Quick Answer
Amazon EC2 pricing uses five models. On-Demand (pay per second), Savings Plans and Reserved Instances (up to 72% off steady use), Spot (up to 90% off interruptible work), and Dedicated Hosts (for licensing and compliance). A t3.micro costs about $0.0104 an hour; an 8-GPU p4d.24xlarge runs about $32.77. For AI, the rate is misleading: what drives margin is cost per inference.
A t3.micro costs about a cent an hour. A p4d.24xlarge costs about $32.77 an hour, roughly $23,900 a month if it never sleeps. Same service. Same console. Same bill that lands on your desk.
For a finance leader, that spread is the problem with Amazon EC2 pricing. EC2 is usually the largest line item on your AWS invoice, the hardest to forecast, and now where your AI spend concentrates. The stakes are no longer just budget variance: 40% of companies now spend more than $10M a year on AI, and most have no idea whether it is worth it, according to a CloudZero report.
This guide is for the person who has to answer for that number. It covers how EC2 pricing works with real rates, why your Cost Explorer bill looks the way it does, what GPU and AI inference do to your margins, and how to prove the return on every AI dollar.
You can model your own numbers in the AWS pricing calculator , and CloudZero is the AI ROI company, so we end on the part finance owns: ROI.
Amazon EC2 pricing at a glance
Amazon EC2 pricing is the per-second or hourly rate you pay to run a virtual server on AWS, set by the instance you choose, the Region, the operating system, and your purchasing model, plus separate charges for storage and data transfer.
Here is the range, On-Demand, Linux, us-east-1:
| Instance | vCPU / accelerator | ~Per hour | ~Per month (730 hrs) |
| t3.micro | 2 vCPU, 1 GiB | $0.0104 | ~$8 |
| t3.large | 2 vCPU, 8 GiB | $0.0832 | ~$61 |
| m5.xlarge | 4 vCPU, 16 GiB | $0.192 | ~$140 |
| g5.xlarge | 1x NVIDIA A10G | $1.006 | ~$734 |
| p4d.24xlarge | 8x NVIDIA A100 | $32.77 | ~$23,922 |
That is a 3,000x range inside one service, which is exactly why a single AI workload can blow up a forecast built on general-purpose assumptions.
Report
Finance needs to prove AI’s return: CloudZero report
260 senior finance leaders (more than half CFOs) told us why the speed of seeing AI spend, not the size of it, separates who pulls ahead on AI from who gets burned.
What is Amazon EC2, and why does it dominate your AWS bill?
Amazon EC2 is the default compute service on AWS. Your engineers rent virtual servers (instances) with a set amount of vCPU, memory, storage, and networking, and you pay for the time they run. It powers web apps, databases, batch jobs, and now AI training and inference on GPU instances.
For finance, the point is concentration and volatility. The same service runs a $8-a-month web server and a $24,000-a-month GPU cluster, and the second kind is multiplying. That is why EC2 drives both your cloud cost and your AI spend, and why the pricing model is the difference between a forecast you can defend and a surprise you cannot.
How does Amazon EC2 pricing work? The five models
AWS bills EC2 five ways, each trading flexibility for price. Billing is per hour or per second (60-second minimum).
On-Demand rates also vary by Region, OS, and size, as the same t3.large shows:
| Instance | Region | OS | On-Demand per hour |
| t3.large (2 vCPU, 8 GiB) | us-east-1 | Linux | ~$0.0832 |
| t3.large (2 vCPU, 8 GiB) | EU (Ireland) | Linux | ~$0.0912 |
| t3.large (2 vCPU, 8 GiB) | US West (N. California) | Windows | ~$0.1268 |
Here are the five models and the discounts AWS offers:
| Model | Max discount | Commitment | Finance read |
| On-Demand | Baseline | None, per second | Maximum flexibility, highest unit price |
| EC2 Instance Savings Plans | Up to 72% | 1 or 3 years | Deepest committed discount, least flexible |
| Compute Savings Plans | Up to 66% | 1 or 3 years | Slightly less discount, flexible across families, Regions, Fargate, Lambda |
| Reserved Instances | Up to 72% | 1 or 3 years | Committed discount plus reserved capacity |
| Spot | Up to 90% | No commitment, but interruptible | Lowest unit price, variable and interruptible |
| Dedicated Hosts | Up to 70% | On-Demand or reserved | Premium for licensing, compliance, isolation |
- On-Demand: flexibility at a premium
On-Demand has no commitment and bills per second. It is the right start for new workloads and the wrong place to leave a steady one, because it carries the highest unit price. EC2 on-demand pricing stops making financial sense once a workload runs predictably.
- Savings Plans: up to 72% for a commitment
Savings Plans trade a one or three-year hourly spend commitment for a discount. EC2 Instance Savings Plans save up to 72% within one family and Region; Compute Savings Plans save up to 66% and stay flexible across EC2, and other Regions. The finance trap is overcommitting, because unused commitment is sunk cost.
Reserved Instances: up to 72%, plus capacity
Reserved Instances save up to 72% (Standard) or 66% (Convertible) and reserve capacity, which Savings Plans do not. AWS adds volume discounts of 5% on $500K to $4M and 10% above $4M.
Spot: up to 90%, with interruptions
Spot Instances use surplus capacity at up to 90% off, with a two-minute reclaim warning. It is a powerful variable-cost lever for fault-tolerant and AI training workloads that can resume.
Dedicated Hosts: the compliance premium
Dedicated Hosts give a full physical server for bring-your-own-license software, compliance, or isolation, with up to 70% off for a commitment.
The model is only part of the bill. The rest is where finance gets surprised.
What drives your EC2 bill, and what is “EC2-Other”?
Your EC2 bill is the instance rate plus a stack of separate charges, and in AWS Cost Explorer most of those separate charges do not show up under “Amazon EC2” at all. They land under a line called EC2-Other, which is why that line so often looks mysteriously large.
EC2-Other bundles the EC2 costs that are not instance compute hours: EBS volumes, EBS snapshots, data transfer, NAT gateway charges, and Elastic IPs.
Here is what a small production web app actually costs in a month (illustrative):
| Line item | Shows up in Cost Explorer as | Example monthly |
| 2x m5.large On-Demand compute | Amazon EC2 | ~$140 |
| 200 GB gp3 EBS storage | EC2-Other | ~$16 |
| EBS snapshots | EC2-Other | ~$5 |
| 100 GB data transfer out | EC2-Other | ~$9 |
| 1 idle Elastic IP | EC2-Other | ~$4 |
| Total | ~$174 |
In that bill, less than a fifth of the “EC2-Other” total is the part most teams budget for. The factors below drive both lines:
| Cost driver | What it does | Finance lever |
| Instance type and family | Biggest factor, cents to tens of dollars per hour | Right-size and pick the right family |
| Processor architecture | AWS Graviton costs 10 to 20% less per hour than equivalent x86, with up to 30% better performance | Move compatible workloads to Graviton, often zero code changes, straight to margin |
| Region | Same instance, different price per Region | Run where it is cheaper if latency allows |
| Operating system | Linux is cheapest; Windows and RHEL add license cost | Use Linux where you can |
| EBS storage | Billed by volume type, size, IOPS, and snapshots (lands in EC2-Other) | Clean up unattached volumes and old snapshots |
| Data transfer | Same-AZ free; cross-AZ, cross-Region, and egress billed (EC2-Other) | Keep traffic in-AZ; watch EC2 to S3 |
| Elastic IP and add-ons | Idle IPs, load balancers, NAT gateways (EC2-Other) | Release idle IPs, audit add-ons |
Because these interact, EC2 cost is hard to forecast and easy to waste: 78% of organizations waste 21 to 50% of their cloud budget, and some lose more than $50,000 a month to oversized and outdated instances (Stacklet, 2024).
Nowhere do these factors hit margin harder than on GPU instances, where AI spend concentrates.
How much do EC2 GPU and AI compute instances cost?
GPU and accelerated instances are priced far above general-purpose, which is why a single AI workload can swing your monthly forecast. These are approximate On-Demand, Linux, us-east-1 rates at publication:
| Instance | Accelerator | Approx per hour | Built for |
| g4dn.xlarge | 1x NVIDIA T4 | ~$0.526 | Entry inference, small models |
| g5.xlarge | 1x NVIDIA A10G | ~$1.006 | Single-GPU inference and training |
| inf2.xlarge | AWS Inferentia2 | ~$0.758 | Production inference at scale |
| p3.2xlarge | 1x NVIDIA V100 | ~$3.06 | Older training and inference |
| p4d.24xlarge | 8x NVIDIA A100 | ~$32.77 | Large training and heavy inference |
| p5.48xlarge | 8x NVIDIA H100 | ~$31 to $55 (cut in 2025) | Foundation model and large-scale training |
A single p4d.24xlarge left running over a weekend costs roughly $1,573. That is the new face of waste, and it does not surface until the invoice.
There is a structural lever to raise with engineering. AWS Trainium and AWS Inferentia are AWS-designed chips built for price-performance. An inf2.xlarge at about $0.758 an hour undercuts a comparable g5.xlarge at about $1.006 an hour, and for supported models it can cut 25 to 40% off cost per inference (Markaicode, 2026), straight to gross margin. The constraint is compatibility through the AWS Neuron SDK, so it is a question to put to engineering, not a switch finance flips.
But the hourly rate on a GPU is the most misleading number in this guide, and the one most likely to wreck a margin assumption.
What does running AI inference on EC2 actually cost?
For AI, the rate card does not tell you cost to serve. The unit that drives margin is cost per million tokens, the hourly rate divided by throughput, and it can swing several times over on the same instance.
GPUs cost more per hour than CPUs, but for the transformer models behind modern AI features they process the work far faster, so the price justifies itself once volume scales.
Batch inference can favor GPUs by 5x to 50x in throughput, and an H100 serves roughly 250 to 300 tokens per second on 13B to 70B models, about double an A100.
Watch what utilization does. Take a g5.xlarge at about $1.006 an hour serving 200 tokens per second (an assumption, your engineers can measure it). Fully loaded, that is 720,000 tokens an hour:
- At 100% utilization: about $1.40 per million tokens.
- At 30% utilization: you still pay $1.006 an hour but produce a third of the tokens, so about $4.66 per million tokens.
Same instance, same rate card, more than 3x the cost to serve, purely from idle time. Utilization, not the sticker price, is the number that protects your margin.
It also decides build versus buy, which finance should price rather than inherit:
| Path | Cheapest when | Margin risk |
| CPU instances (for example c7i) | Small models, low or bursty volume | Falls over on large transformers and latency |
| GPU or Inferentia on EC2 | Sustained, high-volume serving | You own the idle time, and idle GPUs erode margin fast |
| Managed (Amazon SageMaker or a third-party API) | Low, spiky, or early-stage volume | Per-call cost overtakes self-hosting at high sustained volume |
The crossover depends on utilization, and most teams never price it. Compare managed API costs against self-hosting at your real volume using ChatGPT, Claude and OpenAI pricing.
The winner is the lower cost per million tokens at your volume, not the lower sticker price.
If you cannot see cost per inference or per customer today, that blind spot is the problem CloudZero solves. 
.
How to forecast, control, and reduce your EC2 spend
Forecast with the AWS Pricing Calculator, then add interruption risk, idle waste, and cost per inference, which the calculators do not model.
- A steady web app. A t3.large at about $0.0832 an hour over 730 hours is roughly $61 a month before storage and transfer. A one-year Savings Plan cuts the compute by up to about a third.
- An AI training run. Budget cost per run, not always-on: a p4d.24xlarge at about $32.77 an hour is about $786 for a 24-hour run, far less on Spot or on Trainium where compatible.
- A production inference service. Budget by throughput: monthly tokens divided by what one instance serves at target utilization, times the rate.
Then reduce, in rough priority order:
- Right-size and move to Graviton. The fastest wins, no commitment. Graviton cuts 10 to 20% per hour with up to 30% better performance on compatible workloads.
- Cover the baseline with commitments. Coverage, the share of usage running at a discount, is the finance KPI that drives savings. Coverage, the share of usage running at a discount, is the finance KPI that drives savings. Most teams reach 70 to 90% coverage with basic forecasting, though pushing higher gets risky as usage shifts (FinOps Foundation, State of FinOps 2025).
- Use Spot for the right work. Batch, stateless, and checkpointed AI training.
- Make structural AI moves. Shift compatible training to Trainium and inference to Inferentia2, and eliminate idle GPUs. If EC2 still does not fit, weigh EC2 alternatives or AWS alternatives.
All of this controls cost. But in the AI era, controlling cost is only half of finance’s job.
Why controlling cost is only half of finance’s job
Controlling cost is now the floor; proving return is the job. The old playbook (right-size, commit, use Spot) still works, but AI changed the terms finance operates under.
The unit changed: cost per hour is the wrong number for AI, because a high hourly rate can be the cheapest option if it finishes faster and serves more. Waste got brutal: an idle server wastes pennies, but an idle GPU can waste hundreds of dollars over a weekend. And forecasting broke, because AI spend scales with product adoption, not a fixed server count.
The result is a measurement gap. Only about half of organizations (51%) are confident they can measure the ROI of their AI investments (CloudZero, State of AI Costs 2025). The question the board is asking, and the one finance owns, is whether the AI spend is earning its return.
Answering it takes more than a rate card. It takes unit cost intelligence.
How CloudZero turns EC2 and AI spend into provable ROI
CloudZero gives finance the numbers to forecast and defend AI spend, mapped to exactly the problems in this guide.
| The finance problem | What CloudZero does |
| One big EC2 number, no idea what drives it | Cost per customer, feature, team, product, and per AI model and inference |
| A mysterious “EC2-Other” line | Breaks spend down to the resource and team, so storage, transfer, and idle IPs are visible, not buried |
| Surprises and runaway jobs after the fact | Real-time, hourly visibility with anomaly detection that flags a forgotten GPU before the invoice closes |
| Unpredictable AI spend you must forecast | Budgets and forecasting built for spend that scales with adoption |
| Idle and oversized resources | Surfaces idle and oversized instances and idle GPUs; CloudZero Advisor recommends the cheapest fit, including Graviton |
| Commitment decisions made on guesswork | Tracks Savings Plan and Reserved coverage and utilization so you commit with confidence |
| AI spend split across AWS GPUs and managed APIs | AI Hub unifies EC2 and GPU spend with OpenAI, Anthropic, and Amazon Bedrock into total AI spend and cost per model |
| No way to prove return to the board | Ties spend to revenue-driving units, so you can show cost per customer and gross margin, not just a bill |
The results show up on the P&L. Symphony Talent cut AWS costs by 48% after cost-per-client visibility revealed it had been losing money on a client for years, which reshaped its next contract negotiation. Diaceutics reduced its AWS bill by 41% and beat its S3 savings target by 70%. Drift improved COGS by $2.4M. And Upstart reduced cloud costs by $20M through engineering visibility and accountability.
Finance and engineering teams at Toyota, Skyscanner, Grammarly, Duolingo, and Upstart run from the same numbers. .
Frequently Asked Questions about Amazon EC2 Pricing
