Cloud Cost Benchmarks 2026: What Companies Actually Pay
Two companies of similar size can run cloud bills that look nothing alike, and both can be well managed. In Flexera's 2026 State of the Cloud survey of 753 respondents, 69% of small and midsize businesses were concentrated under $50,000 a month in public cloud spend, while 76% of large enterprises spent more than $5 million a month. Those numbers describe survey distributions, not invoices, and they say nothing about discounts, workloads or what each dollar buys. Meaningful benchmarking means separating three things that are usually blurred together: spending surveys, actual billing data and workload economics. This guide shows how to read each one, then turns them into a method you can apply to your own bill.
TL;DR
In Flexera's 2026 survey (753 respondents), 69% of SMBs sat under $50,000 per month and 76% of large enterprises exceeded $5 million per month. These are shares of respondents in spend bands, not averages.
No source reviewed establishes an average cloud bill, a universal cloud-to-revenue ratio or a typical 2026 invoice, so this guide does not publish one.
Respondents estimate that 29% of IaaS and PaaS spend is wasted, the first increase in five years. It is a self-reported estimate, not an audited figure.
Billed cost, amortized (effective) cost and list price can describe the same environment very differently. Choose one basis before comparing anything.
The most defensible benchmark is internal: cost per customer, transaction or AI task, tracked against peers with similar workloads and against your own history.
Quick answer: In 2026, no single average cloud bill exists. In Flexera's survey of 753 respondents, 69% of small and midsize businesses spent under $50,000 per month on public cloud, while 76% of large enterprises spent more than $5 million monthly. These are survey distributions, not audited averages; effective cost depends on discounts, workloads, and unit economics.
Table of Contents
Cloud Spending by Company Size: SMBs, Mid-Market Firms, and Enterprises
Workload-Level Benchmarks: Compute, Kubernetes, Databases, Storage, and Networking
Cloud Cost Unit Economics: The Benchmarks That Actually Matter
Cloud Procurement and Optimization: Where Companies Can Save
Benchmarking Methodology, Data Limitations, and Editorial Notes
Cloud Cost Benchmarks 2026 at a Glance
The strongest verified answer to “what do companies actually pay?” is a distribution, not a number. Flexera's 2026 survey of 753 cloud decision-makers, published March 18, 2026, found that spending rises steeply with organization size. The table converts its thresholds into annual equivalents and states what each figure does and does not measure.
Segment and source | Reported monthly spend | Annualized (calculation) | What it measures |
SMBs, Flexera 2026 | 69% concentrated under $50,000 | Under $600,000 | Share of SMB respondents in one band; not a mean |
Large enterprises, Flexera 2026 | 76% above $5 million | Above $60 million | Share of large-enterprise respondents; not a mean |
Enterprises, Flexera 2021 (context) | 36% above $1 million | Above $12 million | Different segments and thresholds; directional only |
SaaS spend, Flexera 2026 | Largest share at $200,001–$500,000 | About $2.4–$6.0 million | SaaS spending, not public cloud |
FinOps Foundation 2026 | 1,192 respondents representing more than $83 billion a year | Not applicable | Practitioner survey of practices, not typical bills |
Respondent distributions, not averages: every figure above is a share of survey respondents, not a mean, a median or an audited invoice.
What can a decision-maker conclude? Scale differs by orders of magnitude between segments, and most large-enterprise respondents operate in eight-figure annual territory. What cannot be concluded is whether any individual company is over- or under-spending. Dividing the FinOps Foundation's $83 billion by its 1,192 respondents gives roughly $70 million each (calculation), a figure that describes no real segment, which is exactly why this article does not publish an average.
One caution on a figure in circulation: Flexera's published findings attach the $200,001–$500,000 monthly band to SaaS spending, not public cloud, so it is not used here as a cloud benchmark. For market-level forecasts and provider revenue, see Articsledge's cloud spending statistics guide; this article focuses on the bill itself.
What Does “What Companies Actually Pay” Really Mean?
“Actually pay” has at least six meanings, and the same environment produces a different number under each:
List price: the published on-demand rate before any agreement or commitment.
Calculated estimate: a calculator output. The AWS public calculator builds estimates with On-Demand pricing, while the in-console version can include discounts and commitments. The Azure calculator draws on retail prices and can show agreement-specific rates after sign-in.
Gross usage cost: metered usage valued before reductions.
Contracted price: the rate after negotiated or enterprise terms.
Billed cost: what the invoice charges. AWS presents billing data on a cash basis, and the invoice is the final amount owed.
Effective (amortized) cost: upfront or recurring commitment fees spread across the usage they cover.
Adjustments complicate each view. AWS notes that the Bills page excludes refunds while Cost Explorer includes them, so the same month can show different totals depending on where it is read. Taxes, credits, support charges and marketplace purchases add further lines. Beyond the provider invoice sits total cost of ownership, covered in the section on forgotten costs.
Worked example: billed versus amortized cost
Suppose a company buys a one-year, all-upfront Savings Plan for $120,000 in September, the same structure AWS uses to explain amortization. Under each view the twelve-month total is identical, but the monthly story is not.
Month | Billed (cash) view | Amortized (effective) view |
September | $120,000 | $10,000 |
Each of the other 11 months | $0 | $10,000 |
12-month total | $120,000 | $120,000 |
If the plan covers usage that would cost $14,000 a month at on-demand rates, the effective saving is $4,000 a month, or 28.6% (calculation). A benchmark comparing September's billed cost with October's would show a spurious “saving” of 100% on the commitment line. Standards exist for exactly this reason: FOCUS combines billed and amortized cost in one dataset, and Azure's cost analysis lets users switch between actual and amortized views. Choose one basis, document it, and keep it constant across every comparison.
Cloud Spending by Company Size: SMBs, Mid-Market Firms, and Enterprises
Spending rises with organization size, but size is a weak predictor of any one company's bill. Two sources define size differently, so their thresholds must not be mixed. The FinOps Foundation uses employee counts: SMB under 1,000, enterprise 1,000 to 9,999, and large enterprise 10,000 or more. Flexera reports spend by its own SMB, enterprise and large-enterprise subgroups.
Segment | Published signal | Source and caution |
SMB | 69% concentrated under $50,000 per month; SMB public-cloud workloads rose from 55% to 63% year over year | Flexera 2026; subgroup definitions are Flexera's |
Enterprise | Flexera's chart shows enterprises peaking between $100,000 and $500,000 per month | Read from a chart description; no exact percentage verified |
Large enterprise | 76% above $5 million per month | Flexera 2026; share of respondents, not an average |
Survey mix | 47% large enterprise, 33% enterprise, 20% SMB | FinOps Foundation 2026; describes who responded, not who spends what |
Neither survey publishes a “mid-market” spend band, so this article does not invent one. Growing companies sit between the SMB and enterprise rows, and the honest guidance is to benchmark them against peers with similar workloads instead of a size label. Median or percentile claims for any segment would go beyond the evidence.
Headcount misleads because the product often is the workload. A cloud-native B2B SaaS company with a few hundred employees can consume more compute, storage and data transfer than a far larger organization whose systems remain largely on premises. Revenue model, data volume, uptime targets, AI usage and the share of infrastructure in hybrid cloud arrangements usually explain more of the bill than the org chart. Teams also scale differently: the FinOps Foundation reports that organizations managing $100 million or more in annual cloud spend average 8 to 10 FinOps practitioners plus 3 to 10 contractors, which is lean relative to the dollars managed.
Cloud Costs by Business Model and Workload
Business model decides which workloads dominate the bill, and that is why two similar-sized firms diverge. No source reviewed publishes industry-wide average spending by business model, so the table compares cost behavior and the metric each model should track.
Business model | Costs that tend to dominate | Why bills diverge | Metric to track |
Early-stage software startup | Compute, managed database, credits expiring | Free credits mask true run-rate | Cost per active user |
B2B SaaS | Compute, databases, data transfer, observability | Cost scales with tenants and usage patterns | Cost per tenant; share of subscription revenue |
Ecommerce | Compute and CDN at peak, search, databases | Seasonal spikes and autoscaling headroom | Cost per order |
Data-intensive business | Storage, pipelines, analytics platforms | Retention and query habits | Cost per query or per terabyte analyzed |
AI application company | GPUs, inference, vector search, evaluation | Token volume, context length, utilization | Cost per completed task |
Enterprise migrating to cloud | Duplicate on-premises and cloud costs, licensing | Overlap during migration | Cost per migrated workload |
Hybrid or multicloud | Interconnect, duplicated tooling, skills | Complexity and egress | Effective cost by environment |
Evidence for workload mix is partial but instructive. In its 2024 report on AWS cost data from hundreds of Datadog customers, Datadog found that containers made up about 35% of EC2 compute spend and that 83% of container costs were tied to idle resources. That sample covers one provider and Datadog customers only, so it illustrates behavior and does not represent all cloud users. See Articsledge's guides to cloud migration and multicloud for how those two models change cost structure.
AWS vs. Azure vs. Google Cloud: Comparing Real Costs
No provider is universally cheapest. Whether Google Cloud is cheaper than AWS, or Azure cheaper than either, depends on workload, region, discounts, existing agreements and architecture. A valid comparison prices an identical workload at an identical service level, so it must be normalized before it is ranked.
Dimension | Hold constant | Where provider presentation differs | Common pitfall |
Compute | vCPU, memory, CPU architecture, OS, tenancy | Instance families and naming | Comparing different CPU generations |
Storage and databases | Durability, IOPS, replication, backups | Managed service tiers | Ignoring backup and replica cost |
Network | Region pairs, egress volume, cross-zone traffic | Billing grouping differs | Omitting egress |
Support | Response targets | Azure lists Basic, Developer, Standard and Professional Direct | Leaving support out of the estimate |
Discounts | Commitment term and coverage | Savings Plans, reservations, committed use (Google cites up to 57% on Compute Engine) | Treating maximum discounts as typical |
Licensing and agreements | Existing enterprise agreements | Azure calculator can show negotiated rates after sign-in | Comparing list to contracted |
Migration and operations | Skills, tooling, refactoring | Not shown by any calculator | Ignoring labor |
To make a comparison reproducible, define the workload specification, price it the same day in each provider's calculator (AWS, Azure, Google Cloud), apply commitments as a separate step, add data transfer and support, then add migration and operating labor and record the date and region. Calculators produce estimates; Azure's reservation guidance shows that billed and amortized views will still differ later. No live list prices are quoted here because they change and negotiated rates differ. For adoption context, read Articsledge's cloud market share analysis and its Amazon Web Services guide.
Anatomy of a Real Cloud Bill: The Biggest Cost Drivers
A cloud bill is the sum of many separately metered services, and each has its own billing driver. Provider tooling can even group the same charges differently: AWS Cost Explorer splits “EC2 - Instances” from “EC2 - Other” and groups data transfer by service, while the invoice shows a single Data Transfer line. Know which view you are reading before you diagnose a cost.
Cost area | Billing driver | Common surprise | Practical control |
Virtual machines and compute | Size, hours, purchase model | Always-on, oversized instances | Rightsizing, scheduling |
Containers and Kubernetes | Nodes, control plane, add-ons | Idle cluster capacity | Bin packing, resource requests |
Managed databases | Capacity, throughput, replicas | High-availability replicas | Capacity reviews |
Object and block storage | GB stored, tier, requests | Old snapshots and previous-generation volumes | Lifecycle policies |
Networking and egress | GB out of a region or zone | Cross-zone traffic | Service placement, caching |
Load balancers and gateways | Hours plus data processed | Idle gateways | Consolidation |
Serverless | Invocations, duration, memory | Chatty architectures | Batching, memory tuning |
Data platforms and analytics | Compute time, scanned data | Unbounded queries | Quotas, partitioning |
Security services | Events and data scanned | Overlapping tools | Scoped coverage |
Observability and logging | Ingested volume, retention | Verbose logs | Sampling, tiered retention |
Backups and disaster recovery | Replicated data, standby capacity | Over-protecting low-tier systems | Tier recovery objectives |
Support and managed services | Percentage of spend or tier | Growth lifts percentage fees | Annual plan review |
The 2024 Datadog report on AWS data shows why these lines deserve attention: cross-availability-zone traffic made up nearly half of data transfer costs and affected 98% of organizations in its sample, and previous-generation gp2 volumes still represented 58% of the average organization's EBS spend. Treat those as AWS-only observations from Datadog customers, not market-wide rates. Articsledge's guides to cloud networking, cloud telemetry and availability zones explain the mechanics behind several rows.
Workload-Level Benchmarks: Compute, Kubernetes, Databases, Storage, and Networking
Workload-level benchmarks translate a bill into cost per unit of resource or output. The examples below use hypothetical inputs to show the arithmetic, not market rates.
Effective cost per vCPU-hour = effective compute cost ÷ vCPU-hours provisioned. An 8-vCPU instance at a hypothetical $0.40 per hour costs $0.05 per vCPU-hour; at 30% average CPU utilization, each utilized vCPU-hour costs about $0.167 (calculation).
Cost per Kubernetes workload = share of requested node capacity × node cost + allocated idle cost. The Datadog sample attributed 83% of container costs to idle resources, split between cluster idle (54%) and workload idle (29%).
Database cost per 1,000 queries = monthly database cost ÷ queries × 1,000. At a hypothetical $12,000 and 300 million queries, that is $0.04.
Storage cost per usable GB-month = storage cost ÷ usable (not raw) GB. A hypothetical $2,300 for 100,000 raw GB under three-way replication leaves about 33,333 usable GB, or $0.069 per usable GB (calculation).
Data transfer cost per GB = transfer charges ÷ GB moved. A hypothetical $4,500 over 50,000 GB equals $0.09 per GB.
Unit price alone hides utilization, replication, storage class and service level, so only compare units that deliver equal quality. Choose the denominator carefully: the FinOps Foundation warns that if engineers compress data and cut storage by 30%, cost per GB does not change even though the bill falls, whereas cost per stored item would show the improvement. Background reading on Kubernetes, database as a service and the cloud instance can help teams agree on definitions first.
How Commitments and Discounts Change Effective Cloud Costs
Commitments trade flexibility for lower unit rates, and the saving is realized only if usage keeps filling the commitment. The main instruments are AWS Savings Plans and Reserved Instances, Azure reservations and savings plans, Google Cloud committed use discounts, negotiated enterprise agreements, spot or preemptible capacity, and provider credits. Google's pricing page cites savings of up to 57% on Compute Engine committed use, and offers new customers $300 in credits; “up to” figures are ceilings, not typical outcomes.
Adoption evidence is mixed. Flexera reports that fewer than half of organizations use any single commitment discount per provider and that more than half still use on-demand pricing. In Datadog's 2024 AWS sample, 67% of organizations used commitment-based discounts (down from 72%), only 29% covered more than half of eligible spend, and 59% used Savings Plans versus 15% for Reserved Instances. Savings and amortization appear differently by basis: AWS Cost Explorer can spread commitment fees over the usage they cover.
Worked example: theoretical versus realized savings
Assume a hypothetical commitment of $15,000 per month at a 25% discount, which covers $20,000 of on-demand-equivalent usage ($15,000 ÷ 0.75). The fee is owed whether or not usage materializes.
Usage (on-demand equivalent) | Cost paid | Saving versus on-demand |
$20,000 (full use) | $15,000 | $5,000 (25.0%) |
$17,000 (85% use) | $15,000 | $2,000 (11.8%) |
$14,000 (70% use) | $15,000 | −$1,000 (−7.1%) |
Rightsizing after buying a commitment can turn a theoretical 25% saving into a loss. Match commitments only to steady baseline demand, review utilization monthly, and assign ownership under a clear cloud governance model. Classifying a cloud workload as steady or variable before purchase is the cheapest safeguard.
AI, GPUs, and Data Platform Costs in 2026
AI changes cloud spending by adding variable, hard-to-forecast workloads. In Flexera's survey, every respondent used generative AI in some capacity, 45% used it extensively, use of GenAI as a public cloud service rose from 50% to 58%, and 30% cited cost unpredictability as a scaling challenge. The FinOps Foundation found that 98% of respondents now manage AI spend, with visibility, allocation and value measurement as the top difficulties. Broader AI market figures describe vendor revenue or forecasts and must not be read as what any single company pays.
The cost drivers differ from traditional infrastructure. GPU capacity is expensive when idle, so utilization matters as much as price; in Datadog's 2024 AWS sample, GPU instances rose from 10% to 14% of EC2 compute spend for organizations that used them. Training is a project with an end date, while inference recurs with every use. Around the model sit embeddings, retrieval, a vector database, data pipelines, evaluation and monitoring, each billed separately. Demand can also spike, so capacity planning must account for bursts.
API pricing adds its own traps. The FinOps Foundation's token pricing paper notes that output tokens often cost three to five times as much as input tokens, that reasoning tokens are frequently billed as output, and that conversational apps resend history on every call, a pattern it calls context window creep. Caching and batch processing can cut cost, but only when prompts and workloads fit them. It recommends measuring cost per successful outcome as a distribution, not a single average.
Illustrative formula: cost per completed AI task
Cost per task = calls per task × (input tokens × input price + output tokens × output price) × (1 + retry rate) + retrieval and platform cost per task. The inputs below are hypothetical, not any provider's price list.
Assumptions: 3 calls per task; 4,000 input and 600 output tokens per call; $2.00 per million input tokens; $8.00 per million output tokens (4× input); 10% retries; $0.003 retrieval and logging per task.
Per call: 4,000 × $2 ÷ 1,000,000 = $0.0080 input, plus 600 × $8 ÷ 1,000,000 = $0.0048 output, for $0.0128.
Per task: $0.0128 × 3 = $0.0384; × 1.10 = $0.04224; plus $0.003 = $0.04524, about 4.5 cents. At 500,000 tasks a month, that is $22,620.
Sensitivity: if resent history doubles input to 8,000 tokens per call, cost per task rises to $0.07164, a 58% increase, or $35,820 a month.
For deeper context on the technology and its economics, see Articsledge's guides to the context window, AI infrastructure and AI implementation cost.
The Costs Companies Often Forget
The provider invoice is narrower than total cost of ownership. Costs commonly left out include:
SaaS subscriptions, data cloud platforms and AI tools that sit beside the infrastructure bill. The FinOps Foundation reports that 90% of respondents manage or plan to manage SaaS, 64% licensing, 57% private cloud, 48% data center and 28% labor, while Flexera found 8% of respondents are not tracking SaaS costs at all.
Observability and security platforms priced by data volume or endpoints.
Enterprise support, compliance tooling and data-residency requirements.
Bring-your-own licenses and third-party data services.
Backup and resilience capacity held in reserve.
Engineering labor, migration projects and operations.
Avoid double counting by assigning each cost one category and one owner. A marketplace product billed through the cloud invoice is either infrastructure or software, not both, and a logging platform already counted under observability should not reappear under security. Reports on SaaS management software, cloud compliance and backup as a service show where these lines originate, and software maintenance cost covers the labor side.
Cloud Waste in 2026: How Much Spending Is Avoidable?
Respondents to Flexera's 2026 survey estimate that 29% of IaaS and PaaS spend is wasted, reversing a five-year decline; the same publication notes respondents estimated 30% in 2021. It is a self-reported estimate, not a verified share of every bill. As arithmetic, 29% of a hypothetical $1 million monthly bill is $290,000 a month, or $3.48 million a year (calculation), but any single organization may sit far above or below that.
Waste takes recognizable forms: idle or oversized cloud resources, unattached volumes, forgotten snapshots, overprovisioned clusters, inefficient data movement, unused commitments, verbose telemetry and underused GPUs. Datadog notes gp3 volumes cost about 20% less than gp2, yet older volumes still dominated spend in its sample. Not every idle dollar is waste: resilience, redundancy, performance headroom and deliberate experimentation are legitimate. The FinOps Foundation also reports diminishing returns once the biggest waste is removed.
Prioritize with a simple score: Priority = (savings × business impact) ÷ (effort × risk), each rated 1 to 5. The ratings below are illustrative.
Action | Savings | Impact | Effort | Risk | Priority |
Schedule non-production shutdowns | 3 | 3 | 1 | 1 | 9.0 |
Delete unattached volumes | 2 | 2 | 1 | 1 | 4.0 |
Right-size managed databases | 4 | 4 | 3 | 3 | 1.8 |
Redesign cross-zone traffic | 3 | 3 | 4 | 3 | 0.75 |
Quick wins come first, but the later rows may carry larger absolute savings, so review dollar value alongside the score. Cloud automation makes the safe actions repeatable.
Cloud Cost Unit Economics: The Benchmarks That Actually Matter
Unit economics ties cloud cost to something the business sells or serves. The FinOps Foundation's introduction to cloud unit economics describes the practice as relating spend to a business value metric, so a rising bill can be judged against growth instead of in isolation. No source reviewed publishes a universal cloud-to-revenue ratio, so the table below gives formulas, not target values.
Metric | Formula | Best suited to | Main caution |
Cost per customer or tenant | Effective cloud cost ÷ active customers | B2B SaaS | Large accounts distort the average; review the distribution |
Cost per order or transaction | Effective cost ÷ completed transactions | Ecommerce, payments | Peak days skew results; use monthly totals |
Cost per active user | Effective cost ÷ monthly active users | Consumer apps | Firms define “active” differently |
Cost per completed AI task | Token, retrieval and platform cost ÷ successful tasks | AI products | Count successful outcomes only |
Cloud cost as a share of revenue | Effective cloud cost ÷ recognized revenue | Finance reviews | Depends on pricing and margin; no published universal benchmark |
Cost per environment | Allocated cost of production, staging and development | Engineering leaders | Requires consistent tagging |
Choose one numerator (effective, fully allocated cost) and one denominator per product, publish the definition, and watch the trend more than the level. A falling unit cost alongside a rising bill usually signals healthy growth, while a flat bill with falling volume does not. Shared costs need a stated allocation rule, such as usage share, before any division. Revenue-based denominators pair well with subscription billing software data.
Three Worked Cloud-Cost Scenarios
All figures in this section are hypothetical, built to show structure and arithmetic. They are not averages, they are not drawn from any source, and no real company is represented.
Scenario A, early-stage SaaS: a cloud-first strategy startup with 400 customer accounts.
Scenario B, scaled SaaS: a growth-stage product with 38,000 monthly active users and a 4% enterprise discount on spend above a $90,000 monthly commitment.
Scenario C, large enterprise platform: 75 million transactions a month, a 5% enterprise discount on spend above a $5 million monthly commitment, and $100,000 of monthly credits.
Monthly line, gross | A: early-stage SaaS | B: scaled SaaS | C: enterprise platform |
Compute | $9,000 | $62,000 | $3,400,000 |
Databases | $4,500 | $31,000 | $1,100,000 |
Storage | $1,800 | $14,000 | $700,000 |
Networking and egress | $2,200 | $16,000 | $600,000 |
Data platform | $0 | $22,000 | $800,000 |
Observability and other | $2,700 | $19,000 | $450,000 |
AI and GPU | $1,500 | $12,000 | $1,350,000 |
Support and marketplace | $800 | $9,000 | $600,000 |
Total gross | $22,500 | $185,000 | $9,000,000 |
Result | A | B | C |
Commitment discounts | −$1,200 | −$19,800 | −$1,200,000 |
Enterprise discount | None | −$3,800 | −$200,000 |
Credits | None | None | −$100,000 |
Net monthly cost | $21,300 | $161,400 | $7,500,000 |
Annualized (× 12) | $255,600 | $1,936,800 | $90,000,000 |
Unit denominator | 400 customers | 38,000 users | 75 million transactions |
Net cost per unit | $53.25 | $4.25 | $0.10 |
Sensitivity: usage and units +25% | Net $26,550; 500 customers; $53.10 | Net $202,200; 47,500 users; $4.26 | Net $9,495,000; 93.75 million; $0.10 |
Arithmetic: Scenario B's enterprise discount is 4% × ($185,000 − $90,000) = $3,800, and Scenario C's is 5% × ($9,000,000 − $5,000,000) = $200,000. In the sensitivity row, variable lines grow 25% while fixed items stay flat (support, marketplace and a fixed part of other costs: $1,500, $15,000 and $600,000 respectively), commitments are held constant, and enterprise discounts are recalculated on the higher spend.
Two lessons follow. First, net spend rose 24.6%, 25.3% and 26.6% against 25% growth, so fixed commitments can make cost grow slightly faster than volume once usage outruns the committed baseline. Second, unit cost stayed almost flat, which is the signal that matters. Regulated buyers weighing sovereign cloud options should add the premium as its own line instead of burying it in compute.
How to Benchmark Your Own Cloud Bill: A Practical Method
A defensible benchmark compares like with like. The sequence below works for one provider or several, and it starts with your own history because that is the only dataset whose definitions you control.
Fix the cost basis. Use effective (amortized) cost and record billed cost beside it, as described earlier.
Export a full month of usage data in a consistent schema. The FOCUS specification exists to normalize billing fields across providers.
Allocate cost to owners. Group by product, team and environment, and assign an owner to every unallocated dollar. Tagging enforced through infrastructure as code stays more consistent than manual labels.
Add costs outside the invoice, such as SaaS, observability platforms and labor, and give each one category.
Choose denominators from the unit economics table: customers, transactions, users or successful AI tasks.
Compare with your own last three to six months first, then with peers whose workloads, units and commitments resemble yours.
Decompose changes into price and volume. Price variance = (new rate − old rate) × new volume. Volume variance = (new volume − old volume) × old rate.
Document the result and repeat quarterly, with owners from engineering, finance and DevOps teams.
Variance example (hypothetical): compute rises from 100,000 vCPU-hours at $0.050 ($5,000) to 120,000 vCPU-hours at $0.045 ($5,400). Price variance is (0.045 − 0.050) × 120,000 = −$600; volume variance is 20,000 × 0.050 = +$1,000; the net change is +$400 (calculation). A bill that rose 8% hides a 12% rate improvement and a 20% volume increase.
Worksheet field | What to record | Typical source |
Period and basis | Month, billed or amortized, currency | Billing export |
Gross usage cost | Metered usage before reductions | Cost and usage data |
Reductions | Commitments, enterprise discounts, credits | Billing export, contract |
Net effective cost | Gross minus reductions | Calculation |
Unallocated share | Cost with no owner, as % of net | Tagging report |
Units and unit cost | Denominator and net cost per unit | Product analytics |
Trend | Change versus prior three months | Your own history |
Peer comparisons are valid only after checking provider, region, workload type, commitment coverage and service levels. Where any of those differ, treat the peer figure as context, not a target.
Building a Defensible 2027 Cloud Budget From 2026 Costs
A defensible budget starts from effective 2026 cost, not the latest invoice, and states every assumption. One simple structure:
2027 budget = [Base × (1 + g) × (1 − e) + N + A] × (1 + p) × (1 + c)
Here Base is 2026 annual effective cost; g is volume growth; e is planned efficiency savings; N is new non-AI initiatives; A is AI and GPU additions; p is expected price change; and c is a contingency reserve. The table applies the formula to a hypothetical $2,400,000 base.
Input | Low | Base | High |
Volume growth (g) | 10% | 18% | 30% |
Efficiency savings (e) | 10% | 6% | 2% |
New initiatives (N) | $60,000 | $120,000 | $240,000 |
AI and GPU additions (A) | $90,000 | $180,000 | $420,000 |
Price change (p) | 0% | 2% | 5% |
Contingency (c) | 5% | 8% | 12% |
2027 budget | $2,652,300 | $3,263,027 | $4,371,898 |
Change versus base year | +10.5% | +36.0% | +82.2% |
Check on the base case: $2,400,000 × 1.18 × 0.94 = $2,662,080; plus $300,000 = $2,962,080; × 1.02 = $3,021,322; × 1.08 = $3,263,027, or about $271,919 a month (calculation). The spread between cases is the message: AI additions and growth drive uncertainty more than price does, so keep them as separate lines and reforecast each quarter.
Split the budget into committed spend, variable run-rate and project spend, and track cash separately because upfront commitments change timing without changing amortized cost. Finance teams can model the timing with budget planning software and cash flow forecasting software, while AI financial forecasting tools may help with usage-driven lines if their assumptions are visible. Do not commit beyond steady baseline demand, and treat any forecast as a range.
Cloud Procurement and Optimization: Where Companies Can Save
Sequence matters: remove waste and rightsize first, then commit to what remains. Buying commitments before cleanup locks in the inefficiency, which the commitment example earlier showed can turn a saving into a loss. The table lists the main procurement levers and what to verify in each.
Lever | What to verify | Main risk |
Commitments | Coverage of steady baseline only, term, flexibility to change instance family or service | Unused commitment |
Enterprise agreement | Commitment level, shortfall terms, which purchases count toward it | Paying for unmet commitment |
Marketplace purchases | Whether they draw down commitments and who owns the vendor relationship | Duplicate spend |
Credits | Expiry, eligible services, renewal | Run-rate hidden until credits end |
Support plans | Percentage-of-spend pricing and tier fit | Fees rising faster than value |
Data transfer and exit | Egress charges, portability, cross-region design | Lock-in cost at renewal |
Architecture | Placement, caching, storage tiers, scheduling | Savings that need engineering time |
Process supports the savings. A vendor management record and a contract lifecycle calendar keep renewal dates, commitment terms and credit expiry visible. When evaluating a managed service or platform, a build versus buy comparison should include operating labor, not only the sticker price. Organizations choosing between public cloud and platform as a service offerings should compare effective cost per unit, not list rates.
Benchmarking Methodology, Data Limitations, and Editorial Notes
Research cutoff: October 31, 2026. The evidence base is Flexera's 2026 State of the Cloud findings (753 respondents, published March 18, 2026), the FinOps Foundation's 2026 State of FinOps survey (1,192 respondents representing more than $83 billion in annual spend), Datadog's 2024 State of Cloud Costs report (AWS data from Datadog customers only), FinOps Foundation working-group papers, the FOCUS specification and provider documentation for AWS, Azure and Google Cloud.
Sourced: survey percentages, respondent counts, sample descriptions and the Datadog observations, each attributed in the text.
Calculated: every figure marked “calculation,” including annualized spend bands, the billed versus amortized example, the AI task cost and the budget cases.
Hypothetical: all worked examples, scenarios and unit-cost inputs. They illustrate arithmetic and are not market data.
Not used: live list prices, which vary by region and change often, and any claim that could not be tied to a source that was opened and read.
Limitations are real. Surveys are self-reported and use respondent bands, not averages; size categories differ between publishers; Datadog's sample is older and limited to AWS; and converting monthly bands to annual figures is arithmetic, not new data. A figure circulating for a $200,001–$500,000 monthly public cloud band was left out because the published material attaches that band to SaaS spending. For broader market context, see Articsledge's cloud computing statistics and cloud adoption statistics.
Because workloads, discounts and architecture differ, no reliable universal cloud fee, cloud-to-revenue ratio or guaranteed saving exists. This guide is general information, not financial, legal or procurement advice; confirm assumptions with your providers and finance team.
FAQ
What is the average cloud bill in 2026?
No source reviewed establishes one. In Flexera's 2026 survey of 753 respondents, 69% of small and midsize businesses were concentrated under $50,000 per month and 76% of large enterprises spent more than $5 million. Those are shares of respondents, not averages.
What do companies actually pay for cloud?
It depends on the cost basis: list price, contracted price, billed cost or effective (amortized) cost can all describe the same environment. For comparisons, use effective cost, document the basis and keep it constant. See cloud spending statistics for market-level context.
How much cloud spending is wasted?
Respondents to Flexera's 2026 survey estimate that 29% of IaaS and PaaS spend is wasted, the first increase in five years. It is a self-reported estimate. Some idle capacity is deliberate resilience, so measure waste against requirements, not against zero.
Which is cheapest: AWS, Azure or Google Cloud?
None is universally cheapest. Price an identical workload at an identical service level on the same day in each provider's calculator, apply commitments as a separate step, then add data transfer, support and operating labor. Results change with region, discounts and architecture.
What is the difference between billed and amortized cost?
Billed cost follows when cash is charged, so an upfront commitment appears as one large payment. Amortized (effective) cost spreads that fee across the usage it covers. A $120,000 one-year upfront plan is $120,000 billed in one month or $10,000 amortized monthly.
How much do cloud commitments really save?
Google cites savings of up to 57% on Compute Engine committed use, but “up to” is a ceiling. Realized savings depend on utilization: in this guide's hypothetical, a 25% discount falls to 11.8% at 85% use and becomes a 7.1% loss at 70% use.
What is cloud unit economics?
It relates cloud cost to a business output, such as cost per customer, order, active user or completed AI task. The trend matters more than the level: a bill that rises while unit cost falls usually reflects growth, not inefficiency.
How do I benchmark my own cloud bill?
Fix the cost basis, export a full month of usage data, allocate cost to owners, add non-invoice costs, choose denominators, and compare with your own history before peers. Then split changes into price and volume variance and repeat quarterly.
How much does AI cost in the cloud?
No universal figure exists. Compute cost per completed task from calls, tokens, retries and platform costs. In this guide's hypothetical, a task costs about $0.045, and doubling resent input history raises it 58%. Output tokens often cost three to five times input tokens, per the FinOps Foundation. MLOps practices help track it.
What costs are missing from the cloud invoice?
Common omissions are SaaS subscriptions, observability and cloud security platforms, enterprise support, bring-your-own licenses, backup capacity and engineering labor. Assign each cost one category and one owner to avoid double counting.
How should I budget cloud costs for 2027?
Start from 2026 effective cost, apply volume growth and planned efficiency savings, add new initiatives and AI additions as separate lines, then apply price change and contingency. Present low, base and high cases, and reforecast quarterly.
Is there a standard cloud-cost-to-revenue benchmark?
No source reviewed publishes a reliable universal ratio. Revenue model, margins, AI usage and architecture vary too widely. Track your own cloud-cost-to-revenue trend and compare with peers only when workloads and definitions match.
What is FOCUS?
FOCUS, the FinOps Open Cost and Usage Specification, defines a common format for cost and usage data so billing from different providers can be compared. It combines billed and amortized cost in one dataset, and this guide references version 1.4.
Should I buy commitments before optimizing?
Generally no. Remove waste and rightsize first, then commit only to steady baseline demand. Buying first locks in inefficiency, and any unused commitment is owed regardless of usage.
Key Takeaways
No average 2026 cloud bill is established by the sources reviewed; Flexera's figures are shares of respondents in spend bands.
Billed, amortized and list-price views of the same environment can differ sharply, so fix one basis before comparing.
Respondents estimate 29% of IaaS and PaaS spend is wasted, a self-reported number that rose for the first time in five years.
Commitment savings are ceilings until utilization proves them; low use can turn a discount into a loss.
AI spend is driven by tokens, retries, resent history and idle GPUs, so measure cost per successful task.
Unit cost and its trend are more defensible benchmarks than any peer average.
A 2027 budget should show low, base and high cases, with AI and growth on separate lines.
Actionable Next Steps
Export one full month of billing data and record both billed and amortized cost.
Allocate every line to a product, team and environment, and name an owner for unallocated cost.
Add SaaS, observability, support and labor costs outside the invoice.
Choose one unit metric per product and calculate its three-month trend.
Rank waste candidates with the priority score, then verify dollar value.
Review commitment utilization before any purchase or renewal.
Build low, base and high 2027 cases and schedule quarterly reforecasts.
Document your method so future comparisons stay consistent, and fold the findings into any cloud modernization plans.
Glossary
Amortized (effective) cost: commitment fees spread across the usage they cover.
Billed cost: the amount charged on the invoice in a period.
Cloud waste: spending on resources that deliver no needed value.
Commitment: an agreement to a level of usage or spend in return for lower rates, such as a Savings Plan or reservation.
Egress: data leaving a cloud region, zone or provider, usually charged per GB.
Enterprise discount: a negotiated reduction tied to a spend commitment.
FinOps: the practice of managing cloud cost through collaboration between engineering, finance and business teams.
FOCUS: an open specification for cost and usage data.
Idle resource: capacity that is provisioned but little used.
Inference: running a trained model to produce an answer, see AI inference.
On-demand pricing: pay-as-you-go rates with no commitment.
Rightsizing: matching resource size to actual demand.
Tagging: labels that assign cost to owners.
Token: a unit of text processed by a model, central to large language model pricing.
Total cost of ownership: all costs of running a workload, including labor and software.
Unit economics: cloud cost expressed per unit of business output.
Sources & References
Pages without a verifiable date are marked n.d.; all were accessed on October 31, 2026.
Flexera. “Flexera 2026 State of the Cloud Report: The Convergence of Cloud and Value.” Flexera. 2026-03-18. flexera.com
Flexera. “The New Era of Cloud: What 2026 Data Tells Us About Spend, Scale and Strategy.” Flexera. n.d. flexera.com
Flexera. “Cloud Cost Management Trends.” Flexera. n.d. flexera.com
FinOps Foundation. “State of FinOps 2026 Data.” FinOps Foundation. n.d. data.finops.org
FinOps Foundation. “Introduction to Cloud Unit Economics.” FinOps Foundation. n.d. finops.org
FinOps Foundation. “GenAI FinOps: How Token Pricing Really Works.” FinOps Foundation. n.d. finops.org
FinOps Foundation. “What Is FOCUS?” FOCUS project. n.d. focus.finops.org
FinOps Foundation. “FOCUS Specification v1.4.” FOCUS project. n.d. focus.finops.org
Datadog. “State of Cloud Costs” (2024 edition, AWS customer data). Datadog. 2024. datadoghq.com
Amazon Web Services. “Understanding Differences Between Billing Data and Cost Explorer Data.” AWS Billing documentation. n.d. docs.aws.amazon.com
Amazon Web Services. “Exploring Your Data Using Cost Explorer.” AWS Cost Management documentation. n.d. docs.aws.amazon.com
Amazon Web Services. “AWS Pricing Calculator.” AWS documentation. n.d. docs.aws.amazon.com
Microsoft. “Pricing Calculator.” Microsoft Learn, Azure Cost Management. n.d. learn.microsoft.com
Microsoft. “Azure Reservations and Savings Plan FAQ.” Microsoft Learn. n.d. learn.microsoft.com
Google Cloud. “Google Cloud Pricing.” Google. n.d. cloud.google.com
