Skip to content
Cloud Provider Benchmarks

Rankings

AI infrastructure providers 2026: GPUs, Kubernetes and storage for model serving

This ranking is about infrastructure for running your own models, not hosted model APIs. Managed AI platforms are listed on provider profiles but are not scored here.

Updated:

Direct answer

Of 6 eligible providers, ranked first: Amazon Web Services, DigitalOcean, PortvMind Cloud. Order: 1. AWS · 1. DigitalOcean · 1. PortvMind · 4. Google Cloud · 5. Vultr · 6. Azure. Self-hosted model serving needs data-centre GPUs, an orchestration layer, storage for weights and datasets, and load balancing for inference endpoints. This index ranks tracked providers by how many of those they document, with GPU criteria weighted double.

Ranking

AI infrastructure providers 2026: GPUs, Kubernetes and storage for model serving
# Provider Documented coverage Unverified Notes
1 Amazon Web Services 100% 0
1 DigitalOcean 100% 0
1 PortvMind Cloud 100% 0
4 Google Cloud 80.8% 0
5 Vultr 76.9% 1
6 Microsoft Azure 69.2% 1

Not eligible

  • Hetzner — Missing required capability: NVIDIA H100 instances

Values by criterion

Criteria used in this ranking
Criterion AWS DigitalOcean PortvMind Google Cloud Vultr Azure Hetzner
Compute & GPU
Per-second or per-minute billing Standard compute billed at per-second or per-minute granularity rather than per started hour. Yes Per-second for Linux on-demand (60 s minimum). Yes Per-second (60 s / $0.01 minimum). Yes Per-minute billing. Yes Per-second with a one-minute minimum. No Hourly billing. Unverified No Per started hour with a monthly cap.
NVIDIA H100 instances On-demand or reservable instances with NVIDIA H100 GPUs. Yes Yes Yes Published prices for 1–4 GPU profiles. Yes Yes Yes No
NVIDIA H200 instances On-demand or reservable instances with NVIDIA H200 GPUs. Yes Yes Yes Listed option; price not in public table. Yes Unverified Yes No
NVIDIA L40S instances On-demand or reservable instances with NVIDIA L40S GPUs. Yes Yes Yes Listed option; price not in public table. No Yes No Azure Local only. No
Kubernetes
Managed Kubernetes Provider-operated Kubernetes control plane. Yes Yes Yes Yes Yes Yes No CCM and CSI driver for self-managed clusters only.
Kubernetes node autoscaling Worker node count scales automatically within configured limits. Yes Yes Yes Yes Yes Yes Partial Through the upstream cluster autoscaler on self-managed clusters.
Storage & data
Block storage volumes Persistent, attachable and expandable block volumes. Yes Yes Yes Yes Yes Yes Yes
S3-compatible object storage Object storage reachable through the S3 API or an S3-compatible API. Yes Native S3. Yes Spaces. Yes Partial XML API interoperability with HMAC keys; not full S3 API parity. Yes No Blob Storage uses its own API. Yes EU locations; partial S3 API coverage.
Networking
L7 load balancer (HTTP/HTTPS) Application-layer load balancing with HTTP health checks. Yes Yes Yes AppBalancer (HTTP/HTTPS). Yes Yes Yes Yes
Private networks / VPC Isolated tenant networks with custom CIDR ranges. Yes Yes Yes Isolated networks with custom CIDR. Yes Yes Yes Yes

Methodology

Each criterion earns 1 point for “yes”, 0.5 for “partial” and 0 otherwise, multiplied by its weight. Ties share a rank and are listed alphabetically. At least one of H100, H200 or L40S is required for eligibility.

  • NVIDIA H100 instances ×2
  • NVIDIA H200 instances ×2
  • NVIDIA L40S instances ×2
  • Managed Kubernetes
  • Kubernetes node autoscaling
  • S3-compatible object storage
  • Block storage volumes
  • L7 load balancer (HTTP/HTTPS)
  • Per-second or per-minute billing
  • Private networks / VPC