AWS marked 20 years since the launch of the Amazon EC2 beta, which appeared in 2006 in a blog post written by Jeff Barr and introduced resizable virtual Linux servers billed by the hour, with one instance type, m1.small, and one Region, US East. According to AWS, this simple model represented a shift in how computing infrastructure was built and managed.
Since that launch, the core value EC2 provides has not changed: delivering scalable computing capacity within minutes, paying for consumption, and the ability to increase or reduce resources without long-term commitments. But the service’s scope has changed radically; it expanded from one instance type to more than 1,200 types covering general-purpose computing, workloads optimized for processing, memory, or storage, accelerated computing, and high-performance computing. It also grew from one AWS Region to 39 Regions worldwide.
The Foundation on Which the Ecosystem Was Built
AWS links EC2’s evolution to several core services added over time. Amazon Elastic Block Store provided persistent block storage in 2008, while Elastic Load Balancing, Auto Scaling, and Amazon CloudWatch supported scalability and high availability in 2009. That same year, Amazon Virtual Private Cloud introduced logically isolated networks. AWS Nitro System followed in 2017, forming a foundation for accelerating innovation and improving security, followed by the AWS Graviton processor design in 2018 for cost-sensitive, large-scale workloads.
EC2 is no longer confined to traditional AWS Regions. Through AWS Outposts, introduced in 2018, EC2 instances can run locally, while AWS Local Zones have provided locations closer to users since 2019, and AWS Wavelength enables instances to run within the networks of global telecommunications operators using 5G technology.
From General-Purpose Computing to the AI Lifecycle
AWS highlights EC2’s expansion into AI chips. It introduced Inf1 instances equipped with AWS Inferentia chips for large-scale machine learning inference, and Inf2 instances became generally available in April 2023 for large generative inference workloads. In parallel, Trainium instances target training and inference; Trn1 was introduced in November 2021, followed by Trn2, equipped with Trainium2, in December 2024.
Trn2 UltraServers connect up to 64 accelerators through NeuronLink to train foundation models containing trillions of parameters. AWS also says that Trn3 UltraServers, introduced at AWS re:Invent 2025, can connect up to 144 Trainium3 chips for training and serving workloads involving advanced models, agent applications, inference, and video.
What Changed in Practice?
The expansion was not limited to chips. EC2 Mac began in 2020 with Mac mini computers equipped with Intel Core i7 processors, followed by Mac M1 instances in July 2022 as the first Arm-based macOS instances in EC2. M2 Pro instances followed in 2023, then M4 and M4 Pro in 2025, and M3 Ultra and M4 Max in 2026, providing cloud environments for building and testing macOS, iOS, iPadOS, tvOS, watchOS, and visionOS applications.
EC2 Capacity Blocks for ML also introduced a temporary reservation model for GPU capacity; customers can reserve GPU instances, starting with P5, for a future date and only for the required duration. In November 2024, support was added for provisioning within minutes and extending reservations for up to six months. Support later expanded to P6-B300, P6-B200, P5e, P5en, P4d, P4de, Trn1, Trn2, and Trn3, in addition to P5.
The Significance and Limits of the Anniversary
This history shows that EC2 has become an operating layer on which services such as Amazon ECS, Amazon EKS, AWS Lambda, AWS Fargate, AWS Batch, Amazon EMR, Amazon SageMaker AI, and Amazon Bedrock rely. AWS also notes that Graviton5, introduced at AWS re:Invent 2025, has 192 cores, five times more cache, and up to 33% lower inter-core communication latency, with M9g and M9gd instances followed by C9g and C9gd in June 2026.
The latest item AWS mentions is the AWS Nitro Isolation Engine in 2026, a component within the Nitro Hypervisor that uses formal verification to provide a mathematical guarantee of isolating customers’ workloads from one another and from AWS operators. This point, like the performance and token-economics figures cited by the company, remains a description from the source itself rather than an independent assessment. But the overall trajectory is clear: what began as a flexible virtual-server service has become a foundation for running diverse workloads, from web servers to training AI models containing trillions of parameters.