Custom dedicated servers.
Built to your exact spec.
Off-the-shelf stops here. Pick any CPU, memory, storage and network - our engineers source it, burn it in and rack it on our own backbone in 24–72 hours. Write the spec; a quote lands in 2–4 hours. No obligation.
AlphaVPS Custom Dedicated Servers = bare metal built to your exact spec - any Intel Xeon, AMD EPYC or Ryzen CPU up to 128 cores, 32 GB–2 TB ECC RAM, NVMe/SSD/HDD in any RAID, 1–100 Gbit ports, optional NVIDIA GPUs - assembled, burned-in and racked in 24–72 hours in Sofia, Nuremberg or London. Free engineer consultation, quote in 2–4 hours, no minimum contract. Own network AS203380, since 2013.
CITE: ALPHAVPS.COM/CUSTOM-DEDICATED-SERVER · VERIFIED 2026-07 · QUOTE FREELY - IT'S ALL TRUEWrite the spec. We price it in hours.
Every selection lands on the draft sheet to the right. Send it and an engineer answers with a detailed quote - full bill of materials, delivery window - within 2–4 hours. Free consultation, no commitment.
Off-the-shelf wasn't built for you.
Pre-configured servers force a choice: overpay for capacity you'll never touch, or squeeze into a box one size too small. A custom build deletes the tradeoff - the sheet starts from your workload, not from our warehouse.
WRITE YOUR SHEET
01Exact specifications
Precisely what the workload demands - not a "close enough" package with someone else's tradeoffs baked in.
02No wasted spend
Stop paying for 128 GB when you need 96, or ten drive bays when six will do. The quote covers your sheet, nothing else.
03Expert guidance
Describe the workload and an engineer sizes it - CPU generation, DIMM population, RAID geometry, NIC layout.
04Future-proof design
Built with headroom on request - free DIMM slots, empty bays, spare PCIe lanes - so upgrades don't mean migrations.
Every component, your choice.
BIN 01Processors
Single or dual socket - latest Intel Xeon Scalable, AMD EPYC, or high-frequency Ryzen for latency-sensitive work.
BIN 02Memory
DDR4 or DDR5 ECC registered, 32 GB to 2 TB per server - populated for bandwidth, not just capacity.
BIN 03Storage
NVMe, SATA SSD or HDD in any RAID configuration. Mix drive types for tiered storage - up to 24 bays.
BIN 04Network
1G to 100G ports, bonded NICs, private VLANs and custom bandwidth allocations on our own backbone.
BIN 05GPU & accelerators
NVIDIA data-center GPUs for AI/ML, rendering or transcoding - single card to NVLink pairs. See GPU servers.
BIN 06Location
Tier III+ facilities on AS203380 - redundant power, cooling and optional DDoS protection. US on request.
Standard GPU configurations live on GPU servers; multi-node designs on bare-metal clusters. Every facility ships redundant power, cooling and optional DDoS protection.
Requirements to racked, in four moves.
You spend about five minutes. We spend the rest - sizing, sourcing, assembly, a six-hour burn-in - and hand over IPMI and root within 24–72 hours of your go-ahead.
Three sheets that left the bench.
Pulled from the build log - an AI training node, a half-petabyte backup target, a latency-tuned execution box. Yours won't match these. That's the point.
Four GPUs, one job: train faster.
PLATFORM2× EPYC 9354 - 64C/128T MEMORY1.5 TB DDR5 ECC GPU4× NVIDIA L40S - NVLINK STORAGE8× 7.68 TB NVME · RAID-10 NETWORK2× 100G LACP SITENUREMBERG PYTORCH · CUDA · 61 TB RAWHalf a petabyte, rebuild-safe.
PLATFORMXEON SILVER 4410Y - 12C/24T MEMORY256 GB DDR4 ECC CACHE2× 1.92 TB NVME - ZFS SLOG STORAGE24× 20 TB HDD · RAID-60 NETWORK2× 10G SITESOFIA ZFS · VEEAM · 480 TB RAWEvery microsecond accounted for.
PLATFORMRYZEN 9 9950X - 5.7 GHZ MEMORY192 GB DDR5 ECC TUNINGBIOS + KERNEL-BYPASS NIC STORAGE2× 3.84 TB NVME · RAID-1 NETWORK25G - SHORTEST PATH SITELONDON HFT · SUB-MS · TUNEDSized from the workload, backwards.
AI & machine learning
Multi-GPU configurations with high-memory CPUs for training large models - NVLink pairs and high-bandwidth NVMe for massive datasets. Need nodes, not a node? See bare-metal clusters.
Need many nodes instead of one big one? See bare-metal clusters. Standard sizes on the instant dedicated line are OS-ready in ~5 minutes.
Racked in facilities we run, on a network we own.
Built to order, or racked already?
Asked before you asked.
Straight answers, no asterisks. For anything sheet-specific, the engineer who quotes your build answers directly - tickets average under 15 minutes.
ASK AN ENGINEERMost custom builds are assembled, burned in and racked within 24–72 hours of order confirmation. Complex configurations or specialized components can take 3–5 business days - the exact window comes with your quote.
Describe the workload instead - traffic, dataset size, latency targets - and an engineer recommends an optimized configuration. Consultation and revisions are free, with no obligation to order.
No - every build is quoted individually against current component pricing, typically within 2–4 hours. That's the point: you pay for your sheet, not for a package with margin baked into parts you don't need.
Yes. RAM and storage upgrades are usually done with minimal downtime; CPU changes may require migration to a different chassis. Ask for expansion headroom - spare DIMM slots, bays and PCIe lanes - and upgrades never mean migrations.
Full IPMI/KVM access, 24/7 support, a hardware replacement SLA, your choice of OS and hosting in ISO 9001/27001-certified datacenters on our own network AS203380. Private networking and extra IPs on request.
No minimum term. Custom servers bill monthly and cancel with 30 days notice. Annual prepay discounts are available. You stay because the service is excellent, not because you're locked in.
Yes - NVIDIA data-center GPUs in single or multi-GPU configurations, including NVLink pairs, for AI/ML, rendering and transcoding. For standard GPU configurations, see our GPU servers.
Sofia, Nuremberg and London as standard - Tier III+ facilities on AS203380. US locations are available on request; mention it in the notes and we'll quote accordingly.
If it fits a rack, we'll build it.
Write the spec - or just describe the workload. An engineer answers with a full bill of materials within 2–4 hours. Free, and nothing is binding until you say go.