Dedicated isn't dead, it just got narrower
The narrative that "everything's in the cloud" is a marketing simplification. There's a meaningful set of workloads where dedicated servers (rented hardware) and colocation (your hardware in someone's data centre) are still the most cost-effective and operationally sensible answer in 2026.When dedicated wins
- Sustained heavy CPU / GPU workloads — dedicated bare metal at $200-500/month outperforms equivalent cloud at 3-5x the cost over a year.
- Storage-heavy workloads — dedicated boxes with 8x NVMe drives at the cost of cloud "general purpose" SSD provisioned for the same IOPS.
- Network-heavy workloads — egress costs in the cloud are a tax. Dedicated servers with included bandwidth cap at predictable cost.
- Workloads with stable, predictable load — cloud's price advantage is elasticity. If you don't need elasticity, you're paying for it anyway.
- Workloads with strict data sovereignty / compliance — colocation in a specific country / facility is sometimes the only viable answer.
Dedicated server providers in 2026
- Hetzner Server Auction / Robot — German, EU-friendly, the cost leader. Used hardware at very low prices, new Ryzen/Epyc dedicated boxes at competitive rates.
- OVH — French, broad tier coverage, EU regions plus North America.
- Leaseweb / DataPacket / Worldstream — global, broader presence, mid-tier pricing.
- Local TR providers — for TR-presence requirements; quality varies, vet carefully.
Colocation: when it's the right call
Bringing your own hardware to a data centre — colocation — fits when:- You have hardware investment already (existing fleet, specialised hardware)
- Your hardware is custom (GPU farms, FPGA boards, specialised storage)
- You're large enough that dedicated rental margins matter
- Your compliance / control requirements specifically need "your" hardware
For most teams, colocation is more operational overhead than the savings justify. If you're considering it, model the total cost: hardware, networking, power, cooling, hands-and-eyes service, your staff time.
The hidden costs of dedicated
- Hardware failure — drives die, RAM faults, motherboards fail. The provider handles replacement, but downtime during the swap is real.
- No managed services — you run your own database, your own load balancer, your own everything. Cloud's RDS / managed services have no equivalent at the same price point.
- Geographic limitation — you have a server in a place. Multi-region is a separate dedicated server, not a click.
- No autoscaling — you sized for peak; if peak grows, you upgrade or migrate.
The hybrid architecture that works
For mid-size businesses, the pattern that fits:- Application servers on dedicated (high CPU, predictable load)
- Database on dedicated (controlled IOPS, no surprise costs)
- Object storage on cloud (S3-compatible, only pay for what you store)
- CDN on cloud (Cloudflare, BunnyCDN — globally distributed, hard to replicate)
- Backups to a different cloud / region (off-site)
This split tends to optimise total cost while keeping the pieces that benefit from cloud elasticity in the cloud.
Hardware specs that actually matter
For a typical heavy-duty dedicated:- CPU — Ryzen 7950X3D or Epyc for general-purpose; Xeon Scalable or Epyc for high-core-count.
- RAM — ECC. Always. The cost premium over non-ECC is small; the bug-class it eliminates is worth it.
- Storage — NVMe with proper endurance rating (TBW) for your workload. Mirror by default.
- Network — 1 Gbps included is standard; 10 Gbps available. Check the included bandwidth allowance.
- IPMI / out-of-band management — non-negotiable for any serious dedicated.
Operations: the things you now own
- Patching, including kernel reboots
- Hardware monitoring (S.M.A.R.T., RAID controller health, ECC error logs)
- Backup management (off-server)
- DR planning (single-server is single-point-of-failure)
- Capacity planning
This is the work cloud abstracts away. On dedicated, it's yours.
One pattern we'd warn about
Single-dedicated-server architectures for business-critical apps. The hardware will fail. Plan for it: hot standby, replication, fast restore from backup. "We have one big server" is the architecture that produces a multi-day outage.One pattern that always pays off
Out-of-band management. IPMI, iDRAC, iLO — whatever the provider gives you. The day the OS won't boot, the OOB is the only path back. Never deploy without it configured and access-controlled.What's your dedicated stack? And — for the Hetzner-server-auction folks — what's the secondary-market hardware that has been most reliable for you?
