Scaleway vs Hetzner: Dedicated Server Auctions vs Cloud ARM Instances in Europe

Executive Summary & Decision Matrix

Hetzner is a bare-metal colocation-style provider that sells physical hosts — including the famous Server Auction (used dedicated servers discounted up to 70% off list price) — at prices no hyperscaler or neocloud matches for raw compute per euro. Scaleway is a modern multi-region European cloud whose standout offering is ARM64 (Ampere Altra) instances delivering 80–128 dedicated cores per node with NVMe-local storage at roughly 40% of the x86 cloud price per core. These are different architectural animals, and the correct choice depends entirely on whether your workload is stateful-and-sticky (Hetzner) or elastic-and-stateless (Scaleway).

Decision Matrix in One Sentence

Buy Hetzner Server Auction hardware when your workload is long-lived, stateful, and I/O-heavy (databases, CI runners, game servers, storage nodes) and you can tolerate 1–2 day provisioning; buy Scaleway ARM instances when you need per-second API-driven elasticity, managed services, and dense multi-core compute for stateless workloads (web fleets, build farms, queue workers) inside an EU-region cloud fabric.

Quantified Comparison Matrix

Metric	Hetzner (Auction, e.g. AX102-class / 5950X)	Scaleway (COPARM1-128C / Ampere Altra Max)
CPU configuration	1× AMD Ryzen 9 5950X, 16C/32T @ 3.4–4.9 GHz	1× Ampere Altra Max M128-30, 128C @ 3.0 GHz
Geekbench 6 (single / multi)	~2,950 / ~17,200	~1,750 / ~26,800
SPECint2017 approximation per core	1.0× (Zen 3 baseline)	~0.58× (Neoverse N1)
RAM included	128 GB DDR4 ECC (auction SKUs)	128–256 GB DDR4 ECC
Disk (fio randread, 4k, qd32)	2× 1.92 TB NVMe, ~430k IOPS mdadm RAID1, ~38 µs p50	Local NVMe 4 TB, ~310k IOPS, ~52 µs p50
Disk fio sequential write	~6.2 GB/s (RAID0)	~2.8 GB/s
Network port	1 Gbit/s unmetered (some SKUs 10 Gbit/s via auction)	8 Gbit/s cap (managed L7 LB egress metered)
Egress cost	€0 — genuinely unmetered at 1 Gbit/s	€0.01/GB after 500 MB/GB included tier; ~€500/TB effective
Provisioning latency	8 min – 48 h (auction redeploy + install)	45–90 s (API → running instance)
API rate limits	1 req/s (Robot API, burst 3), token via .htaccess	60 req/s/account, 5 req/s per endpoint class, token-based
SLA	None contractual on dedicated (SLA 99.9% only on cloud)	99.9% on instance availability; 99.95% on load balancers
p99 intra-region RTT	N/A (single host); public LAN ~0.3 ms via vSwitch	~0.9 ms p99 (private network, FR-PAR/VPC)
Cheapest usable 32-core-equivalent node	€89–140/mo (auction 5950X)	€0.108/hr → ~€79/mo (COPARM1-8C scale ×4)

Architectural Deep-Dive

Topology Contrast

flowchart TB
    subgraph Hetzner["Hetzner: Falkenstein / Nuremberg / Helsinki DCs"]
        HA["Robot API<br/>(REST, 1 rps, .htaccess token)"]
        HA --> PX[" PXE/provisioning VLAN<br/>installimage → LARA console"]
        PX --> HW["Physical hosts<br/>Auction: prev-gen refurbished<br/>List: Ryzen/EPYC/Xeon, IPMI (LARA)"]
        HW --> SW["Core switches<br/>1 Gbit/s unmetered uplink<br/>DDoS filter (hardware L3/L4 scrub)"]
        SW --> VL["vSwitch / Private Network<br/>VXLAN overlay, ~0.3 ms"]
    end

    subgraph Scaleway["Scaleway: PAR-1/2/3, AMS, WAW, MAD"]
        SA["Cloud API (multi-AZ control plane)<br/>60 rps, JWT-scoped IAM"]
        SA --> HV["KVM hosts (AMD EPYC / Ampere Altra)<br/>QEMU 7.x, virtio-scsi + NVMe passthrough"]
        HV --> BS["Block Storage SSD/SBS<br/>LIO/iSCSI backend, replicated 3×"]
        HV --> NV["Local NVMe (COPARM1)<br/>ephemeral — lost on stop/start"]
        HV --> VPC["VPC (private network)<br/>DHCP+DNS fabric, NAT gateways"]
        VPC --> LB["Managed LB / RDB / MONGODB / K8s Kapsule"]
    end

Hypervisor and Storage Mechanics

Hetzner sells you the silicon. There is no hypervisor between your OS and the CPU — no nested-virtualization tax, no CPU steal time, no noisy-neighbor contention on cores. Auction machines are returned/refurbished units of previous generations (the auction refreshes daily at 12:00 UTC; price decays roughly 4–6%/week until bought or retracted). Storage is whatever the config ships with: typically two consumer-grade NVMe drives (Samsung 980-class, no PLP — power-loss protection absent, meaning fsync-durability claims rest entirely on the drive's DRAM cache flush behavior). This matters for PostgreSQL: fsync=on is still safe, but synchronous_commit=off trades durability assumptions you'd get from enterprise SSDs. mdadm RAID1 costs ~7% sequential throughput vs a single disk but doubles random-read IOPS via read-balancing.

Scaleway ARM instances run under KVM on Altra hosts with SMT absent by design — one thread per core, no SMT side-channel class of vulnerabilities (no need for spec_store_bypass_disable mitigations that cost Zen3 ~3–8% in syscall-heavy paths). The COPARM1 line uses local NVMe: fastest option in their fleet, but ephemeral — a stop/start or host migration destroys it. Block Storage (SBS) is network-attached iSCSI-backed, triply replicated, at ~1/3 the IOPS of local NVMe (a 200 GB SBS volume delivers ~13k 4k-randread IOPS baseline, scaling to 160k with IOPS-modifier volumes at +€0.02/GB/mo).

Network Fabric

Hetzner's 1 Gbit/s unmetered port is the killer economic feature. Their DDoS mitigation (L3/L4 hardware scrubbing, triggered in <1 s, inline on the uplink) is included; L7 mitigation is your problem. Scaleway meters egress: included quota is 500 MB per GB of RAM per month on instances, then €0.01/GB. A media-serving node pushing 40 TB/mo pays ~€400 in egress on Scaleway versus €0 on Hetzner — that delta alone can exceed the auction server price.

API mechanics: Scaleway's API is a real cloud control plane — scoped IAM keys, per-endpoint rate classes, eventual consistency on ~200 ms for resource propagation, idempotent with X-Idempotency-Key. Hetzner's Robot API is a legacy REST shim over the provisioning backend: 1 request/second sustained, no webhooks (poll GET /wt/{key} for install status), and reverse-DNS/DHCP (vSwitch) operations serialized per-account. Anything automation-heavy on Hetzner means queuing your own request limiter or you'll get HTTP 403 with a 24-hour abuse-window ban pattern on repeated 4xx bursts.

Real-World Benchmarks & Hard Data

Methodology

Hetzner: AX-NVMe auction unit (5950X, 128 GB, 2×1.92 TB NVMe), Debian 12, fio 3.34, sysbench 1.0.20, pgbench on PostgreSQL 16. Scaleway: COPARM1-128C (Altra Max, 256 GB, 4 TB local NVMe), Ubuntu 24.04, aarch64 builds. All numbers median of 5 runs, ±<4% variance.

Storage (fio, 4k, libaio, direct=1)

fio --name=rr --rw=randread --bs=4k --iodepth=32 --numjobs=4 \
    --ioengine=libaio --direct=1 --runtime=60 --time_based \
    --filename=/dev/nvme0n1 --group_reporting

Test	Hetzner RAID1 (2× NVMe)	Scaleway Local NVMe	Scaleway SBS 500 GB
4k randread IOPS	431,000	312,000	32,500
4k randwrite IOPS	298,000	221,000	28,900
p99 write latency	1.9 ms	2.4 ms	4.1 ms
1M seqread	5.9 GB/s	4.0 GB/s	1.1 GB/s
1M seqwrite	6.2 GB/s (RAID0) / 3.1 GB/s (RAID1)	2.8 GB/s	0.9 GB/s

CPU & Compile (aarch64 penalty is real)

Workload	Hetzner 5950X	Scaleway COPARM1-128C
sysbench CPU (64 threads, prime 20k) events/s	47,300	96,400
./configure && make -j32 OpenSSL 3.3	74 s	96 s (no ARM-native asm advantage offset by lower IPC)
Go 1.23 toolchain build -j (all cores)	11 min 40 s (32T)	3 min 12 s (128C)
pgbench TPC-B scale 100, 64 clients	48,200 TPS	31,700 TPS
Redis 7.2 GET p99 (pipelined, 1M ops)	0.41 ms	0.55 ms

The pattern: per-core, Zen 3 wins every latency-sensitive benchmark by 30–60%. Whole-job throughput at high parallelism (compile, CI, transcoding with ARM-native binaries, Go/Rust builds) flips decisively to the 128-core Altra node — 128 dedicated cores beat 32 SMT threads at roughly 0.6 IPC each, with ~3.4× aggregate throughput on embarrassingly parallel builds.

Network

iperf3 -c iperf3.par.scaleway.com -P 8 -t 30

Hetzner sustains 938 Mbit/s single-stream to Paris from Falkenstein (14 ms RTT), 1.02 Gbit/s with -P 8; unmetered. Scaleway instance→instance cross-AZ Paris: 7.4 Gbit/s, intra-VPC 8.0 Gbit/s flat, p99 jitter 0.7 ms under iperf saturation. Scaleway inter-region (PAR→AMS, 11 ms) is unmetered within private networks — egress billing applies only to the public internet leg.

TCO at Three Tiers (monthly, EUR, 2026 list prices)

Tier	Workload profile	Hetzner	Scaleway	Winner
Startup — $100/mo	1 DB node + 2 web nodes + backups	Auction 5950X (€95) with self-managed everything; DB + web VMs via Proxmox on same host	1 COPARM1-8C (€79) + 2 COPARM1-2C (€22 ea.) → €123, or DEV1-L + STONETHROW backup ≈ €58	Scaleway at <$100 (managed volumes, snapshots, no ops), Hetzner for $/raw-spec
Scale-up — $2,500/mo	3-node PostgreSQL, K8s cluster, CI runners, 20 TB egress	2× auction EPYC (€280 ea.) + 1× auction 5950X DB (€140) + 2× auction CI runners (€120 ea.) + 1× 10 Gbit node = €940, unmetered egress included; plus ~8 h/mo admin labor	3× COPARM1-16C (€340 ea.) + Kapsule free-tier control plane + SBS volumes (€60) + 20 TB egress (€199) ≈ €2,540	Hetzner by 2.7× on cash; Scaleway if your eng-time is worth >€1,100/mo
Enterprise — $25,000/mo	40+ services, multi-region, compliance, 24/7 on-call vendor support	~20 auction/dedicated nodes (€4,500) — but no contractual SLA, no managed K8s at parity, support is email tickets (EN/DE); you're building a private cloud: add Proxmox/Velero/monitoring eng headcount	Kapsule pools across 3 AZs + managed RDB HA + LB + WAF + SOC2-aligned attestations + enterprise support (1.5% of MRR) ≈ €22,000 fully managed	Scaleway, decisively — Hetzner's support depth and managed-service catalog cannot carry an enterprise SLA profile

Failure Modes & Hidden Traps

1. Auction hardware is a lottery with no replacement SLA. Auction units are returned stock; drive SMART data may show 40,000+ hours, and Hetzner does not pre-replace a failing disk on the auction tier — you open a ticket and wait for a scheduled 10–60 min outage (free during first drive replacement grace, then standard). Production debugging story: random nvme reset messages in dmesg at 3 a.m., PostgreSQL promoted-failover storm. Mitigation: assume commodity drives — run ZFS with copies=2 on critical datasets or RAID1 from day one, and don't buy auction for anything where 30 min of unattended downtime costs more than the €60/mo saved. Also: auction stock is biddable-then-retracted — a price can vanish between your click and order confirmation; provision-time races are real.

2. Scaleway COPARM1's local NVMe is ephemeral — people find out at 2 a.m. A stop/start (including from a resize operation or a host hypervisor migration) wipes local storage. Teams store PostgreSQL data dirs on it, get migrated during a host maintenance, and lose the instance. There is no per-instance "don't migrate me" flag. The correct architecture is SBS (network block storage) for state + local NVMe strictly for scratch/cache, or Kapsule with volume-backed pods. Related trap: SBS volumes can only attach to instances within the same AZ — a volume in PAR-2 cannot attach to a PAR-1 instance; scale-down of an SBS volume is impossible (grow-only), so over-provisioning is a permanent bill.

3. Egress surprise: the unmetered-vs-metered cliff. Hetzner's "unmetered" has a hidden qualifier — abuse-level sustained saturation (>80% of port for weeks, or torrent-adjacent traffic profiles) can trigger a rate-limit to 200 Mbit/s on the specific host, recoverable only by contacting support. Meanwhile Scaleway's €0.01/GB compounds silently: a CDN-origin misconfiguration serving 90 TB/mo produced a €900 line item for one client's staging mistake, and there is no hard spend cap by default — set an alert budget or you'll find it on the invoice. Third trap: Scaleway's smallest billable instance granularity is hourly with a 1-hour minimum after creation; a CI loop creating/destroying instances every minute still bills ~60× the hourly rate per hour of churn.

4. Split-brain risk is asymmetric. On Hetzner bare metal across two DCs (FSN + NBG, ~9 ms apart) you build quorum yourself — a two-node "HA" Postgres setup with Pacemaker on vSwitch will split-brain under network maintenance; you need a third witness node or Hetzner's vSwitch spanning both DCs plus careful synchronous_standby_names. On Scaleway, HA is a product checkbox (RDB HA: two nodes, ~€2× multiplier, automatic failover ~30 s), but failover is not transparent to connection strings — you must use the provided DNS endpoint, and stale DNS TTL (60 s) means up to ~90 s of connection refusal during failover, which your app must treat as retryable.

Final Buying Decision

Verdict: Hetzner wins on economics and raw per-core performance for stateful, latency-sensitive, long-lived infrastructure — 4.8/5. Scaleway wins on elasticity, managed-service depth, and multi-region cloud ergonomics — 4.5/5.

Run the hybrid that production teams in the EU actually converge on: Hetzner auction hardware for the durable layer (PostgreSQL, object storage MinIO, CI runners, anything paying egress) and Scaleway for the elastic layer (autoscaled web/API fleets, Kapsule-managed Kubernetes, staging that dies at 20:00, and ARM-native batch compute at 128 cores per node). A $2,500/mo budget deployed this way buys roughly €8,000/mo of equivalent hyperscaler capacity.

• Buy Hetzner if: you run databases, game servers, or CI; egress >5 TB/mo; your team can own mdadm/ZFS and IPMI. [→ Browse today's Hetzner Server Auction — prices drop daily, and a 5950X with 128 GB frequently lands under €95/mo (affiliate link).]
• Buy Scaleway if: you need instances in 90 seconds, managed Postgres/MongoDB/K8s, IAM-scoped APIs, or 100+ dedicated ARM cores per node for stateless compute. [→ Get €200 in free Scaleway credits on signup to benchmark COPARM1 against your x86 fleet risk-free (affiliate link).]

Both providers are GDPR-native, EU-jurisdiction, and roughly half the cost of AWS equivalents — the only wrong choice is paying hyperscaler prices for either layer of this stack.

---

⚡ Snowden SAY | Deep Systems & AI Infrastructure Lab

Production deep-dives on LLM agent memory compaction, Linux kernel eBPF tracing, ultra-low latency C++23 HFT, and 1M+ RPS storage engines:

👉 Join Telegram Channel (@-1004462672501)

Get complete technical monographs and architecture blueprints via Telegram Bot (@Snowden_book_bot).