Amazon EC2 marks 20 years of cloud computing growth

Amazon Web Services has marked the 20th anniversary of Amazon EC2, which has grown from a single instance type in one region to more than 1,200 instance types across 39 regions.

EC2, or Elastic Compute Cloud, launched with Linux virtual servers billed by the hour. That model helped establish on-demand infrastructure rental as an alternative to buying and maintaining physical servers.

Over the past two decades, the service has become the base layer for much of AWS's broader compute portfolio. Services including Amazon ECS, Amazon EKS, AWS Lambda, AWS Fargate, AWS Batch, Amazon EMR, Amazon SageMaker AI and Amazon Bedrock ultimately run on EC2 capacity.

Early building blocks

AWS highlighted several early milestones that shaped the service. Amazon Elastic Block Store added persistent block storage, while Elastic Load Balancing, Auto Scaling and Amazon CloudWatch introduced tools for scaling applications and monitoring workloads.

Amazon Virtual Private Cloud gave customers isolated network environments within AWS. Later, the AWS Nitro System reshaped EC2's underlying design by moving virtualisation and other infrastructure functions into dedicated hardware and software components.

That architecture became the foundation for a broader push into in-house chip design. AWS launched Graviton processors for general-purpose computing and later introduced chips for machine learning training and inference.

AI and silicon

Over the past five years, AWS has placed greater emphasis on EC2 products for artificial intelligence workloads. Inferentia chips were introduced for machine learning inference, followed by Inf2 instances for larger generative AI inference tasks.

Trainium was introduced for AI training, with later versions expanding the scale of those systems. AWS said Trn2 UltraServers link 64 Trainium2 accelerators, while Trn3 UltraServers interconnect up to 144 Trainium3 chips for training and serving large models.

AWS also expanded access models for AI computing through EC2 Capacity Blocks for ML. The product lets customers reserve GPU- or Trainium-based capacity for a defined period at a future date. Supported instances now include several Nvidia-based systems as well as Trn1, Trn2 and Trn3.

Another recent area of expansion is custom Arm-based processors. AWS said Graviton5, previewed ahead of the latest M9g, M9gd, C9g and C9gd instances, includes 192 cores, a larger cache and lower inter-core latency than earlier generations.

Beyond the cloud region

EC2 has also moved beyond its original model of running only inside standard AWS regions. AWS pointed to Outposts, Local Zones and Wavelength as examples of how the service has been extended into customer sites, metropolitan locations and telecoms networks.

That shift reflects broader enterprise demand for applications that require lower latency or data handling closer to end users and devices. By extending EC2 into those settings, AWS has sought to keep its core compute platform relevant across a wider range of deployments.

AWS also highlighted EC2 Mac instances, which brought Apple hardware into the service for developers building and testing software for Apple operating systems. Since the first Mac instances, the lineup has expanded to systems based on multiple Apple chip generations, including M1, M2 Pro, M4, M4 Pro, M3 Ultra and M4 Max.

Security focus

One of the platform's latest additions is the Nitro Isolation Engine. AWS said the feature is designed to provide formal proof of workload isolation in the Nitro Hypervisor and to show that customer workloads are separated from one another and from AWS operators.

The focus on isolation and verification comes as cloud customers face greater scrutiny over where workloads run and how they are separated in shared infrastructure. AWS linked that work to the sixth generation of the Nitro System, underscoring how a design introduced several years ago continues to underpin new EC2 features.

Despite the scale of change, AWS said EC2's central model remains the same: customers can obtain computing capacity quickly, pay for what they use and scale without long-term commitments. That approach now supports workloads ranging from basic web applications to large AI model training clusters.

Across the architectural patterns customers have built over the past 20 years, from simple web servers to trillion-parameter foundation model training clusters, each begins with a decision to launch an instance.