AMD Turin Destroys 44 Cloud VMs — Hetzner Costs 80% Less Than AWS
AMD Turin Destroys 44 Cloud VMs — Hetzner Costs 80% Less Than AWSHonestly, someone finally did the homework nobody else wants to do — and the results are brutal for your AWS bill.
7 cloud providers. 44 VM types. 3 CPU architectures. 1 clear winner your boss won’t let you use.
A solo engineer just benchmarked every major cloud provider head-to-head across DKbench, Geekbench, Phoronix, and FFmpeg tests — on Debian 13 and Ubuntu 24.04, across multiple regions. The price/performance gaps are… not subtle.
🧩 Dumb Mode Dictionary
| Term | Translation |
|---|---|
| vCPU | A virtual CPU core — but not all vCPUs are equal (some share physical cores, some don’t) |
| SMT | Simultaneous Multi-Threading — cramming two threads onto one core. Sometimes great, sometimes a lie |
| Turin / EPYC | AMD’s newest server CPU. The one making Intel sweat |
| Granite Rapids | Intel’s latest server chip. Trying very hard |
| ARM (Axion/Cobalt/Graviton) | Non-Intel/AMD chips. Each cloud provider has their own flavor now |
| Spot/Preemptible | Cheap VMs that can be killed anytime. Like an apartment with a 0-day eviction clause |
| On-demand pricing | Full retail. The price you pay when you don’t plan ahead |
| FinOps | The art of not going bankrupt running cloud servers |
📖 Backstory: Why This Benchmark Exists
Most cloud benchmarks are marketing garbage — sponsored by the provider, testing one workload, ignoring price. This one’s different. A solo dev ran DKbench (custom suite), Geekbench 5, Phoronix (7-zip, nginx, OpenSSL RSA 4096-bit with AVX512), and FFmpeg/libx264 transcoding across all seven providers.
44 VM types. Three x86 families plus ARM. Multiple regions to catch “noisy neighbor” variance. No sponsorship. No affiliate links. Just raw numbers.
The new CPUs tested include AMD EPYC Turin, Intel Granite Rapids, Google Axion, Azure Cobalt 100, and Ampere AmpereOne M. And honestly, the gap between the best and worst value is so wide you could park a data center in it.
🏆 Performance Rankings: Who's Actually Fastest
Single-Thread King: AMD EPYC Turin. The author literally wrote this is “the first time in my series where a CPU had this clear of a performance lead.” AWS’s C8a instance topped the charts.
Multi-Thread Beast: Non-SMT Turin instances (again, AWS C8a) dominated. Google Axion — their custom ARM chip — matched previous-gen leaders but couldn’t touch Turin.
Scalability: ARM systems hit near-100% scaling efficiency. AMD generally beat Intel on SMT. Intel’s Emerald Rapids on GCP showed inconsistent performance thanks to “boost behavior + node contention.” Translation: noisy neighbors are real and Intel suffers more from them.
One weird anomaly: an Akamai (Linode) Turin instance showed 71.9% scalability with lower single-thread scores. The author verified it multiple times. Still unexplained. Spooky.
💰 Price/Performance: Where Your Money Actually Goes
On-Demand Value (Single-Thread):
| Rank | Provider | Notes |
|---|---|---|
| 1 |  Hetzner |
Prices that seem fake. They’re not |
| 2 | Oracle Cloud | Free tier is absurd (4 vCPU ARM, permanently free) |
| 3 | Akamai/Linode | Solid but verify your CPU generation |
| 4 | GCP n4d | Turin at Google scale |
| … | AWS | Dead last. Fastest chips, highest bills |
On-Demand Value (Multi-Thread):
| Rank | Provider | Notes |
|---|---|---|
| 1 | Oracle ARM | AmpereOne M is a beast for the price |
| 2 | Hetzner |
Again |
| 3 | Google Axion | Custom ARM doing work |
| 4 | Azure Cobalt 100 | Microsoft’s sleeper hit |
AWS ranked last in on-demand value. Again. Like it does every year. But okay, nobody’s paying on-demand at scale, right? Right?
📊 Reserved & Spot: The Real Pricing Game
Here’s where it gets interesting.
1-Year Reserved: GCP’s Turin matched Oracle’s value. That’s notable — Oracle has always been the “cheap but weird” option.
3-Year Reserved: Azure’s Cobalt 100 topped the multi-thread price/performance charts. The author called it “surprising.” Honestly, I’d call it shocking. Azure winning a value competition is like seeing a unicorn at a DMV.
Spot/Preemptible VMs: This is where the real deals live. Oracle does a flat 50% discount (predictable, boring, good). GCP and Azure offer variable discounts that can get deep — roughly 2x the performance per dollar compared to 3-year reservations. If your workload can handle interruptions, spot is the play.
🚨 The Gotchas Nobody Mentions
- DigitalOcean’s fleet hasn’t changed in ages. Older CPU generations showing oversubscription. You’re paying more for less.
- Linode shared-CPU instances might give you Milan or might not. You need to verify. Roll the dice.
- Hetzner shared-core instances are EU-only with limited availability. Good luck getting one when you need it.
- “vCPU” means different things. An AWS vCPU, a Hetzner vCPU, and an Oracle vCPU are not the same thing. Stop comparing them 1:1.
- Old CPU generations cost MORE and perform WORSE. Providers don’t always drop prices on last-gen hardware. Check what silicon you’re actually getting.
🗣️ Community Reactions
The discussion has been… predictable in the best way:
- Hetzner fans: “We’ve been telling you for years.” They’re not wrong.
- AWS defenders: “But the ecosystem! The integrations! The 47 different managed services!” Also not wrong, but that’s not what this benchmark measures.
- Oracle skeptics: “Yeah but have you tried their console?” Fair point. The free tier is insane but the UX is a war crime.
- Self-hosters: Already forwarding this to their bosses with “SEE?” highlighted in yellow.
- FinOps teams: Quietly updating their spreadsheets and booking meetings nobody asked for.
Cool. So Cloud Providers Have Been Overcharging You. Now What the Hell Do We Do? ಠ_ಠ
💰 Hustle #1: Cloud Cost Optimization Consulting
Companies waste an estimated $44.5 billion per year on unused or oversized cloud resources. If you can read a benchmark and a billing dashboard, you can save companies 30-50% on their cloud spend — and take a cut.
Example: Germán Neironi, a solo full-stack engineer in Argentina, built CloudPruneAI — an AWS cost optimization tool using Python/FastAPI and Claude API that scans accounts for waste and generates infrastructure-as-code fixes. Running cost: $27/month.
Timeline: 2-3 weeks to learn AWS Cost Explorer + Trusted Advisor. Start with small businesses overpaying on RDS and EC2. Charge 20% of first-year savings.
📊 Hustle #2: Build a Server Comparison Site (Affiliate Revenue)
Benchmark data is gold when paired with affiliate links. Every cloud provider has a referral program ($50-200 per signup). Build a comparison tool, drive SEO traffic, collect checks.
Example: Gergely Daróczi’s team in Hungary built Spare Cores — a free benchmarking database tracking 2,000+ server types across 130 regions with 275,000 pricing points. Secured €150K/year in EU grant funding and is now building a paid container deployment layer on top.
🔧 Hustle #3: Migration-as-a-Service (Move Companies Off AWS)
This benchmark proves companies are overpaying by 2-5x on AWS. Offer to migrate them to Hetzner, Oracle, or GCP — and pocket a project fee plus ongoing management.
Example: A senior DevOps consultant in Pakistan runs cloud migration projects (on-prem to AWS/Azure, or AWS-to-cheaper-provider) through Upwork and direct outreach. Using Terraform and Kubernetes, they handle 2-3 clients at $100-125/hour — pulling $10K-16K/month from LATAM and EU clients.
🛠️ Hustle #4: Managed Hosting Reseller (Buy Cheap, Sell Managed)
Buy Hetzner VPS at $3-5/month. Add automated backups, monitoring, managed WordPress/cPanel. Sell for $25-50/month. The margin is absurd because your customers are paying for support, not compute.
Example: A full-stack engineer in India resells DigitalOcean droplets as managed WordPress hosting to small Indian businesses. Uses WHMCS for billing, offers WhatsApp support. 60+ clients at $30-50/month each — $2,500-4,000/month with near-zero support overhead thanks to automation.
💡 Hustle #5: Cloud Arbitrage for AI Workloads
GPU cloud pricing is even more insane than CPU pricing. Buy cheap compute from Hetzner/OCI, package it with a simple API layer, resell to AI startups who don’t want to manage infrastructure.
Example: An infrastructure engineer in Mexico buys Hetzner dedicated servers and Oracle ARM instances, packages them with a managed Kubernetes layer using xCloud, and sells to LATAM startups at 3-5x markup. $3K-7K/month from ~50 customers who’d rather pay more for “it just works.”
🛠️ Follow-Up Actions
| Step | Action | Tool/Resource |
|---|---|---|
| 1 | Read the full benchmark | devblog.ecuadors.net |
| 2 | Audit your own cloud spend | AWS Cost Explorer / GCP Billing / Azure Cost Management |
| 3 | Spin up a Hetzner VPS and test your workload | Hetzner Cloud Console (€3.29/month for 2 vCPU) |
| 4 | Grab Oracle’s free tier | 4 vCPU ARM instance, permanently free. Not a trial |
| 5 | Learn Terraform basics | HashiCorp Learn (free) — makes multi-cloud migration trivial |
| 6 | Check spot pricing for your region | spot.io or provider-native dashboards |
Quick Hits
| Want… | Do… |
|---|---|
 Raw single-thread speed |
AWS C8a (AMD Turin, non-SMT) |
 Best bang-for-buck (on-demand) |
Hetzner or Oracle ARM free tier |
 Best reserved value (3-year) |
Azure Cobalt 100 (yes, really) |
 Maximum savings if you can handle interruptions |
GCP/Azure spot instances (~2x value vs reserved) |
 Avoid at all costs |
Any provider still selling you last-gen CPUs at full price |
Honestly, the cloud pricing game is simple: everyone charges what they think you’re too lazy to compare. This benchmark is the spreadsheet your CFO never made.
!