Cloud VM benchmarks 2026

Original Article ↗ Hacker News Discussion ↗

AMD CPUs and AI positioning

  • Commenters praise AMD’s EPYC line: Genoa was a big jump over Milan, and Turin is seen as an even larger leap, with very strong performance in cloud VMs and on-prem.
  • Some criticize AMD for “dropping the ball” on AI accelerators and software compared to NVIDIA’s ecosystem, arguing that software stack matters more than hardware specs.
  • Others counter that AMD does produce competitive AI chips and is landing major partnerships, suggesting the AI story isn’t as dire.
  • Several participants say they “don’t care” about accelerators as much as they care that AMD broke Intel’s CPU dominance and brought core counts and prices back to sanity.

Cloud vs self‑host / colocation economics

  • Many argue big cloud (especially AWS/GCP) is dramatically more expensive than owning/racking hardware or using dedicated servers/colo, particularly for steady workloads like CI or databases.
  • Counterarguments emphasize total cost of ownership: datacenter space, power, cooling, hardware failures, hands-on ops, provisioning delays, redundancy, and staff time.
  • Cloud is seen as paying for instant provisioning, autoscaling, multi-region, managed services, and powerful APIs; it shines for spiky/unpredictable load and fast iteration.
  • Opinions differ on when self-host becomes cheaper: some cite $1–3M/month cloud spend as the tipping point; others report big savings even at modest scale by moving to colo or smaller providers.
  • Hybrid approaches are suggested: run base load on owned/dedicated hardware and burst to cloud for peaks.

Oracle Cloud: cheap but distrusted

  • Several are wary due to Oracle’s reputation for aggressive licensing and lock‑in; one calls Oracle broadly “predatory.”
  • Some still use Oracle Cloud for very small, portable projects because compute can be extremely cheap, as long as no proprietary services (especially Oracle DBaaS) are used.
  • Reports of poor UX: confusing signup, free tier instances reclaimed for “idleness,” trials abruptly shut down, and inconsistent account handling.
  • A large migration case (hundreds of VMs) reports ~40% cost savings vs AWS/GCP by using only basic primitives (compute, storage, load balancers), free egress, flexible CPU/RAM combos, and flat regional pricing—but warns that many higher-level managed services are unreliable and the platform UX is “nightmarish.”

Hetzner, OVH, and other non‑hyperscale options

  • Hetzner is repeatedly highlighted as having outstanding performance per dollar, especially for dedicated servers; many say a cheap Hetzner box would top the benchmark charts.
  • Caveats: more initial work (no-frills bare metal, you handle DR, hardware failures, encryption), slower provisioning than cloud VMs, and at least one warning not to use Hetzner for “anything actually important” due to past incidents.
  • OVH is frequently mentioned as a similarly priced or cheaper alternative with strong dedicated offerings; experiences with support are mixed but some report very reliable use in production and testing.
  • Other providers surfaced include Vultr, HostHatch, netcup, and various “low-end” hosts; opinion on Vultr is split between “great for years” and “expensive with annoying limits.”
  • Several tools and comparison sites exist to navigate fragmented pricing, benchmark low-end hosts (e.g., via common scripts), and automate cross‑cloud/right‑sizing choices.

Benchmarking nuances and missing dimensions

  • Multiple comments stress that VM benchmarks measure not just CPU but hypervisor behavior, CPU pinning, storage backend, network virtualization, and noisy neighbors; these can shift results 20–30%.
  • vCPUs often don’t match bare-metal CPU capabilities or exposed feature sets; cloud providers sometimes mask features to enable live migration.
  • Some note that consumer/gaming CPUs (e.g., high‑end Ryzen/desktop chips) can significantly outperform many cloud VMs in single-thread and memory performance.
  • Participants want additional benchmarks: network throughput and egress pricing, SAN/non‑local storage performance, GPU training/inference workloads, and AWS burstable instance families (t4g vs competitors).

Debates on writing style and AI-generated content

  • A side thread critiques a popular “we moved from AWS to Hetzner” article as likely AI‑generated “slop,” citing repetitive marketing phrasing.
  • Others argue repetition is a normal rhetorical device and that dismissing content solely for being AI‑assisted is unhelpful; accuracy and effort should matter more.
  • Several express frustration that AI‑style, overlong, promotional writing wastes readers’ time and erodes diversity of style, even when technically correct.