The State of Cloud Waste 2026: $100B+ in Unnecessary Spend

AWS

AWS has the largest market share (31% per Synergy Research Q4 2024) and the most complex pricing structure. The dominant waste categories on AWS are:

• EC2 on-demand vs. Savings Plans gap: Datadog 2024 found that 42% of EC2 hours were purchased on-demand despite running continuously for 30+ days — a strong signal for Savings Plan coverage. The average on-demand premium over equivalent Savings Plans pricing is 35–40%.
• NAT Gateway data processing: NAT Gateway charges $0.045/GB for data processed. In microservice architectures, internal traffic that routes unnecessarily through NAT can represent significant cost. AWS introduced NAT Gateway free-tier for same-AZ traffic in 2024, but only for new routes configured correctly.
• Unoptimized S3 storage classes: The default S3 storage class (Standard) costs $0.023/GB/month. Objects older than 30 days that are rarely accessed belong in Infrequent Access ($0.0125/GB/month) or Glacier ($0.004/GB/month). Without lifecycle policies, objects accumulate in Standard indefinitely.

Azure

Azure has approximately 25% cloud market share and is the dominant choice for Microsoft-ecosystem enterprises. Azure waste patterns reflect its enterprise customer base:

• Orphaned managed disks: Azure managed disks persist after VM deletion by default. The Azure portal provides warnings but does not enforce cleanup. Enterprise environments with hundreds or thousands of VMs accumulate significant orphaned disk inventory over time.
• Dev/test subscriptions used for production: Azure offers significantly reduced pricing through Dev/Test offer subscriptions, but these have SLA restrictions. Some teams inadvertently run production workloads on Dev/Test pricing, which can create compliance exposure, though the pricing itself represents a legitimate saving.
• Unused Azure Hybrid Benefit assignments: Azure Hybrid Benefit allows organizations with existing Windows Server and SQL Server licenses to apply them to Azure VMs for significant discounts. Flexera 2025 found that 38% of eligible Azure workloads had not claimed Hybrid Benefit — leaving substantial discounts unclaimed.

GCP

Google Cloud Platform has roughly 11% market share but is growing fastest (29% YoY per Synergy Q4 2024), driven heavily by AI workloads. GCP's waste patterns include:

• Unattached persistent disks: Like AWS EBS, GCP persistent disks persist after instance deletion. GCP Recommender surfaces these, but the recommendations require active review.
• BigQuery on-demand scan costs: BigQuery on-demand pricing charges per byte scanned ($6.25/TB). Queries without partition filters or with broad wildcard scans can generate unexpected costs. Harness 2025 flagged BigQuery as the fastest-growing unexpected cost category among GCP customers.
• Committed Use Discount gaps: GCP's Committed Use Discounts (CUDs) offer up to 57% savings for 1- or 3-year commitments. Flexera 2025 found GCP customers had the lowest CUD coverage rate of the three major providers, likely because GCP's Sustained Use Discounts (automatic discounts for consistent usage) reduce the urgency of committing.

1. No ownership

Cloud resources are provisioned by engineering teams but costs are paid centrally. The engineer who provisions a dev server has no direct financial incentive to terminate it. The finance team that pays the bill cannot identify which engineer owns which resource. Without clear ownership at the team or individual level, no one acts.

Only 44% of organizations have implemented chargeback or showback — where teams are shown or charged their own cloud costs (CNCF FinOps Survey, 2024). In organizations without chargeback, the waste rate is consistently higher. The mechanism is simple: when someone else pays the bill, the bill grows.

2. Fear of breaking things

Right-sizing and terminating idle resources carries perceived risk. Terminating an instance that turns out to be needed causes immediate consequences. Leaving an idle instance running causes no immediate consequences. The asymmetric downside creates rational inaction.

58% of respondents cited "fear of production impact" as the most common barrier to acting on cost recommendations (Harness, 2025). The solution is better observability — tools that confirm an instance has had zero inbound connections for 90 days before recommending deletion, not just that CPU is low.

3. Lack of visibility

Native cloud cost dashboards show what was spent, not what should have been spent. AWS Cost Explorer, Azure Cost Management, and GCP Billing do not surface idle resources, cross-reference utilization with cost data, or produce actionable recommendations sorted by dollar savings. Engineers must manually correlate Cost Explorer with CloudWatch metrics — significant work requiring SQL-level comfort with billing exports.

61% of engineering teams cannot attribute more than 80% of their cloud costs to a specific team, service, or product line (Datadog, 2024). If you cannot attribute costs, you cannot hold anyone accountable for reducing them.

4. Vendor incentives point the wrong way

Cloud vendors profit from overprovisioning. Their cost management tools are useful but deliberately conservative — AWS Trusted Advisor recommendations avoid false positives that could cause customer incidents. No cloud vendor has a business incentive to surface the most aggressive cost reductions available.

Engineering performance reviews rarely include cloud cost as a metric. The developer who ships a feature that overprovisions by 10x is not penalized. The developer who optimizes infrastructure often gets no recognition. Without changing incentives, behavior does not change.

IPv4 charges became a structural cost line

AWS began charging $0.005/hour for all public IPv4 addresses on February 1, 2024. At $3.65/month per IP, the charge is modest in isolation — but organizations with hundreds of load balancers and EC2 instances saw meaningful bill increases immediately. Azure followed with similar IPv4 charges in July 2025. GCP extended charges to in-use IPs on standard-tier VMs in 2025. IPv4 exhaustion-driven pricing is now a structural cost line across all three major providers, not just a single-provider anomaly.

GPU idle waste emerged as a high-dollar new category

GPU instances (AWS P4, P5, G5; Azure NCv4 A100; GCP A3) are 10–20x more expensive than equivalent CPU compute. AI teams provision large GPU clusters for training runs and leave them running between experiments. A single p4d.24xlarge on AWS costs $32.77/hour on-demand. An idle GPU cluster over a long weekend generates $5,000–$20,000 in unnecessary spend before anyone notices.

GPU idle time is the fastest-growing waste category (Harness, 2025), with average GPU utilization at 23% — meaning 77% of provisioned GPU capacity is wasted. Fewer than 20% of organizations have implemented automatic shutdown policies for GPU instances (CNCF FinOps Survey, 2024). This is where the next wave of cloud cost savings will come from.

GCP eliminated internet egress charges

In March 2024, Google Cloud eliminated internet egress charges for data leaving GCP in most regions (announced January 2024). Egress had been a major cost driver and lock-in mechanism for years. AWS reduced cross-region data transfer pricing in 2024 but did not eliminate internet egress charges ($0.09/GB still applies). Azure made targeted reductions in specific geographies.

For organizations with significant egress spend, GCP's change creates a real pricing asymmetry. Workloads with high outbound traffic currently running on AWS have a legitimate case for a total cost comparison with GCP — one that was harder to justify before 2024.

Spot instance viability expanded

AWS Spot interruption rates declined through 2024 into 2025 for most instance families, making Spot more viable across a wider range of workloads. Organizations using Spot for at least 30% of non-production compute saved an average of 22% on total EC2 spend (Harness, 2025). Spot adoption remains below 30% of eligible workloads for most organizations — real savings left on the table.

Step 1: Establish your cost baseline

Pull your last 90 days of cloud spend by service. Use AWS Cost Explorer, Azure Cost Management, or GCP Billing export. Tag coverage matters here — if more than 20% of spend is untagged, your waste identification will be incomplete because you cannot attribute resources to teams or workloads. Improve tagging first if needed.

Step 2: Identify idle resources (low-effort waste)

Filter all compute instances with average CPU below 5% over the past 30 days. Cross-reference with network I/O: an instance with low CPU but high network traffic may be a proxy or load balancer, not idle. Flag storage volumes not attached to any instance. Flag snapshots older than your retention policy. This category is addressable without performance risk.

Dollar estimate: multiply identified resource costs by 100%. Everything in this category is waste, not buffer.

Step 3: Quantify overprovisioning (higher-effort waste)

For instances with CPU 5–40% average utilization, calculate the right-sized equivalent. If an instance averages 15% CPU utilization and your target headroom is 60% (i.e., peak at 40% average), the right-sized instance should be roughly half the current size. The cost difference between current and right-sized is your overprovisioning waste estimate.

Apply a 50% confidence discount: not all of this waste is actionable (some workloads have spiky patterns that require headroom), so halve your estimate for planning purposes.

Step 4: Calculate discount gap

Identify resources that have run continuously for 30+ days on on-demand pricing where Savings Plans or Reserved Instances would apply. The delta between on-demand cost and equivalent committed pricing is your discount gap. This is waste of a different kind — you are not using unnecessary resources, but you are paying more than necessary for resources you will certainly use.

Step 5: Compute your waste rate

Add idle resource cost (100%) + overprovisioning estimate (discounted 50%) + discount gap, then divide by total spend. Compare against the 27% benchmark. If you are above 27%, you have above-average waste. If you are below 15%, you are in the top quartile of cost efficiency.

Where does the 27% waste figure come from?

The 27% figure comes from Flexera's 2025 State of the Cloud Report, which surveys 750+ cloud decision-makers annually. The same survey has reported waste rates in the 27–32% range every year since 2019. Harness (2025) found 28% average waste among its customer base; Datadog (2024) found 25–30% depending on organization maturity. We use 27% as the consensus mid-point from three independent sources.

Is the $100B figure really accurate?

It depends on how you define waste. Applying Flexera's 27% rate to Gartner's $675B 2025 cloud market gives a gross figure of $182B. The $100B figure uses a conservative 15% "actionable waste" threshold — resources that could be terminated or right-sized with low risk, based on Harness 2025 methodology. Both figures are defensible. We use $100B+ as the headline because it is more conservative and therefore harder to dispute.

Why hasn't cloud waste decreased over five years?

Cloud spend grows rapidly, generating new waste faster than existing waste is cleared. Organizational complexity increases as companies scale — more teams, more accounts, more services. Incentive structures do not reward optimization. The fastest-growing workloads (AI, real-time data pipelines) are the hardest to right-size because demand is highly variable. The result is a treadmill: organizations optimize what they can see, but new waste accumulates faster than cleanup keeps pace.

How does cloud waste compare to other enterprise software waste?

Cloud waste at 27% is broadly comparable to SaaS license waste — Flexera's IT asset management data shows 30–35% of SaaS seats go unused. The mechanisms are the same: provisioning is easy and immediate; deprovisioning requires deliberate action; no one with budget authority is directly harmed by the waste. Cloud waste is arguably more tractable. Resources are programmatically discoverable, and savings are quantifiable before you even start.

What's the fastest way to reduce cloud waste for a small team?

Start with idle and unattached resources. They require no performance judgment and carry essentially no risk. Run a scan using SpendArk or AWS Trusted Advisor to identify them. Terminate or snapshot-and-delete everything with zero activity for 30+ days. This step typically recovers 10–15% of total spend for teams that have never audited. Then move to right-sizing — carefully, with utilization data in hand.