Cloud Cost Benchmark Report 2026: What Teams Pay

Company headcount is the strongest single predictor of cloud spend — stronger than funding stage, geography, or technology stack. A 50-person company spends roughly 18x more on cloud than a 10-person company, even with the same revenue (Flexera 2025, SpendArk data). The ranges below are compiled from Flexera 2025, CNCF annual survey data, and anonymized SpendArk user billing across AWS, Azure, and GCP.

1–10 employees ($100–$500/month): Most early-stage teams run a managed database, one or two compute instances, object storage, and a CDN. This tier is comfortably served by AWS free tier, GCP's Always Free products, or Azure's startup credits. Average waste at this size is lower than the industry mean — not because small teams are disciplined, but because there is less to forget.

11–50 employees ($500–$3,000/month): Engineering teams of this size typically add staging environments, monitoring stacks (Datadog, Grafana Cloud), CI/CD pipelines, and the first container orchestration. The bill doubles not because traffic doubled but because operational maturity costs money. This is where the first serious waste appears: dev databases that run 24/7, CI runners that never terminate, and load balancers serving zero traffic. These "invisible" costs are rarely reviewed because no single line item is large enough to trigger concern.

51–200 employees ($3,000–$15,000/month): Multi-service architectures, dedicated data warehouses, and the first compliance tooling arrive. According to the CNCF 2024 Annual Survey, 67% of organizations at this size run production Kubernetes, which introduces new cost surface area: node pool sizing, persistent volume waste, and idle namespace resources.

201–1,000 employees ($15,000–$80,000/month): At this scale, cloud spend is a board-level line item. Savings Plans and Reserved Instances become financially material. Organizations in this range that have not committed any Reserved Instances typically overpay by 30–40% compared to peers with equivalent workloads (Flexera 2025). FinOps becomes a dedicated function rather than a part-time responsibility.

For a per-employee breakdown of these numbers, see our cloud cost per user benchmarks.

SaaS companies spend 8–15% of revenue on cloud infrastructure; fintech companies spend 10–20%, driven largely by compliance tooling (Gartner Infrastructure Benchmark, 2025; Harness State of Cloud Cost, 2025). Raw dollar amounts only tell half the story. The ratio of cloud spend to revenue reveals whether infrastructure is scaling efficiently — or whether it grew faster than revenue. These ranges are compiled from Gartner infrastructure benchmarks, publicly available SaaS COGS disclosures, and Harness's State of Cloud Cost 2025 report.

SaaS (8–15% of revenue): Early-stage SaaS companies cluster near the top of this range. At $1M ARR, 15% means $150K/year on cloud — feasible but tight. Mature SaaS companies (post Series B) typically compress to 8–10% as revenue scales faster than infrastructure. Public SaaS companies with efficient architecture often report 4–6% gross margin impact from cloud (Gartner 2025 Infrastructure Benchmark).

E-commerce (3–8% of revenue): E-commerce benefits from relatively predictable traffic patterns and high revenue volume relative to compute. Most spend goes to CDN, image processing, and search. The low end (3%) is achievable for high-margin, lower-SKU catalog businesses. Marketplaces with real-time pricing or high-frequency inventory updates trend toward 8%.

Fintech (10–20% of revenue): The widest range of any vertical. Compliance requirements (PCI-DSS, SOC 2, encryption-at-rest across every datastore) add overhead that purely product-driven companies avoid. Real-time fraud detection and transaction processing at scale pushes the top of this range. Companies in this vertical that can 't explain why they're above 15% should audit their compliance tooling spend specifically — this is where the most common overprovisioning occurs.

Media and streaming (5–12% of revenue): Egress costs dominate this vertical. A streaming platform delivering HD video spends 30–40% of its cloud bill on data transfer alone. CDN optimization has the highest ROI of any cost reduction available to media companies. All three major providers reduced egress pricing in 2024–2025 following regulatory pressure — though internal transfer costs (cross-AZ, NAT Gateway) were largely unaffected (see the trends section below).

Compute accounts for 35% of all wasted cloud dollars — the single largest category, driven by overprovisioned instance sizes chosen at launch and never revisited (Flexera 2025, Harness 2025). The 27% overall waste figure is an average across all resource types. The distribution matters more than the headline number. Some categories waste far more than others. The highest-waste categories are also the most expensive, which means fixing them first compounds savings.

Compute: 35% of wasted dollars

Overprovisioned instances are the primary culprit. Engineers select instance types at launch based on worst-case projections, and those selections rarely get revisited. The average EC2 instance runs at 6–8% CPU utilization (AWS, 2024). An m5.2xlarge at $0.384/hour running at 7% CPU costs roughly 14x what an equivalently utilized t3.medium would cost. That gap is pure margin left on the table. Most teams have three to five instances matching this profile right now.

Kubernetes amplifies this problem. Resource requests that are set too high cause nodes to appear full before they are, leading teams to add nodes rather than rightsize requests. The CNCF 2024 survey found that 53% of Kubernetes users identified resource optimization as their top operational challenge.

Storage: 25% of wasted dollars

Storage waste is the most invisible category. Unattached EBS volumes accumulate silently at $0.10/GB/month. Automated snapshots run for years after the underlying resource is deleted. S3 buckets grow without lifecycle policies. A company that has been running AWS for three years without a storage audit typically has 15–25% of its S3 storage in objects not accessed in over a year — which could move to Glacier Instant Retrieval at one-third the cost (AWS pricing, 2025). The audit takes two hours. Most teams never do it.

Data transfer: 20% of wasted dollars

Cross-AZ data transfer ($0.01/GB each direction) is consistently underestimated. A microservices architecture where services are spread across availability zones can generate thousands of dollars per month in inter-AZ traffic that appears as "EC2-Other" on the AWS bill. NAT Gateway processing charges ($0.045/GB) compound this: every byte routed through a NAT Gateway to reach an AWS service that has a VPC endpoint available is pure waste.

Unused/forgotten resources: 15% of wasted dollars

Unused resources are the most embarrassing category — these are services that no longer do anything. Load balancers running in front of terminated instances. RDS clusters spun up for a feature that shipped six months ago. Elastic IPs that cost $3.65/month each since AWS began charging for them in 2024. A quarterly cleanup audit typically recovers 8–12% of monthly spend from this category alone.

The median cloud cost per employee across industries is $380/month for SaaS, $620/month for fintech, and $750/month for media and streaming companies (Gartner 2025 Infrastructure Benchmark, Harness 2025). Cost per employee — total monthly cloud spend divided by headcount — is the most practical normalizing metric for comparing organizations of different sizes. It also surfaces structural problems fast: a 50-person company spending $1,200/employee/month on cloud should be asking hard questions.

Source: Compiled from Gartner 2025 Infrastructure Benchmark, Harness State of Cloud Cost 2025, and SpendArk anonymized user data. Ranges reflect 10th–90th percentile.

If your cost per employee exceeds the "High" benchmark for your industry, the most common causes are: no Reserved Instance or Savings Plan coverage, Kubernetes resource requests set to worst-case rather than actual usage, and storage that has never been audited. SpendArk's guide to cloud waste and overprovisioning covers the tactical remediation steps for each.

Not sure whether your spend is abnormal? Our guide on how much cloud should cost a startup and the State of Cloud Waste 2026 report provide additional context for benchmarking your bill.

AWS customers spend more on average at every company size tier — not because AWS is more expensive, but because AWS attracts a broader, less-optimized customer base including many early-stage teams still on on-demand pricing (Flexera 2025). Azure is strong in enterprise and Microsoft-ecosystem shops. GCP punches above its weight in data and ML workloads and applies sustained-use discounts automatically, which reduces waste for teams that have not adopted Reserved Instances.

AWS: The default choice for startups means AWS has the broadest customer base, including many early-stage companies that are not yet cost-optimized. Average spend per customer is the highest of the three providers, but this reflects customer mix as much as pricing. AWS's Reserved Instance and Savings Plan ecosystem is the most mature — companies that commit properly can reduce effective rates by 30–60% versus on-demand.

Azure: Enterprise and Microsoft-ecosystem companies skew Azure's numbers. Average Azure spend per customer at the enterprise tier is 10–15% lower than AWS equivalents, partly because Azure's hybrid benefit (BYOL for Windows and SQL Server licenses) provides meaningful discounts unavailable on AWS for Microsoft workloads.

GCP: Google Cloud attracts data-heavy and ML-heavy workloads. Average spend is lower than AWS and Azure, but per-workload cost for BigQuery and Vertex AI pipelines can exceed equivalent AWS costs. GCP's sustained use discounts apply automatically (no commitment required), which reduces waste for teams that have not adopted Reserved Instances. For a detailed price comparison, see our AWS vs Azure pricing breakdown.

Three structural shifts define the 2025-to-2026 transition. IPv4 charges are now a material budget line. AI/GPU spend grew 62% year-over-year in 2025 (Harness, 2025). Egress pricing declined modestly — but internal transfer costs were unaffected. Each shift affects your bill differently depending on workload type.

IPv4 address charges are now a real budget line

AWS began charging $0.005/hour ($3.65/month) for every public IPv4 address in February 2024. For a company running 50 EC2 instances and load balancers each with a public IP, that is $182/month that did not exist before. Multiply across elastic IPs, NAT gateways, and RDS instances with public accessibility enabled, and the number climbs quickly. A 2026 bill that looks 8–12% higher than a comparable 2022 baseline without any new resources often traces entirely to IPv4 charges. The fix is straightforward: audit public IPs, migrate internal services to private addressing, and enable IPv6 where possible.

AI and GPU workloads: the fastest-growing spend category

Harness reported that AI/GPU cloud spend grew 62% year-over-year in 2025. For organizations experimenting with LLM inference, fine-tuning, or embedding pipelines, GPU compute is often the single largest and most volatile bill item. An A100 instance on AWS (p4d.24xlarge) costs $32.77/hour. A model evaluation job left running overnight costs $786. The pattern is consistent across SpendArk users: GPU waste comes almost entirely from jobs that finished but whose instances were not terminated.

For teams running regular AI workloads, Spot instances reduce GPU costs by 60–90% for batch/offline tasks. For production inference, reserved capacity pricing has improved significantly on all three major providers in 2025–2026.

Egress costs are declining — but still significant

All three major cloud providers reduced or eliminated certain egress charges in 2024–2025 following regulatory pressure in the EU and competitive dynamics. AWS eliminated egress charges for customers switching providers (effective May 2024). GCP expanded its free egress tier. Azure lowered inter-region transfer costs by 30–50% for European customers.

Despite these reductions, data transfer remains 15–25% of the average cloud bill. The categories that declined (provider-to-internet egress for migrating customers) are not the main sources of waste. Cross-AZ traffic, NAT Gateway processing, and CDN origin pull costs were largely unaffected by the 2024–2025 pricing changes. For a detailed breakdown, see our cloud egress costs guide.

FinOps maturity is improving — slowly

82% of organizations now have a formal cloud cost management initiative, up from 68% in 2022 (Flexera, 2025). But 59% still describe their FinOps practice as "crawl" or "walk" maturity — meaning they have visibility but not optimization, or optimization without governance. Organizations that have reached "run" maturity report 35–45% lower waste rates, achieving effective waste of 14–18% versus the 27% mean. The gap between crawl and run is not tooling. It is process. Teams that review waste monthly achieve run-maturity results faster than teams that buy better dashboards without reviewing them.

This report compiles data from the following sources:

• Flexera 2025 State of the Cloud Report — 750+ IT decision-makers surveyed, primary source for waste percentages and FinOps maturity data.
• Gartner 2025 Infrastructure Benchmark — Cost-per-employee and revenue percentage benchmarks across verticals.
• CNCF Annual Survey 2024 — Kubernetes adoption rates and resource optimization challenges.
• Harness State of Cloud Cost 2025 — AI/GPU spend growth and workload-level cost trend data.
• SpendArk anonymized user data — Aggregated, anonymized billing data from SpendArk users across AWS, Azure, and GCP, covering companies from 1 to 500 employees. No individual company data is identifiable or disclosed.

Where ranges are provided rather than point estimates, they reflect the 25th to 75th percentile of observed values. Outliers (top and bottom 10%) are excluded from reported ranges to avoid skewing benchmarks with atypical workloads. All figures are in USD and reflect calendar year 2025 data unless otherwise noted.

This report is updated annually. If you have data to contribute or corrections to suggest, contact us at [email protected].