A guided path · 8 modules · ~15 minutes
Start at zero and finish able to read an instance name like m7g.large, explain when a cheap burstable box quietly becomes the expensive one, and estimate your own monthly cost. Nothing to sign up for.
Module 1 of 8
It's renting a computer by the hour instead of buying one. That's genuinely most of it.
When you launch an EC2 instance, AWS carves a slice out of a physical server sitting in one of their data centres and hands it to you. You choose how much CPU, RAM, storage and network capacity that slice has. You pay for every hour (usually every second) it's switched on, and nothing once you shut it down.
Instance — in other words: a virtual server — a computer you control, running on hardware someone else owns.
vCPU — in other words: one virtual processor core; two vCPUs can do roughly twice as much work at once as one.
You
rent →
AWS
Module 2 of 8
Every instance name follows one pattern: [Family][Generation][Attributes].[Size]. Once you can read it, the catalogue stops being intimidating.
g Graviton/ARM, i Intel, a AMD, d local NVMe disk attached..large → .xlarge → .2xlarge. Each step roughly doubles vCPU, RAM and price.Click any name below to break it apart.
t4g.large
A 4th-gen burstable instance on AWS's own ARM chip. The classic cheap dev/low-traffic box.
Module 3 of 8
The first letter tells you what the machine is shaped for. Get this right and everything else is fine-tuning.
Bursts above a CPU baseline using earned "credits"; cheapest idle-heavy option.
Balanced CPU:RAM (roughly 1:4), fixed performance, no credit system.
More CPU per GB of RAM.
Roughly 1:8 vCPU:RAM ratio.
Very high RAM per core.
Fast local NVMe disks attached directly to the instance.
GPU or custom accelerator hardware.
Real Apple hardware in an AWS data center.
Module 4 of 8
This is the one beginners get wrong. T-series instances are cheap right up until they aren't — and the switch happens quietly.
A T-series instance has a CPU baseline — 30% for a t4g.large. Run below it and you bank credits. Run above it and you spend them. A t4g.large earns 36 credits/hour (864/day), and one credit buys one vCPU-minute at 100%.
CPU baseline — in other words: the amount of processor you've actually paid for; anything above it is borrowed against your credit balance.
When the credits run out, the instance is throttled back to baseline. No surprise charges — just slower performance until credits rebuild.
The instance keeps running at full speed past the credit limit, billed at a surplus rate of $0.04 per extra vCPU-hour. No throttling — but sustained use gets expensive fast.
| t4g.large (Unlimited, 100% CPU) | m7g.large (fixed) | |
|---|---|---|
| Base rate | $0.0672/hr | $0.0816/hr |
| Surplus at 100% CPU | +$0.056/hr (1.4 vCPU-hrs deficit × $0.04) | $0.00 |
| Total at sustained 100% CPU | $0.1232/hr (~$89.94/mo) | $0.0816/hr (~$59.57/mo) |
Prices shown are illustrative examples as of 2026 and change over time — check the AWS Pricing Calculator for current rates.
Pick a workload to watch the credit tank fill or drain.
Credit balance
100%
Avg CPU
8%
baseline 30%
No surplus charges
Idles most of the day, occasional spikes when you deploy or run a test.
Well under the 30% baseline. The credit tank stays full (capped), spikes are paid for out of earned credits, and you never see a surplus charge. t4g.large at $0.0672/hr — about $49/mo.
Takeaway
Below roughly 30% average utilization, T-series is the cheapest option available. Above it, the "cheap" burstable instance quietly becomes the expensive one.
Module 5 of 8
Four ways to pay for the exact same machine. The difference between them is how much flexibility you're willing to trade for a discount.
Pay as you go.
Billed per hour (per second for most Linux instances), no commitment, cancel whenever. It is the most expensive per-hour rate — you are paying for total flexibility.
Commit to a machine, get a discount.
You commit to a specific instance type in a specific region for 1 or 3 years. Discounts reach roughly 60% on a 3-year, all-upfront commitment.
Commit to a spend, not a machine.
You promise a dollars-per-hour spend instead of a specific instance type. More flexible than Reserved if your instance size might change as you grow.
Cheap leftovers, can be taken back.
You bid on AWS's spare capacity for steep discounts, but AWS can reclaim the instance with short notice. Only for work that can be interrupted and restarted safely.
An average month has about 730 hours. That's the only arithmetic you need:
$0.0672/hr × 730 = ~$49/mo
$0.0816/hr × 730 = ~$59.57/mo
Wherever you see an hourly rate on this site, the monthly figure sits next to it — get in the habit of reading both.
Takeaway
If an instance will run more than roughly 40–50% of the year, even the cheapest Reserved or Savings commitment beats On-Demand.
Prices shown are illustrative examples as of 2026 and change over time — check the AWS Pricing Calculator for current rates.
Module 6 of 8
AWS designs its own ARM-based chip, Graviton. It's typically 15–40% cheaper per unit of compute than the equivalent Intel or AMD instance.
The g in t4g.large or m7g.large means Graviton. Same instance shape, lower price. There is exactly one real catch: your Docker images and dependencies need linux/arm64 builds. Most mainstream stacks — Postgres, Redis, RabbitMQ, Node, Python, Go, Java — ship ARM64 images out of the box. Some older or specialized packages don't.
linux/arm64 — in other words: a version of a program compiled for ARM chips; x86 builds simply won't run on Graviton.
Choose Graviton if
You're comfortable checking your Docker images for ARM64 support. You get the same performance for less money.
Choose x86 if
You want zero compatibility thinking. A small premium buys universal compatibility with anything you might install.
Module 7 of 8
You can run Postgres or Redis in a container on the EC2 instance you already pay for, or pay AWS to run it for you via RDS, ElastiCache or Amazon MQ.
Self-hosting typically cuts database costs 40–70% versus the fully managed equivalent. In exchange you take on your own backups, patching and failover — the unglamorous work the managed service was doing quietly in the background.
Failover — in other words: automatically switching to a standby copy of the database when the main one dies.
Self-host if
Use managed (RDS/etc.) if
Module 8 of 8
Four situations you're likely to actually be in. Flip between them and see the reasoning, not just the answer.
The situation
You're one person running a Dockerized FastAPI backend with Postgres on the same box. A handful of requests per minute.
Why
Average CPU sits far below the 30% baseline, so credits keep accumulating and your spikes are free. Graviton keeps the base rate low, and both FastAPI and Postgres have first-class ARM64 images.
Recommended
t4g.large
Estimated cost
~$49/mo
Prices shown are illustrative examples as of 2026 and change over time — check the AWS Pricing Calculator for current rates.
The cheat sheet consolidates the family table, the naming decoder, the credit formula and the pricing models onto one page. That's the one to bookmark.
Open the cheat sheet