AWS EC2

What AWS EC2 actually is

Amazon EC2 is the original public cloud compute service — the 2006 launch that started the IaaS market and still the reference implementation of "a virtual server by the hour." You pick an instance type (AWS now offers over 1,000 of them across Intel, AMD and Arm processors), an Amazon Machine Image, a region and availability zone, and you get a virtual machine with root access on the Nitro System. Per-second billing with a 60-second minimum, no long-term commitment on On-Demand, and roughly every hardware shape anyone has ever asked for: general purpose, compute-, memory-, storage- and accelerator-optimized, bare metal, Mac, and GPU instances up to the p5en.48xlarge with 192 vCPUs and 2 TB of RAM.

The instance catalog is the real product. Burstable T-series machines for spiky small workloads (t4g.nano at $0.0042/hour, t3.nano at $0.0052/hour, t3.micro at $0.0104/hour for 2 vCPU and 1 GiB), Graviton Arm instances for price/performance (m7g.large: 2 vCPU, 8 GiB at $0.0816/hour), compute-optimized C-series, memory-optimized R and X series, storage-optimized I and D families, and accelerator instances (g5.8xlarge at $2.448/hour for 32 vCPU and 128 GiB with NVIDIA GPUs) for ML training and inference. Bare metal is available too — a c6in.metal runs $7.2576/hour in us-east-1 — which matters for licensing, hypervisor-level isolation and specialized hardware.

Pricing levers sit on top of the on-demand rate: Spot Instances offer up to 90% discount on spare capacity (with the interruption caveat), and Savings Plans reduce costs by up to 72% in exchange for a one- or three-year hourly spend commitment. Reserved Instances and dedicated hosts add further shapes of commitment. In our view, no one serious pays list price forever — the on-demand rate is the ceiling, not the budget.

Pricing: what you really pay

The unit economics are public and granular. In US East (N. Virginia), Linux, On-Demand, from AWS's own pricing feed: t4g.nano $0.0042/hour (2 vCPU, 0.5 GiB, EBS-only), t3.micro $0.0104/hour (2 vCPU, 1 GiB), t3.medium $0.0416/hour (2 vCPU, 4 GiB), m7g.large $0.0816/hour (2 vCPU, 8 GiB), c7g.2xlarge $0.29/hour (8 vCPU, 16 GiB), t3.2xlarge $0.3328/hour (8 vCPU, 32 GiB). Windows and other licensed AMIs add per-hour license fees on top.

The instance price is never the bill. EBS volumes bill per GB-month plus IOPS and throughput (gp3 at roughly $0.08/GB-month in us-east-1), snapshots bill per GB-month stored, elastic IPv4 addresses now bill $0.005/hour when attached, data transfer out starts free-ish and climbs per GB, and load balancers, NAT gateways and CloudWatch all meter separately. This is the consistent surprise with EC2: a $0.0104/hour instance is rarely a $7.50/month bill.

New accounts receive up to $200 in AWS Free Tier credits applicable to eligible services including EC2, with the free plan available for 6 months after account creation and credits usable within 12 months; the pricing page also still carries the classic Free Tier terms footnote (monthly usage across regions, excluding GovCloud and IBM AMIs). For committed workloads, the arithmetic that matters: On-Demand for spiky dev/test, Spot for interruption-tolerant batch and CI, Savings Plans for steady production. We have not run these instances through our /benchmarks suite, so no throughput claims from us — but for anyone building their own comparisons, the per-second billing makes short test runs genuinely cheap. One practical note on comparisons: measure on the instance family you will actually run. Graviton (T4g, M7g, C7g) and Intel/AMD instances of the same nominal size behave differently on workloads that are not Arm-clean, and the per-core price gap is large enough that porting effort often pays back inside a quarter.

Performance & infrastructure

AWS runs 124 Availability Zones across 39 geographic regions, with more announced (Saudi Arabia and Chile next). EC2 spans all of them, plus Local Zones and Wavelength for single-digit-millisecond edge placement. Each AZ is a physically separate failure domain with independent power and cooling; the Nitro System underpins the hypervisor, offloads networking and storage, and has materially reduced the virtualization overhead of recent generations. Placement groups, ENA networking up to 100+ Gbps on bigger types, and NVMe instance storage on the I and D families cover the performance-critical cases. The Nitro pace of iteration also shows up as steady generational gains — each new family generation has improved price/performance, which is one reason we recommend re-pricing workloads annually rather than assuming the instance type chosen three years ago is still the right one.

For teams with data-residency constraints, the region map covers EU (Frankfurt, Ireland, London, Paris, Stockholm, Milan, Spain, Zurich), US, and APAC (Tokyo, Seoul, Singapore, Sydney, Mumbai, Jakarta and more), and AWS commits contractually to region-scoped services. Sovereignty-focused customers can also look at dedicated instances and, in Europe, at the sovereignty controls AWS documents per region — worth reading before assuming anything about where data physically lives.

Reliability commitments are contractual: the Region-Level SLA is 99.99% monthly uptime when your running instances are deployed across two or more AZs, and the Instance-Level SLA is 99.5% for a single instance, each with service credits of 10%, 30% or 100% of affected charges. Note the structure — the four-nines number only applies if you do the multi-AZ work. Single-AZ deployments are contractually a 99.5% product.

Security posture: security groups and network ACLs at the VPC layer, IAM instance roles, EBS encryption, Nitro-based isolation, and AWS Shield Standard DDoS protection at no additional charge. IPv6 is supported end-to-end in VPC. Patching, however, is yours — EC2 is unmanaged IaaS.

Where it falls short

Cost control is the perennial complaint, and it is structural. The pricing page for one region/OS combination contains over a thousand instance sizes; the bill adds EBS, snapshots, IPv4 addresses, data transfer, NAT, and per-service meters. Teams without a FinOps habit routinely overspend by 30%+ — idle instances, oversized EBS volumes, forgotten elastic IPs. AWS gives you the tools (Cost Explorer, budgets, Compute Optimizer); it does not stop you.

Expertise required is real. VPC subnet design, security groups, AMI hygiene, IAM least privilege, and patch management are all on you. A misconfigured security group is the classic AWS breach vector, and the console will happily let you create one. Compare this with a managed panel where the safe defaults are the only defaults.

The SLA caveat again: 99.99% is multi-AZ only, and achieving it costs double the infrastructure. A single instance is contractually 99.5%, which allows over three hours of downtime a month. And Spot's 90% discount comes with interruptions — great for CI and batch, wrong for anything with a long-lived in-memory state.

One more complaint from our own experience: the console's default choices are not cost defaults. A launch wizard will happily attach an 8 GB gp2 volume, an elastic IP and a NAT gateway path to a $0.0104/hour instance, and the resulting bill is dominated by everything except the instance. We treat "launch it with Terraform from a reviewed module" as the only reliable way to keep the defaults honest.

Finally: egress. Data transfer out to the internet is metered per GB and is one of the most reliable surprise costs in the AWS bill. Moving large datasets out of AWS can cost more than storing them did.

Who should pick this

Our pick for teams that need the widest hardware and location selection in the market and have the engineering discipline to use it. Enterprise workloads with compliance requirements, ML training on GPU clusters, HPC, Windows workloads with license mobility, and any architecture that must run identically in twelve countries — EC2's catalog answers all of it. The Spot + Savings Plans combination is also our standard advice for cost-conscious production shops willing to do the math.

We would not pick it for a personal blog or a single small service — a fixed-price VPS at a fraction of the mental overhead is the better trade there (see /category/vps-providers). We would also look elsewhere for teams without AWS expertise and without budget for a managed layer: the hourly rate is only cheap if the architecture around it is cheap too.

Key takeaways

• The original cloud compute service (2006), now over 1,000 instance types across Intel, AMD and Arm, including bare metal and GPU families up to 192 vCPUs / 2 TB RAM.
• On-Demand in us-east-1 (Linux): t4g.nano $0.0042/hour, t3.micro $0.0104/hour (2 vCPU/1 GiB), m7g.large $0.0816/hour (2 vCPU/8 GiB) — per second, 60-second minimum.
• Spot offers up to 90% off spare capacity; Savings Plans up to 72% off for a 1–3 year hourly commitment; new accounts get up to $200 in Free Tier credits (6-month free plan).
• SLA: 99.99% region-level (only with instances across 2+ AZs), 99.5% for a single instance — both with 10/30/100% service credits.
• 124 Availability Zones in 39 regions; Nitro System virtualization; Shield Standard DDoS protection included.
• The instance price is the ceiling, not the bill: EBS, snapshots, elastic IPv4, data transfer out and supporting services all meter separately.

Sources

• https://aws.amazon.com/ec2/
• https://aws.amazon.com/ec2/pricing/on-demand/
• https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20(N.%20Virginia)/Linux/index.json
• https://aws.amazon.com/ec2/sla/
• https://aws.amazon.com/about-aws/global-infrastructure/
• https://aws.amazon.com/shield/pricing/
• https://en.wikipedia.org/wiki/Amazon_Web_Services