build pulled 198W at the wall with PBO enabled, 1.05A idle with C-states on.

Been through three pallets of these now. The 9900X is the sleeper pick if you don't need all 16 cores. Better thermals, same memory controller, easier to keep under 1.5A in a 1U. The 9950X wants a 2U heatsink minimum unless you like 92C.

Idle power matters more than people think at scale. 0.45A vs 0.8A over a year is 300+ kWh difference. At $0.12/kWh that's $36/year per box. Run 100 boxes and you're eating a whole extra rack of power budget.

Shipping pallets: watch out for the Supermicro H13SSL-NT with the older BMC firmware. Fan curves are broken, sounds like a jet engine until you flash 2.5 or newer. Lost a whole weekend to that last month.

Anyone seen 9950X pricing actually drop? Still paying $520 tray.

Hetzner 9950X in IAD, fresh run

# ## ## ## ## ## ## ## ## ## ## ## ## ## ## ##
#              Yet-Another-Bench-Script              #
#                     v2026-07-15                  #
# ## ## ## ## ## ## ## ## ## ## ## ## ## ## ##

CPU : AMD Ryzen 9 9950X (16 Cores, 32 Threads)
RAM : 128 GB DDR5-5600
Disk: Micron 7450 3.84TB (xfs, 4k)
OS  : Debian 12

fio Disk Speed Tests (Mixed R/W 50/50):
---------------------------------
Block Size | 4k            (IOPS) | 64k           (IOPS)
  ------   | ---            ----  | ----           ----
Read       | 1.2 GB/s    (300k)   | 5.8 GB/s     (91k)
Write      | 1.2 GB/s    (300k)   | 5.8 GB/s     (91k)

iperf3 Network Speed Tests (IPv4):
---------------------------------
Provider    | Location        | Send Speed    | Recv Speed
------------|-----------------|---------------|------------
Clouvider   | London, UK      | 9.42 Gbit/s   | 9.38 Gbit/s
Wifx        | Paris, FR       | 9.15 Gbit/s   | 9.11 Gbit/s
Uztelecom   | Tashkent, UZ    | 4.22 Gbit/s   | 4.18 Gbit/s

Decent disk IO, meh network (expected on shared 25GbE)

Geekbench 6:

Single: 3,342
Multi:  21,847

Comparison to my other Zen5 boxes:
- OVHcloud SJC 9950X: single 3,298, multi 21,102 (older AGESA, probably)
- RackNerd MIA 9900X: single 3,105, multi 15,440
- my own colo 9950X with PBO: single 3,410, multi 22,600

Hetzner's conservative on power limits. Fine for hosting, leaves performance on table.

Ngl running inference on these is tempting lol

9950X isn't the play for LLMs though, you want VRAM. 128GB system RAM lets you run 70B Q4_0 at like 2 tok/sec which is... bad. Fine for embeddings or small classifier models.

I got a Hetzner 9900X doing Whisper transcription, 8 concurrent streams, uses like 40% CPU. The real win is when you pair with a cheap used GPU off ebay but then you're in "is this datacenter allowed" territory

Contabo has some 9950X + A4000 bundles now I think — https://contabo.com/en/vps/ ? VRAM math: 16GB gets you 7B at Q8 or 13B at Q4 with vLLM. Still not competitive vs cloud GPU pricing unless you're running 24/7.

LichunLars for game servers tho yeah clock speed is king. Srcds is basically single-threaded with a side of "please don't stutter". 9950X boost to 5.7GHz is legit, seen it hold 5.5GHz all-core in good cooling.

What's the PPT limit on the Hetzner 9950X? I don't see it listed